CUTLASS 3.5.0 (#1411)
This commit is contained in:
Vijay Thakkar
@@ -59,11 +59,11 @@ class TestEVTMixed(EVTTestCaseBase):
|
||||
return D, F, F_row_max, E_col_max
|
||||
|
||||
if device_cc() == 80:
|
||||
aligments = [2, 4, 8]
|
||||
alignments = [2, 4, 8]
|
||||
else:
|
||||
# Sm90 EVT currently only supports 128-bit alignment
|
||||
aligments = [8,]
|
||||
for align in aligments:
|
||||
alignments = [8,]
|
||||
for align in alignments:
|
||||
for m, n, k, l in self.get_problem_sizes(align):
|
||||
example_inputs = {
|
||||
"accum": self.fake_tensor(self.element, (l, m, n)),
|
||||
|
||||
@@ -117,10 +117,12 @@ void FilterArchitecture() {
|
||||
{ "SM70*", 70, 75},
|
||||
{ "SM75*", 75, kMaxDevice},
|
||||
{ "SM80*", 80, kMaxDevice},
|
||||
{ "SM90*", 90, 90 },
|
||||
{ "SM89*", 89, 89},
|
||||
{ "SM90*", 90, 90},
|
||||
{ 0, 0, false }
|
||||
};
|
||||
|
||||
|
||||
// Set negative test filters
|
||||
std::stringstream ss;
|
||||
ss << "-";
|
||||
|
||||
@@ -47,3 +47,4 @@ foreach(SUBDIR
|
||||
|
||||
endforeach()
|
||||
|
||||
add_subdirectory(device_3x)
|
||||
|
||||
@@ -122,14 +122,19 @@ inline std::ostream &operator<<(std::ostream &out, CachedTestKey const &result)
|
||||
|
||||
struct CachedTestResult {
|
||||
uint32_t D;
|
||||
|
||||
uint32_t sum;
|
||||
uint32_t sum_of_square;
|
||||
uint32_t second_sum_of_square;
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CachedTestResult(): D() { }
|
||||
CachedTestResult(): D(), sum(), sum_of_square(), second_sum_of_square() { }
|
||||
|
||||
CachedTestResult(uint32_t D): D(D) { }
|
||||
CachedTestResult(uint32_t D): D(D), sum(), sum_of_square(), second_sum_of_square() { }
|
||||
|
||||
CachedTestResult(uint32_t D, uint32_t sum, uint32_t sum_of_square, uint32_t second_sum_of_square):
|
||||
D(D), sum(sum), sum_of_square(sum_of_square), second_sum_of_square(second_sum_of_square) { }
|
||||
|
||||
operator bool() const {
|
||||
return bool(D);
|
||||
@@ -325,6 +330,34 @@ inline std::ostream &EncodeProblemSize(
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Encode 3.x ConvNd ProblemShape
|
||||
template <class ProblemShape>
|
||||
inline std::ostream &EncodeProblemSize(
|
||||
std::ostream &out,
|
||||
ProblemShape const& problem_shape) {
|
||||
|
||||
out << problem_shape.shape_A << "_";
|
||||
out << problem_shape.shape_B << "_";
|
||||
|
||||
out << "padl" << problem_shape.lower_padding << "_";
|
||||
out << "padu" << problem_shape.upper_padding << "_";
|
||||
out << "str" << problem_shape.traversal_stride << "_";
|
||||
out << "dil" << problem_shape.dilation << "_";
|
||||
|
||||
switch (problem_shape.mode) {
|
||||
case cutlass::conv::Mode::kCrossCorrelation:
|
||||
out << "corr";
|
||||
break;
|
||||
case cutlass::conv::Mode::kConvolution:
|
||||
out << "conv";
|
||||
break;
|
||||
}
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename Element>
|
||||
inline std::string ElementTypeName() {
|
||||
return std::string(typeid(Element).name());
|
||||
@@ -822,6 +855,56 @@ inline CachedTestKey CreateCachedConv3dTestKey(
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
class ProblemShape,
|
||||
typename ElementA,
|
||||
typename ElementB,
|
||||
typename ElementC,
|
||||
typename ElementD
|
||||
>
|
||||
inline CachedTestKey CreateCachedConvNd3xTestKey(
|
||||
cutlass::conv::Operator conv_operator,
|
||||
ProblemShape const& problem_shape,
|
||||
double alpha,
|
||||
double beta,
|
||||
thrust::universal_vector<ElementA> A,
|
||||
thrust::universal_vector<ElementB> B,
|
||||
thrust::universal_vector<ElementC> C
|
||||
) {
|
||||
|
||||
CachedTestKey key;
|
||||
|
||||
// Encode convNd operator and problem sizes
|
||||
std::stringstream ss_op;
|
||||
ss_op << "conv" << ProblemShape::RankS << "d";
|
||||
key.op = ss_op.str();
|
||||
|
||||
std::stringstream ss_problem;
|
||||
ss_problem << EncodeOperator(conv_operator) << "_";
|
||||
EncodeProblemSize(ss_problem, problem_shape);
|
||||
ss_problem << "_alpha" << EncodeScalar(alpha) << "_beta" << EncodeScalar(beta);
|
||||
key.problem = ss_problem.str();
|
||||
|
||||
// Encode problem data types
|
||||
std::stringstream ss_types;
|
||||
EncodeTypes<
|
||||
ElementA,
|
||||
ElementB,
|
||||
ElementC,
|
||||
ElementD>(ss_types);
|
||||
key.types = ss_types.str();
|
||||
|
||||
// Encode problem data
|
||||
CRC32 crc_hash;
|
||||
key.A = TensorHash(A, crc_hash);
|
||||
key.B = TensorHash(B, crc_hash);
|
||||
key.C = TensorHash(C, crc_hash);
|
||||
|
||||
return key;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace test::conv::device
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -93,6 +93,20 @@ if (CUTLASS_NVCC_MAX_ARCH GREATER_EQUAL 80)
|
||||
|
||||
endif()
|
||||
|
||||
if (CUTLASS_NVCC_MAX_ARCH GREATER_EQUAL 89)
|
||||
|
||||
add_dependencies(
|
||||
cutlass_test_unit_conv_device
|
||||
cutlass_test_unit_conv_device_tensorop_f8_sm89
|
||||
)
|
||||
|
||||
add_dependencies(
|
||||
test_unit_conv_device
|
||||
test_unit_conv_device_tensorop_f8_sm89
|
||||
)
|
||||
|
||||
endif()
|
||||
|
||||
#
|
||||
# OpClassSimt (CUDA cores)
|
||||
#
|
||||
@@ -126,6 +140,14 @@ if (CUTLASS_NVCC_MAX_ARCH GREATER_EQUAL 80)
|
||||
conv2d_fprop_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm80.cu
|
||||
conv2d_dgrad_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm80.cu
|
||||
conv2d_wgrad_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm80.cu
|
||||
|
||||
conv2d_fprop_with_broadcast_simt_sm80.cu
|
||||
|
||||
conv3d_fprop_implicit_gemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32_sm80.cu
|
||||
conv3d_dgrad_implicit_gemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32_sm80.cu
|
||||
conv3d_wgrad_implicit_gemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32_sm80.cu
|
||||
|
||||
conv3d_fprop_with_broadcast_simt_sm80.cu
|
||||
)
|
||||
|
||||
endif()
|
||||
@@ -245,3 +267,14 @@ if (CUTLASS_NVCC_MAX_ARCH GREATER_EQUAL 75)
|
||||
|
||||
endif()
|
||||
|
||||
if (CUTLASS_NVCC_MAX_ARCH GREATER_EQUAL 89)
|
||||
|
||||
# Conv - F8 input, F8 output, F32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f8_sm89
|
||||
|
||||
conv2d_fprop_implicit_gemm_f8nhwc_f8nhwc_f8nhwc_tensor_op_f32_sm89.cu
|
||||
)
|
||||
|
||||
endif()
|
||||
|
||||
|
||||
+368
@@ -0,0 +1,368 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide Conv2d fprop interface with:
|
||||
A: NHWC, of type FE4M4 or FE5M2
|
||||
B: NHWC, of type FE4M3 or FE5M2
|
||||
C: NHWC, of FE4M3 or FE5M2
|
||||
Accum: F32
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/epilogue/thread/activation.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_generic_with_scaling.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d_fprop_with_absmax.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "conv2d_with_absmax_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM89_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Analytic_ImplicitGemm_fe4m3nhwc_fe4mnhwc_fe4mnhwc_tensor_op_f32,
|
||||
identity_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Analytic_ImplicitGemm_fe5m2nhwc_fe4m3nhwc_fe4m3nhwc_tensor_op_f32,
|
||||
identity_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Analytic_ImplicitGemm_fe5m2nhwc_fe4m3nhwc_fe5m2nhwc_tensor_op_f32,
|
||||
identity_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e5m2_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Optimized_ImplicitGemm_fe4m3nhwc_fe4mnhwc_fe4mnhwc_tensor_op_f32,
|
||||
identity_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Optimized_ImplicitGemm_fe4m3nhwc_fe4mnhwc_fe4mnhwc_tensor_op_f32,
|
||||
relu_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::ReLu,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::ReLu>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Optimized_ImplicitGemm_fe4m3nhwc_fe4mnhwc_fe4mnhwc_tensor_op_f32,
|
||||
identity_fastacc_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAddFastAccum,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Conv2d_Fprop_Optimized_ImplicitGemm_fe4m3nhwc_fe4mnhwc_fe4mnhwc_tensor_op_f32,
|
||||
identity_noScale_128x256x64_64x3_64x64x64) {
|
||||
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithAbsMax<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementOutput, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithAbsmax<Conv2dFprop, cutlass::epilogue::thread::Identity>(
|
||||
/* scaleA = */false,
|
||||
/* scaleB = */false,
|
||||
/* scaleC = */false
|
||||
);
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM89_SUPPORTED
|
||||
@@ -0,0 +1,171 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide Implicit GEMM interface
|
||||
*/
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/array.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_bias_elementwise.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_residual_block.h"
|
||||
#include "cutlass/epilogue/thread/activation.h"
|
||||
#include "cutlass/conv/kernel/default_conv2d_fprop_with_broadcast.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_with_broadcast_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_With_Broadcast_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementCompute = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationBiasElementwise<
|
||||
ElementC,
|
||||
ElementAccumulator,
|
||||
ElementCompute,
|
||||
ElementC,
|
||||
ElementC,
|
||||
1,
|
||||
cutlass::epilogue::thread::ReLu<float>
|
||||
>;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithBroadcast<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2dWithBroadcast<Conv2dFprop>());
|
||||
}
|
||||
|
||||
// Test residual block fusion: UnaryOp(BinaryOp(ActivationOp(Conv2d(X) + bias), residual))
|
||||
// LinearCombinationResidualBlock does not support the split-k mode unless ActivationOp is Identity.
|
||||
// This is because the activation needs to be applied to the fully accumulated output of the Conv2d op,
|
||||
// which only the last thread block would have an access to, before applying BinaryOp.
|
||||
// The epilogue functor in the last thread block would have to be given three inputs, namely
|
||||
// partial outputs, bias, and residual, but this is not supported in the current interface.
|
||||
// Set TestSplitK = false to skip split-k tests with non-trivial ActivationOp.
|
||||
template <
|
||||
template<typename T> class ActivationOp,
|
||||
template<typename T> class BinaryOp,
|
||||
template<typename T> class UnaryOp,
|
||||
bool TestSplitK = true
|
||||
>
|
||||
void TestResidaulBlock() {
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementD = ElementC;
|
||||
using ElementCompute = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationResidualBlock<
|
||||
ElementD,
|
||||
ElementAccumulator,
|
||||
ElementCompute,
|
||||
ElementC,
|
||||
1,
|
||||
ActivationOp,
|
||||
BinaryOp,
|
||||
UnaryOp
|
||||
>;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFpropWithBroadcast<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
struct ReferenceOp {
|
||||
using OutputOp = typename Conv2dFprop::EpilogueOutputOp;
|
||||
using ElementZ = typename OutputOp::ElementZ;
|
||||
|
||||
ActivationOp<ElementCompute> activation;
|
||||
BinaryOp<ElementCompute> binary_op;
|
||||
UnaryOp<ElementCompute> unary_op;
|
||||
|
||||
void operator()(ElementZ &Z, ElementZ&, ElementCompute conv2d, ElementCompute residual) {
|
||||
Z = ElementZ(unary_op(binary_op(activation(conv2d), residual)));
|
||||
}
|
||||
};
|
||||
|
||||
bool passed = test::conv::device::TestAllConv2dWithBroadcast<Conv2dFprop, ReferenceOp, true, TestSplitK>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_With_Residual_Block_Plus_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
// Resnet
|
||||
TestResidaulBlock<cutlass::epilogue::thread::Identity, cutlass::plus, cutlass::epilogue::thread::ReLu>();
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -231,6 +231,75 @@ struct TestbedConv2dProblemSizes {
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Small input size x stride (1,1) asymmetric paddings (1, 0, 1, 0)
|
||||
// C < CTA::K and non-multiples of CTA::K. Typical CTA::K = {32, 64}
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 1, 1, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 1, 1, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 1, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 1, 3, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 7, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 3, 3, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 7, 9, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 4, 4, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{2, 7, 9, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 5, 5, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{3, 7, 9, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 6, 5, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{3, 7, 9, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 6, 6, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{3, 7, 9, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 7, 7, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 0, 1, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Small input size x stride (2,2)
|
||||
// C < CTA::K and non-multiples of CTA::K. Typical CTA::K = {32, 64}
|
||||
|
||||
@@ -192,7 +192,7 @@ public:
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv2d::UnderlyingKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Conv2d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
|
||||
@@ -191,7 +191,7 @@ public:
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv2d::UnderlyingKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Conv2d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
|
||||
@@ -0,0 +1,622 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Testbed for running device-level Conv2Ds with absolute maximum calculation and scaling
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
#include <sstream>
|
||||
|
||||
#include "conv2d_problems.h"
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "../../gemm/device/testbed_utils.h"
|
||||
|
||||
#include "cutlass/matrix_coord.h"
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/layout/matrix.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/distribution.h"
|
||||
#include "cutlass/util/reference/host/convolution.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_reduce.h"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
typename Conv,
|
||||
template<typename T> class ActivationFunctor
|
||||
>
|
||||
struct TestbedConv2dWithAbsMax {
|
||||
|
||||
using ElementAccumulator = typename Conv::ElementAccumulator;
|
||||
using ElementCompute = typename Conv::UnderlyingKernel::Epilogue::OutputOp::ElementCompute;
|
||||
using ElementScalingFactor = typename Conv::EpilogueOutputOp::ElementScalingFactor;
|
||||
using ElementAbsmax = typename Conv::EpilogueOutputOp::ElementAbsmax;
|
||||
static cutlass::conv::Operator const kConvolutionalOperator = Conv::kConvolutionalOperator;
|
||||
|
||||
static bool const kScaleAux = Conv::EpilogueOutputOp::kIsScalingAndAmaxAuxOutputNeeded;
|
||||
static bool const kScaleOutput = Conv::EpilogueOutputOp::kIsScalingAndAmaxOutputNeeded;
|
||||
bool doScaleA;
|
||||
bool doScaleB;
|
||||
bool doScaleC;
|
||||
|
||||
/// Initialization
|
||||
cutlass::Distribution::Kind init_A;
|
||||
cutlass::Distribution::Kind init_B;
|
||||
cutlass::Distribution::Kind init_C;
|
||||
uint64_t seed;
|
||||
|
||||
cutlass::HostTensor<typename Conv::ElementA, typename Conv::LayoutA> tensor_A;
|
||||
cutlass::HostTensor<typename Conv::ElementB, typename Conv::LayoutB> tensor_B;
|
||||
cutlass::HostTensor<typename Conv::ElementC, typename Conv::LayoutC> tensor_C;
|
||||
cutlass::HostTensor<typename Conv::EpilogueOutputOp::ElementAuxOutput, typename Conv::LayoutC> tensor_Aux;
|
||||
cutlass::HostTensor<typename Conv::EpilogueOutputOp::ElementOutput, typename Conv::LayoutC> tensor_D;
|
||||
cutlass::HostTensor<typename Conv::ElementC, typename Conv::LayoutC> tensor_Vector;
|
||||
cutlass::HostTensor<ElementAccumulator, typename Conv::LayoutC> tmp_D;
|
||||
cutlass::HostTensor<typename Conv::EpilogueOutputOp::ElementOutput, typename Conv::LayoutC> reference_D;
|
||||
cutlass::HostTensor<typename Conv::EpilogueOutputOp::ElementAuxOutput, typename Conv::LayoutC> reference_Aux;
|
||||
cutlass::HostTensor<ElementScalingFactor, typename Conv::LayoutC> scale_A;
|
||||
cutlass::HostTensor<ElementScalingFactor, typename Conv::LayoutC> scale_B;
|
||||
cutlass::HostTensor<ElementScalingFactor, typename Conv::LayoutC> scale_C;
|
||||
cutlass::HostTensor<ElementScalingFactor, typename Conv::LayoutC> scale_D;
|
||||
cutlass::HostTensor<ElementScalingFactor, typename Conv::LayoutC> scale_Aux;
|
||||
cutlass::HostTensor<ElementAbsmax, typename Conv::LayoutC> abs_max_Aux;
|
||||
cutlass::HostTensor<ElementAbsmax, typename Conv::LayoutC> abs_max_D;
|
||||
cutlass::HostTensor<ElementAbsmax, typename Conv::LayoutC> reference_abs_max_Aux;
|
||||
cutlass::HostTensor<ElementAbsmax, typename Conv::LayoutC> reference_abs_max_D;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
TestbedConv2dWithAbsMax(
|
||||
bool scaleA = true,
|
||||
bool scaleB = true,
|
||||
bool scaleC = true,
|
||||
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = 2080
|
||||
):
|
||||
doScaleA(scaleA), doScaleB(scaleB), doScaleC(scaleC),
|
||||
init_A(init_A_), init_B(init_B_), init_C(init_C_), seed(seed_) { }
|
||||
|
||||
/// Helper to initialize scaling factors
|
||||
template <typename Element, typename Layout>
|
||||
bool initialize_scale_factor(cutlass::TensorView<Element, Layout> view, uint64_t seed, int bits=0) {
|
||||
cutlass::reference::host::TensorFillRandomUniform(view, seed, double(1.), double(0.), bits);
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Helper to initialize a tensor view
|
||||
template <typename Element, typename Layout>
|
||||
bool initialize_tensor(
|
||||
cutlass::TensorView<Element, Layout> view,
|
||||
cutlass::Distribution::Kind dist_kind,
|
||||
uint64_t seed) {
|
||||
|
||||
if (dist_kind == cutlass::Distribution::Uniform) {
|
||||
|
||||
double scope_max, scope_min;
|
||||
int bits_input = cutlass::sizeof_bits<Element>::value;
|
||||
int bits_output = cutlass::sizeof_bits<typename Conv::ElementC>::value;
|
||||
|
||||
if (bits_input == 1) {
|
||||
scope_max = 2;
|
||||
scope_min = 0;
|
||||
} else if (bits_input <= 8) {
|
||||
scope_max = 2;
|
||||
scope_min = -2;
|
||||
} else if (bits_output == 16) {
|
||||
scope_max = 5;
|
||||
scope_min = -5;
|
||||
} else {
|
||||
scope_max = 8;
|
||||
scope_min = -8;
|
||||
}
|
||||
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
view, seed, scope_max, scope_min, 0);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Identity) {
|
||||
|
||||
cutlass::reference::host::TensorFillIdentity(view);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Gaussian) {
|
||||
|
||||
cutlass::reference::host::TensorFillRandomGaussian(view, seed, 0, 0.5);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Sequential) {
|
||||
|
||||
cutlass::reference::host::BlockFillSequential(
|
||||
view.data(), view.capacity());
|
||||
}
|
||||
else {
|
||||
EXPECT_TRUE(false) << "Not implemented";
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Initializes data structures
|
||||
void initialize(cutlass::conv::Conv2dProblemSize const &problem_size) {
|
||||
//
|
||||
// Allocate the GEMM workspace
|
||||
//
|
||||
|
||||
tensor_A.resize(implicit_gemm_tensor_a_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_B.resize(implicit_gemm_tensor_b_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_C.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_D.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_Vector.resize({1, 1, 1, implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size).c()});
|
||||
reference_D.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size), false);
|
||||
tmp_D.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size), false);
|
||||
|
||||
EXPECT_TRUE(initialize_tensor(tensor_A.host_view(), init_A, seed + 2019));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_B.host_view(), init_B, seed + 2018));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_C.host_view(), init_C, seed + 2017));
|
||||
EXPECT_TRUE(initialize_tensor(tensor_Vector.host_view(), init_C, seed + 2020));
|
||||
|
||||
// It is possible to randomly initialize to all zeros, so override this with non-zeros
|
||||
// in the upper left corner of each operand.
|
||||
cutlass::Coord<4> origin(0);
|
||||
tensor_A.host_view().at(origin) = typename Conv::ElementA(1);
|
||||
tensor_B.host_view().at(origin) = typename Conv::ElementB(1);
|
||||
tensor_C.host_view().at(origin) = typename Conv::ElementC(1);
|
||||
tensor_Vector.host_view().at(origin) = typename Conv::ElementC(1);
|
||||
|
||||
cutlass::reference::host::TensorFill(tensor_D.host_view());
|
||||
cutlass::reference::host::TensorCopy(reference_D.host_view(), tensor_C.host_view());
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_D.sync_device();
|
||||
tensor_Vector.sync_device();
|
||||
|
||||
int scale_bits = 2;
|
||||
if (doScaleA) {
|
||||
scale_A.resize({1, 1, 1, 1});
|
||||
EXPECT_TRUE(initialize_scale_factor(scale_A.host_view(), seed + 2021, scale_bits));
|
||||
scale_A.sync_device();
|
||||
}
|
||||
|
||||
if (doScaleB) {
|
||||
scale_B.resize({1, 1, 1, 1});
|
||||
EXPECT_TRUE(initialize_scale_factor(scale_B.host_view(), seed + 2022, scale_bits));
|
||||
scale_B.sync_device();
|
||||
}
|
||||
|
||||
if (doScaleC) {
|
||||
scale_C.resize({1, 1, 1, 1});
|
||||
EXPECT_TRUE(initialize_scale_factor(scale_C.host_view(), seed + 2023, scale_bits));
|
||||
scale_C.sync_device();
|
||||
}
|
||||
|
||||
if (kScaleOutput) {
|
||||
scale_D.resize({1, 1, 1, 1});
|
||||
EXPECT_TRUE(initialize_scale_factor(scale_D.host_view(), seed + 2024, scale_bits));
|
||||
scale_D.sync_device();
|
||||
|
||||
abs_max_D.resize({1, 1, 1, 1});
|
||||
cutlass::reference::host::TensorFill(abs_max_D.host_view());
|
||||
abs_max_D.sync_device();
|
||||
|
||||
reference_abs_max_D.resize({1, 1, 1, 1});
|
||||
}
|
||||
|
||||
if (kScaleAux) {
|
||||
tensor_Aux.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
cutlass::reference::host::TensorFill(tensor_Aux.host_view());
|
||||
tensor_Aux.sync_device();
|
||||
|
||||
scale_Aux.resize({1, 1, 1, 1});
|
||||
EXPECT_TRUE(initialize_scale_factor(scale_Aux.host_view(), seed + 2025, scale_bits));
|
||||
scale_Aux.sync_device();
|
||||
|
||||
abs_max_Aux.resize({1, 1, 1, 1});
|
||||
cutlass::reference::host::TensorFill(abs_max_Aux.host_view());
|
||||
abs_max_Aux.sync_device();
|
||||
|
||||
reference_Aux.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size), false);
|
||||
reference_abs_max_Aux.resize({1, 1, 1, 1});
|
||||
}
|
||||
}
|
||||
|
||||
/// Compares computed reference with device reference and outputs to a file if incorrect
|
||||
bool compare_reference(
|
||||
cutlass::conv::Conv2dProblemSize const &problem_size,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
tensor_D.sync_host();
|
||||
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_A.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_B.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_C.host_view()), 0);
|
||||
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_D.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(reference_D.host_view()), 0);
|
||||
bool passed = cutlass::reference::host::TensorEquals(reference_D.host_view(), tensor_D.host_view());
|
||||
|
||||
if (kScaleAux) {
|
||||
tensor_Aux.sync_host();
|
||||
abs_max_Aux.sync_host();
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_Aux.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(abs_max_Aux.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(reference_Aux.host_view()), 0);
|
||||
passed &= cutlass::reference::host::TensorEquals(reference_Aux.host_view(), tensor_Aux.host_view());
|
||||
passed &= cutlass::reference::host::TensorEquals(abs_max_Aux.host_view(), reference_abs_max_Aux.host_view());
|
||||
}
|
||||
|
||||
if (kScaleOutput) {
|
||||
abs_max_D.sync_host();
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(abs_max_D.host_view()), 0);
|
||||
passed &= cutlass::reference::host::TensorEquals(abs_max_D.host_view(), reference_abs_max_D.host_view());
|
||||
}
|
||||
|
||||
EXPECT_TRUE(passed) << " mismatched reference";
|
||||
|
||||
if (!passed) {
|
||||
|
||||
std::ofstream file0("conv_testbed_with_amax_errors_reference.txt");
|
||||
std::ofstream file1("conv_testbed_with_amax_errors_computed.txt");
|
||||
|
||||
std::ofstream file("conv_testbed_with_amax_errors.txt");
|
||||
|
||||
file
|
||||
<< "problem: " << problem_size
|
||||
<< ", alpha: " << alpha << ", beta: " << beta << "\n\n";
|
||||
|
||||
file
|
||||
<< "A =\n" << tensor_A.host_view()
|
||||
<< "\nB =\n" << tensor_B.host_view()
|
||||
<< "\nC =\n" << tensor_C.host_view()
|
||||
<< "\nVector =\n" << tensor_Vector.host_view()
|
||||
<< "\nScaleA = " << scale_A.host_view()
|
||||
<< "\nScaleB = " << scale_B.host_view()
|
||||
<< "\nScaleC = " << scale_C.host_view()
|
||||
<< "\nScaleD = " << scale_D.host_view()
|
||||
<< "\nScaleAux = " << scale_Aux.host_view()
|
||||
<< std::endl;
|
||||
|
||||
file0 << "\n\nReference D =\n" << reference_D.host_view() << std::endl;
|
||||
file1 << "\n\nComputed D =\n" << tensor_D.host_view() << std::endl;
|
||||
if (kScaleAux) {
|
||||
file0 << "\n\nReference Aux =\n" << reference_Aux.host_view() << std::endl;
|
||||
file1 << "\n\nComputed Aux =\n" << tensor_Aux.host_view() << std::endl;
|
||||
file0 << "\n\nReference Absmax Aux = " << reference_abs_max_Aux.host_view() << std::endl;
|
||||
file1 << "\n\nComputed Absmax Aux = " << abs_max_Aux.host_view() << std::endl;
|
||||
}
|
||||
if (kScaleOutput) {
|
||||
file0 << "\n\nReference Absmax D = " << reference_abs_max_D.host_view() << std::endl;
|
||||
file1 << "\n\nComputed Absmax D = " << abs_max_D.host_view() << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
/// Verifies the result is a GEMM
|
||||
bool verify(
|
||||
cutlass::conv::Conv2dProblemSize const &problem_size,
|
||||
ElementCompute alpha,
|
||||
ElementCompute beta) {
|
||||
|
||||
cutlass::Coord<4> origin(0);
|
||||
ElementCompute scaled_alpha = alpha;
|
||||
if (doScaleA) {
|
||||
scaled_alpha *= scale_A.host_view().at(origin);
|
||||
}
|
||||
if (doScaleB) {
|
||||
scaled_alpha *= scale_B.host_view().at(origin);
|
||||
}
|
||||
|
||||
ElementCompute scaled_beta = beta;
|
||||
if (doScaleC) {
|
||||
scaled_beta *= scale_C.host_view().at(origin);
|
||||
}
|
||||
|
||||
//
|
||||
// Verify
|
||||
//
|
||||
|
||||
cutlass::reference::host::Conv2d<
|
||||
typename Conv::ElementA, typename Conv::LayoutA,
|
||||
typename Conv::ElementB, typename Conv::LayoutB,
|
||||
typename Conv::ElementC, typename Conv::LayoutC,
|
||||
ElementCompute, ElementAccumulator, ElementAccumulator
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.host_ref(),
|
||||
tensor_B.host_ref(),
|
||||
tensor_C.host_ref(),
|
||||
tmp_D.host_ref(),
|
||||
scaled_alpha,
|
||||
scaled_beta
|
||||
);
|
||||
|
||||
ElementCompute tmp_abs_max_Aux(0.);
|
||||
ElementCompute tmp_abs_max_D(0.);
|
||||
|
||||
cutlass::NumericConverter<ElementCompute, typename Conv::ElementC> cvt_c_to_compute;
|
||||
cutlass::NumericConverter<ElementCompute, ElementAccumulator> cvt_accum_to_compute;
|
||||
cutlass::NumericConverter<ElementAbsmax, ElementCompute> cvt_compute_to_absmax;
|
||||
cutlass::NumericConverter<typename Conv::EpilogueOutputOp::ElementOutput, ElementCompute> cvt_compute_to_d;
|
||||
cutlass::NumericConverter<typename Conv::EpilogueOutputOp::ElementAuxOutput, ElementCompute> cvt_compute_to_aux;
|
||||
|
||||
cutlass::absolute_value_op<ElementCompute> abs;
|
||||
cutlass::maximum_with_nan_propogation<ElementCompute> max;
|
||||
ActivationFunctor<ElementCompute> act;
|
||||
|
||||
ElementScalingFactor d_scale = kScaleOutput ? scale_D.host_view().at(origin) : ElementScalingFactor(1.);
|
||||
|
||||
for (int n = 0; n < problem_size.N; ++n) {
|
||||
for (int p = 0; p < problem_size.P; ++p) {
|
||||
for (int q = 0; q < problem_size.Q; ++q) {
|
||||
for (int k = 0; k < problem_size.K; ++k) {
|
||||
ElementCompute intermediate = cvt_accum_to_compute(tmp_D.host_view().at({n, p, q, k}));
|
||||
ElementCompute bias = cvt_c_to_compute(tensor_Vector.host_view().at({0, 0, 0, k}));
|
||||
ElementCompute aux = intermediate + bias;
|
||||
ElementCompute d = act(aux);
|
||||
tmp_abs_max_Aux = max(abs(aux), tmp_abs_max_Aux);
|
||||
tmp_abs_max_D = max(abs(d), tmp_abs_max_D);
|
||||
reference_D.host_view().at({n, p, q, k}) = cvt_compute_to_d(d * d_scale);
|
||||
|
||||
if (kScaleAux) {
|
||||
reference_Aux.host_view().at({n, p, q, k}) = cvt_compute_to_aux(aux * scale_Aux.host_view().at(origin));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
if (kScaleAux) {
|
||||
reference_abs_max_Aux.host_view().at(origin) = cvt_compute_to_absmax(tmp_abs_max_Aux);
|
||||
}
|
||||
|
||||
if (kScaleOutput) {
|
||||
reference_abs_max_D.host_view().at(origin) = cvt_compute_to_absmax(tmp_abs_max_D);
|
||||
}
|
||||
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
/// Returns true if the CUDA device is sufficient to execute the kernel.
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
size_t smem_size = sizeof(typename Conv::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
cudaError_t result = cudaGetDevice(&device_idx);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaGetDevice() API call failed.");
|
||||
}
|
||||
|
||||
result = cudaGetDeviceProperties(&properties, device_idx);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaGetDeviceProperties() failed");
|
||||
}
|
||||
|
||||
if (properties.sharedMemPerBlockOptin < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::conv::Conv2dProblemSize const &problem_size,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0))
|
||||
{
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
if (CUTLASS_TEST_UNIT_ENABLE_WARNINGS) {
|
||||
std::cerr << "Test waived due to insufficient CUDA device." << std::endl;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
this->initialize(problem_size);
|
||||
|
||||
//
|
||||
// Initialize the GEMM operator
|
||||
//
|
||||
|
||||
typename Conv::EpilogueOutputOp::Params::ActivationParams activation_params{alpha, beta};
|
||||
typename Conv::EpilogueOutputOp::Params epilogue_params{
|
||||
activation_params,
|
||||
scale_A.device_data(),
|
||||
scale_B.device_data(),
|
||||
scale_C.device_data(),
|
||||
scale_D.device_data(),
|
||||
scale_Aux.device_data(),
|
||||
abs_max_Aux.device_data(),
|
||||
abs_max_D.device_data()
|
||||
};
|
||||
|
||||
typename Conv::Arguments arguments{
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D.device_ref(),
|
||||
tensor_Aux.device_ref(),
|
||||
epilogue_params,
|
||||
cutlass::conv::SplitKMode::kSerial,
|
||||
tensor_Vector.device_data(),
|
||||
0
|
||||
};
|
||||
|
||||
Conv conv2d_op;
|
||||
|
||||
cutlass::Status status = conv2d_op.can_implement(arguments);
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
size_t workspace_size = Conv::get_workspace_size(arguments);
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
status = conv2d_op.initialize(arguments, workspace.get());
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
//
|
||||
// Run the GEMM
|
||||
//
|
||||
|
||||
status = conv2d_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
|
||||
cudaError_t cuda_error = cudaDeviceSynchronize();
|
||||
EXPECT_TRUE(cuda_error == cudaSuccess) << cudaGetErrorString(cuda_error);
|
||||
|
||||
//
|
||||
// Verify
|
||||
//
|
||||
|
||||
bool passed = this->verify(problem_size, alpha, beta);
|
||||
|
||||
if (!passed) {
|
||||
std::cout << "Failed" << std::endl;
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
typename ImplicitGemm,
|
||||
template<typename T> class ActivationFunctor = cutlass::epilogue::thread::Identity
|
||||
>
|
||||
bool TestAllConv2dWithAbsmax(bool scaleA=true, bool scaleB=true, bool scaleC=true) {
|
||||
const Conv2dProblemVector &conv_test_sizes = Conv2dProblemVector();
|
||||
const Conv2dProblemVector &conv_blacklist_sizes = Conv2dProblemVector();
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv2dWithAbsMax<ImplicitGemm, ActivationFunctor> testbed(scaleA, scaleB, scaleC);
|
||||
|
||||
//
|
||||
// Get conv problem sizes to run conv operator
|
||||
//
|
||||
TestbedConv2dProblemSizes conv_problems(128/cutlass::sizeof_bits<typename ImplicitGemm::ElementA>::value);
|
||||
|
||||
// Vector of conv2d problem sizes to avoid duplicate runs
|
||||
Conv2dProblemVector conv_tested_sizes;
|
||||
|
||||
Conv2dProblemVector const *problem_vectors[] = {
|
||||
&conv_test_sizes, // run user specified sizes
|
||||
&conv_problems.conv2d_default_sizes, // run default and cudnn bug sizes
|
||||
&conv_problems.conv2d_resnet50_sizes, // run resnet50 sizes
|
||||
#if CUTLASS_CONV_UNIT_TEST_RIGOROUS_SIZE_ENABLED
|
||||
&conv_problems.conv2d_rigorous_sizes, // run large and rigorous sizes if enabled
|
||||
#endif
|
||||
};
|
||||
|
||||
bool passed = true;
|
||||
|
||||
// Sweep conv2d problem sizes (split-k-mode=kSerial, split-k-slice=1, alpha=1.0, beta=0.0)
|
||||
for (Conv2dProblemVector const * problem_vector : problem_vectors) {
|
||||
|
||||
// Prune all problems with channels that aren't divisible by the number of elements accessed per
|
||||
// load for operands A and B. This is meant to align with the requirements of iterators used for
|
||||
// fprop kernels.
|
||||
ChannelDivisibilitySpecification channel_spec(128 / cutlass::sizeof_bits<typename ImplicitGemm::ElementA>::value);
|
||||
auto pruned_problem_vector = prune(*problem_vector, channel_spec);
|
||||
|
||||
// Run conv testbed on default convolution sizes
|
||||
for(auto conv_problem : pruned_problem_vector) {
|
||||
|
||||
// Skip blacklist and avoid duplicate problem sizes
|
||||
if (std::find(conv_blacklist_sizes.begin(), conv_blacklist_sizes.end(), conv_problem) != conv_blacklist_sizes.end() ||
|
||||
std::find(conv_tested_sizes.begin(), conv_tested_sizes.end(), conv_problem) != conv_tested_sizes.end()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
//
|
||||
// Test
|
||||
//
|
||||
// push back tested problem size to avoid re-running duplicates
|
||||
conv_tested_sizes.push_back(conv_problem);
|
||||
|
||||
// test mode = xcross
|
||||
passed &= testbed.run(conv_problem);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// test mode = convolution
|
||||
passed &= testbed.run(conv_problem.reset_mode(cutlass::conv::Mode::kConvolution));
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -254,7 +254,7 @@ public:
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv2d::UnderlyingKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Conv2d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
@@ -408,11 +408,11 @@ public:
|
||||
for (int q = 0; q < problem_size.Q; ++q) {
|
||||
for (int k = 0; k < problem_size.K; ++k) {
|
||||
|
||||
ElementZ z;
|
||||
ElementT t;
|
||||
ElementZ z{};
|
||||
ElementT t{};
|
||||
|
||||
ElementCompute accum = tensor_Y_reference.at({n, p, q, k});
|
||||
ElementCompute bias = ElementCompute(tensor_Broadcast.at({0, 0, 0, k}));
|
||||
ElementCompute bias = ElementCompute(tensor_Broadcast.at({0, 0, 0, k}));
|
||||
|
||||
|
||||
if (kAddBroadcastFirst) {
|
||||
@@ -499,6 +499,52 @@ public:
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename ImplicitGemm,
|
||||
typename ReferenceOp = Conv2dWithBroadcastReferenceOp<ImplicitGemm>,
|
||||
bool AddBroadcastFirst = false>
|
||||
bool TestSpecificConv2dWithBroadcast(
|
||||
const Conv2dProblemVector & problem_sizes) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv2dWithBroadcast<ImplicitGemm, ReferenceOp, AddBroadcastFirst> testbed;
|
||||
|
||||
// Sweep conv2d problem sizes (split-k-mode=kSerial, split-k-slice=1, alpha=1.0, beta=0.0)
|
||||
for(auto conv_problem : problem_sizes) {
|
||||
|
||||
//
|
||||
// Test
|
||||
//
|
||||
|
||||
// test mode = xcross
|
||||
passed = testbed.run(
|
||||
conv_problem,
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// test mode = convolution
|
||||
passed = testbed.run(
|
||||
conv_problem.reset_mode(cutlass::conv::Mode::kConvolution),
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// TestAllConv: Runs cutlass::conv::device::ImplicitGemmConvolution operator and compares it with reference
|
||||
// TestAllConv runs conv operator on default conv problem sizes from test::conv::device::TestbedConv2dProblemSizes
|
||||
|
||||
@@ -182,7 +182,7 @@ public:
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv2d::UnderlyingKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Conv2d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
|
||||
+142
@@ -0,0 +1,142 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide Implicit GEMM interface
|
||||
*/
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/conv/kernel/default_conv3d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv3d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Dgrad_Analytic_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dDgradKernel = typename cutlass::conv::kernel::DefaultConv3dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic,
|
||||
cutlass::conv::StrideSupport::kStrided
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv3dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Dgrad_Optimized_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dDgradKernel = typename cutlass::conv::kernel::DefaultConv3dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv3dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+137
@@ -0,0 +1,137 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide Implicit GEMM interface
|
||||
*/
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
|
||||
#include "cutlass/conv/kernel/default_conv3d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv3d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Fprop_Analytic_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dFpropKernel = typename cutlass::conv::kernel::DefaultConv3dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv3dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Fprop_Optimized_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dFpropKernel = typename cutlass::conv::kernel::DefaultConv3dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv3dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
@@ -0,0 +1,171 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide Implicit GEMM interface
|
||||
*/
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/array.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_bias_elementwise.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_residual_block.h"
|
||||
#include "cutlass/epilogue/thread/activation.h"
|
||||
#include "cutlass/conv/kernel/default_conv3d_fprop_with_broadcast.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv3d_with_broadcast_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
|
||||
TEST(SM80_Device_Conv3d_Fprop_With_Broadcast_Analytic_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementCompute = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationBiasElementwise<
|
||||
ElementC,
|
||||
ElementAccumulator,
|
||||
ElementCompute,
|
||||
ElementC,
|
||||
ElementC,
|
||||
1,
|
||||
cutlass::epilogue::thread::ReLu<float>
|
||||
>;
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dFpropKernel = typename cutlass::conv::kernel::DefaultConv3dFpropWithBroadcast<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv3dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3dWithBroadcast<Conv3dFprop>());
|
||||
}
|
||||
|
||||
// Test residual block fusion: UnaryOp(BinaryOp(ActivationOp(Conv3d(X) + bias), residual))
|
||||
// LinearCombinationResidualBlock does not support the split-k mode unless ActivationOp is Identity.
|
||||
// This is because the activation needs to be applied to the fully accumulated output of the Conv3d op,
|
||||
// which only the last thread block would have an access to, before applying BinaryOp.
|
||||
// The epilogue functor in the last thread block would have to be given three inputs, namely
|
||||
// partial outputs, bias, and residual, but this is not supported in the current interface.
|
||||
// Set TestSplitK = false to skip split-k tests with non-trivial ActivationOp.
|
||||
template <
|
||||
template<typename T> class ActivationOp,
|
||||
template<typename T> class BinaryOp,
|
||||
template<typename T> class UnaryOp,
|
||||
bool TestSplitK = true
|
||||
>
|
||||
void TestResidaulBlock() {
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementD = ElementC;
|
||||
using ElementCompute = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationResidualBlock<
|
||||
ElementD,
|
||||
ElementAccumulator,
|
||||
ElementCompute,
|
||||
ElementC,
|
||||
1,
|
||||
ActivationOp,
|
||||
BinaryOp,
|
||||
UnaryOp
|
||||
>;
|
||||
|
||||
using Conv3dFpropKernel = typename cutlass::conv::kernel::DefaultConv3dFpropWithBroadcast<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv3dFpropKernel>;
|
||||
|
||||
struct ReferenceOp {
|
||||
using OutputOp = typename Conv3dFprop::EpilogueOutputOp;
|
||||
using ElementZ = typename OutputOp::ElementZ;
|
||||
|
||||
ActivationOp<ElementCompute> activation;
|
||||
BinaryOp<ElementCompute> binary_op;
|
||||
UnaryOp<ElementCompute> unary_op;
|
||||
|
||||
void operator()(ElementZ &Z, ElementZ&, ElementCompute conv3d, ElementCompute residual) {
|
||||
Z = ElementZ(unary_op(binary_op(activation(conv3d), residual)));
|
||||
}
|
||||
};
|
||||
|
||||
bool passed = test::conv::device::TestAllConv3dWithBroadcast<Conv3dFprop, ReferenceOp, true, TestSplitK>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Conv3d_Fprop_With_Residual_Block_Plus_Analytic_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
// Resnet
|
||||
TestResidaulBlock<cutlass::epilogue::thread::Identity, cutlass::plus, cutlass::epilogue::thread::ReLu>();
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -117,6 +117,17 @@ struct TestbedConv3dProblemSizes {
|
||||
{8, 1, 1, 3, minimum_channel_size}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({1, 1, 1}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({1, 1, 1}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 1, 1, 8, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 1, 1, 3, minimum_channel_size}, // filter size (KTRSC)
|
||||
CUTLASS_STL_NAMESPACE::make_tuple(
|
||||
cutlass::Coord<3>({1, 1, 1}), // near padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({0, 0, 0}) // far padding (pad_d, pad_h, pad_w)
|
||||
),
|
||||
cutlass::Coord<3>({1, 1, 1}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
@@ -128,6 +139,17 @@ struct TestbedConv3dProblemSizes {
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 8, 8, 8, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 3, 3, 3, minimum_channel_size}, // filter size (KTRSC)
|
||||
CUTLASS_STL_NAMESPACE::make_tuple(
|
||||
cutlass::Coord<3>({1, 1, 1}), // near padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({0, 0, 0}) // far padding (pad_d, pad_h, pad_w)
|
||||
),
|
||||
cutlass::Coord<3>({1, 1, 1}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 16, 16, 16, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 3, 3, 3, minimum_channel_size}, // filter size (KTRSC)
|
||||
|
||||
@@ -184,7 +184,7 @@ public:
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv3d::UnderlyingKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Conv3d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
@@ -662,6 +662,47 @@ bool TestAllConv3d(
|
||||
return passed;
|
||||
}
|
||||
|
||||
template <typename ImplicitGemm>
|
||||
bool TestSpecificConv3d(
|
||||
const Conv3dProblemVector & problem_sizes) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv3d<ImplicitGemm> testbed;
|
||||
|
||||
// Sweep conv3d problem sizes (split-k-mode=kSerial, split-k-slice=1, alpha=1.0, beta=0.0)
|
||||
for(auto conv_problem : problem_sizes) {
|
||||
|
||||
//
|
||||
// Test
|
||||
//
|
||||
|
||||
// test mode = xcross
|
||||
passed = testbed.run(
|
||||
conv_problem,
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// test mode = convolution
|
||||
passed = testbed.run(
|
||||
conv_problem.reset_mode(cutlass::conv::Mode::kConvolution),
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
|
||||
+136
@@ -0,0 +1,136 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide Implicit GEMM interface
|
||||
*/
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/conv/kernel/default_conv3d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv3d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Wgrad_Analytic_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv3dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv3d_Wgrad_Optimized_ImplicitGemm_f32ndhwc_f32ndhwc_f32ndhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = float;
|
||||
using ElementB = float;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
|
||||
/// Device-level Conv3d instance
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv3dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv3dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv3d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv3d<Conv3dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
@@ -0,0 +1,716 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Implicit GEMM for fused epilogue broadcast testbed
|
||||
|
||||
Parallel split-k is not tested because we can just use regular conv kernel
|
||||
when we need to use parallel-splitk. Broadcast can happen in the reduction
|
||||
kernel.
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#include <fstream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
#include "cutlass/reduction/device/reduce_split_k.h"
|
||||
#include "cutlass/reduction/thread/reduction_operators.h"
|
||||
|
||||
#include "conv3d_problems.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/device/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
|
||||
#include "cutlass/util/reference/host/convolution.h"
|
||||
#include "cutlass/util/reference/device/convolution.h"
|
||||
|
||||
#include "cutlass/core_io.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "../cache_testbed_output.h"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename Conv3d>
|
||||
struct Conv3dWithBroadcastReferenceOp {
|
||||
|
||||
using OutputOp = typename Conv3d::EpilogueOutputOp;
|
||||
|
||||
using ElementCompute = typename OutputOp::ElementCompute;
|
||||
using ElementZ = typename OutputOp::ElementZ;
|
||||
using ElementT = typename OutputOp::ElementT;
|
||||
|
||||
typename OutputOp::BinaryOp binary_op;
|
||||
typename OutputOp::ElementwiseOp elementwise_op;
|
||||
|
||||
Conv3dWithBroadcastReferenceOp() { }
|
||||
|
||||
void operator()(ElementZ &Z, ElementT &T, ElementCompute conv3d, ElementCompute bias) {
|
||||
ElementCompute t_full = binary_op(conv3d, bias);
|
||||
T = ElementT(t_full);
|
||||
|
||||
ElementCompute z_full = elementwise_op(t_full);
|
||||
Z = ElementZ(z_full);
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// Fused testbed
|
||||
//
|
||||
// Y = CONV(AB, C)
|
||||
//
|
||||
// T[n, o, p, q, k] = ReductionOp(Y[n, o, p, q, k], Broadcast[k])
|
||||
//
|
||||
// Z[n, o, p, q, k] = Elementwise(T[n, o, p, q, k])
|
||||
//
|
||||
|
||||
template <
|
||||
typename Conv3d,
|
||||
typename ReferenceOp,
|
||||
bool AddBroadcastFirst = false
|
||||
>
|
||||
class TestbedConv3dWithBroadcast {
|
||||
public:
|
||||
|
||||
using ElementA = typename Conv3d::ElementA;
|
||||
using LayoutA = typename Conv3d::LayoutA;
|
||||
using ElementB = typename Conv3d::ElementB;
|
||||
using LayoutB = typename Conv3d::LayoutB;
|
||||
using ElementC = typename Conv3d::ElementC;
|
||||
using LayoutC = typename Conv3d::LayoutC;
|
||||
using ElementAccumulator = typename Conv3d::ElementAccumulator;
|
||||
using ElementCompute = typename Conv3d::ElementCompute;
|
||||
using EpilogueOutputOp = typename Conv3d::EpilogueOutputOp;
|
||||
using ElementZ = typename EpilogueOutputOp::ElementZ;
|
||||
using ElementT = typename EpilogueOutputOp::ElementT;
|
||||
using ElementVector = typename EpilogueOutputOp::ElementVector;
|
||||
|
||||
static cutlass::conv::Operator const kConvolutionalOperator = Conv3d::kConvolutionalOperator;
|
||||
static const bool kAddBroadcastFirst = AddBroadcastFirst;
|
||||
static const bool kStoreT = EpilogueOutputOp::kStoreT;
|
||||
|
||||
public:
|
||||
|
||||
/// Initialization
|
||||
cutlass::Distribution::Kind init_A;
|
||||
cutlass::Distribution::Kind init_B;
|
||||
cutlass::Distribution::Kind init_C;
|
||||
uint64_t seed;
|
||||
|
||||
cutlass::HostTensor<ElementA, LayoutA> tensor_A;
|
||||
cutlass::HostTensor<ElementB, LayoutB> tensor_B;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_C;
|
||||
cutlass::HostTensor<ElementAccumulator, LayoutC> tensor_C_reference;
|
||||
cutlass::HostTensor<ElementZ, LayoutC> tensor_Z_computed;
|
||||
cutlass::HostTensor<ElementZ, LayoutC> tensor_Z_reference;
|
||||
cutlass::HostTensor<ElementT, LayoutC> tensor_T_computed;
|
||||
cutlass::HostTensor<ElementT, LayoutC> tensor_T_reference;
|
||||
cutlass::HostTensor<ElementAccumulator, LayoutC> tensor_Y_reference;
|
||||
cutlass::HostTensor<ElementVector, LayoutC> tensor_Broadcast; // Input Broadcast
|
||||
|
||||
public:
|
||||
|
||||
TestbedConv3dWithBroadcast(
|
||||
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = 2080
|
||||
):
|
||||
init_A(init_A_), init_B(init_B_), init_C(init_C_), seed(seed_) {
|
||||
|
||||
}
|
||||
|
||||
/// Helper to initialize a tensor view
|
||||
template <typename Element, typename Layout>
|
||||
void initialize_tensor(
|
||||
cutlass::TensorView<Element, Layout> view,
|
||||
cutlass::Distribution::Kind dist_kind,
|
||||
uint64_t seed) {
|
||||
|
||||
if (dist_kind == cutlass::Distribution::Uniform) {
|
||||
|
||||
int scope;
|
||||
int bits = cutlass::sizeof_bits<Element>::value;
|
||||
|
||||
if (bits <= 8) {
|
||||
scope = 2;
|
||||
}
|
||||
else if (bits == 16) {
|
||||
if (cutlass::sizeof_bits<ElementAccumulator>::value <= 16) {
|
||||
scope = 3;
|
||||
}
|
||||
else {
|
||||
scope = 5;
|
||||
}
|
||||
}
|
||||
else {
|
||||
scope = 8;
|
||||
}
|
||||
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
view, seed, scope, -scope, 0);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Identity) {
|
||||
|
||||
cutlass::reference::host::TensorFillIdentity(view);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Gaussian) {
|
||||
|
||||
cutlass::reference::host::TensorFillRandomGaussian(view, seed, 0, 0.5);
|
||||
}
|
||||
else if (dist_kind == cutlass::Distribution::Sequential) {
|
||||
|
||||
cutlass::reference::host::BlockFillSequential(view.data(), view.capacity());
|
||||
}
|
||||
else {
|
||||
}
|
||||
}
|
||||
|
||||
void initialize(
|
||||
cutlass::conv::Conv3dProblemSize const &problem_size, uint64_t seed = 2019) {
|
||||
|
||||
tensor_A.resize(implicit_gemm_tensor_a_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_B.resize(implicit_gemm_tensor_b_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_C.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_C_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_Z_computed.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_Z_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_T_computed.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_T_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_Y_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_Broadcast.resize({
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size).c(),
|
||||
});
|
||||
|
||||
initialize_tensor(tensor_A.host_view(), init_A, seed);
|
||||
initialize_tensor(tensor_B.host_view(), init_B, seed * 17);
|
||||
initialize_tensor(tensor_C.host_view(), init_C, seed * 39);
|
||||
initialize_tensor(tensor_Broadcast.host_view(), init_C, seed * 39);
|
||||
|
||||
for (int n = 0; n < tensor_C_reference.extent().n(); ++n) {
|
||||
for (int o = 0; o < tensor_C_reference.extent().d(); ++o) {
|
||||
for (int p = 0; p < tensor_C_reference.extent().h(); ++p) {
|
||||
for (int q = 0; q < tensor_C_reference.extent().w(); ++q) {
|
||||
for (int k = 0; k < tensor_C_reference.extent().c(); ++k) {
|
||||
tensor_C_reference.at({n, o, p, q, k}) = ElementAccumulator(tensor_C.at({n, o, p, q, k}));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_Broadcast.sync_device();
|
||||
tensor_C_reference.sync_device();
|
||||
tensor_Z_computed.sync_device();
|
||||
tensor_Z_reference.sync_device();
|
||||
tensor_T_computed.sync_device();
|
||||
tensor_T_reference.sync_device();
|
||||
tensor_Y_reference.sync_device();
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
size_t smem_size = sizeof(typename Conv3d::UnderlyingKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
cudaError_t result = cudaGetDevice(&device_idx);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaGetDevice() API call failed.");
|
||||
}
|
||||
|
||||
result = cudaGetDeviceProperties(&properties, device_idx);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaGetDeviceProperties() failed");
|
||||
}
|
||||
|
||||
if (properties.sharedMemPerBlockOptin < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::conv::Conv3dProblemSize const &problem_size,
|
||||
cutlass::conv::SplitKMode const &split_k_mode = cutlass::conv::SplitKMode::kSerial,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(1)) {
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
if (CUTLASS_TEST_UNIT_ENABLE_WARNINGS) {
|
||||
std::cerr << "Test waived due to insufficient CUDA device." << std::endl;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
#if 0 //display conv3d problem size for debugging
|
||||
std::cout << problem_size << std::endl
|
||||
<< "alpha, beta: (" << alpha << ", " << beta << ")" << std::endl
|
||||
<< "split_k_mode: " << ((split_k_mode == cutlass::conv::SplitKMode::kSerial) ? "(serial)" : "(parallel)") << std::endl
|
||||
<< std::endl;
|
||||
#endif
|
||||
|
||||
initialize(problem_size);
|
||||
|
||||
// configure the operator
|
||||
Conv3d conv3d_op;
|
||||
typename Conv3d::Arguments conv3d_args(
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_Z_computed.device_ref(),
|
||||
{alpha, beta},
|
||||
split_k_mode,
|
||||
tensor_Broadcast.device_data(),
|
||||
kStoreT ? tensor_T_computed.device_data() : nullptr,
|
||||
0, // This must be zero
|
||||
implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size).c()
|
||||
);
|
||||
|
||||
// initialize the kernel
|
||||
size_t workspace_size = Conv3d::get_workspace_size(conv3d_args);
|
||||
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
cutlass::Status status = conv3d_op.initialize(conv3d_args, workspace.get());
|
||||
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
std::cerr << "This test is not supported: " << cudaGetErrorString(error) << "\n";
|
||||
return true;
|
||||
}
|
||||
|
||||
// run conv3d operator
|
||||
status = conv3d_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
bool passed = false;
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
EXPECT_EQ(result, cudaSuccess) << " device reference error: "
|
||||
<< cudaGetErrorString(result);
|
||||
|
||||
tensor_T_computed.sync_host();
|
||||
tensor_Z_computed.sync_host();
|
||||
|
||||
//
|
||||
// Reference check
|
||||
//
|
||||
|
||||
// When kAddBroadcastFirst is true, add bias on the host
|
||||
ElementCompute beta_ref = kAddBroadcastFirst ? ElementCompute(0) : beta;
|
||||
|
||||
#if CUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED
|
||||
|
||||
cutlass::reference::device::Conv3d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementAccumulator,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C_reference.device_ref(),
|
||||
tensor_Y_reference.device_ref(),
|
||||
alpha,
|
||||
beta_ref);
|
||||
|
||||
// sync host (copy device data to host) for dumping error output in case of mismatches
|
||||
tensor_Y_reference.sync_host();
|
||||
|
||||
#else
|
||||
|
||||
cutlass::reference::host::Conv3d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementAccumulator,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.host_ref(),
|
||||
tensor_B.host_ref(),
|
||||
tensor_C_reference.host_ref(),
|
||||
tensor_Y_reference.host_ref(),
|
||||
alpha,
|
||||
beta_ref);
|
||||
|
||||
#endif
|
||||
ReferenceOp reference_op;
|
||||
|
||||
// compute tensor Z and tensor T
|
||||
for (int n = 0; n < problem_size.N; ++n) {
|
||||
for (int o = 0; o < problem_size.Z; ++o) {
|
||||
for (int p = 0; p < problem_size.P; ++p) {
|
||||
for (int q = 0; q < problem_size.Q; ++q) {
|
||||
for (int k = 0; k < problem_size.K; ++k) {
|
||||
|
||||
ElementZ z{};
|
||||
ElementT t{};
|
||||
|
||||
ElementCompute accum = tensor_Y_reference.at({n, o, p, q, k});
|
||||
ElementCompute bias = ElementCompute(tensor_Broadcast.at({0, 0, 0, 0, k}));
|
||||
|
||||
|
||||
if (kAddBroadcastFirst) {
|
||||
reference_op(z, t, accum + bias,
|
||||
beta * ElementCompute(tensor_C_reference.at({n, o, p, q, k})));
|
||||
} else {
|
||||
reference_op(z, t, accum, bias);
|
||||
}
|
||||
|
||||
tensor_Z_reference.at({n, o, p, q, k}) = z;
|
||||
tensor_T_reference.at({n, o, p, q, k}) = t;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (kStoreT) {
|
||||
passed = cutlass::reference::host::TensorEquals(
|
||||
tensor_T_computed.host_view(),
|
||||
tensor_T_reference.host_view());
|
||||
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
passed = cutlass::reference::host::TensorEquals(
|
||||
tensor_Z_computed.host_view(),
|
||||
tensor_Z_reference.host_view());
|
||||
|
||||
EXPECT_TRUE(passed);
|
||||
|
||||
if (!passed) {
|
||||
std::stringstream fname;
|
||||
|
||||
fname << "error_Conv3d_ImplicitGemm_device_"
|
||||
<< (split_k_mode == cutlass::conv::SplitKMode::kSerial ? "serial_reduction_" : "parallel_reduction_")
|
||||
<< (Conv3d::kConvolutionalOperator == cutlass::conv::Operator::kFprop ? "fprop_" :
|
||||
(Conv3d::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ? "dgrad_" : "wgrad_"))
|
||||
<< "nnhwc_"
|
||||
<< problem_size.N << "x"
|
||||
<< problem_size.D << "x"
|
||||
<< problem_size.H << "x"
|
||||
<< problem_size.W << "x"
|
||||
<< problem_size.C
|
||||
<< "_krsc_"
|
||||
<< problem_size.K << "x"
|
||||
<< problem_size.T << "x"
|
||||
<< problem_size.R << "x"
|
||||
<< problem_size.S << "x"
|
||||
<< problem_size.C
|
||||
<< "_padding_"
|
||||
<< problem_size.pad_d << "x"
|
||||
<< problem_size.pad_h << "x"
|
||||
<< problem_size.pad_w
|
||||
<< "_stride_"
|
||||
<< problem_size.stride_d << "x"
|
||||
<< problem_size.stride_h << "x"
|
||||
<< problem_size.stride_w
|
||||
<< "_dilation_"
|
||||
<< problem_size.dilation_d << "x"
|
||||
<< problem_size.dilation_h << "x"
|
||||
<< problem_size.dilation_w << "_"
|
||||
<< (problem_size.mode == cutlass::conv::Mode::kCrossCorrelation ? "xcorr_" : "conv_")
|
||||
<< Conv3d::ThreadblockShape::kM << "x"
|
||||
<< Conv3d::ThreadblockShape::kN << "x"
|
||||
<< Conv3d::ThreadblockShape::kK << "_"
|
||||
<< Conv3d::WarpShape::kM << "x"
|
||||
<< Conv3d::WarpShape::kN << "x"
|
||||
<< Conv3d::WarpShape::kK << ".txt";
|
||||
|
||||
std::cout << fname.str() << std::endl;
|
||||
|
||||
std::ofstream results(fname.str());
|
||||
|
||||
results << problem_size << std::endl;
|
||||
|
||||
results
|
||||
<< "\nA:\n" << tensor_A.host_view() << "\n"
|
||||
<< "\nB:\n" << tensor_B.host_view() << "\n"
|
||||
<< "\nC:\n" << tensor_C.host_view() << "\n"
|
||||
<< "\nBroadcast:\n" << tensor_Broadcast.host_view() << "\n"
|
||||
<< "\nY reference:\n" << tensor_Y_reference.host_view() << "\n"
|
||||
<< "\nT reference:\n" << tensor_T_reference.host_view() << "\n"
|
||||
<< "\nT computed:\n" << tensor_T_computed.host_view() << "\n"
|
||||
<< "\nZ reference:\n" << tensor_Z_reference.host_view() << "\n"
|
||||
<< "\nZ computed:\n" << tensor_Z_computed.host_view() << "\n";
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// TestAllConv: Runs cutlass::conv::device::ImplicitGemmConvolution operator and compares it with reference
|
||||
// TestAllConv runs conv operator on default conv problem sizes from test::conv::device::TestbedConv3dProblemSizes
|
||||
// Additionally, each conv3d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
|
||||
// (conv_blacklist_sizes)
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
template <typename ImplicitGemm,
|
||||
typename ReferenceOp = Conv3dWithBroadcastReferenceOp<ImplicitGemm>,
|
||||
bool AddBroadcastFirst = false,
|
||||
bool TestSplitK = true
|
||||
>
|
||||
bool TestAllConv3dWithBroadcast(
|
||||
const Conv3dProblemVector &conv_test_sizes = Conv3dProblemVector(),
|
||||
const Conv3dProblemVector &conv_blacklist_sizes = Conv3dProblemVector()) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv3dWithBroadcast<ImplicitGemm, ReferenceOp, AddBroadcastFirst> testbed;
|
||||
|
||||
//
|
||||
// Get conv problem sizes to run conv operator
|
||||
//
|
||||
TestbedConv3dProblemSizes conv3d_problems(128/cutlass::sizeof_bits<typename ImplicitGemm::ElementA>::value);
|
||||
|
||||
// Vector of conv3d problem sizes to avoid duplicate runs
|
||||
Conv3dProblemVector conv_tested_sizes;
|
||||
|
||||
Conv3dProblemVector const *problem_vectors[] = {
|
||||
&conv3d_problems.conv3d_default_sizes,
|
||||
&conv3d_problems.conv3d_vnet_medical_sizes,
|
||||
&conv_test_sizes
|
||||
};
|
||||
|
||||
// Sweep conv3d problem sizes (split-k-mode=kSerial, split-k-slice=1, alpha=1.0, beta=0.0)
|
||||
for (Conv3dProblemVector const * problem_vector : problem_vectors) {
|
||||
|
||||
// Run conv testbed on default convolution sizes
|
||||
for(auto conv_problem : *problem_vector) {
|
||||
|
||||
// Skip blacklist and avoid duplicate problem sizes
|
||||
if (std::find(conv_blacklist_sizes.begin(), conv_blacklist_sizes.end(), conv_problem) != conv_blacklist_sizes.end() ||
|
||||
std::find(conv_tested_sizes.begin(), conv_tested_sizes.end(), conv_problem) != conv_tested_sizes.end()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
//
|
||||
// Procedurally disable certain cases
|
||||
//
|
||||
|
||||
// CUTLASS DGRAD's *unity* stride specialization only support stride {1, 1}
|
||||
if ((ImplicitGemm::kConvolutionalOperator ==
|
||||
cutlass::conv::Operator::kDgrad) &&
|
||||
(ImplicitGemm::UnderlyingKernel::Mma::IteratorA::kStrideSupport ==
|
||||
cutlass::conv::StrideSupport::kUnity)) {
|
||||
if (!((conv_problem.stride_d == 1) &&
|
||||
(conv_problem.stride_h == 1) &&
|
||||
(conv_problem.stride_w == 1))
|
||||
) {
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
#if 0 // relax restrictions on analytic strided dgrad
|
||||
// CUTLASS DGRAD's *strided* specialization only support stride >= {2, 2}
|
||||
if ((ImplicitGemm::kConvolutionalOperator ==
|
||||
cutlass::conv::Operator::kDgrad) &&
|
||||
(ImplicitGemm::UnderlyingKernel::Mma::IteratorA::kStrideSupport ==
|
||||
cutlass::conv::StrideSupport::kStrided)) {
|
||||
if (((conv_problem.stride_d == 1) && (conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
|
||||
continue;
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
//
|
||||
// Test
|
||||
//
|
||||
// push back tested problem size to avoid re-running duplicates
|
||||
conv_tested_sizes.push_back(conv_problem);
|
||||
|
||||
// test mode = xcross
|
||||
passed = testbed.run(
|
||||
conv_problem,
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// test mode = convolution
|
||||
passed = testbed.run(
|
||||
conv_problem.reset_mode(cutlass::conv::Mode::kConvolution),
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (!TestSplitK)
|
||||
return passed;
|
||||
|
||||
// Sweep split-k-slice using serial and prallel reduction with non-unity alpha and non-zero beta for
|
||||
// a single conv3d problem size. Convolution unit tests take a long time to run so only sweep parameters
|
||||
// which are abolutely necessary to catch functional bugs. The below code does provide option to sweep
|
||||
// alpha and beta for local testing, but only runs one value for alpha and beta.
|
||||
cutlass::conv::Conv3dProblemSize conv3d_split_k_test_size (
|
||||
{1, 8, 8, 8, 32}, // input size (NDHWC)
|
||||
{32, 3, 3, 3, 32}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({1, 1, 1}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
);
|
||||
|
||||
cutlass::conv::SplitKMode split_k_modes [] = {
|
||||
cutlass::conv::SplitKMode::kSerial
|
||||
};
|
||||
|
||||
int split_k_slices[] = {
|
||||
1, 2, 3, 4, 201
|
||||
};
|
||||
|
||||
double problem_alpha[] = {
|
||||
2.0
|
||||
};
|
||||
|
||||
double problem_beta[] = {
|
||||
2.0
|
||||
};
|
||||
|
||||
for (auto split_k_mode : split_k_modes) {
|
||||
for (auto split_k_slice : split_k_slices) {
|
||||
for (auto alpha : problem_alpha) {
|
||||
for (auto beta : problem_beta) {
|
||||
|
||||
passed = testbed.run(
|
||||
conv3d_split_k_test_size.reset_split_k_slices(split_k_slice),
|
||||
split_k_mode,
|
||||
cutlass::from_real<typename ImplicitGemm::ElementCompute>(alpha),
|
||||
cutlass::from_real<typename ImplicitGemm::ElementCompute>(beta));
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
template <typename ImplicitGemm,
|
||||
typename ReferenceOp = Conv3dWithBroadcastReferenceOp<ImplicitGemm>,
|
||||
bool AddBroadcastFirst = false>
|
||||
bool TestSpecificConv3dWithBroadcast(
|
||||
const Conv3dProblemVector & problem_sizes) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv3dWithBroadcast<ImplicitGemm, ReferenceOp, AddBroadcastFirst> testbed;
|
||||
|
||||
// Sweep conv3d problem sizes (split-k-mode=kSerial, split-k-slice=1, alpha=1.0, beta=0.0)
|
||||
for(auto conv_problem : problem_sizes) {
|
||||
|
||||
//
|
||||
// Test
|
||||
//
|
||||
|
||||
// test mode = xcross
|
||||
passed = testbed.run(
|
||||
conv_problem,
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// test mode = convolution
|
||||
passed = testbed.run(
|
||||
conv_problem.reset_mode(cutlass::conv::Mode::kConvolution),
|
||||
cutlass::conv::SplitKMode::kSerial);
|
||||
|
||||
if (!passed) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
@@ -165,7 +165,7 @@ class TestbedDepthwiseDirectConv2d {
|
||||
throw std::runtime_error("cudaGetDeviceProperties() failed");
|
||||
}
|
||||
|
||||
if (properties.sharedMemPerBlockOptin < smem_size) {
|
||||
if (properties.sharedMemPerBlockOptin < static_cast<size_t>(smem_size)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,31 @@
|
||||
# Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: BSD-3-Clause
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright notice, this
|
||||
# list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
# this list of conditions and the following disclaimer in the documentation
|
||||
# and/or other materials provided with the distribution.
|
||||
#
|
||||
# 3. Neither the name of the copyright holder nor the names of its
|
||||
# contributors may be used to endorse or promote products derived from
|
||||
# this software without specific prior written permission.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
add_subdirectory(fprop)
|
||||
add_subdirectory(wgrad)
|
||||
add_subdirectory(dgrad)
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,49 @@
|
||||
# Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: BSD-3-Clause
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright notice, this
|
||||
# list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
# this list of conditions and the following disclaimer in the documentation
|
||||
# and/or other materials provided with the distribution.
|
||||
#
|
||||
# 3. Neither the name of the copyright holder nor the names of its
|
||||
# contributors may be used to endorse or promote products derived from
|
||||
# this software without specific prior written permission.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
add_custom_target(
|
||||
cutlass_test_unit_conv_dgrad_device
|
||||
DEPENDS
|
||||
cutlass_test_unit_conv_dgrad_device_tensorop_sm90
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_dgrad_device_tensorop_sm90
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 1
|
||||
|
||||
sm90_conv1d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv2d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv3d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
|
||||
sm90_conv1d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv2d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv3d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
)
|
||||
|
||||
+406
@@ -0,0 +1,406 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide CONV interface
|
||||
*/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+405
@@ -0,0 +1,405 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide CONV interface
|
||||
*/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_dgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+413
@@ -0,0 +1,413 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide CONV interface
|
||||
*/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_dgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+411
@@ -0,0 +1,411 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_dgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorNDHWC, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kDgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
@@ -0,0 +1,75 @@
|
||||
# Copyright (c) 2013 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: BSD-3-Clause
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright notice, this
|
||||
# list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
# this list of conditions and the following disclaimer in the documentation
|
||||
# and/or other materials provided with the distribution.
|
||||
#
|
||||
# 3. Neither the name of the copyright holder nor the names of its
|
||||
# contributors may be used to endorse or promote products derived from
|
||||
# this software without specific prior written permission.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
add_custom_target(
|
||||
cutlass_test_unit_conv_fprop_device
|
||||
DEPENDS
|
||||
cutlass_test_unit_conv1d_fprop_device_tensorop_sm90
|
||||
cutlass_test_unit_conv2d_fprop_device_tensorop_sm90
|
||||
cutlass_test_unit_conv3d_fprop_device_tensorop_sm90
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv1d_fprop_device_tensorop_sm90
|
||||
|
||||
# No batching of source to control compiler memory usage
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 1
|
||||
|
||||
sm90_conv1d_fprop_implicit_gemm_s8_s8_s32_tensorop_s32.cu
|
||||
sm90_conv1d_fprop_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv1d_fprop_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv1d_fprop_implicit_gemm_tf32_tf32_f32_tensorop_f32.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv2d_fprop_device_tensorop_sm90
|
||||
|
||||
# No batching of source to control compiler memory usage
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 1
|
||||
|
||||
sm90_conv2d_fprop_implicit_gemm_s8_s8_s32_tensorop_s32.cu
|
||||
sm90_conv2d_fprop_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv2d_fprop_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv2d_fprop_implicit_gemm_tf32_tf32_f32_tensorop_f32.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv3d_fprop_device_tensorop_sm90
|
||||
|
||||
# No batching of source to control compiler memory usage
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 1
|
||||
|
||||
sm90_conv3d_fprop_implicit_gemm_s8_s8_s32_tensorop_s32.cu
|
||||
sm90_conv3d_fprop_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv3d_fprop_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv3d_fprop_implicit_gemm_tf32_tf32_f32_tensorop_f32.cu
|
||||
)
|
||||
|
||||
+403
@@ -0,0 +1,403 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+402
@@ -0,0 +1,402 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+403
@@ -0,0 +1,403 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 64x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 64x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 64x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 64x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 128x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 128x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 128x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_s8nwc_s8nwc_s32nwc_tensor_op_s32, 128x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNWC, 16,
|
||||
int32_t, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+402
@@ -0,0 +1,402 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 64x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 64x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 64x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 64x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 128x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 128x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 128x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_fprop_implicitgemm_tf32nwc_tf32nwc_f32nwc_tensor_op_f32, 128x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
float, cutlass::layout::TensorNWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 64x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 64x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 64x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 64x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 128x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 128x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 128x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32, 128x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 64x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 64x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 64x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 64x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 128x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 128x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 128x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_fprop_implicitgemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32, 128x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
float, cutlass::layout::TensorNHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementAct>::value,
|
||||
ElementOut, cutlass::layout::TensorNDHWC, 128 / cutlass::sizeof_bits<ElementOut>::value,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 64x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 128x64x64_1x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 128x64x64_2x1x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 128x64x64_1x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_s8ndhwc_s8ndhwc_s32ndhwc_tensor_op_s32, 128x64x64_2x2x1) {
|
||||
using ElementAct = int8_t;
|
||||
using ElementFlt = int8_t;
|
||||
using ElementOut = int32_t;
|
||||
using ElementAcc = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
int8_t, cutlass::layout::TensorNDHWC, 16,
|
||||
int32_t, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementAct),
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 16 / sizeof(ElementFlt),
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 64x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 64x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 64x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 64x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x32
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 128x64x32_1x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 128x64x32_2x1x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 128x64x32_1x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_fprop_implicitgemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32, 128x64x32_2x2x1) {
|
||||
using ElementAct = float;
|
||||
using ElementFlt = float;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, _64, Shape<_32>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
float, cutlass::layout::TensorNDHWC, 4,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kFprop,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 4,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
@@ -0,0 +1,551 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Implicit GEMM testbed for 3.x API
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cutlass/kernel_hardware_info.hpp"
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/conv/convnd_problem_shape.hpp"
|
||||
|
||||
#include "thrust/universal_vector.h"
|
||||
#include "cutlass/util/distribution.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/packed_stride.hpp"
|
||||
#include "cutlass/util/reference/host/conv.hpp"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_norm.h"
|
||||
#include "cutlass/util/reference/device/tensor_fill.h"
|
||||
#include "cutlass/util/reference/device/tensor_compare.h"
|
||||
|
||||
#include "conv_problem_sizes.hpp"
|
||||
#include "../cache_testbed_output.h"
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cute/layout.hpp"
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace test::conv::device {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Initializes a flat device buffer
|
||||
template <typename Element>
|
||||
static void
|
||||
initialize_values(
|
||||
thrust::universal_vector<Element>& dst_ptr,
|
||||
cutlass::Distribution::Kind dist_kind,
|
||||
uint64_t seed) {
|
||||
if (cutlass::Distribution::Uniform == dist_kind) {
|
||||
int scope;
|
||||
int bits = cutlass::sizeof_bits<Element>::value;
|
||||
|
||||
if (bits <= 8) {
|
||||
scope = 2;
|
||||
}
|
||||
else if (bits == 16) {
|
||||
scope = 4;
|
||||
}
|
||||
else {
|
||||
scope = 8;
|
||||
}
|
||||
cutlass::reference::host::BlockFillRandomUniform(
|
||||
dst_ptr.data().get(), dst_ptr.size(), seed, scope, -scope, 0);
|
||||
}
|
||||
else if (cutlass::Distribution::Identity == dist_kind) {
|
||||
cutlass::reference::host::BlockFillRandomUniform(
|
||||
dst_ptr.data().get(), dst_ptr.size(), seed, 0, 0, 0);
|
||||
}
|
||||
else if (cutlass::Distribution::Gaussian == dist_kind) {
|
||||
cutlass::reference::host::BlockFillRandomGaussian(dst_ptr.data().get(), dst_ptr.size(), seed, 0, 0.5);
|
||||
}
|
||||
else if (cutlass::Distribution::Sequential == dist_kind) {
|
||||
cutlass::reference::host::BlockFillSequential(dst_ptr.data().get(), dst_ptr.size());
|
||||
}
|
||||
else {
|
||||
std::cerr << "Invalid distribution kind!\n.";
|
||||
exit(1);
|
||||
}
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
template <class Conv>
|
||||
struct ConvTestbed {
|
||||
// Kernel data types
|
||||
using ElementA = typename Conv::ConvKernel::ElementA;
|
||||
using ElementB = typename Conv::ConvKernel::ElementB;
|
||||
using ElementC = cute::conditional_t<cute::is_void_v<typename Conv::ConvKernel::ElementC>,
|
||||
typename Conv::ConvKernel::ElementD, typename Conv::ConvKernel::ElementC>;
|
||||
using ElementD = typename Conv::ConvKernel::ElementD;
|
||||
using ElementAccumulator = typename Conv::ConvKernel::ElementAccumulator;
|
||||
|
||||
//
|
||||
// FusionOperation derived types/queries
|
||||
//
|
||||
using FusionOp = typename Conv::EpilogueOutputOp;
|
||||
|
||||
// fusion types are potentially void if the fusion is not supported
|
||||
// helper so we don't try to construct HostTensor with void type
|
||||
template <typename T, typename U = uint8_t>
|
||||
using non_void_t = cute::conditional_t<cute::is_void_v<T>, U, T>;
|
||||
using ElementScalar = typename FusionOp::ElementScalar;
|
||||
using ElementCompute = typename FusionOp::ElementCompute;
|
||||
using BiasType = typename cutlass::epilogue::collective::detail::IsThreadEpilogueOpWithBias<FusionOp>::type;
|
||||
using ElementBias = non_void_t<BiasType>;
|
||||
using ActivationType = non_void_t<typename cutlass::epilogue::collective::detail::IsThreadEpilogueOpWithActivation<FusionOp>::type,
|
||||
cutlass::epilogue::thread::Identity<ElementCompute>>;
|
||||
static constexpr bool IsActivationEnabled = cutlass::epilogue::collective::detail::IsThreadEpilogueOpWithActivation<FusionOp>::value;
|
||||
using ActivationFunctor = cute::conditional_t<IsActivationEnabled, ActivationType, cutlass::epilogue::thread::Identity<ElementCompute>>;
|
||||
|
||||
static constexpr bool IsBiasEnabled = cutlass::epilogue::collective::detail::IsThreadEpilogueOpWithBias<FusionOp>::value &&
|
||||
!cute::is_same_v<BiasType, void>;
|
||||
using StrideC = typename Conv::ConvKernel::StrideC;
|
||||
using StrideD = typename Conv::ConvKernel::StrideD;
|
||||
using ThreadEpilogueOp = typename Conv::ConvKernel::CollectiveEpilogue::ThreadEpilogueOp;
|
||||
|
||||
static constexpr cutlass::conv::Operator ConvOp = Conv::DispatchPolicy::ConvOp;
|
||||
static constexpr int NumSpatialDimensions = Conv::NumSpatialDimensions;
|
||||
using ProblemShape = cutlass::conv::ConvProblemShape<ConvOp, NumSpatialDimensions>;
|
||||
|
||||
using Schedule = typename Conv::DispatchPolicy::Schedule;
|
||||
/// Initialization
|
||||
cutlass::Distribution::Kind init_A = cutlass::Distribution::Uniform;
|
||||
cutlass::Distribution::Kind init_B = cutlass::Distribution::Uniform;
|
||||
cutlass::Distribution::Kind init_C = cutlass::Distribution::Uniform;
|
||||
cutlass::Distribution::Kind init_bias = cutlass::Distribution::Uniform;
|
||||
uint64_t seed = 6090;
|
||||
float epsilon = 0.0f;
|
||||
int split_p_slices = 1;
|
||||
thrust::universal_vector<ElementA> tensor_A;
|
||||
thrust::universal_vector<ElementB> tensor_B;
|
||||
thrust::universal_vector<ElementC> tensor_C;
|
||||
thrust::universal_vector<ElementD> tensor_D_computed;
|
||||
thrust::universal_vector<ElementD> tensor_D_reference;
|
||||
thrust::universal_vector<ElementBias> tensor_bias;
|
||||
thrust::universal_vector<ElementScalar> tensor_alpha;
|
||||
thrust::universal_vector<ElementScalar> tensor_beta;
|
||||
|
||||
void initialize(ProblemShape const& problem_shape, uint64_t seed = 6090) {
|
||||
tensor_A.resize(sizeof(ElementA) * problem_shape.size_A());
|
||||
tensor_B.resize(sizeof(ElementB) * problem_shape.size_B());
|
||||
tensor_C.resize(sizeof(ElementC) * problem_shape.size_C());
|
||||
tensor_D_computed.resize(sizeof(ElementD) * problem_shape.size_C());
|
||||
tensor_D_reference.resize(sizeof(ElementD) * problem_shape.size_C());
|
||||
tensor_bias.resize(sizeof(ElementBias) * cute::size(cute::get<0>(problem_shape.get_shape_B())));
|
||||
initialize_values(tensor_A, init_A, seed);
|
||||
initialize_values(tensor_B, init_B, seed * 11);
|
||||
initialize_values(tensor_C, init_C, seed * 17);
|
||||
initialize_values(tensor_bias, init_bias, seed * 19);
|
||||
}
|
||||
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
bool sufficient() const {
|
||||
int device_idx;
|
||||
cudaError_t result = cudaGetDevice(&device_idx);
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaGetDevice() API call failed.");
|
||||
}
|
||||
|
||||
int max_smem_size;
|
||||
result = cudaDeviceGetAttribute(&max_smem_size, cudaDevAttrMaxSharedMemoryPerBlockOptin, device_idx);
|
||||
if (result != cudaSuccess) {
|
||||
throw std::runtime_error("cudaDeviceGetAttribute() failed");
|
||||
}
|
||||
|
||||
return max_smem_size >= Conv::ConvKernel::SharedStorageSize;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
ProblemShape const& problem_shape,
|
||||
ElementScalar alpha = ElementScalar(1),
|
||||
ElementScalar beta = ElementScalar(0)) {
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
if (CUTLASS_TEST_UNIT_ENABLE_WARNINGS) {
|
||||
std::cerr << "Test waived due to insufficient CUDA device.\n";
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
initialize(problem_shape);
|
||||
|
||||
cutlass::KernelHardwareInfo hw_info;
|
||||
cudaGetDevice(&hw_info.device_id);
|
||||
hw_info.sm_count = cutlass::KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
|
||||
|
||||
// configure the operator
|
||||
Conv conv_op;
|
||||
auto stride_C = StrideC{};
|
||||
auto stride_D = StrideD{};
|
||||
if constexpr (ConvOp == cutlass::conv::Operator::kWgrad) {
|
||||
stride_C = cutlass::make_cute_packed_stride(
|
||||
StrideC{}, problem_shape.shape_C, problem_shape.stride_C, ConvOp);
|
||||
stride_D = cutlass::make_cute_packed_stride(
|
||||
StrideD{}, problem_shape.shape_C, problem_shape.stride_C, ConvOp);
|
||||
}
|
||||
// Need to support non-packed output strides for fprop and dgrad kernel.
|
||||
else {
|
||||
cute::for_each(cute::make_seq<cute::rank<0>(StrideC{})>{}, [&](auto i) {
|
||||
cute::get<0, i>(stride_C) = problem_shape.stride_C[ProblemShape::RankT-2-i];
|
||||
});
|
||||
cute::for_each(cute::make_seq<cute::rank<0>(StrideD{})>{}, [&](auto i) {
|
||||
cute::get<0, i>(stride_D) = problem_shape.stride_C[ProblemShape::RankT-2-i];
|
||||
});
|
||||
}
|
||||
typename Conv::ConvKernel::TileScheduler::Arguments scheduler_args{};
|
||||
auto args = typename Conv::Arguments {
|
||||
{
|
||||
problem_shape,
|
||||
tensor_A.data().get(),
|
||||
tensor_B.data().get(),
|
||||
}, // MainloopArguments
|
||||
{
|
||||
{},
|
||||
tensor_C.data().get(),
|
||||
stride_C,
|
||||
tensor_D_computed.data().get(),
|
||||
stride_D,
|
||||
}, // EpilogueArguments
|
||||
hw_info,
|
||||
scheduler_args
|
||||
};
|
||||
|
||||
auto &fusion_args = args.epilogue.thread;
|
||||
|
||||
// some fused patterns have no linear combination
|
||||
if constexpr (IsBiasEnabled) {
|
||||
fusion_args.bias_ptr = tensor_bias.data().get();
|
||||
}
|
||||
|
||||
// Clamp bound
|
||||
if constexpr (cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::Clamp<ElementCompute>>) {
|
||||
fusion_args.activation.lower_bound = ElementCompute{0};
|
||||
fusion_args.activation.upper_bound = CUTLASS_STL_NAMESPACE::numeric_limits<ElementCompute>::max();
|
||||
}
|
||||
|
||||
// Scale
|
||||
if constexpr (cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledGELU_taylor<ElementCompute>> ||
|
||||
cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledGELU<ElementCompute>>) {
|
||||
fusion_args.activation.scale = ElementCompute{1};
|
||||
}
|
||||
|
||||
cutlass::Status status = cutlass::Status::kInvalid;
|
||||
|
||||
status = conv_op.can_implement(args);
|
||||
EXPECT_EQ(conv_op.can_implement(args), cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
std::cerr << "can_implement failed for the given problem_shape: \n";
|
||||
print(problem_shape);
|
||||
return false;
|
||||
}
|
||||
|
||||
// find workspace requirement for parallel split-k reduction
|
||||
size_t workspace_size = Conv::get_workspace_size(args);
|
||||
thrust::universal_vector<uint8_t> workspace(workspace_size);
|
||||
|
||||
status = conv_op.initialize(args, workspace.data().get());
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
std::cerr << "This test is not supported: " << cudaGetErrorString(error) << "\n";
|
||||
return true;
|
||||
}
|
||||
|
||||
// run conv3d operator
|
||||
status = conv_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
bool passed = false;
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
EXPECT_EQ(result, cudaSuccess) << " Kernel execution error: "
|
||||
<< cudaGetErrorString(result);
|
||||
|
||||
// Create cute::Tensors using the logical rank-3 MNK multi-mode shapes the mainloop gives us
|
||||
auto shape_mA = cute::reverse(problem_shape.shape_A);
|
||||
auto shape_mB = cute::reverse(problem_shape.shape_B);
|
||||
auto shape_mC = cute::reverse(problem_shape.shape_C);
|
||||
auto shape_mBias = cute::make_shape(cute::size(cute::get<0>(problem_shape.get_shape_B())));
|
||||
|
||||
auto stride_mA = cute::reverse(problem_shape.stride_A);
|
||||
auto stride_mB = cute::reverse(problem_shape.stride_B);
|
||||
auto stride_mC = cute::reverse(problem_shape.stride_C);
|
||||
|
||||
auto mA = make_tensor(tensor_A.data().get(), make_layout(shape_mA, stride_mA));
|
||||
auto mB = make_tensor(tensor_B.data().get(), make_layout(shape_mB, stride_mB));
|
||||
auto mC = make_tensor(tensor_C.data().get(), make_layout(shape_mC, stride_mC));
|
||||
auto mD_ref = make_tensor(tensor_D_reference.data().get(), make_layout(shape_mC, stride_mC));
|
||||
auto mD_computed = make_tensor(tensor_D_computed.data().get(), make_layout(shape_mC, stride_mC));
|
||||
auto mBias = make_tensor(tensor_bias.data().get(), make_layout(shape_mBias));
|
||||
auto mAlpha = make_tensor(tensor_alpha.data().get(), make_layout(shape_mBias));
|
||||
auto mBeta = make_tensor(tensor_beta.data().get(), make_layout(shape_mBias));
|
||||
|
||||
cutlass::reference::host::ConvEpilogueFusionParams<
|
||||
ElementAccumulator,
|
||||
ElementScalar,
|
||||
ElementCompute,
|
||||
ElementC,
|
||||
ElementD,
|
||||
decltype(mAlpha),
|
||||
decltype(mBeta),
|
||||
decltype(mBias),
|
||||
ActivationFunctor>
|
||||
epilogue_fusion_params{};
|
||||
|
||||
epilogue_fusion_params.alpha = alpha;
|
||||
epilogue_fusion_params.beta = beta;
|
||||
|
||||
if constexpr (IsBiasEnabled) {
|
||||
epilogue_fusion_params.tensor_bias = mBias;
|
||||
}
|
||||
|
||||
auto padding = cute::reverse(problem_shape.lower_padding);
|
||||
auto tstride = cute::reverse(problem_shape.traversal_stride);
|
||||
auto dilation = cute::reverse(problem_shape.dilation);
|
||||
|
||||
cutlass::reference::host::ConvReferenceImpl<
|
||||
ConvOp,
|
||||
NumSpatialDimensions,
|
||||
decltype(mA),
|
||||
decltype(mB),
|
||||
decltype(mC),
|
||||
decltype(mD_ref),
|
||||
decltype(padding),
|
||||
decltype(tstride),
|
||||
decltype(dilation),
|
||||
decltype(epilogue_fusion_params)>
|
||||
reference_impl(mA, mB, mC, mD_ref, padding, tstride, dilation, epilogue_fusion_params);
|
||||
|
||||
//
|
||||
// Reference check - support caching results
|
||||
//
|
||||
|
||||
CachedTestKey cached_test_key = CreateCachedConvNd3xTestKey<
|
||||
ProblemShape,
|
||||
ElementA,
|
||||
ElementB,
|
||||
ElementC,
|
||||
ElementD
|
||||
>(
|
||||
ConvOp,
|
||||
problem_shape,
|
||||
alpha,
|
||||
beta,
|
||||
tensor_A,
|
||||
tensor_B,
|
||||
tensor_C
|
||||
);
|
||||
|
||||
//
|
||||
// Look for the cached key
|
||||
//
|
||||
|
||||
bool cached_result_loaded = false;
|
||||
CachedTestResult cached_test_result;
|
||||
|
||||
std::string convnd_result_cache_name =
|
||||
std::string("cached_results_") + CUTLASS_TARGET_NAME + ".txt";
|
||||
|
||||
#if (CUTLASS_TEST_ENABLE_CACHED_RESULTS)
|
||||
CachedTestResultListing cached_results(convnd_result_cache_name);
|
||||
|
||||
auto cached = cached_results.find(cached_test_key);
|
||||
|
||||
cached_result_loaded = cached.first;
|
||||
if (cached_result_loaded) {
|
||||
cached_test_result = cached.second;
|
||||
}
|
||||
#endif
|
||||
|
||||
if (!cached_result_loaded) {
|
||||
// Compute reference
|
||||
reference_impl.compute_reference();
|
||||
|
||||
#if (CUTLASS_TEST_ENABLE_CACHED_RESULTS)
|
||||
cached_test_result.D = TensorHash(tensor_D_reference);
|
||||
CachedTestResultListing cached_results(convnd_result_cache_name);
|
||||
|
||||
cached_results.append(cached_test_key, cached_test_result);
|
||||
cached_results.write(convnd_result_cache_name);
|
||||
#endif
|
||||
} // if (!cached_result_loaded)
|
||||
|
||||
#if (CUTLASS_TEST_ENABLE_CACHED_RESULTS)
|
||||
uint32_t tensor_D_computed_hash = TensorHash(tensor_D_computed);
|
||||
passed = (tensor_D_computed_hash == cached_test_result.D);
|
||||
// If hash fails, double check against reference implementation.
|
||||
if(!passed) {
|
||||
std::cerr << "Hash-based comparison unsuccessful for key:" << "\n" << cached_test_key
|
||||
<< ", comparing with reference implementation now.\n";
|
||||
if (cached_result_loaded) {
|
||||
// Compute reference
|
||||
reference_impl.compute_reference();
|
||||
}
|
||||
// Validate kernel against reference
|
||||
passed = compare_reference(
|
||||
mD_ref, mD_computed, mA, mB, mAlpha,
|
||||
mBeta, mBias,
|
||||
this->epsilon);
|
||||
}
|
||||
#else
|
||||
// Validate kernel against reference
|
||||
passed = compare_reference(
|
||||
mD_ref, mD_computed, mA, mB, mAlpha,
|
||||
mBeta, mBias,
|
||||
this->epsilon);
|
||||
#endif
|
||||
|
||||
EXPECT_TRUE(passed);
|
||||
return passed;
|
||||
}
|
||||
|
||||
template<
|
||||
class Engine, class Layout,
|
||||
class EngineA, class LayoutA,
|
||||
class EngineB, class LayoutB,
|
||||
class EngineAlpha, class LayoutAlpha,
|
||||
class EngineBeta, class LayoutBeta,
|
||||
class EngineBias, class LayoutBias
|
||||
>
|
||||
static constexpr bool
|
||||
compare_reference(
|
||||
cute::Tensor<Engine, Layout> const& reference,
|
||||
cute::Tensor<Engine, Layout> const& computed,
|
||||
cute::Tensor<EngineA, LayoutA> const& A,
|
||||
cute::Tensor<EngineB, LayoutB> const& B,
|
||||
cute::Tensor<EngineAlpha, LayoutAlpha> const& tensor_alpha,
|
||||
cute::Tensor<EngineBeta, LayoutBeta> const& tensor_beta,
|
||||
cute::Tensor<EngineBias, LayoutBias> const& tensor_bias,
|
||||
float epsilon = 0.0f) {
|
||||
if (size(reference) != size(computed)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
bool passed = true;
|
||||
if (epsilon == 0.0f) {
|
||||
// fast refcheck w/o epsilon
|
||||
for (size_t i = 0; i < size_t(size(reference)); ++i) {
|
||||
if (reference(i) != computed(i)) {
|
||||
passed = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
// refcheck with epsilon
|
||||
for (size_t i = 0; i < size_t(size(reference)); ++i) {
|
||||
auto ref = static_cast<float>(reference(i));
|
||||
auto act = static_cast<float>(computed(i));
|
||||
auto abs_error = std::abs(act - ref);
|
||||
auto rel_error = abs_error / (std::max(std::abs(act), std::abs(ref)) + 0.00001f);
|
||||
if (std::isnan(abs_error) || std::isnan(rel_error) ||
|
||||
std::min(abs_error, rel_error) > epsilon) {
|
||||
passed = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
#if CUTLASS_DEBUG_TRACE_LEVEL > 1
|
||||
if (not passed) {
|
||||
cute::print("Reference:");
|
||||
cute::print_tensor(reference);
|
||||
cute::print("\nComputed:");
|
||||
cute::print_tensor(computed);
|
||||
cute::print("\n");
|
||||
|
||||
for (size_t i = 0; i < size_t(size(A)); ++i) {
|
||||
printf("[%ld]: A = %f\n", i, float(A(i)));
|
||||
}
|
||||
for (size_t i = 0; i < size_t(size(B)); ++i) {
|
||||
printf("[%ld]: B = %f\n", i, float(B(i)));
|
||||
}
|
||||
if constexpr (IsBiasEnabled) {
|
||||
for (size_t i = 0; i < size_t(size(tensor_bias)); ++i) {
|
||||
printf("[%ld]: bias = %f\n", i, float(tensor_bias(i)));
|
||||
}
|
||||
}
|
||||
|
||||
for (size_t i = 0; i < size_t(size(reference)); ++i) {
|
||||
printf("[%ld]: ref = %f, computed = %f\n", i, float(reference(i)), float(computed(i)));
|
||||
}
|
||||
}
|
||||
#endif
|
||||
return passed;
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename Conv>
|
||||
bool TestAllConv(double alpha = 1.0, double beta = 0.0, float epsilon = 0.0f) {
|
||||
using ElementScalar = typename Conv::EpilogueOutputOp::ElementScalar;
|
||||
|
||||
bool passed = true;
|
||||
ConvTestbed<Conv> testbed;
|
||||
testbed.epsilon = epsilon;
|
||||
auto problem_vector = get_conv_problem_vector<
|
||||
Conv::NumSpatialDimensions, Conv::DispatchPolicy::ConvOp>();
|
||||
|
||||
for (auto conv_problem : problem_vector) {
|
||||
#if CUTLASS_DEBUG_TRACE_LEVEL > 0
|
||||
print(conv_problem);
|
||||
#endif
|
||||
|
||||
passed = testbed.run(
|
||||
conv_problem,
|
||||
cutlass::from_real<ElementScalar>(alpha),
|
||||
cutlass::from_real<ElementScalar>(beta));
|
||||
|
||||
if (!passed) {
|
||||
printf("Failed test for "); print(conv_problem);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace test::conv::device
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,46 @@
|
||||
# Copyright (c) 2013 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
# SPDX-License-Identifier: BSD-3-Clause
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions are met:
|
||||
#
|
||||
# 1. Redistributions of source code must retain the above copyright notice, this
|
||||
# list of conditions and the following disclaimer.
|
||||
#
|
||||
# 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
# this list of conditions and the following disclaimer in the documentation
|
||||
# and/or other materials provided with the distribution.
|
||||
#
|
||||
# 3. Neither the name of the copyright holder nor the names of its
|
||||
# contributors may be used to endorse or promote products derived from
|
||||
# this software without specific prior written permission.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
add_custom_target(
|
||||
cutlass_test_unit_conv_wgrad_device
|
||||
DEPENDS
|
||||
cutlass_test_unit_conv_wgrad_device_tensorop_sm90
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_wgrad_device_tensorop_sm90
|
||||
|
||||
sm90_conv1d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv2d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
sm90_conv3d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu
|
||||
|
||||
sm90_conv1d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv2d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
sm90_conv3d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f32.cu
|
||||
)
|
||||
|
||||
+403
@@ -0,0 +1,403 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+402
@@ -0,0 +1,402 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
TEST(SM90_device_conv1d_wgrad_implicitgemm_f16nwc_f16nwc_f32nwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCS, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv2d_wgrad_implicitgemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSR, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+411
@@ -0,0 +1,411 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::TmaWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f16, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAcc, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
+410
@@ -0,0 +1,410 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cute/tensor.hpp"
|
||||
#include "cute/atom/mma_atom.hpp"
|
||||
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/conv/device/conv_universal_adapter.hpp"
|
||||
#include "cutlass/conv/kernel/conv_universal.hpp"
|
||||
#include "cutlass/conv/collective/collective_builder.hpp"
|
||||
#include "cutlass/epilogue/collective/collective_builder.hpp"
|
||||
|
||||
#include "../testbed_conv.hpp"
|
||||
using namespace cute;
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 64x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 64x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_64, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Tile shape 128x64x64
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Cluster 1x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x1x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x1x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_1,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 1x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_1x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_1,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
//
|
||||
// Cluster 2x2x1
|
||||
//
|
||||
|
||||
TEST(SM90_device_conv3d_wgrad_implicitgemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32, 128x64x64_2x2x1) {
|
||||
using ElementAct = cutlass::half_t;
|
||||
using ElementFlt = cutlass::half_t;
|
||||
using ElementOut = float;
|
||||
using ElementAcc = float;
|
||||
using ElementCompute = float;
|
||||
using TileShapeMNK = Shape<_128, Shape<_64>, Shape<_64>>;
|
||||
using ClusterShapeMNK = Shape<_2,_2,_1>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::epilogue::collective::EpilogueTileAuto,
|
||||
ElementAcc, ElementCompute,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::half_t, cutlass::layout::TensorKCSRT, 8,
|
||||
cutlass::epilogue::NoSmemWarpSpecialized
|
||||
>::CollectiveOp;
|
||||
|
||||
using CollectiveMainloop = typename cutlass::conv::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
cutlass::conv::Operator::kWgrad,
|
||||
ElementAct, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementFlt, cutlass::layout::TensorNDHWC, 8,
|
||||
ElementAcc,
|
||||
TileShapeMNK, ClusterShapeMNK,
|
||||
cutlass::conv::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::conv::collective::KernelScheduleAuto
|
||||
>::CollectiveOp;
|
||||
|
||||
using ConvKernel = cutlass::conv::kernel::ConvUniversal<
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue
|
||||
>;
|
||||
|
||||
using Conv = cutlass::conv::device::ConvUniversalAdapter<ConvKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv<Conv>());
|
||||
}
|
||||
|
||||
#endif // defined(CUTLASS_ARCH_MMA_SM90_SUPPORTED)
|
||||
@@ -117,8 +117,8 @@ public:
|
||||
void run() {
|
||||
|
||||
/// Device memory containing output
|
||||
cutlass::device_memory::allocation< ArrayTy > output(kThreads);
|
||||
std::vector< ArrayTy > output_host(kThreads);
|
||||
cutlass::device_memory::allocation< ArrayTy > output(static_cast<size_t>(kThreads));
|
||||
std::vector< ArrayTy > output_host(static_cast<size_t>(kThreads));
|
||||
|
||||
dim3 grid(1,1);
|
||||
dim3 block(kThreads, 1, 1);
|
||||
@@ -138,7 +138,7 @@ public:
|
||||
ASSERT_EQ(result, cudaSuccess) << "CUDA error: " << cudaGetErrorString(result);
|
||||
|
||||
char const *ptr_host = reinterpret_cast<char const *>(output_host.data());
|
||||
for (int i = 0; i < sizeof(ArrayTy) * kThreads; ++i) {
|
||||
for (size_t i = 0; i < sizeof(ArrayTy) * kThreads; ++i) {
|
||||
EXPECT_FALSE(ptr_host[i]);
|
||||
}
|
||||
|
||||
|
||||
@@ -49,7 +49,7 @@ TEST(complex, f64_to_f32_conversion) {
|
||||
|
||||
cutlass::complex<float> dest = cutlass::complex<float>(source); // explicit conversion
|
||||
|
||||
EXPECT_TRUE(source.real() == 1.5 && source.imag() == -1.25 &&
|
||||
EXPECT_TRUE(source.real() == 1.5 && source.imag() == -1.25 &&
|
||||
dest.real() == 1.5f && dest.imag() == -1.25f);
|
||||
}
|
||||
|
||||
@@ -61,7 +61,7 @@ TEST(complex, f32_to_f64_conversion) {
|
||||
|
||||
cutlass::complex<double> dest = source; // implicit conversion
|
||||
|
||||
EXPECT_TRUE(source.real() == -1.5f && source.imag() == 1.25f &&
|
||||
EXPECT_TRUE(source.real() == -1.5f && source.imag() == 1.25f &&
|
||||
dest.real() == -1.5 && dest.imag() == 1.25);
|
||||
}
|
||||
|
||||
@@ -73,7 +73,7 @@ TEST(complex, s32_to_f64_conversion) {
|
||||
|
||||
cutlass::complex<double> dest = source; // implicit conversion
|
||||
|
||||
EXPECT_TRUE(source.real() == -2 && source.imag() == 1 &&
|
||||
EXPECT_TRUE(source.real() == -2 && source.imag() == 1 &&
|
||||
dest.real() == -2 && dest.imag() == 1);
|
||||
}
|
||||
|
||||
@@ -86,14 +86,14 @@ TEST(complex, f16_to_f32_conversion) {
|
||||
|
||||
cutlass::complex<float> dest = cutlass::complex<float>(source); // explicit conversion
|
||||
|
||||
EXPECT_TRUE(source.real() == 1.5_hf && source.imag() == -1.25_hf &&
|
||||
EXPECT_TRUE(source.real() == 1.5_hf && source.imag() == -1.25_hf &&
|
||||
dest.real() == 1.5f && dest.imag() == -1.25f);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(complex, exp_f32) {
|
||||
|
||||
|
||||
cutlass::complex<float> Z[] = {
|
||||
{1, 1},
|
||||
{2 , cutlass::constants::pi<float>()/2.0f },
|
||||
@@ -103,16 +103,16 @@ TEST(complex, exp_f32) {
|
||||
};
|
||||
|
||||
cutlass::complex<double> Expected[] = {
|
||||
{1.4686939399158851, 2.2873552871788423},
|
||||
{1.4686939399158851, 2.2873552871788423},
|
||||
{4.524491950137825e-16, 7.38905609893065},
|
||||
{-1.6487212707001282, 2.019101226849069e-16},
|
||||
{-1.6487212707001282, 2.019101226849069e-16},
|
||||
{-0.9079430793557842, 0.9079430793557843},
|
||||
{1, 0}
|
||||
};
|
||||
|
||||
double tolerance = 0.00001;
|
||||
|
||||
for (int i = 0; cutlass::real(Z[i]); ++i) {
|
||||
for (int i = 0; cutlass::real(Z[i]) != 0.0f; ++i) {
|
||||
double e_r = cutlass::real(Expected[i]);
|
||||
double e_i = cutlass::real(Expected[i]);
|
||||
|
||||
|
||||
@@ -31,3 +31,8 @@ cutlass_test_unit_add_executable(
|
||||
cp_async.cu
|
||||
ldsm.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_cute_ampere_tiled_cp_async
|
||||
tiled_cp_async.cu
|
||||
)
|
||||
|
||||
@@ -0,0 +1,134 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include <iostream>
|
||||
#include <iomanip>
|
||||
#include <utility>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
#include <numeric>
|
||||
|
||||
#include <thrust/host_vector.h>
|
||||
#include <thrust/device_vector.h>
|
||||
|
||||
#include <cute/tensor.hpp>
|
||||
|
||||
#include "tiled_cp_async_testbed.hpp"
|
||||
|
||||
using namespace cute;
|
||||
|
||||
TEST(SM80_CuTe_tiled_cp_async, no_swizzle_mn_single_tile)
|
||||
{
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16, _8>, Stride< _1,_16>>{});
|
||||
using val_layout = decltype(Layout<Shape<_2,_1>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using smem_layout_atom = decltype(Layout<Shape <_16, _4>, Stride< _1,_16>>{});
|
||||
using gmem_stride_type = decltype(LayoutLeft{});
|
||||
test_cp_async_no_swizzle<double, cute::Int<64>, cute::Int<16>, gmem_stride_type, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16, _8>, Stride< _1,_16>>{});
|
||||
using val_layout = decltype(Layout<Shape<_2,_1>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using smem_layout_atom = decltype(Layout<Shape <_16, _4>, Stride< _1,_16>>{});
|
||||
using gmem_stride_type = decltype(LayoutLeft{});
|
||||
test_cp_async_no_swizzle<double, cute::Int<128>, cute::Int<16>, gmem_stride_type, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
}
|
||||
|
||||
TEST(SM80_CuTe_tiled_cp_async, no_swizzle_k_single_tile)
|
||||
{
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16, _8>, Stride< _8,_1>>{});
|
||||
using val_layout = decltype(Layout<Shape<_1,_2>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using smem_layout_atom = decltype(make_ordered_layout(Shape<_128,_16>{}, Step <_2, _1>{}));
|
||||
using gmem_stride_type = decltype(LayoutRight{});
|
||||
test_cp_async_no_swizzle<double, cute::Int<128>, cute::Int<16>, gmem_stride_type, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
}
|
||||
|
||||
TEST(SM80_CuTe_tiled_cp_async, swizzle_mn_single_tile)
|
||||
{
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16, _8>, Stride< _1,_16>>{});
|
||||
using val_layout = decltype(Layout<Shape<_2,_1>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using swizzle_atom = decltype(Swizzle<2,2,2>{});
|
||||
using smem_layout_atom = decltype(Layout<Shape <_16, _4>, Stride< _1,_16>>{});
|
||||
using gmem_stride_type = decltype(LayoutLeft{});
|
||||
test_cp_async_with_swizzle<double, cute::Int<64>, cute::Int<16>, gmem_stride_type, swizzle_atom, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16, _8>, Stride< _1,_16>>{});
|
||||
using val_layout = decltype(Layout<Shape<_2,_1>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using swizzle_atom = decltype(Swizzle<2,2,2>{});
|
||||
using smem_layout_atom = decltype(Layout<Shape <_16, _4>, Stride< _1,_16>>{});
|
||||
using gmem_stride_type = decltype(LayoutLeft{});
|
||||
test_cp_async_with_swizzle<double, cute::Int<128>, cute::Int<16>, gmem_stride_type, swizzle_atom, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
}
|
||||
|
||||
TEST(SM80_CuTe_tiled_cp_async, swizzle_k_single_tile)
|
||||
{
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<double>, double>{});
|
||||
using thr_layout = decltype(Layout<Shape < _8,_16>, Stride<_16, _1>>{});
|
||||
using val_layout = decltype(Layout<Shape<_1,_1>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using swizzle_atom = decltype(Swizzle<2,0,4>{});
|
||||
using smem_layout_atom = decltype(Layout<Shape <_4,_16>, Stride<_1, _4>>{});
|
||||
using gmem_stride_type = decltype(LayoutRight{});
|
||||
test_cp_async_with_swizzle<double, cute::Int<128>, cute::Int<16>, gmem_stride_type, swizzle_atom, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
|
||||
{
|
||||
using copy_atom = decltype(Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<cute::uint128_t>, tfloat32_t>{});
|
||||
using thr_layout = decltype(Layout<Shape <_16,_8>, Stride< _8,_1>>{});
|
||||
using val_layout = decltype(Layout<Shape < _1,_4>>{});
|
||||
using tiled_copy = decltype(make_tiled_copy(copy_atom{}, thr_layout{}, val_layout{}));
|
||||
using swizzle_atom = decltype(Swizzle<3,2,3>{});
|
||||
using smem_layout_atom = decltype(Layout<Shape < _8,_32>, Stride<_32, _1>>{});
|
||||
using gmem_stride_type = decltype(LayoutRight{});
|
||||
test_cp_async_with_swizzle<tfloat32_t, cute::Int<128>, cute::Int<32>, gmem_stride_type, swizzle_atom, smem_layout_atom, tiled_copy>();
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,158 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include <iostream>
|
||||
#include <iomanip>
|
||||
#include <utility>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
#include <numeric>
|
||||
|
||||
#include <thrust/host_vector.h>
|
||||
#include <thrust/device_vector.h>
|
||||
|
||||
#include <cute/tensor.hpp>
|
||||
|
||||
using namespace cute;
|
||||
|
||||
template <class ElementType, class SmemLayout>
|
||||
struct SharedStorage
|
||||
{
|
||||
cute::ArrayEngine<ElementType, cute::cosize_v<SmemLayout>> smem;
|
||||
};
|
||||
|
||||
template <class T, class TiledCopy, class GmemLayout, class SmemLayout>
|
||||
__global__ void
|
||||
test_tiled_cp_async_device_cute(T const* g_in, T* g_out,
|
||||
TiledCopy const tiled_copy,
|
||||
GmemLayout gmem_layout, SmemLayout smem_layout)
|
||||
{
|
||||
using namespace cute;
|
||||
|
||||
extern __shared__ char shared_memory[];
|
||||
using SharedStorage = SharedStorage<T, SmemLayout>;
|
||||
SharedStorage& shared_storage = *reinterpret_cast<SharedStorage*>(shared_memory);
|
||||
|
||||
auto thr_copy = tiled_copy.get_slice(threadIdx.x);
|
||||
Tensor gA = make_tensor(make_gmem_ptr(g_in), gmem_layout);
|
||||
Tensor gB = make_tensor(make_gmem_ptr(g_out), gmem_layout);
|
||||
|
||||
// Construct SMEM tensor
|
||||
Tensor sA = make_tensor(make_smem_ptr(shared_storage.smem.begin()), smem_layout);
|
||||
|
||||
auto tAgA = thr_copy.partition_S(gA);
|
||||
auto tAsA = thr_copy.partition_D(sA);
|
||||
|
||||
#if 0
|
||||
if (thread0()) {
|
||||
print("gA : "); print(gA.layout()); print("\n");
|
||||
print("sA : "); print(sA.layout()); print("\n");
|
||||
print("tAgA: "); print(tAgA.layout()); print("\n");
|
||||
print("tAsA: "); print(tAsA.layout()); print("\n");
|
||||
}
|
||||
#endif
|
||||
|
||||
copy(tiled_copy, tAgA, tAsA);
|
||||
|
||||
cp_async_fence();
|
||||
cp_async_wait<0>();
|
||||
__syncthreads();
|
||||
|
||||
// Store trivially smem -> gmem
|
||||
|
||||
if (thread0()) {
|
||||
copy(sA, gB);
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
template <class T, class TiledCopy, class GMEM_Layout, class SMEM_Layout>
|
||||
void
|
||||
test_tiled_cp_async(
|
||||
TiledCopy const tiled_copy,
|
||||
GMEM_Layout const& gmem_layout,
|
||||
SMEM_Layout const& smem_layout)
|
||||
{
|
||||
using namespace cute;
|
||||
|
||||
// Allocate and initialize host test data
|
||||
size_t N = ceil_div(cosize(gmem_layout) * sizeof_bits<T>::value, 8);
|
||||
thrust::host_vector<T> h_in(N);
|
||||
Tensor hA_in = make_tensor(recast_ptr<T>(h_in.data()), gmem_layout);
|
||||
for (int i = 0; i < size(hA_in); ++i) { hA_in(i) = static_cast<T>(i % 13); }
|
||||
|
||||
// Allocate and initialize device test data
|
||||
thrust::device_vector<T> d_in = h_in;
|
||||
thrust::device_vector<T> d_out(h_in.size(), T(-1));
|
||||
|
||||
// Launch
|
||||
int smem_size = int(sizeof(SharedStorage<T, decltype(smem_layout)>));
|
||||
test_tiled_cp_async_device_cute<<<1, 128, smem_size>>>(
|
||||
reinterpret_cast<T const*>(raw_pointer_cast(d_in.data())),
|
||||
reinterpret_cast<T*> (raw_pointer_cast(d_out.data())),
|
||||
tiled_copy,
|
||||
gmem_layout,
|
||||
smem_layout);
|
||||
|
||||
// Copy results back to host
|
||||
thrust::host_vector<T> h_out = d_out;
|
||||
Tensor hA_out = make_tensor(recast_ptr<T>(h_out.data()), gmem_layout);
|
||||
|
||||
// Validate the results. Print only the first 3 errors.
|
||||
int count = 3;
|
||||
for (int i = 0; i < size(hA_out) && count > 0; ++i) {
|
||||
EXPECT_EQ(hA_in(i), hA_out(i));
|
||||
if (hA_in(i) != hA_out(i)) {
|
||||
--count;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T, typename M, typename N, typename GMEM_STRIDE_TYPE, typename SMEM_LAYOUT, typename TILED_COPY>
|
||||
void test_cp_async_no_swizzle() {
|
||||
using namespace cute;
|
||||
auto smem_atom = SMEM_LAYOUT{};
|
||||
auto smem_layout = tile_to_shape(smem_atom, Shape<M, N>{});
|
||||
auto gmem_layout = make_layout(make_shape(M{}, N{}), GMEM_STRIDE_TYPE{});
|
||||
test_tiled_cp_async<T>(TILED_COPY{}, gmem_layout, smem_layout);
|
||||
}
|
||||
|
||||
template <typename T, typename M, typename N, typename GMEM_STRIDE_TYPE, typename SWIZZLE_ATOM, typename SMEM_LAYOUT, typename TILED_COPY>
|
||||
void test_cp_async_with_swizzle() {
|
||||
using namespace cute;
|
||||
auto swizzle_atom = SWIZZLE_ATOM{};
|
||||
auto smem_atom = composition(swizzle_atom, SMEM_LAYOUT{});
|
||||
auto smem_layout = tile_to_shape(smem_atom, Shape<M, N>{});
|
||||
auto gmem_layout = make_layout(make_shape(M{}, N{}), GMEM_STRIDE_TYPE{});
|
||||
test_tiled_cp_async<T>(TILED_COPY{}, gmem_layout, smem_layout);
|
||||
}
|
||||
@@ -37,32 +37,33 @@
|
||||
|
||||
#include <cute/container/array_subbyte.hpp>
|
||||
#include <cute/tensor.hpp>
|
||||
#include <cute/numeric/numeric_types.hpp>
|
||||
|
||||
TEST(CuTe_core, ArraySubbyte)
|
||||
{
|
||||
using namespace cute;
|
||||
{
|
||||
array_subbyte<int4_t, 10> array0;
|
||||
array_subbyte<int4_t, 5> array1;
|
||||
array_subbyte<int4_t, 10> array0{};
|
||||
array_subbyte<int4_t, 5> array1{};
|
||||
fill(array0, int4_t(0));
|
||||
fill(array1, int4_t(1));
|
||||
|
||||
for (int i = 0; i < array1.size(); ++i) {
|
||||
for (size_t i = 0; i < array1.size(); ++i) {
|
||||
array0[i+5] = array1[i];
|
||||
}
|
||||
|
||||
EXPECT_EQ(int4_t(array0.back()), int4_t(1));
|
||||
|
||||
for (int i = 0; i < array1.size(); ++i) {
|
||||
EXPECT_EQ(int4_t(array0[i]), int4_t(i / 5));
|
||||
for (size_t i = 0; i < array1.size(); ++i) {
|
||||
EXPECT_EQ(int4_t(array0[i]), int4_t(int(i) / 5));
|
||||
}
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<uint8_t, 14> a;
|
||||
array_subbyte<uint8_t, 14> a{};
|
||||
|
||||
//std::cout << sizeof_bits<decltype(a)>::value << std::endl;
|
||||
EXPECT_EQ(sizeof_bits<decltype(a)>::value, 14*8);
|
||||
EXPECT_EQ(cute::sizeof_bits_v<decltype(a)>, 14*8);
|
||||
|
||||
fill(a, uint8_t(13));
|
||||
for (int i = 0; i < int(a.size()); ++i) {
|
||||
@@ -77,10 +78,10 @@ TEST(CuTe_core, ArraySubbyte)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<int4_t, 14> a;
|
||||
array_subbyte<int4_t, 14> a{};
|
||||
|
||||
//std::cout << sizeof_bits<decltype(a)>::value << std::endl;
|
||||
EXPECT_EQ(sizeof_bits<decltype(a)>::value, 14/2*8);
|
||||
EXPECT_EQ(cute::sizeof_bits_v<decltype(a)>, 14/2*8);
|
||||
|
||||
fill(a, int4_t(-5));
|
||||
for (int i = 0; i < int(a.size()); ++i) {
|
||||
@@ -95,10 +96,10 @@ TEST(CuTe_core, ArraySubbyte)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<uint2_t, 14> a;
|
||||
array_subbyte<uint2_t, 14> a{};
|
||||
|
||||
//std::cout << sizeof_bits<decltype(a)>::value << std::endl;
|
||||
EXPECT_EQ(sizeof_bits<decltype(a)>::value, 4*8);
|
||||
EXPECT_EQ(cute::sizeof_bits_v<decltype(a)>, 4*8);
|
||||
|
||||
fill(a, uint2_t(-5));
|
||||
for (int i = 0; i < int(a.size()); ++i) {
|
||||
@@ -113,10 +114,10 @@ TEST(CuTe_core, ArraySubbyte)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<bool, 14> a;
|
||||
array_subbyte<bool, 14> a{};
|
||||
|
||||
//std::cout << sizeof_bits<decltype(a)>::value << std::endl;
|
||||
EXPECT_EQ(sizeof_bits<decltype(a)>::value, 2*8);
|
||||
EXPECT_EQ(cute::sizeof_bits_v<decltype(a)>, 2*8);
|
||||
|
||||
fill(a, bool(1));
|
||||
for (int i = 0; i < int(a.size()); ++i) {
|
||||
@@ -135,7 +136,7 @@ TEST(CuTe_core, Subbyte_iterator)
|
||||
using namespace cute;
|
||||
|
||||
{
|
||||
array_subbyte<uint8_t, 15> a;
|
||||
array_subbyte<uint8_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<uint8_t>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, uint8_t(13));
|
||||
@@ -148,7 +149,7 @@ TEST(CuTe_core, Subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<int4_t, 15> a;
|
||||
array_subbyte<int4_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<int4_t>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, int4_t(-5));
|
||||
@@ -161,7 +162,7 @@ TEST(CuTe_core, Subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<uint2_t, 15> a;
|
||||
array_subbyte<uint2_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<uint2_t>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, uint2_t(-5));
|
||||
@@ -174,7 +175,7 @@ TEST(CuTe_core, Subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<bool, 15> a;
|
||||
array_subbyte<bool, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<bool>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, bool(1));
|
||||
@@ -191,7 +192,7 @@ TEST(CuTe_core, Const_subbyte_iterator)
|
||||
using namespace cute;
|
||||
|
||||
{
|
||||
array_subbyte<uint8_t, 15> a;
|
||||
array_subbyte<uint8_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<uint8_t const>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, uint8_t(13));
|
||||
@@ -204,7 +205,7 @@ TEST(CuTe_core, Const_subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<int4_t, 15> a;
|
||||
array_subbyte<int4_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<int4_t const>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, int4_t(-5));
|
||||
@@ -217,7 +218,7 @@ TEST(CuTe_core, Const_subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<uint2_t, 15> a;
|
||||
array_subbyte<uint2_t, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<uint2_t const>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, uint2_t(-5));
|
||||
@@ -230,7 +231,7 @@ TEST(CuTe_core, Const_subbyte_iterator)
|
||||
}
|
||||
|
||||
{
|
||||
array_subbyte<bool, 15> a;
|
||||
array_subbyte<bool, 15> a{};
|
||||
auto tensor = make_tensor(subbyte_iterator<bool const>(a.raw_data()), make_shape(15));
|
||||
|
||||
fill(a, bool(1));
|
||||
|
||||
@@ -124,8 +124,8 @@ void run_and_validate(GLayout gmem_layout,
|
||||
SLayout smem_layout)
|
||||
{
|
||||
thrust::host_vector<T> h_in(cosize(gmem_layout));
|
||||
for (int32_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = T(i);
|
||||
for (size_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = static_cast<T>(int(i));
|
||||
}
|
||||
|
||||
thrust::device_vector<T> d_in = h_in;
|
||||
|
||||
@@ -106,8 +106,8 @@ void run_and_validate(GLayout gmem_layout,
|
||||
SLayout smem_layout)
|
||||
{
|
||||
thrust::host_vector<T> h_in(cosize(gmem_layout));
|
||||
for (int32_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = T(i);
|
||||
for (size_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = static_cast<T>(int(i));
|
||||
}
|
||||
|
||||
thrust::device_vector<T> d_in = h_in;
|
||||
|
||||
@@ -34,6 +34,7 @@
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include <iostream>
|
||||
#include <cstdint>
|
||||
|
||||
#include <thrust/host_vector.h>
|
||||
#include <thrust/device_vector.h>
|
||||
@@ -46,7 +47,7 @@ template <class ElementType, class SmemLayout>
|
||||
struct SharedStorage
|
||||
{
|
||||
cute::ArrayEngine<ElementType, cute::cosize_v<SmemLayout>> smem;
|
||||
cute::uint64_t tma_load_mbar[1];
|
||||
alignas(16) cute::uint64_t tma_load_mbar[1];
|
||||
};
|
||||
|
||||
#if CUDA_12_0_SM90_FEATURES_SUPPORTED
|
||||
@@ -146,14 +147,6 @@ tma_test_device_cute(T const* g_in, T* g_out,
|
||||
// Write out trivially smem -> gmem
|
||||
//
|
||||
|
||||
//if (thread0()) {
|
||||
// print_tensor(sA);
|
||||
//}
|
||||
|
||||
// for (int i = threadIdx.x; i < size(sA); i += blockDim.x) {
|
||||
// tBgB(i,stage) = sA(i);
|
||||
// }
|
||||
|
||||
// Subbyte elements could cause race conditions, so be even more conservative
|
||||
if (thread0()) {
|
||||
copy(sA, tBgB(_,stage));
|
||||
@@ -174,13 +167,15 @@ test_tma_load(CopyOp const& copy_op,
|
||||
|
||||
// Allocate and initialize host test data
|
||||
size_t N = ceil_div(cosize(gmem_layout) * sizeof_bits<T>::value, 8);
|
||||
thrust::host_vector<char> h_in(N);
|
||||
thrust::host_vector<uint8_t> h_in(N);
|
||||
for (size_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = uint8_t(i % 13);
|
||||
}
|
||||
Tensor hA_in = make_tensor(recast_ptr<T>(h_in.data()), gmem_layout);
|
||||
for (int i = 0; i < size(hA_in); ++i) { hA_in(i) = static_cast<T>(i % 13); }
|
||||
|
||||
// Allocate and initialize device test data
|
||||
thrust::device_vector<char> d_in = h_in;
|
||||
thrust::device_vector<char> d_out(h_in.size(), char(-1));
|
||||
thrust::device_vector<uint8_t> d_in = h_in;
|
||||
thrust::device_vector<uint8_t> d_out(h_in.size(), uint8_t(-1)); // overflow uint
|
||||
|
||||
// Create TMA for this device Tensor
|
||||
Tensor gA = make_tensor(make_gmem_ptr<T>(raw_pointer_cast(d_in.data())), gmem_layout);
|
||||
@@ -197,12 +192,12 @@ test_tma_load(CopyOp const& copy_op,
|
||||
smem_layout);
|
||||
|
||||
// Copy results back to host
|
||||
thrust::host_vector<char> h_out = d_out;
|
||||
thrust::host_vector<uint8_t> h_out = d_out;
|
||||
Tensor hA_out = make_tensor(recast_ptr<T>(h_out.data()), gmem_layout);
|
||||
|
||||
// Validate the results. Print only the first 3 errors.
|
||||
int count = 3;
|
||||
for (int i = 0; i < size(hA_out) && count > 0; ++i) {
|
||||
for (int i = 0; i < int(size(hA_out)) && count > 0; ++i) {
|
||||
EXPECT_EQ(hA_in(i), hA_out(i));
|
||||
if (hA_in(i) != hA_out(i)) {
|
||||
--count;
|
||||
|
||||
@@ -34,6 +34,7 @@
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include <iostream>
|
||||
#include <cstdint>
|
||||
|
||||
#include <thrust/host_vector.h>
|
||||
#include <thrust/device_vector.h>
|
||||
@@ -122,11 +123,6 @@ tma_test_device_cute(T const* g_in, T* g_out,
|
||||
//
|
||||
// Read in trivially gmem -> smem
|
||||
//
|
||||
|
||||
// for (int i = threadIdx.x; i < size(sB); i += blockDim.x) {
|
||||
// sB(i) = tAgA(i,stage);
|
||||
// }
|
||||
|
||||
// Subbyte elements could cause race conditions, so be even more conservative
|
||||
if (thread0()) {
|
||||
copy(tAgA(_,stage), sB);
|
||||
@@ -159,13 +155,15 @@ test_tma_store(CopyOp const& copy_op,
|
||||
|
||||
// Allocate and initialize host test data
|
||||
size_t N = ceil_div(cosize(gmem_layout) * sizeof_bits<T>::value, 8);
|
||||
thrust::host_vector<char> h_in(N);
|
||||
thrust::host_vector<uint8_t> h_in(N);
|
||||
for (size_t i = 0; i < h_in.size(); ++i) {
|
||||
h_in[i] = uint8_t(i % 13);
|
||||
}
|
||||
Tensor hA_in = make_tensor(recast_ptr<T>(h_in.data()), gmem_layout);
|
||||
for (int i = 0; i < size(hA_in); ++i) { hA_in(i) = static_cast<T>(i % 13); }
|
||||
|
||||
// Allocate and initialize device test data
|
||||
thrust::device_vector<char> d_in = h_in;
|
||||
thrust::device_vector<char> d_out(h_in.size(), char(-1));
|
||||
thrust::device_vector<uint8_t> d_in = h_in;
|
||||
thrust::device_vector<uint8_t> d_out(h_in.size(), uint8_t(-1)); // overflow uint
|
||||
|
||||
// Create TMA for this device Tensor
|
||||
Tensor gA = make_tensor(make_gmem_ptr<T>(raw_pointer_cast(d_out.data())), gmem_layout);
|
||||
@@ -182,12 +180,12 @@ test_tma_store(CopyOp const& copy_op,
|
||||
smem_layout);
|
||||
|
||||
// Copy results back to host
|
||||
thrust::host_vector<char> h_out = d_out;
|
||||
thrust::host_vector<uint8_t> h_out = d_out;
|
||||
Tensor hA_out = make_tensor(recast_ptr<T>(h_out.data()), gmem_layout);
|
||||
|
||||
// Validate the results. Print only the first 3 errors.
|
||||
int count = 3;
|
||||
for (int i = 0; i < size(hA_out) && count > 0; ++i) {
|
||||
for (int i = 0; i < int(size(hA_out)) && count > 0; ++i) {
|
||||
EXPECT_EQ(hA_in(i), hA_out(i));
|
||||
if (hA_in(i) != hA_out(i)) {
|
||||
--count;
|
||||
|
||||
@@ -29,4 +29,5 @@
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_cute_volta
|
||||
vectorization_auto.cu
|
||||
cooperative_copy.cu
|
||||
)
|
||||
|
||||
@@ -0,0 +1,438 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
#include "cutlass_unit_test.h"
|
||||
|
||||
#include <iostream>
|
||||
#include <iomanip>
|
||||
#include <utility>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
#include <numeric>
|
||||
#include <tuple>
|
||||
|
||||
#include <thrust/host_vector.h>
|
||||
#include <thrust/device_vector.h>
|
||||
|
||||
#include <cute/tensor.hpp>
|
||||
#include <cute/numeric/numeric_types.hpp>
|
||||
|
||||
using namespace cute;
|
||||
|
||||
namespace cooperative_copy_mode {
|
||||
struct global_shared {};
|
||||
struct global_global {};
|
||||
struct shared_shared {};
|
||||
}
|
||||
|
||||
// gs --> global to/from shared
|
||||
template <int MaxVecBits, class GMemLayout, class SMemLayout, uint32_t ThreadBlockSize, class T>
|
||||
__device__ void
|
||||
cooperative_copy_default_gs(T const* g_in, T* g_out)
|
||||
{
|
||||
using namespace cute;
|
||||
extern __shared__ float4 smem_buf[];
|
||||
// Cast smem_buf to smem_uint8_ptr and move it by MaxVecBits bits
|
||||
// This is to make sure tests pass on pointer aligned to MaxVecBits bits
|
||||
uint8_t* smem_uint8_ptr = reinterpret_cast<uint8_t*>(smem_buf) + (MaxVecBits/8);
|
||||
T* smem = reinterpret_cast<T*>(smem_uint8_ptr);
|
||||
|
||||
Tensor g_in_tensor = make_tensor(make_gmem_ptr(g_in), GMemLayout{});
|
||||
Tensor g_out_tensor = make_tensor(make_gmem_ptr(g_out), GMemLayout{});
|
||||
Tensor s_tensor = make_tensor(make_smem_ptr(smem), SMemLayout{});
|
||||
|
||||
cooperative_copy<ThreadBlockSize, MaxVecBits>(threadIdx.x, g_in_tensor, s_tensor);
|
||||
__syncthreads();
|
||||
|
||||
if(thread0()) {
|
||||
for(int i = 0; i < size(s_tensor); ++i) {
|
||||
s_tensor(i) += T(i);
|
||||
}
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
cooperative_copy<ThreadBlockSize, MaxVecBits>(threadIdx.x, s_tensor, g_out_tensor);
|
||||
}
|
||||
|
||||
// ss --> shared to shared
|
||||
template <int MaxVecBits, class Layout1, class Layout2, uint32_t ThreadBlockSize, class T>
|
||||
__device__ void
|
||||
cooperative_copy_default_ss(T const* g_in, T* g_out)
|
||||
{
|
||||
using namespace cute;
|
||||
extern __shared__ float4 smem_buf[];
|
||||
// Cast smem_buf to smem_uint8_ptr and move it by MaxVecBits bits
|
||||
// This is to make sure tests pass on pointer aligned to MaxVecBits bits
|
||||
T* smem1 = reinterpret_cast<T*>(smem_buf);
|
||||
uint8_t* smem2_uint8_ptr = reinterpret_cast<uint8_t*>(smem_buf) + (MaxVecBits/8);
|
||||
T* smem2 = reinterpret_cast<T*>(smem2_uint8_ptr) + cute::cosize(Layout2{});
|
||||
|
||||
Tensor g_in_tensor = make_tensor(make_gmem_ptr(g_in), Layout1 {});
|
||||
Tensor g_out_tensor = make_tensor(make_gmem_ptr(g_out), Layout2 {});
|
||||
|
||||
Tensor s1_tensor = make_tensor(make_smem_ptr(smem1), Layout2 {});
|
||||
Tensor s2_tensor = make_tensor(make_smem_ptr(smem2), Layout1 {});
|
||||
|
||||
cooperative_copy<ThreadBlockSize, cute::sizeof_bits_v<T>>(threadIdx.x, g_in_tensor, s1_tensor);
|
||||
__syncthreads();
|
||||
|
||||
if(thread0()) {
|
||||
for(int i = 0; i < size(s1_tensor); ++i) {
|
||||
s1_tensor(i) += T(i);
|
||||
}
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
cooperative_copy<ThreadBlockSize, MaxVecBits>(threadIdx.x, s1_tensor, s2_tensor);
|
||||
__syncthreads();
|
||||
|
||||
cooperative_copy<ThreadBlockSize, cute::sizeof_bits_v<T>>(threadIdx.x, s2_tensor, g_out_tensor);
|
||||
}
|
||||
|
||||
// gg --> global to global
|
||||
template <int MaxVecBits, class Layout1, class Layout2, uint32_t ThreadBlockSize, class T>
|
||||
__device__ void
|
||||
cooperative_copy_default_gg(T const* g_in, T* g_out)
|
||||
{
|
||||
using namespace cute;
|
||||
|
||||
Tensor g_in_tensor = make_tensor(make_gmem_ptr(g_in), Layout1{});
|
||||
Tensor g_out_tensor = make_tensor(make_gmem_ptr(g_out), Layout2{});
|
||||
|
||||
cooperative_copy<ThreadBlockSize, MaxVecBits>(threadIdx.x, g_in_tensor, g_out_tensor);
|
||||
}
|
||||
|
||||
template <class Mode, int MaxVecBits, class Layout1, class Layout2, uint32_t ThreadBlockSize, class T>
|
||||
__global__ void
|
||||
cooperative_copy_default_kernel(T const* g_in, T* g_out)
|
||||
{
|
||||
if constexpr(std::is_same_v<Mode, cooperative_copy_mode::global_shared>) {
|
||||
cooperative_copy_default_gs<MaxVecBits, Layout1, Layout2, ThreadBlockSize>(g_in, g_out);
|
||||
} else if constexpr (std::is_same_v<Mode, cooperative_copy_mode::global_global>) {
|
||||
cooperative_copy_default_gg<MaxVecBits, Layout1, Layout2, ThreadBlockSize>(g_in, g_out);
|
||||
} else if constexpr (std::is_same_v<Mode, cooperative_copy_mode::shared_shared>) {
|
||||
cooperative_copy_default_ss<MaxVecBits, Layout1, Layout2, ThreadBlockSize>(g_in, g_out);
|
||||
}
|
||||
}
|
||||
|
||||
// Mode - defines memory types of src and dst in cooperative_copy operation
|
||||
// MaxVecBits - defines max vectorization in cooperative_copy operation, and enforces that
|
||||
// alignment on used pointers to ensure correct testing
|
||||
template <class Mode, int MaxVecBits, class Layout1, class Layout2, uint32_t ThreadBlockSize, class T>
|
||||
void test_cooperative_copy_default()
|
||||
{
|
||||
using value_type = T;
|
||||
static_assert(cute::size(Layout1{}) == cute::size(Layout2{}));
|
||||
|
||||
using gmem_layout_in = Layout1;
|
||||
using gmem_layout_out = std::conditional_t<std::is_same_v<Mode, cooperative_copy_mode::global_shared>, Layout1, Layout2>;
|
||||
|
||||
#if 0
|
||||
print(" "); print("layout1: "); print(Layout1{}); print("\n");
|
||||
print(" "); print("layout2: "); print(Layout2{}); print("\n");
|
||||
print(" "); print("threads: "); print(ThreadBlockSize); print("\n");
|
||||
#endif
|
||||
|
||||
if constexpr (MaxVecBits < cute::sizeof_bits_v<value_type>) {
|
||||
GTEST_SKIP() << "Skipping test since MaxVecBits (=" << MaxVecBits
|
||||
<< ") < cute::sizeof_bits_v<value_type> (=" << cute::sizeof_bits_v<value_type> << ")";
|
||||
} else {
|
||||
constexpr auto max_vec_bytes = MaxVecBits / 8;
|
||||
static_assert((max_vec_bytes % sizeof(T)) == 0);
|
||||
|
||||
constexpr uint32_t count = cute::cosize(gmem_layout_in {});
|
||||
// Extra elements to force MaxVecBits alignment in global memory
|
||||
constexpr uint32_t extra_elements = max_vec_bytes / sizeof(value_type);
|
||||
|
||||
// Allocate
|
||||
thrust::host_vector<value_type> h_in(count + extra_elements);
|
||||
thrust::host_vector<value_type> h_out(count + extra_elements);
|
||||
|
||||
// Initialize
|
||||
Tensor h_in_tensor = make_tensor((h_in.data() + extra_elements), gmem_layout_in {});
|
||||
Tensor h_out_tensor = make_tensor((h_out.data() + extra_elements), gmem_layout_out {});
|
||||
for (int i = 0; i < cute::size(h_in_tensor); ++i) {
|
||||
h_in_tensor(i) = value_type(float(i));
|
||||
// For global-to-global copy need to compare against the same value
|
||||
h_out_tensor(i) = std::is_same_v<Mode, cooperative_copy_mode::global_global> ? value_type(float(i)) : value_type(float(2 * i));
|
||||
}
|
||||
|
||||
// To GPU
|
||||
thrust::device_vector<value_type> d_in = h_in;
|
||||
thrust::device_vector<value_type> d_out(d_in.size(), value_type(float(-2)));
|
||||
|
||||
// Adds (MaxVecBits/8) bytes to shared memory as we'll move pointer by that many bytes inside the kernel to enforce
|
||||
// alignment to (MaxVecBits/8) bytes
|
||||
size_t shared_memory_bytes = (sizeof(value_type) * count) + max_vec_bytes;
|
||||
shared_memory_bytes += std::is_same_v<Mode, cooperative_copy_mode::shared_shared> * (sizeof(value_type) * count);
|
||||
|
||||
// Launch
|
||||
auto coop_copy = cooperative_copy_default_kernel<Mode, MaxVecBits, Layout1, Layout2, ThreadBlockSize, value_type>;
|
||||
ASSERT_EQ(cudaFuncSetAttribute(coop_copy, cudaFuncAttributeMaxDynamicSharedMemorySize, static_cast<int>(shared_memory_bytes)), cudaSuccess);
|
||||
|
||||
auto d_in_ptr = thrust::raw_pointer_cast(d_in.data() + extra_elements);
|
||||
auto d_out_ptr = thrust::raw_pointer_cast(d_out.data() + extra_elements);
|
||||
coop_copy<<<1, ThreadBlockSize, shared_memory_bytes>>>(d_in_ptr, d_out_ptr);
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
if (result != cudaSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
FAIL() << "Error at kernel sync: " << cudaGetErrorString(error) << "\n";
|
||||
}
|
||||
|
||||
// Validate
|
||||
thrust::host_vector<value_type> h_result = d_out;
|
||||
Tensor h_result_tensor = make_tensor((h_result.data() + extra_elements), gmem_layout_out {});
|
||||
for (int i = 0; i < cute::size(h_in_tensor); ++i) {
|
||||
ASSERT_EQ(h_result_tensor(i), h_out_tensor(i))
|
||||
<< i << " - result:" << h_result_tensor(i) << " expected:" << h_out_tensor(i);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template<class T>
|
||||
class SM70_CuTe_Volta;
|
||||
|
||||
template<class Mode, class MaxVecBits>
|
||||
class SM70_CuTe_Volta<std::tuple<Mode, MaxVecBits>>: public testing::Test
|
||||
{
|
||||
public:
|
||||
using mode = Mode;
|
||||
static constexpr int max_vec_bits = MaxVecBits::value;
|
||||
};
|
||||
|
||||
typedef testing::Types<
|
||||
std::tuple<cooperative_copy_mode::global_shared, cute::Int<128>>,
|
||||
std::tuple<cooperative_copy_mode::global_shared, cute::Int<64>>,
|
||||
std::tuple<cooperative_copy_mode::global_shared, cute::Int<32>>,
|
||||
std::tuple<cooperative_copy_mode::global_shared, cute::Int<16>>,
|
||||
|
||||
std::tuple<cooperative_copy_mode::global_global, cute::Int<128>>,
|
||||
std::tuple<cooperative_copy_mode::global_global, cute::Int<64>>,
|
||||
std::tuple<cooperative_copy_mode::global_global, cute::Int<32>>,
|
||||
std::tuple<cooperative_copy_mode::global_global, cute::Int<16>>,
|
||||
|
||||
std::tuple<cooperative_copy_mode::shared_shared, cute::Int<128>>,
|
||||
std::tuple<cooperative_copy_mode::shared_shared, cute::Int<64>>,
|
||||
std::tuple<cooperative_copy_mode::shared_shared, cute::Int<32>>,
|
||||
std::tuple<cooperative_copy_mode::shared_shared, cute::Int<16>>,
|
||||
> CooperativeCopyModeMaxVecBitsList;
|
||||
|
||||
TYPED_TEST_SUITE(SM70_CuTe_Volta, CooperativeCopyModeMaxVecBitsList);
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefault1D)
|
||||
{
|
||||
using value_type = float;
|
||||
constexpr uint32_t count = 512;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<count>{})));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<count>{})));
|
||||
constexpr uint32_t thread_block_size = 64;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSG2D)
|
||||
{
|
||||
using value_type = float;
|
||||
constexpr uint32_t x = 32;
|
||||
constexpr uint32_t y = 32;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{})));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{})));
|
||||
constexpr uint32_t thread_block_size = 64;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSG2DCustomStride)
|
||||
{
|
||||
using value_type = float;
|
||||
constexpr uint32_t x = 16;
|
||||
constexpr uint32_t y = 16;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{}), make_stride(Int<y>{}, Int<1>{})));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{}), make_stride(Int<1>{}, Int<x>{})));
|
||||
constexpr uint32_t thread_block_size = 64;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSG3D)
|
||||
{
|
||||
using value_type = cute::half_t;
|
||||
constexpr uint32_t x = 8;
|
||||
constexpr uint32_t y = 8;
|
||||
constexpr uint32_t z = 16;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{}, Int<z>{})));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y>{}, Int<z>{})));
|
||||
constexpr uint32_t thread_block_size = 64;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSG2Dto3D)
|
||||
{
|
||||
using value_type = double;
|
||||
constexpr uint32_t x = 16;
|
||||
constexpr uint32_t y = 16;
|
||||
constexpr uint32_t z = 4;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<x>{}, Int<y*z>{})));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<z>{}, Int<y>{}, Int<x>{})));
|
||||
constexpr uint32_t thread_block_size = 64;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGCustom1)
|
||||
{
|
||||
using value_type = double;
|
||||
using gmem_layout_t = decltype(make_layout(
|
||||
make_shape(Int<8>{}, make_shape(Int<2>{}, Int<2>{})),
|
||||
make_stride(Int<2>{}, make_shape(Int<1>{}, Int<16>{}))
|
||||
));
|
||||
using smem_layout_t = decltype(make_layout(
|
||||
make_shape(Int<8>{}, Int<4>{}),
|
||||
make_stride(Int<4>{}, Int<1>{})
|
||||
));
|
||||
constexpr uint32_t thread_block_size = 8;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGCustom2)
|
||||
{
|
||||
using value_type = float;
|
||||
using gmem_layout_t = decltype(make_layout(
|
||||
make_shape(make_shape(Int<4>{}, Int<2>{}), make_shape(Int<2>{}, Int<2>{})),
|
||||
make_stride(make_shape(Int<4>{}, Int<1>{}), make_shape(Int<16>{}, Int<2>{}))
|
||||
));
|
||||
using smem_layout_t = decltype(make_layout(
|
||||
make_shape(make_shape(Int<2>{}, Int<2>{}, Int<2>{}), make_shape(Int<2>{}, Int<2>{})),
|
||||
make_stride(make_shape(Int<16>{}, Int<4>{}, Int<1>{}), make_shape(Int<8>{}, Int<2>{}))
|
||||
));
|
||||
constexpr uint32_t thread_block_size = 16;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGSwizzle1)
|
||||
{
|
||||
using value_type = float;
|
||||
using gmem_layout_t = Layout<Shape<_8, _64>, Stride<_64, _1>>;
|
||||
using smem_layout_t = decltype(composition(Swizzle<3, 3, 3>{}, Layout<Shape<_8, _64>, Stride<_64, _1>>{}));
|
||||
constexpr uint32_t thread_block_size = 128;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGSwizzle2)
|
||||
{
|
||||
using value_type = cute::half_t;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<64>{}, Int<64>{})));
|
||||
using smem_atom_layout_t = decltype(composition(Swizzle<3, 2, 3> {}, Layout<Shape<_8, _32>, Stride<_32, _1>>{}));
|
||||
using smem_layout_t = decltype(tile_to_shape(
|
||||
smem_atom_layout_t{},
|
||||
make_shape(shape<0>(gmem_layout_t{}), shape<1>(gmem_layout_t{})))
|
||||
);
|
||||
constexpr uint32_t thread_block_size = 128;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGSwizzle3)
|
||||
{
|
||||
using value_type = cute::half_t;
|
||||
using gmem_layout_t = decltype(make_layout(make_shape(Int<64>{}, Int<64>{})));
|
||||
using smem_atom_layout_t = decltype(composition(Swizzle<2, 4, 3> {}, Layout<Shape<_16, _64>, Stride<_64, _1>>{}));
|
||||
using smem_layout_t = decltype(tile_to_shape(
|
||||
smem_atom_layout_t{},
|
||||
make_shape(shape<0>(gmem_layout_t{}), shape<1>(gmem_layout_t{})))
|
||||
);
|
||||
constexpr uint32_t thread_block_size = 128;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
|
||||
TYPED_TEST(SM70_CuTe_Volta, CooperativeCopyDefaultGSSGSwizzle4)
|
||||
{
|
||||
using value_type = cute::half_t;
|
||||
using gmem_atom_layout_t = decltype(composition(Swizzle<3, 2, 3> {}, Layout<Shape<_8, _32>, Stride<_32, _1>>{}));
|
||||
using smem_layout_t = decltype(make_layout(make_shape(Int<64>{}, Int<64>{})));
|
||||
using gmem_layout_t = decltype(tile_to_shape(
|
||||
gmem_atom_layout_t{},
|
||||
make_shape(shape<0>(smem_layout_t{}), shape<1>(smem_layout_t{})))
|
||||
);
|
||||
constexpr uint32_t thread_block_size = 128;
|
||||
test_cooperative_copy_default<typename TestFixture::mode,
|
||||
TestFixture::max_vec_bits,
|
||||
gmem_layout_t,
|
||||
smem_layout_t,
|
||||
thread_block_size,
|
||||
value_type>();
|
||||
}
|
||||
+21
@@ -0,0 +1,21 @@
|
||||
conv1d dgrad_(1,8,64)_(64,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2903348270 3426051534 1688878159 1937506630
|
||||
conv1d dgrad_(1,8,64)_(64,1,16)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2903348270 4139873761 674035563 2969043798
|
||||
conv1d dgrad_(2,8,64)_(64,1,96)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 1021860383 1204062629 320546932 1467678853
|
||||
conv1d dgrad_(7,8,64)_(64,1,256)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 1658584634 3542030567 2895019148 3285129185
|
||||
conv1d dgrad_(2,6,64)_(64,3,256)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2768091490 3498596809 1298679529 150156903
|
||||
conv1d dgrad_(2,8,32)_(32,3,256)_padl(1)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2903348270 1317709417 1298679529 2499651710
|
||||
conv1d dgrad_(2,6,64)_(64,4,256)_padl(0)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2768091490 2047019225 1298679529 2646233918
|
||||
conv1d dgrad_(2,13,256)_(256,3,64)_padl(0)_padu(1)_str(1)_dil(2)_corr_alpha1_beta0 h_h_h_h 2818970884 3498596809 626177484 3228881313
|
||||
conv2d dgrad_(1,8,8,64)_(64,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1260538290 3426051534 1298679529 885253342
|
||||
conv2d dgrad_(1,8,8,64)_(64,1,1,16)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1260538290 4139873761 2377544542 1796715273
|
||||
conv2d dgrad_(2,8,8,64)_(64,1,1,96)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 913918101 1204062629 2917884534 3849387550
|
||||
conv2d dgrad_(7,8,8,64)_(64,1,1,256)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1794462925 3542030567 2009277002 1494590621
|
||||
conv2d dgrad_(2,6,6,64)_(64,3,3,256)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 3783509692 1508460534 184863710 2539209373
|
||||
conv2d dgrad_(2,8,8,32)_(32,3,3,256)_padl(1,1)_padu(1,1)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1260538290 19124920 184863710 310844165
|
||||
conv2d dgrad_(2,8,5,64)_(64,2,5,256)_padl(1,1)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1829721897 2657646585 184863710 1493692231
|
||||
conv2d dgrad_(2,15,5,256)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(1,1)_dil(2,3)_corr_alpha1_beta0 h_h_h_h 2554054572 2657646585 184863710 732073700
|
||||
conv3d dgrad_(1,1,8,8,64)_(64,1,1,1,16)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 1260538290 4139873761 2377544542 2164020490
|
||||
conv3d dgrad_(2,1,8,8,64)_(64,1,1,1,96)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 913918101 1204062629 2917884534 3849387550
|
||||
conv3d dgrad_(2,3,4,6,64)_(64,3,4,5,96)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 3168426837 2541149106 3798069020 2802887403
|
||||
conv3d dgrad_(2,2,4,5,64)_(64,3,4,5,96)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 1829721897 2541149106 3798069020 2514979496
|
||||
conv3d dgrad_(2,13,6,5,64)_(64,3,4,5,96)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(2,2,3)_corr_alpha1_beta0 h_h_h_h 3720598515 2541149106 2799255748 1382638415
|
||||
+87
@@ -0,0 +1,87 @@
|
||||
conv1d fprop_(1,8,64)_(64,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 2230797104 291412341 2967010883 3359645502
|
||||
conv1d fprop_(1,8,64)_(16,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 2230797104 2455330259 2158037455 380099506
|
||||
conv1d fprop_(2,8,64)_(96,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 3337558967 3397548888 1959374308 4192617409
|
||||
conv1d fprop_(7,8,64)_(256,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 2412446234 659604944 754742037 743464550
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 3337558967 2907034004 968493294 303241124
|
||||
conv1d fprop_(2,8,32)_(256,3,32)_padl(1)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 2230797104 684772342 2456226145 4051186022
|
||||
conv1d fprop_(2,8,64)_(256,4,64)_padl(0)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 3337558967 2257694519 968493294 2992278890
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(2)_dil(1)_corr_alpha1_beta0 s8_s8_s8_i 3337558967 2907034004 3833157308 197395388
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(1)_dil(2)_corr_alpha1_beta0 s8_s8_s8_i 3337558967 2907034004 2965816260 2146305550
|
||||
conv2d fprop_(1,8,8,64)_(64,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 3696905065 291412341 2456226145 222634564
|
||||
conv2d fprop_(1,8,8,64)_(16,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 3696905065 2455330259 2674438810 1247569549
|
||||
conv2d fprop_(2,8,8,64)_(96,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 589562497 3397548888 3067871195 4124984848
|
||||
conv2d fprop_(7,8,8,64)_(256,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 3050405131 659604944 217342478 2681355190
|
||||
conv2d fprop_(2,8,8,64)_(256,3,3,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 589562497 1759432517 1946286282 4189881018
|
||||
conv2d fprop_(2,8,8,32)_(256,3,3,32)_padl(1,1)_padu(1,1)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 3696905065 3664146697 1273468542 3134139003
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(2,2)_str(1,1)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 589562497 917721283 2404731639 151314389
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(1,1)_corr_alpha1_beta0 s8_s8_s8_i 589562497 917721283 2456226145 1566464786
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(1,1)_dil(2,3)_corr_alpha1_beta0 s8_s8_s8_i 2290515358 917721283 1403142369 3451674023
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(2,3)_corr_alpha1_beta0 s8_s8_s8_i 2290515358 917721283 1240515274 3524645563
|
||||
conv3d fprop_(1,1,8,8,64)_(64,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 3696905065 291412341 2456226145 222634564
|
||||
conv3d fprop_(1,1,8,8,64)_(16,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 3696905065 2455330259 2674438810 1247569549
|
||||
conv3d fprop_(2,1,8,8,64)_(96,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 589562497 3397548888 3067871195 4124984848
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,3,3,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 2665613442 3049451099 844780218 2240615929
|
||||
conv3d fprop_(2,3,5,8,32)_(96,3,3,3,32)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 2957859809 2989051488 3910298589 2663735744
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 2665613442 1144053949 2036439885 2424293561
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 2665613442 1144053949 3855271504 1232523257
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(1,1,1)_corr_alpha1_beta0 s8_s8_s8_i 3404775028 1144053949 1403142369 2248157699
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(2,2,3)_corr_alpha1_beta0 s8_s8_s8_i 3404775028 1144053949 374107293 1044176262
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(2,2,3)_corr_alpha1_beta0 s8_s8_s8_i 3404775028 1144053949 3659031110 1946240511
|
||||
conv1d fprop_(1,8,64)_(64,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 2903348270 3426051534 1688878159 1544142545
|
||||
conv1d fprop_(1,8,64)_(16,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 2903348270 4139873761 674035563 451236625
|
||||
conv1d fprop_(2,8,64)_(96,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 1021860383 1204062629 320546932 3992938435
|
||||
conv1d fprop_(7,8,64)_(256,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 1658584634 3542030567 2895019148 2225783091
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 1021860383 3498596809 1715902861 2177803653
|
||||
conv1d fprop_(2,8,32)_(256,3,32)_padl(1)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 2903348270 1317709417 1298679529 4040627173
|
||||
conv1d fprop_(2,8,64)_(256,4,64)_padl(0)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_f 1021860383 2047019225 1715902861 2671813939
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(2)_dil(1)_corr_alpha1_beta0 h_h_h_f 1021860383 3498596809 626177484 2917692677
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(1)_dil(2)_corr_alpha1_beta0 h_h_h_f 1021860383 3498596809 2486710374 3994207567
|
||||
conv2d fprop_(1,8,8,64)_(64,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 1260538290 3426051534 1298679529 99551275
|
||||
conv2d fprop_(1,8,8,64)_(16,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 1260538290 4139873761 2377544542 1503522773
|
||||
conv2d fprop_(2,8,8,64)_(96,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 913918101 1204062629 2917884534 2239833744
|
||||
conv2d fprop_(7,8,8,64)_(256,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 1794462925 3542030567 2009277002 189498806
|
||||
conv2d fprop_(2,8,8,64)_(256,3,3,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 913918101 1508460534 3907625767 968673523
|
||||
conv2d fprop_(2,8,8,32)_(256,3,3,32)_padl(1,1)_padu(1,1)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 1260538290 19124920 184863710 656273881
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(2,2)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 913918101 2657646585 1969539841 646972978
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(1,1)_corr_alpha1_beta0 h_h_h_f 913918101 2657646585 1298679529 3394877772
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(1,1)_dil(2,3)_corr_alpha1_beta0 h_h_h_f 1788526765 2657646585 3112324813 3072386268
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(2,3)_corr_alpha1_beta0 h_h_h_f 1788526765 2657646585 1911791555 3769534326
|
||||
conv3d fprop_(1,1,8,8,64)_(64,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 1260538290 3426051534 1298679529 99551275
|
||||
conv3d fprop_(1,1,8,8,64)_(16,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 1260538290 4139873761 2377544542 1503522773
|
||||
conv3d fprop_(2,1,8,8,64)_(96,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 913918101 1204062629 2917884534 2239833744
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,3,3,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 2536233239 3623022813 3183260694 2444914935
|
||||
conv3d fprop_(2,3,5,8,32)_(96,3,3,3,32)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 3533222980 3319027099 3798069020 1596009643
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 2536233239 2541149106 3816338811 3343435781
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 2536233239 2541149106 23174956 4214553739
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_f 438040955 2541149106 3112324813 184240726
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(2,2,3)_corr_alpha1_beta0 h_h_h_f 438040955 2541149106 1729700443 1845241883
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(2,2,3)_corr_alpha1_beta0 h_h_h_f 438040955 2541149106 2215856509 1015378858
|
||||
conv1d fprop_(1,8,64)_(64,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 3457276160 1628467649 1540949872 2136373177
|
||||
conv1d fprop_(1,8,64)_(16,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 3457276160 3090183904 3899451983 3023937644
|
||||
conv1d fprop_(2,8,64)_(96,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 1880214374 3677150560 310726324 4149087957
|
||||
conv1d fprop_(7,8,64)_(256,1,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 11721734 2875614321 876330871 1062881547
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 1880214374 354521127 3224870611 2936608625
|
||||
conv1d fprop_(2,8,32)_(256,3,32)_padl(1)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 3457276160 3940958820 592544256 281505365
|
||||
conv1d fprop_(2,8,64)_(256,4,64)_padl(0)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 f_f_f_f 1880214374 2370127693 3224870611 3916432513
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(2)_dil(1)_corr_alpha1_beta0 f_f_f_f 1880214374 354521127 3565144715 3108552041
|
||||
conv1d fprop_(2,8,64)_(256,3,64)_padl(0)_padu(1)_str(1)_dil(2)_corr_alpha1_beta0 f_f_f_f 1880214374 354521127 1319667750 2636132047
|
||||
conv2d fprop_(1,8,8,64)_(64,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 3453982556 1628467649 592544256 1519969106
|
||||
conv2d fprop_(1,8,8,64)_(16,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 3453982556 3090183904 930148554 3904697107
|
||||
conv2d fprop_(2,8,8,64)_(96,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 854950536 3677150560 1408652759 1475171676
|
||||
conv2d fprop_(7,8,8,64)_(256,1,1,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 1066744638 2875614321 2252791511 3400916910
|
||||
conv2d fprop_(2,8,8,64)_(256,3,3,64)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 854950536 795739645 1545596299 4043117388
|
||||
conv2d fprop_(2,8,8,32)_(256,3,3,32)_padl(1,1)_padu(1,1)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 3453982556 2561845304 1899218706 1185670554
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(2,2)_str(1,1)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 854950536 4046266669 893484124 4189127440
|
||||
conv2d fprop_(2,8,8,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(1,1)_corr_alpha1_beta0 f_f_f_f 854950536 4046266669 592544256 2322571289
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(1,1)_dil(2,3)_corr_alpha1_beta0 f_f_f_f 191074901 4046266669 1081688479 1745547693
|
||||
conv2d fprop_(2,16,16,64)_(256,2,5,64)_padl(1,1)_padu(0,0)_str(2,3)_dil(2,3)_corr_alpha1_beta0 f_f_f_f 191074901 4046266669 1031384344 1016941631
|
||||
conv3d fprop_(1,1,8,8,64)_(64,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 3453982556 1628467649 592544256 1519969106
|
||||
conv3d fprop_(1,1,8,8,64)_(16,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 3453982556 3090183904 930148554 3904697107
|
||||
conv3d fprop_(2,1,8,8,64)_(96,1,1,1,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 854950536 3677150560 1408652759 1475171676
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,3,3,64)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 1304210401 3269133702 2815004824 1363667010
|
||||
conv3d fprop_(2,3,5,8,32)_(96,3,3,3,32)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 1689793983 2561137151 3947985101 3903813688
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 1304210401 2276316058 1262249481 1546938459
|
||||
conv3d fprop_(2,3,5,8,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 1304210401 2276316058 288977156 2430714665
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(1,1,1)_corr_alpha1_beta0 f_f_f_f 4289052901 2276316058 1081688479 1271132150
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(2,2,3)_corr_alpha1_beta0 f_f_f_f 4289052901 2276316058 3305425723 749517833
|
||||
conv3d fprop_(2,16,10,16,64)_(96,3,4,5,64)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(2,2,3)_corr_alpha1_beta0 f_f_f_f 4289052901 2276316058 304972512 1722195827
|
||||
+23
@@ -0,0 +1,23 @@
|
||||
conv1d wgrad_(1,8,64)_(1,8,128)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2903348270 4139873761 1911791555 2094311858
|
||||
conv1d wgrad_(1,8,16)_(1,8,128)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 3472074346 4139873761 626177484 4014899436
|
||||
conv1d wgrad_(2,8,96)_(2,8,128)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 2790635024 2105873592 2917884534 4063149047
|
||||
conv1d wgrad_(7,8,256)_(7,8,128)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 37621077 58406568 184863710 3825831917
|
||||
conv1d wgrad_(2,6,256)_(2,8,128)_padl(0)_padu(0)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 3284272669 2105873592 1352114402 170044063
|
||||
conv1d wgrad_(2,8,256)_(2,8,128)_padl(1)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 1260538290 2105873592 1352114402 1054386746
|
||||
conv1d wgrad_(2,6,256)_(2,8,128)_padl(0)_padu(1)_str(1)_dil(1)_corr_alpha1_beta0 h_h_h_h 3284272669 2105873592 4149848551 3241398662
|
||||
conv1d wgrad_(2,4,256)_(2,8,128)_padl(0)_padu(1)_str(2)_dil(1)_corr_alpha1_beta0 h_h_h_h 2501680779 2105873592 1352114402 2277261527
|
||||
conv1d wgrad_(2,5,256)_(2,8,128)_padl(0)_padu(1)_str(1)_dil(2)_corr_alpha1_beta0 h_h_h_h 3472351235 2105873592 1352114402 1619048523
|
||||
conv2d wgrad_(1,8,8,64)_(1,8,8,128)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1260538290 2039847788 1911791555 1010376807
|
||||
conv2d wgrad_(1,8,8,16)_(1,8,8,128)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1021860383 2039847788 626177484 4023661355
|
||||
conv2d wgrad_(2,8,8,96)_(2,8,8,128)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1661538138 3542030567 2917884534 798509586
|
||||
conv2d wgrad_(7,8,8,256)_(7,8,8,128)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 854475261 2837606078 184863710 2833795023
|
||||
conv2d wgrad_(2,6,6,256)_(2,8,8,128)_padl(0,0)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 269005094 3542030567 587374223 722686340
|
||||
conv2d wgrad_(2,8,8,256)_(2,8,8,128)_padl(1,1)_padu(1,1)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1788526765 3542030567 587374223 3596796540
|
||||
conv2d wgrad_(2,8,5,256)_(2,8,8,128)_padl(1,1)_padu(0,0)_str(1,1)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1986808706 3542030567 2799255748 3793023440
|
||||
conv2d wgrad_(2,8,5,256)_(2,16,16,128)_padl(1,1)_padu(0,0)_str(2,3)_dil(1,1)_corr_alpha1_beta0 h_h_h_h 1986808706 2047019225 2799255748 1869700538
|
||||
conv2d wgrad_(2,15,5,256)_(2,16,16,128)_padl(1,1)_padu(0,0)_str(1,1)_dil(2,3)_corr_alpha1_beta0 h_h_h_h 2554054572 2047019225 2799255748 1061991327
|
||||
conv2d wgrad_(2,8,2,256)_(2,16,16,128)_padl(1,1)_padu(0,0)_str(2,3)_dil(2,3)_corr_alpha1_beta0 h_h_h_h 913918101 2047019225 2799255748 1723667339
|
||||
conv3d wgrad_(1,1,8,8,16)_(1,1,8,8,128)_padl(0,0,0)_padu(0,0,0)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 1021860383 2039847788 626177484 1510320256
|
||||
conv3d wgrad_(2,3,4,6,96)_(2,3,5,8,128)_padl(1,1,1)_padu(1,1,1)_str(1,1,1)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 2885338846 4129534099 955078878 481515417
|
||||
conv3d wgrad_(2,8,5,5,96)_(2,16,10,16,128)_padl(1,0,1)_padu(0,2,0)_str(2,2,3)_dil(1,1,1)_corr_alpha1_beta0 h_h_h_h 2554054572 1879264901 955078878 3349975418
|
||||
conv3d wgrad_(2,13,6,5,96)_(2,16,10,16,128)_padl(1,0,1)_padu(0,2,0)_str(1,1,1)_dil(2,2,3)_corr_alpha1_beta0 h_h_h_h 698054658 1879264901 955078878 4113825945
|
||||
@@ -133,29 +133,11 @@ __global__ void epilogue_with_reduction_threadblock(
|
||||
// For debugging, enable this block of code to fill each accumulator element with its
|
||||
// source thread ID.
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulators.size(); ++i) {
|
||||
for (size_t i = 0; i < accumulators.size(); ++i) {
|
||||
typename Epilogue::WarpMmaOperator::ElementC x(threadIdx.x);
|
||||
//typename Epilogue::WarpMmaOperator::ElementC x(i);
|
||||
accumulators[i] = x;
|
||||
}
|
||||
|
||||
/*
|
||||
#pragma unroll 1
|
||||
for (int tid = 0; tid < 32; ++tid) {
|
||||
if (tid == thread_idx) {
|
||||
printf("\nT%d: ", thread_idx);
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulators.size(); ++i) {
|
||||
printf("%d ", int(accumulators[i]));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (thread_idx == 0) {
|
||||
printf("\n\n");
|
||||
}
|
||||
*/
|
||||
|
||||
__syncthreads();
|
||||
|
||||
#endif
|
||||
|
||||
@@ -122,29 +122,11 @@ __global__ void epilogue_threadblock(
|
||||
// For debugging, enable this block of code to fill each accumulator element with its
|
||||
// source thread ID.
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulators.size(); ++i) {
|
||||
for (size_t i = 0; i < accumulators.size(); ++i) {
|
||||
typename Epilogue::WarpMmaOperator::ElementC x(threadIdx.x);
|
||||
//typename Epilogue::WarpMmaOperator::ElementC x(i);
|
||||
accumulators[i] = x;
|
||||
}
|
||||
|
||||
/*
|
||||
#pragma unroll 1
|
||||
for (int tid = 0; tid < 32; ++tid) {
|
||||
if (tid == thread_idx) {
|
||||
printf("\nT%d: ", thread_idx);
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulators.size(); ++i) {
|
||||
printf("%d ", int(accumulators[i]));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (thread_idx == 0) {
|
||||
printf("\n\n");
|
||||
}
|
||||
*/
|
||||
|
||||
__syncthreads();
|
||||
|
||||
#endif
|
||||
|
||||
@@ -93,8 +93,8 @@ TEST(SM70_Epilogue_warp_FragmentIterator, mma_f16_64x64x4) {
|
||||
typename MmaTensorOp::IteratorC::Fragment accumulator_tile;
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulator_tile.size(); ++i) {
|
||||
accumulator_tile[i] = ElementC(i);
|
||||
for (size_t i = 0; i < accumulator_tile.size(); ++i) {
|
||||
accumulator_tile[i] = static_cast<ElementC>(int(i));
|
||||
}
|
||||
|
||||
using FragmentIterator = cutlass::epilogue::warp::FragmentIteratorVoltaTensorOp<
|
||||
@@ -114,7 +114,7 @@ TEST(SM70_Epilogue_warp_FragmentIterator, mma_f16_64x64x4) {
|
||||
|
||||
#if 0
|
||||
std::cout << "T" << tid << ": ";
|
||||
for (int i = 0; i < frag.size(); ++i) {
|
||||
for (size_t i = 0; i < frag.size(); ++i) {
|
||||
std::cout << " " << frag[i];
|
||||
}
|
||||
std::cout << std::endl;
|
||||
@@ -169,8 +169,8 @@ TEST(SM70_Epilogue_warp_FragmentIterator, mma_f32_64x64x4) {
|
||||
typename MmaTensorOp::IteratorC::Fragment accumulator_tile;
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < accumulator_tile.size(); ++i) {
|
||||
accumulator_tile[i] = ElementC(i);
|
||||
for (size_t i = 0; i < accumulator_tile.size(); ++i) {
|
||||
accumulator_tile[i] = static_cast<ElementC>(i);
|
||||
}
|
||||
|
||||
typename MmaTensorOp::IteratorC iterator_C(accumulator_tensor.host_ref(), tid);
|
||||
|
||||
@@ -131,7 +131,6 @@ cutlass_test_unit_add_executable(
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
|
||||
sm50_gemm_f32_f32_f32_simt.cu
|
||||
sm80_gemm_f32_f32_f32_simt.cu
|
||||
sm50_gemm_f64_f64_f64_simt.cu
|
||||
@@ -146,7 +145,6 @@ cutlass_test_unit_add_executable(
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
|
||||
gemm_f16n_f16n_f32t_volta_tensor_op_f32_sm70.cu
|
||||
gemm_f16n_f16t_f32t_volta_tensor_op_f32_sm70.cu
|
||||
gemm_f16t_f16n_f32t_volta_tensor_op_f32_sm70.cu
|
||||
@@ -170,6 +168,7 @@ cutlass_test_unit_add_executable(
|
||||
gemm_universal_f16n_f16t_f32t_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_sm75.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f32_sm75.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_sm75.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_slicedk_sm75.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_slicedk_sm75.cu
|
||||
@@ -224,8 +223,10 @@ cutlass_test_unit_add_executable(
|
||||
gemm_f16n_f16n_f32n_tensor_op_f32_sm80.cu
|
||||
gemm_f16n_f16n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f32_sm80.cu
|
||||
gemm_f16n_f16t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16t_f32n_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16t_f32t_tensor_op_f32_sm80.cu
|
||||
@@ -290,6 +291,7 @@ cutlass_test_unit_add_executable(
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
sm90_gemm_f16_f16_f16_alignx_tensor_op_f32.cu
|
||||
sm90_gemm_f16_f16_f16_alignx_tensor_op_f32_warpspecialized.cu
|
||||
sm90_gemm_f16_f16_f16_alignx_tensor_op_f32_warpspecialized_cooperative.cu
|
||||
@@ -314,6 +316,7 @@ cutlass_test_unit_add_executable(
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
sm90_gemm_f16_f16_f16_tensor_op_f32_tensor_broadcast.cu
|
||||
sm90_gemm_f32_f32_f32_tensor_op_f32_tensor_broadcast.cu
|
||||
sm90_gemm_s8_s8_s8_tensor_op_s32_tensor_broadcast.cu
|
||||
@@ -328,6 +331,10 @@ cutlass_test_unit_add_executable(
|
||||
sm90_gemm_f16_f16_f16_tensor_op_f32_cluster_warpspecialized_cooperative_dag.cu
|
||||
sm90_gemm_f16_f16_f16_tensor_op_f32_cluster_warpspecialized_pingpong_dag.cu
|
||||
sm90_gemm_f16_f16_f16_tensor_op_f32_cluster_warpspecialized_cooperative_aux_store.cu
|
||||
# Fp8
|
||||
sm90_gemm_f8_f8_f8_tensor_op_fp32_evt.cu
|
||||
sm90_gemm_f8_f8_bf16_tensor_op_fp32_evt.cu
|
||||
sm90_gemm_f8_f8_f32_tensor_op_f32_cluster_warpspecialized_cooperative_evt.cu
|
||||
)
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_cluster_multicast_sm90
|
||||
@@ -341,6 +348,7 @@ cutlass_test_unit_add_executable(
|
||||
sm90_gemm_f16_f16_f16_tensor_op_f32_cluster_warpspecialized_cooperative.cu
|
||||
sm90_gemm_f8_f8_f32_tensor_op_f32_cluster_warpspecialized_cooperative.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_gmma_rs_warpspecialized_sm90
|
||||
|
||||
@@ -490,6 +498,19 @@ cutlass_test_unit_add_executable(
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm75.cu
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm80.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_sm89
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_f8t_f8n_f32t_tensor_op_f32_sm89.cu
|
||||
gemm_f8t_f8n_f32t_tensor_op_f32_sparse_sm89.cu
|
||||
gemm_f8t_f8n_f8t_tensor_op_f32_sm89.cu
|
||||
gemm_f8t_f8n_f8t_tensor_op_f32_sparse_sm89.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_grouped
|
||||
|
||||
@@ -538,7 +559,6 @@ cutlass_test_unit_add_executable(
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sparse_sm80.cu
|
||||
)
|
||||
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemv_device
|
||||
|
||||
@@ -784,3 +804,4 @@ cutlass_test_unit_add_executable(
|
||||
)
|
||||
|
||||
endif()
|
||||
|
||||
|
||||
@@ -238,6 +238,253 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 32x128x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 32x256x32_32x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 256, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x32x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x32x32_64x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x128x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x256x32_32x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 256, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 128x32x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 256x32x32_64x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif
|
||||
|
||||
@@ -37,7 +37,7 @@
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/gemm/device/gemm_universal.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
@@ -54,10 +54,10 @@
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmUniversal_f16n_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
@@ -68,14 +68,239 @@ TEST(SM80_Device_GemmUniversal_f16n_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32)
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 10>;
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x128x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x256x32_32x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 128x32x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 256x32x32_64x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x128x64_32x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 32x256x64_32x64x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 128x32x64_32x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 256x32x64_64x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 16x128x64_16x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 128, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 16x256x64_16x64x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 256, 64>,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 128x16x64_32x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 16, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f32, 256x16x64_64x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 16, 64>,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 5>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
@@ -0,0 +1,398 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x32x32_16x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x32_16x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x32_16x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<16, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x32_32x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x256x32_32x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 256, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x32_32x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x32_64x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x32_16x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 256x32x32_64x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif
|
||||
@@ -0,0 +1,924 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for device-wide GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 16x64x64_16x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 16x128x64_16x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 128, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 16x256x64_16x64x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 256, 64>,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x16x64_16x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x16x64_32x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 16, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 256x16x64_64x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 16, 64>,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x32x32_16x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x32_16x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x32_16x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<16, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x32_32x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x256x32_32x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x32_32x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x32_64x16x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 16, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x32_16x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x32_32x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 256x32x32_64x32x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x32x64_16x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x64_16x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x64_16x64x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x64x64_32x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x128x64_32x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 32x256x64_32x64x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x64_32x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x64_64x16x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 64x32x64_16x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 128x32x64_32x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f32, 256x32x64_64x32x64) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif
|
||||
@@ -0,0 +1,154 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide GEMM interface with:
|
||||
A: row major, of type FE4M4 or FE5M2
|
||||
B: column major, of type FE4M3 or FE5M2
|
||||
C: row major, of type F32
|
||||
Accum: F32
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM89_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe5m2n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe5m2t_fe4m3n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe5m2t_fe5m2n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM89_SUPPORTED
|
||||
@@ -0,0 +1,154 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide sparse GEMM interface with:
|
||||
A: row major, of type FE4M4 or FE5M2
|
||||
B: column major, of type FE4M3 or FE5M2
|
||||
C: row major, of type F32
|
||||
Accum: F32
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_sparse.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_sparse.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM89_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_f32t_tensor_op_f32, 128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllSparseGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe5m2n_f32t_tensor_op_f32, 128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllSparseGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe5m2t_fe4m3n_f32t_tensor_op_f32, 128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllSparseGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe5m2t_fe5m2n_f32t_tensor_op_f32, 128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemm<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllSparseGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM89_SUPPORTED
|
||||
@@ -0,0 +1,430 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide GEMM interface with:
|
||||
A: row major, of type FE4M4 or FE5M2
|
||||
B: column major, of type FE4M3 or FE5M2
|
||||
C: row major, of FE4M3 or FE5M2
|
||||
Accum: F32
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/epilogue/thread/activation.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_generic_with_scaling.h"
|
||||
#include "cutlass/gemm/device/gemm_universal_with_absmax.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
#include "testbed_with_absmax.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM89_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_fastacc_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
static int const kAlignment = 16;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages,
|
||||
kAlignment, kAlignment, cutlass::arch::OpMultiplyAddFastAccum
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, relu_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::ReLu,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::ReLu>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe5m2n_fe4m3t_tensor_op_f32, identity_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe5m2t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe5m2t_fe5m2n_fe4m3t_tensor_op_f32, identity_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe5m2t_tensor_op_f32, identity_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e5m2_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe5m2t_fe5m2n_fe5m2t_tensor_op_f32, identity_diff_aux_output_types_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = cutlass::float_e5m2_t;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x128x64_32x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>, cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_noScale_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>(
|
||||
/* scaleA = */false,
|
||||
/* scaleB = */false,
|
||||
/* scaleC = */false
|
||||
);
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_noAux_128x256x64_64x64x64) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalWithAbsMax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>, cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::Testbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM89_SUPPORTED
|
||||
@@ -0,0 +1,430 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
||||
* SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice, this
|
||||
* list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
||||
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
||||
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
||||
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
||||
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
||||
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/*! \file
|
||||
\brief Tests for device-wide sparse GEMM interface with:
|
||||
A: row major, of type FE4M4 or FE5M2
|
||||
B: column major, of type FE4M3 or FE5M2
|
||||
C: row major, of FE4M3 or FE5M2
|
||||
Accum: F32
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/epilogue/thread/activation.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination_generic_with_scaling.h"
|
||||
#include "cutlass/gemm/device/gemm_sparse_with_absmax.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_sparse.h"
|
||||
#include "testbed_with_absmax.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM89_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_fastacc_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
static int const kAlignment = 16;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages,
|
||||
kAlignment, kAlignment, false, cutlass::arch::OpMultiplyAddFastAccum
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, relu_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::ReLu,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::ReLu>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe5m2n_fe4m3t_tensor_op_f32, identity_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe5m2t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe5m2t_fe5m2n_fe4m3t_tensor_op_f32, identity_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe5m2t_tensor_op_f32, identity_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e5m2_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe5m2t_fe5m2n_fe5m2t_tensor_op_f32, identity_diff_aux_output_types_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e5m2_t;
|
||||
using ElementB = cutlass::float_e5m2_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = cutlass::float_e5m2_t;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_128x64x128_32x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 64, 128>, cutlass::gemm::GemmShape<32, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_noScale_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = ElementOutput;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>(
|
||||
/* scaleA = */false,
|
||||
/* scaleB = */false,
|
||||
/* scaleC = */false
|
||||
);
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM89_Device_Sparse_Gemm_fe4m3t_fe4m3n_fe4m3t_tensor_op_f32, identity_noAux_128x128x128_64x64x128) {
|
||||
using ElementA = cutlass::float_e4m3_t;
|
||||
using ElementB = cutlass::float_e4m3_t;
|
||||
using ElementOutput = cutlass::float_e4m3_t;
|
||||
using ElementAuxOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
using LayoutC = cutlass::layout::RowMajor;
|
||||
static int const kStages = 3;
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationGenericWithScalingAndAbsMax<
|
||||
cutlass::epilogue::thread::Identity,
|
||||
ElementOutput,
|
||||
ElementAuxOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::SparseGemmWithAbsmax<
|
||||
ElementA, LayoutA, ElementB, LayoutB, ElementOutput, LayoutC,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm89,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>, cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
EpilogueOutputOp, cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, kStages
|
||||
>;
|
||||
|
||||
bool passed = test::gemm::device::TestAllGemmWithAbsmax<Gemm, test::gemm::device::SparseTestbed<Gemm>, cutlass::epilogue::thread::Identity>();
|
||||
EXPECT_TRUE(passed);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM89_SUPPORTED
|
||||
@@ -339,5 +339,229 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x64x128_32x32x128) {
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x64x128_16x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x128_16x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<16, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x128_32x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x256x128_32x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 256, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x32x128_16x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 128x32x128_32x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 256x32x128_64x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 32, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmBasic<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
#endif
|
||||
|
||||
@@ -369,6 +369,74 @@ CUTLASS_TEST_L0(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 128x64x128_64x32x128
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 16x128x256_16x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 128, 256>,
|
||||
cutlass::gemm::GemmShape<16, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 16x256x256_16x64x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 256, 256>,
|
||||
cutlass::gemm::GemmShape<16, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x64x128_32x32x128, {
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
@@ -389,6 +457,481 @@ CUTLASS_TEST_L0(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x64x128_32x32x128,
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x64x128_16x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x128_16x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<16, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x128_32x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x256x128_32x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x32x128_16x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 128x32x128_32x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 256x32x128_64x32x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x64x256_16x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x256_16x64x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 256>,
|
||||
cutlass::gemm::GemmShape<16, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x128x256_32x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 256>,
|
||||
cutlass::gemm::GemmShape<32, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 32x256x256_32x64x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 256>,
|
||||
cutlass::gemm::GemmShape<32, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 64x32x256_16x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 128x32x256_32x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 256>,
|
||||
cutlass::gemm::GemmShape<32, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s4t_tensor_op_s32, 256x32x256_64x32x256) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 256>,
|
||||
cutlass::gemm::GemmShape<64, 32, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // #if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
|
||||
@@ -389,6 +389,26 @@ CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x64x128_64x64x128_align4, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 4, int32_t, float>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
|
||||
@@ -184,5 +184,357 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x32x64_16x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 16, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x64_16x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x64_16x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x64_32x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x64_32x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x256x64_32x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<32, 256, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x64_32x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x64_64x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x64_16x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x64_32x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM75_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 256x32x64_64x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
#endif
|
||||
|
||||
@@ -389,6 +389,956 @@ CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 16x64x128_16x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 16x128x128_16x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 128, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 16x256x128_16x64x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<16, 256, 128>,
|
||||
cutlass::gemm::GemmShape<16, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x16x128_16x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x16x128_32x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 16, 128>,
|
||||
cutlass::gemm::GemmShape<32, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 256x16x128_64x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 16, 128>,
|
||||
cutlass::gemm::GemmShape<64, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x32x64_16x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x64_16x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x64_16x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<16, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x64_32x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x64_32x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x256x64_32x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x64_32x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x64_64x16x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 16, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x64_16x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x64_32x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 256x32x64_64x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x32x128_16x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x128_16x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x128_16x64x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<16, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x64x128_32x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<32, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x128x128_32x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 32x256x128_32x64x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 256, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x128_32x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<32, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x128_64x16x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 128>,
|
||||
cutlass::gemm::GemmShape<64, 16, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
32 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 64x32x128_16x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 128x32x128_32x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 32, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s8t_s8n_s8t_tensor_op_s32, 256x32x128_64x32x128) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::RowMajor,
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 32, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // #if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
|
||||
@@ -41,7 +41,6 @@
|
||||
#include <random>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/distribution.h"
|
||||
@@ -54,7 +53,6 @@
|
||||
#include "cutlass/epilogue/collective/default_epilogue.hpp"
|
||||
#include "cutlass/epilogue/fusion/operations.hpp"
|
||||
#include "cutlass/complex.h"
|
||||
|
||||
#include "testbed_utils.h"
|
||||
|
||||
#include "cutlass/kernel_hardware_info.hpp"
|
||||
@@ -64,6 +62,7 @@
|
||||
|
||||
#include "cute/int_tuple.hpp"
|
||||
#include "cute/layout.hpp"
|
||||
#include "cute/numeric/int.hpp"
|
||||
|
||||
namespace test {
|
||||
namespace gemm {
|
||||
@@ -122,7 +121,7 @@ public:
|
||||
|
||||
template<class IntegralNotBool,
|
||||
__CUTE_REQUIRES((std::is_integral_v<IntegralNotBool> &&
|
||||
!std::is_same_v<IntegralNotBool, bool>)) >
|
||||
!cute::is_same_v<IntegralNotBool, bool>)) >
|
||||
explicit MaxSwizzleSize(IntegralNotBool max_swizzle_size) : max_swizzle_size_(max_swizzle_size) {}
|
||||
explicit operator int() const { return max_swizzle_size_; }
|
||||
private:
|
||||
@@ -132,7 +131,7 @@ private:
|
||||
template <typename T>
|
||||
auto make_iterator(T* ptr) {
|
||||
using namespace cute;
|
||||
if constexpr (is_subbyte_v<T>) {
|
||||
if constexpr (cute::is_subbyte_v<T>) {
|
||||
return subbyte_iterator<T>(ptr);
|
||||
}
|
||||
else {
|
||||
@@ -174,7 +173,7 @@ public:
|
||||
|
||||
template<class IntegralNotBool,
|
||||
__CUTE_REQUIRES((std::is_integral_v<IntegralNotBool> &&
|
||||
!std::is_same_v<IntegralNotBool, bool>)) >
|
||||
!cute::is_same_v<IntegralNotBool, bool>)) >
|
||||
explicit Splits(IntegralNotBool splits) : splits_(splits) {}
|
||||
explicit operator int() const { return splits_; }
|
||||
private:
|
||||
@@ -192,7 +191,7 @@ public:
|
||||
|
||||
template<class IntegralNotBool,
|
||||
__CUTE_REQUIRES((std::is_integral_v<IntegralNotBool> &&
|
||||
!std::is_same_v<IntegralNotBool, bool>)) >
|
||||
!cute::is_same_v<IntegralNotBool, bool>)) >
|
||||
explicit Iterations(IntegralNotBool iterations) : iterations_(iterations) {}
|
||||
explicit operator int() const { return iterations_; }
|
||||
private:
|
||||
@@ -214,12 +213,12 @@ bool initialize_tensor(
|
||||
scope_min = 0;
|
||||
}
|
||||
else if (bits_input <= 8) {
|
||||
scope_max = 2;
|
||||
scope_min = -2;
|
||||
scope_max = 1;
|
||||
scope_min = -1;
|
||||
}
|
||||
else{
|
||||
scope_max = 5;
|
||||
scope_min = -5;
|
||||
scope_max = 4;
|
||||
scope_min = -4;
|
||||
}
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
view, seed, scope_max, scope_min, 0);
|
||||
@@ -263,12 +262,16 @@ static constexpr bool is_row_or_col_major(){
|
||||
//
|
||||
// Default MMA input Operands : A , B
|
||||
//
|
||||
template<class ScheduleType_, class Gemm>
|
||||
template<
|
||||
class ScheduleType_,
|
||||
class Gemm,
|
||||
class ElementA_ = typename Gemm::GemmKernel::ElementA,
|
||||
class ElementB_ = typename Gemm::GemmKernel::ElementB>
|
||||
struct HostCollectiveMainloop {
|
||||
// Kernel data types
|
||||
using ElementA = typename Gemm::GemmKernel::ElementA;
|
||||
using ElementA = ElementA_;
|
||||
using StrideA = typename Gemm::GemmKernel::StrideA;
|
||||
using ElementB = typename Gemm::GemmKernel::ElementB;
|
||||
using ElementB = ElementB_;
|
||||
using StrideB = typename Gemm::GemmKernel::StrideB;
|
||||
using ScheduleType = typename Gemm::GemmKernel::CollectiveMainloop::DispatchPolicy::Schedule;
|
||||
using LayoutTagA = cutlass::detail::StrideToLayoutTagA_t<StrideA>;
|
||||
@@ -295,6 +298,8 @@ struct HostCollectiveMainloop {
|
||||
|
||||
cutlass::HostTensor<ElementA, LayoutTagA> tensor_A;
|
||||
cutlass::HostTensor<ElementB, LayoutTagB> tensor_B;
|
||||
// Whether to use relative equality checks
|
||||
CheckEquality check_relative_equality = CheckEquality::EXACT;
|
||||
|
||||
uint64_t seed;
|
||||
static constexpr uint64_t kDefaultSeed = 4096;
|
||||
@@ -306,6 +311,7 @@ struct HostCollectiveMainloop {
|
||||
"ERROR : B Layout is neither Row / Column Major)");
|
||||
|
||||
HostCollectiveMainloop(
|
||||
CheckEquality check_relative_equality_ = CheckEquality::EXACT,
|
||||
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = kDefaultSeed,
|
||||
@@ -314,10 +320,11 @@ struct HostCollectiveMainloop {
|
||||
):
|
||||
stride_factor_A(stride_factor_A_),
|
||||
stride_factor_B(stride_factor_B_),
|
||||
init_A(init_A_), init_B(init_B_), seed(seed_) { }
|
||||
init_A(init_A_), init_B(init_B_), seed(seed_),
|
||||
check_relative_equality(check_relative_equality_) { }
|
||||
|
||||
template<class ProblemShapeType>
|
||||
void initialize(ProblemShapeType problem_size) {
|
||||
bool initialize(ProblemShapeType problem_size) {
|
||||
//
|
||||
// Allocate the GEMM workspace
|
||||
//
|
||||
@@ -350,13 +357,17 @@ struct HostCollectiveMainloop {
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
Arguments to_args() {
|
||||
return {
|
||||
tensor_A.device_data(), stride_a,
|
||||
tensor_B.device_data(), stride_b
|
||||
|
||||
Arguments arguments =
|
||||
{
|
||||
tensor_A.device_data(), stride_a, tensor_B.device_data(), stride_b
|
||||
};
|
||||
return arguments;
|
||||
}
|
||||
|
||||
auto to_host_args(ProblemShapeType problem_size) {
|
||||
@@ -374,7 +385,16 @@ struct HostCollectiveMainloop {
|
||||
auto B = make_tensor(make_iterator(tensor_B.host_data()),
|
||||
make_layout(make_shape(N, K, L), stride_b));
|
||||
|
||||
cutlass::reference::host::GettMainloopParams<ElementAccumulator, decltype(A), decltype(B)> mainloop_params{A, B, TransformA, TransformB};
|
||||
cutlass::reference::host::GettMainloopParams<ElementAccumulator,
|
||||
decltype(A),
|
||||
decltype(B)
|
||||
> mainloop_params{};
|
||||
|
||||
mainloop_params.A = A;
|
||||
mainloop_params.B = B;
|
||||
mainloop_params.transform_A = TransformA;
|
||||
mainloop_params.transform_B = TransformB;
|
||||
|
||||
return mainloop_params;
|
||||
}
|
||||
|
||||
@@ -383,17 +403,45 @@ struct HostCollectiveMainloop {
|
||||
<< "\nB =\n" << tensor_B.host_view();
|
||||
}
|
||||
|
||||
template <
|
||||
class Element,
|
||||
class Layout
|
||||
>
|
||||
bool equality_check(
|
||||
cutlass::TensorView<Element, Layout> const& lhs,
|
||||
cutlass::TensorView<Element, Layout> const& rhs) const {
|
||||
|
||||
// Factors used for calculating relative equality. CUTLASS's relative-equality
|
||||
// checks in include/cutlass/relatively_equal.h are inspired by
|
||||
// https://floating-point-gui.de/errors/comparison/. This reference suggests using
|
||||
// the minimum normal value of a given type as the nonzero_floor.
|
||||
Element epsilon(static_cast<Element>(0.1f));
|
||||
Element nonzero_floor(std::numeric_limits<Element>::min());
|
||||
|
||||
if constexpr (!cutlass::is_complex<Element>::value) {
|
||||
if (check_relative_equality == CheckEquality::RELATIVE) {
|
||||
return cutlass::reference::host::TensorRelativelyEquals(
|
||||
lhs, rhs, epsilon, nonzero_floor);
|
||||
}
|
||||
else {
|
||||
return cutlass::reference::host::TensorEquals(lhs, rhs);
|
||||
}
|
||||
}
|
||||
else {
|
||||
return cutlass::reference::host::TensorEquals(lhs, rhs);
|
||||
}
|
||||
}
|
||||
|
||||
bool compare_reference(
|
||||
cute::Shape<int,int,int,int> problem_shape_MNKL) {
|
||||
auto [M, N, K, L] = problem_shape_MNKL;
|
||||
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_A.host_view()), 0);
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_B.host_view()), 0);
|
||||
return true;
|
||||
|
||||
bool passed = true;
|
||||
return passed;
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
template<class Gemm>
|
||||
struct HostCollectiveDefaultEpilogue {
|
||||
// fusion types are potentially void if the fusion is not supported
|
||||
@@ -474,7 +522,7 @@ struct HostCollectiveDefaultEpilogue {
|
||||
check_relative_equality(check_relative_equality_),
|
||||
use_device_scalars(use_device_scalars_){ }
|
||||
|
||||
void initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
bool initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
// Initialize Epilogue tensors
|
||||
auto problem_shape_MNKL = cute::append<4>(problem_size, 1);
|
||||
auto [M, N, K, L] = problem_shape_MNKL;
|
||||
@@ -496,6 +544,8 @@ struct HostCollectiveDefaultEpilogue {
|
||||
|
||||
alpha = alpha_;
|
||||
beta = beta_;
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
template <
|
||||
@@ -510,7 +560,7 @@ struct HostCollectiveDefaultEpilogue {
|
||||
// checks in include/cutlass/relatively_equal.h are inspired by
|
||||
// https://floating-point-gui.de/errors/comparison/. This reference suggests using
|
||||
// the minimum normal value of a given type as the nonzero_floor.
|
||||
Element epsilon(0.1f);
|
||||
Element epsilon(static_cast<Element>(0.1f));
|
||||
Element nonzero_floor(std::numeric_limits<Element>::min());
|
||||
|
||||
if constexpr (!cutlass::is_complex<Element>::value) {
|
||||
@@ -559,7 +609,6 @@ struct HostCollectiveDefaultEpilogue {
|
||||
}
|
||||
|
||||
Arguments to_args(ProblemShapeType problem_size) {
|
||||
auto coord_0 = cutlass::make_Coord(0);
|
||||
Arguments arguments =
|
||||
{
|
||||
{alpha, beta},
|
||||
@@ -693,6 +742,7 @@ struct HostCollectiveEpilogue {
|
||||
cutlass::HostTensor<ElementScalar, LayoutTagScalar> scale_Aux;
|
||||
cutlass::HostTensor<ElementBias , LayoutTagVector> bias;
|
||||
cutlass::HostTensor<ElementC, LayoutTagC> tensor_C;
|
||||
cutlass::HostTensor<ElementCompute, LayoutTagScalar> norm_constant;
|
||||
|
||||
// Outputs
|
||||
cutlass::HostTensor<ElementAmax, LayoutTagScalar> abs_max_Aux;
|
||||
@@ -738,10 +788,13 @@ struct HostCollectiveEpilogue {
|
||||
check_relative_equality(check_relative_equality_),
|
||||
use_device_scalars(use_device_scalars_){ }
|
||||
|
||||
void initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
bool initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
// Initialize Epilogue tensors
|
||||
auto problem_shape_MNKL = cute::append<4>(problem_size, 1);
|
||||
auto [M, N, K, L] = problem_shape_MNKL;
|
||||
auto M = cute::size<0>(problem_shape_MNKL);
|
||||
auto N = cute::size<1>(problem_shape_MNKL);
|
||||
auto K = cute::size<2>(problem_shape_MNKL);
|
||||
auto L = cute::size<3>(problem_shape_MNKL);
|
||||
|
||||
stride_c = cutlass::make_cute_packed_stride(StrideC{}, cute::make_shape(M, N, L));
|
||||
stride_d = cutlass::make_cute_packed_stride(StrideD{}, cute::make_shape(M, N, L));
|
||||
@@ -854,6 +907,7 @@ struct HostCollectiveEpilogue {
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
template <
|
||||
@@ -868,7 +922,7 @@ struct HostCollectiveEpilogue {
|
||||
// checks in include/cutlass/relatively_equal.h are inspired by
|
||||
// https://floating-point-gui.de/errors/comparison/. This reference suggests using
|
||||
// the minimum normal value of a given type as the nonzero_floor.
|
||||
Element epsilon(0.1f);
|
||||
Element epsilon(static_cast<Element>(0.1f));
|
||||
Element nonzero_floor(std::numeric_limits<Element>::min());
|
||||
|
||||
if constexpr (!cutlass::is_complex<Element>::value) {
|
||||
@@ -889,8 +943,6 @@ struct HostCollectiveEpilogue {
|
||||
cute::Shape<int,int,int,int> problem_shape_MNKL,
|
||||
ElementScalar alpha,
|
||||
ElementScalar beta) {
|
||||
auto [M, N, K, L] = problem_shape_MNKL;
|
||||
|
||||
tensor_D.sync_host();
|
||||
EXPECT_GT(cutlass::reference::host::TensorNorm(tensor_C.host_view()), 0);
|
||||
|
||||
@@ -1169,13 +1221,15 @@ struct HostCollectiveEpilogue {
|
||||
template <
|
||||
typename Gemm,
|
||||
template <class T> class ActivationFunctor_ = cutlass::epilogue::thread::Identity,
|
||||
bool force_legacy_epilogue = false
|
||||
bool force_legacy_epilogue = false,
|
||||
typename ElementA = typename Gemm::GemmKernel::ElementA,
|
||||
typename ElementB = typename Gemm::GemmKernel::ElementB
|
||||
>
|
||||
struct TestbedImpl {
|
||||
// Kernel data types
|
||||
using ScheduleType = typename Gemm::GemmKernel::CollectiveMainloop::DispatchPolicy::Schedule;
|
||||
// All Collective MMA operands are defined by HostCollectiveMainloopType based on the schedule type
|
||||
using HostCollectiveMainloopType = HostCollectiveMainloop<ScheduleType, Gemm>;
|
||||
using HostCollectiveMainloopType = HostCollectiveMainloop<ScheduleType, Gemm, ElementA, ElementB>;
|
||||
using CollectiveEpilogue = cute::conditional_t<IsDefaultEpilogue<typename Gemm::GemmKernel::CollectiveEpilogue>::value || force_legacy_epilogue,
|
||||
HostCollectiveDefaultEpilogue<Gemm>,
|
||||
HostCollectiveEpilogue<Gemm>>;
|
||||
@@ -1215,7 +1269,7 @@ struct TestbedImpl {
|
||||
cutlass::Distribution::Kind init_scale_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_bias_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = kDefaultSeed
|
||||
): collective_mma_inputs(HostCollectiveMainloopType(init_A_, init_B_, seed_)),
|
||||
): collective_mma_inputs(HostCollectiveMainloopType(check_relative_equality_, init_A_, init_B_, seed_)),
|
||||
collective_epilogue(CollectiveEpilogue(check_relative_equality_, use_device_scalars_, disable_vector_beta_, init_C_, init_scale_, init_bias_, seed_)) { }
|
||||
|
||||
TestbedImpl(
|
||||
@@ -1232,13 +1286,15 @@ struct TestbedImpl {
|
||||
cutlass::Distribution::Kind init_scale_ = cutlass::Distribution::Uniform,
|
||||
cutlass::Distribution::Kind init_bias_ = cutlass::Distribution::Uniform,
|
||||
uint64_t seed_ = kDefaultSeed
|
||||
): collective_mma_inputs(HostCollectiveMainloopType(stride_factor_A_, stride_factor_B_, init_A_, init_B_, seed_)),
|
||||
): collective_mma_inputs(HostCollectiveMainloopType(check_relative_equality_, stride_factor_A_, stride_factor_B_, init_A_, init_B_, seed_)),
|
||||
collective_epilogue(CollectiveEpilogue(check_relative_equality_, use_device_scalars_, disable_vector_beta_, init_C_, init_scale_, init_bias_, seed_)) { }
|
||||
|
||||
/// Initializes data structures
|
||||
void initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
bool initialize(ProblemShapeType problem_size, ElementScalar alpha_=1.f, ElementScalar beta_=0.f) {
|
||||
collective_mma_inputs.initialize(problem_size);
|
||||
collective_epilogue.initialize(problem_size, alpha_, beta_);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Compares computed reference with device reference and outputs to a file if incorrect
|
||||
@@ -1280,7 +1336,6 @@ struct TestbedImpl {
|
||||
{
|
||||
using namespace cute;
|
||||
auto problem_shape_MNKL = cute::append<4>(problem_size, 1);
|
||||
auto [M, N, K, L] = problem_shape_MNKL;
|
||||
|
||||
auto mainloop_params = collective_mma_inputs.to_host_args(problem_size);
|
||||
auto epilogue_params = collective_epilogue.to_host_args(problem_size);
|
||||
@@ -1297,7 +1352,7 @@ struct TestbedImpl {
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = Gemm::GemmKernel::SharedStorageSize;
|
||||
size_t smem_size = static_cast<size_t>(Gemm::GemmKernel::SharedStorageSize);
|
||||
|
||||
int device_idx;
|
||||
cudaError_t result = cudaGetDevice(&device_idx);
|
||||
@@ -1371,7 +1426,8 @@ struct TestbedImpl {
|
||||
RasterOrderOptions raster_order = RasterOrderOptions::Heuristic,
|
||||
detail::MaxSwizzleSize max_swizzle = detail::MaxSwizzleSize{},
|
||||
detail::Splits splits = detail::Splits{},
|
||||
DecompositionMode decomposition_mode = DecompositionMode::Heuristic)
|
||||
DecompositionMode decomposition_mode = DecompositionMode::Heuristic
|
||||
)
|
||||
{
|
||||
|
||||
// Fail test if insufficient CUDA device
|
||||
@@ -1380,7 +1436,10 @@ struct TestbedImpl {
|
||||
return false;
|
||||
}
|
||||
|
||||
this->initialize(problem_size, alpha, beta);
|
||||
if (!this->initialize(problem_size, alpha, beta)) {
|
||||
std::cerr << "Initialization failed \n";
|
||||
return false;
|
||||
}
|
||||
|
||||
//
|
||||
// Initialize the GEMM operator
|
||||
@@ -1399,16 +1458,21 @@ struct TestbedImpl {
|
||||
}
|
||||
|
||||
typename Gemm::GemmKernel::TileScheduler::Arguments scheduler_args;
|
||||
if constexpr (std::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>) {
|
||||
if constexpr (cute::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>) {
|
||||
scheduler_args = { static_cast<int>(splits), static_cast<int>(max_swizzle), raster_order, decomposition_mode };
|
||||
}
|
||||
else {
|
||||
scheduler_args = { static_cast<int>(max_swizzle), raster_order };
|
||||
}
|
||||
arguments = {
|
||||
typename HostCollectiveMainloopType::Arguments mainloop_args;
|
||||
|
||||
mainloop_args = collective_mma_inputs.to_args();
|
||||
|
||||
arguments =
|
||||
{
|
||||
cutlass::gemm::GemmUniversalMode::kGemm,
|
||||
problem_size,
|
||||
collective_mma_inputs.to_args(),
|
||||
mainloop_args,
|
||||
collective_epilogue.to_args(problem_size),
|
||||
hw_info,
|
||||
scheduler_args
|
||||
@@ -1471,11 +1535,19 @@ struct TestbedImpl {
|
||||
template <
|
||||
typename Gemm,
|
||||
template <class T> class ActivationFunctor = cutlass::epilogue::thread::Identity,
|
||||
bool force_legacy_epilogue = false
|
||||
bool force_legacy_epilogue = false,
|
||||
typename ElementA = typename Gemm::GemmKernel::ElementA,
|
||||
typename ElementB = typename Gemm::GemmKernel::ElementB
|
||||
>
|
||||
struct Testbed3x {
|
||||
|
||||
using TestBedImpl = typename detail::TestbedImpl<Gemm, ActivationFunctor, force_legacy_epilogue>;
|
||||
using TestBedImpl = typename detail::TestbedImpl<
|
||||
Gemm,
|
||||
ActivationFunctor,
|
||||
force_legacy_epilogue,
|
||||
ElementA,
|
||||
ElementB
|
||||
>;
|
||||
using Kernel = typename Gemm::GemmKernel;
|
||||
using Epilogue = typename Gemm::GemmKernel::CollectiveEpilogue;
|
||||
|
||||
@@ -1514,7 +1586,8 @@ struct Testbed3x {
|
||||
detail::Splits splits = detail::Splits{},
|
||||
DecompositionMode decomposition_mode = DecompositionMode::Heuristic,
|
||||
bool profiling = false,
|
||||
detail::Iterations iterations = detail::Iterations{})
|
||||
detail::Iterations iterations = detail::Iterations{}
|
||||
)
|
||||
{
|
||||
return impl_.run(
|
||||
problem_size, alpha, beta, profiling, iterations, raster_order, max_swizzle, splits, decomposition_mode
|
||||
@@ -1582,7 +1655,7 @@ bool TestAll(double alpha = 1.0, double beta = 0.0, CheckEquality check_relative
|
||||
std::vector<int> problem_size_m = {max_alignment, 512 - 3 * max_alignment};
|
||||
std::vector<int> problem_size_n = {max_alignment, 512 - 2 * max_alignment};
|
||||
|
||||
if constexpr (std::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
if constexpr (cute::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong>) {
|
||||
problem_size_m.push_back(768);
|
||||
problem_size_n.push_back(768);
|
||||
@@ -1596,7 +1669,7 @@ bool TestAll(double alpha = 1.0, double beta = 0.0, CheckEquality check_relative
|
||||
using DecompositionMode = typename cutlass::gemm::kernel::detail::PersistentTileSchedulerSm90StreamKParams::DecompositionMode;
|
||||
std::vector<DecompositionMode> decomposition_modes = {DecompositionMode::Heuristic};
|
||||
std::vector problem_splits = {detail::Splits{1}};
|
||||
static constexpr bool UsesStreamKScheduler = std::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>;
|
||||
static constexpr bool UsesStreamKScheduler = cute::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>;
|
||||
if constexpr (UsesStreamKScheduler) {
|
||||
problem_splits.push_back(detail::Splits{2});
|
||||
problem_splits.push_back(detail::Splits{3});
|
||||
|
||||
@@ -163,14 +163,15 @@ public:
|
||||
using ElementCompute = typename Base::ElementCompute;
|
||||
|
||||
struct Arguments {
|
||||
ElementCompute scalar[BroadcastCount];
|
||||
ElementCompute const* scalar_ptrs[BroadcastCount];
|
||||
cute::Stride<cute::_0,cute::_0,cute::_0> dScalar;
|
||||
ElementCompute scalar[BroadcastCount] = {0};
|
||||
ElementCompute const* scalar_ptrs[BroadcastCount] = { nullptr };
|
||||
cute::Stride<cute::_0,cute::_0,cute::_0> dScalar{};
|
||||
};
|
||||
private:
|
||||
ElementCompute _scalar;
|
||||
ElementCompute _scalar{};
|
||||
public:
|
||||
HostScalarBroadcast(){}
|
||||
|
||||
template<typename ProblemShapeType, typename TestBedImpl>
|
||||
HostScalarBroadcast(ProblemShapeType problem_size, TestBedImpl impl, bool check_relative_equality=false)
|
||||
: Base(check_relative_equality), _scalar(ElementCompute(Value)) {}
|
||||
@@ -1313,7 +1314,7 @@ public:
|
||||
}
|
||||
|
||||
typename Gemm::GemmKernel::TileScheduler::Arguments scheduler_args;
|
||||
if constexpr (std::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>) {
|
||||
if constexpr (cute::is_same_v<typename Gemm::GemmKernel::TileSchedulerTag, cutlass::gemm::StreamKScheduler>) {
|
||||
scheduler_args = { splits };
|
||||
}
|
||||
|
||||
@@ -1399,7 +1400,7 @@ bool TestAllEVT(bool check_relative_equality=false) {
|
||||
std::vector<int> problem_size_m = {max_alignment, 512 - 3 * max_alignment};
|
||||
std::vector<int> problem_size_n = {max_alignment, 512 - 2 * max_alignment};
|
||||
|
||||
if constexpr (std::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
if constexpr (cute::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong>) {
|
||||
problem_size_m.push_back(768);
|
||||
problem_size_n.push_back(768);
|
||||
|
||||
@@ -439,7 +439,7 @@ bool TestAllTensorBroadcast(bool use_bias=true) {
|
||||
std::vector<int> problem_size_m = {max_alignment, 512 - 3 * max_alignment};
|
||||
std::vector<int> problem_size_n = {max_alignment, 512 - 2 * max_alignment};
|
||||
|
||||
if constexpr (std::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
if constexpr (cute::is_same_v<typename Gemm::GemmKernel::DispatchPolicy::Schedule,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong>) {
|
||||
problem_size_m.push_back(768);
|
||||
problem_size_n.push_back(768);
|
||||
|
||||
@@ -457,6 +457,38 @@ TEST(SM80_Device_GemmWithBroadcast_RELU_f16n_f16n_f16n_tensor_op_f32, 128x128_32
|
||||
test::gemm::device::TestAllGemmWithBroadcast<Gemm, GemmWithBiasReluReferenceOp<Gemm> >();
|
||||
}
|
||||
|
||||
TEST(SM80_Device_GemmWithBroadcast_RELU_f32n_f32n_f32n_tensor_op_f32, 64x64_16x10_32x32x16_16x8x8) {
|
||||
|
||||
using EpilogueOutputOp = cutlass::epilogue::thread::LinearCombinationBiasRelu<
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
4,
|
||||
false
|
||||
>;
|
||||
|
||||
using GemmKernel =
|
||||
typename cutlass::gemm::kernel::DefaultGemmWithBroadcast<
|
||||
float, cutlass::layout::RowMajor, cutlass::ComplexTransform::kNone, 4, // transposed B operand
|
||||
float, cutlass::layout::RowMajor, cutlass::ComplexTransform::kNone, 4, // transposed A operand
|
||||
float, cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
EpilogueOutputOp,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<8>,
|
||||
10,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
|
||||
|
||||
test::gemm::device::TestAllGemmWithBroadcast<Gemm, GemmWithBiasReluReferenceOp<Gemm> >();
|
||||
}
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif
|
||||
|
||||
@@ -113,7 +113,7 @@ struct MultistageTestbed {
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Gemm::GemmKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
|
||||
@@ -118,7 +118,7 @@ struct MultistageInterleavedTestbed {
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::SharedStorage));
|
||||
size_t smem_size = sizeof(typename Gemm::GemmKernel::SharedStorage);
|
||||
|
||||
cudaDeviceProp properties;
|
||||
int device_idx;
|
||||
|
||||
@@ -194,12 +194,12 @@ public:
|
||||
// Aux = scale_aux * Z
|
||||
// else
|
||||
// Aux = Z
|
||||
template <class Gemm, template <class> class ActivationFn, class ElementD>
|
||||
template <class Gemm, template <class> class ActivationFn, class ElementD, class ElementAux = ElementD>
|
||||
class HostScaledLinCombPerRowBiasEltActAmaxAux {
|
||||
public:
|
||||
template <typename T>
|
||||
using amax = cutlass::maximum_absolute_value_reduction<T, true>;
|
||||
using EVTModule = HEVT<
|
||||
using EVTModuleAuxFp8 = HEVT<
|
||||
HostAuxStore<Gemm, true>,
|
||||
HST<Gemm,
|
||||
// Z = scale_a * scale_b * alpha * acc + scale_c * beta * C + per-row bias
|
||||
@@ -228,9 +228,9 @@ public:
|
||||
>,
|
||||
// Aux = Z * scale_aux, amax_aux = max(abs(elements in Aux))
|
||||
HEVT<
|
||||
HostAuxStore<Gemm, false, ElementD, cutlass::layout::RowMajor>,
|
||||
HostAuxStore<Gemm, false, ElementAux, cutlass::layout::RowMajor>,
|
||||
HEVT<
|
||||
HostCompute<Gemm, cutlass::epilogue::fusion::detail::ScaleOutOp<ElementD>::template Op>,
|
||||
HostCompute<Gemm, cutlass::multiplies>,
|
||||
HEVT<
|
||||
HostScalarReduce<Gemm, amax, float>,
|
||||
HostAccumulator<Gemm>
|
||||
@@ -240,6 +240,40 @@ public:
|
||||
>
|
||||
>
|
||||
>;
|
||||
|
||||
using EVTModuleAuxNotFp8 = HEVT<
|
||||
// D = activation(Z) * scaled_d, amax_d = max(abs(elements in D))
|
||||
HostAuxStore<Gemm, true>,
|
||||
HEVT<
|
||||
HostCompute<Gemm, cutlass::epilogue::fusion::detail::ScaleOutOp<ElementD>::template Op>,
|
||||
HEVT<
|
||||
HostScalarReduce<Gemm, amax, float>,
|
||||
HEVT<
|
||||
HostCompute<Gemm, ActivationFn>, //activation(Z) * scaled_d
|
||||
HEVT<
|
||||
// Aux = Z
|
||||
HostAuxStore<Gemm, false, ElementAux, cutlass::layout::RowMajor>,
|
||||
// Z = scale_a * scale_b * alpha * acc + scale_c * beta * C + per-row bias
|
||||
HEVT<
|
||||
HostCompute<Gemm, cutlass::homogeneous_multiply_add>,
|
||||
HostScalarBroadcast<Gemm, 1, 2>, // scale_c * beta
|
||||
HostAuxLoad<Gemm, true>, // C
|
||||
HEVT<
|
||||
HostCompute<Gemm, cutlass::homogeneous_multiply_add>,
|
||||
HostScalarBroadcast<Gemm, 1, 3>, // scale_a * scale_b * alpha
|
||||
HostAccumulator<Gemm>,
|
||||
HostColBroadcast<Gemm, ElementD>
|
||||
>
|
||||
>
|
||||
>
|
||||
>
|
||||
>,
|
||||
HostScalarBroadcast<Gemm, 1> // scale_d
|
||||
>
|
||||
>;
|
||||
|
||||
using EVTModule = cute::conditional_t<cutlass::epilogue::fusion::detail::is_fp8_v<ElementAux>, EVTModuleAuxFp8, EVTModuleAuxNotFp8>;
|
||||
|
||||
};
|
||||
} // namespace test::gemm::device
|
||||
|
||||
@@ -400,7 +434,7 @@ template<
|
||||
FloatRoundStyle RoundStyle = FloatRoundStyle::round_to_nearest
|
||||
>
|
||||
using Sm90LinCombPerColumnReduce =
|
||||
Sm90EVT<Sm90RowReduction<RegReduceFn, GmemReduceFn, 0, CtaTileShapeMNK, ElementReduce, ElementCompute, RoundStyle>, // per column reduce
|
||||
Sm90EVT<Sm90RowReduction<RegReduceFn, RegReduceFn, GmemReduceFn, 0, CtaTileShapeMNK, ElementReduce, ElementCompute, RoundStyle>, // per column reduce
|
||||
Sm90EVT<Sm90Compute<homogeneous_multiply_add, ElementOutput, ElementCompute, RoundStyle>, // beta * C + alpha * acc
|
||||
Sm90ScalarBroadcast<ElementScalar>, // beta
|
||||
Sm90SrcFetch<ElementOutput>, // C
|
||||
@@ -425,7 +459,7 @@ template<
|
||||
FloatRoundStyle RoundStyle = FloatRoundStyle::round_to_nearest
|
||||
>
|
||||
using Sm90LinCombPerRowReduce =
|
||||
Sm90EVT<Sm90ColReduction<RegReduceFn, GmemReduceFn, 0, CtaTileShapeMNK, ElementReduce, ElementCompute, RoundStyle>, // per column reduce
|
||||
Sm90EVT<Sm90ColReduction<RegReduceFn, RegReduceFn, GmemReduceFn, 0, CtaTileShapeMNK, ElementReduce, ElementCompute, RoundStyle>, // per column reduce
|
||||
Sm90EVT<Sm90Compute<homogeneous_multiply_add, ElementOutput, ElementCompute, RoundStyle>, // beta * C + alpha * acc
|
||||
Sm90ScalarBroadcast<ElementScalar>, // beta
|
||||
Sm90SrcFetch<ElementOutput>, // C
|
||||
|
||||
+4
-4
@@ -796,7 +796,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -833,7 +833,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -870,7 +870,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -907,7 +907,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+3
-3
@@ -98,7 +98,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -155,7 +155,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -212,7 +212,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+29
-29
@@ -30,7 +30,7 @@
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief Tests for Sm90 f16_f16_f16 with cooperative EVT epilogue
|
||||
D = alpha * acc + beta * c + aux_load
|
||||
D = alpha * acc + beta * c + aux_load
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
@@ -72,15 +72,15 @@ static constexpr auto select_evt_d() {
|
||||
RoundStyle>, // alpha * acc
|
||||
Sm90ScalarBroadcast<ElementAccumulator>, // alpha
|
||||
Sm90AccFetch // acc
|
||||
>;
|
||||
>;
|
||||
if constexpr (IsCNeed) {
|
||||
using EVT_D = Sm90EVT<Sm90Compute<cutlass::homogeneous_multiply_add, ElementCompute, ElementCompute, RoundStyle>,
|
||||
Sm90ScalarBroadcast<ElementAccumulator>, // beta
|
||||
Sm90SrcFetch<ElementCompute>, // C
|
||||
BinaryCompute0>;
|
||||
return *(EVT_D *)(nullptr);
|
||||
return EVT_D{};
|
||||
} else {
|
||||
return *(BinaryCompute0 *)(nullptr);
|
||||
return BinaryCompute0{};
|
||||
}
|
||||
}
|
||||
|
||||
@@ -152,7 +152,7 @@ bool testEVTAuxStoreWithoutD() {
|
||||
D_block.get(), stride_D,
|
||||
}, // Epilogue arguments end
|
||||
/*hw_info=*/{},
|
||||
/*scheduler_args=*/{}
|
||||
/*scheduler_args=*/{}
|
||||
};
|
||||
|
||||
// check without D aux store
|
||||
@@ -275,7 +275,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::RowMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -292,7 +292,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -310,7 +310,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -319,7 +319,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
@@ -345,7 +345,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::ColumnMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -362,7 +362,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -380,7 +380,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -389,7 +389,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
@@ -415,7 +415,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::RowMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -432,7 +432,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -450,7 +450,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -459,7 +459,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
@@ -485,7 +485,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::RowMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -502,7 +502,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -520,7 +520,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -529,7 +529,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
@@ -555,7 +555,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::ColumnMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -572,7 +572,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -590,7 +590,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -599,7 +599,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
@@ -625,7 +625,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
|
||||
using EpilogueDescriptor = cutlass::epilogue::collective::detail::EpilogueDescriptor<
|
||||
TileShape_MNK, EpilogueTileType, cutlass::half_t, cutlass::half_t, EpilogueSchedule
|
||||
>;
|
||||
>;
|
||||
using AuxStoreDescriptor = cutlass::epilogue::collective::detail::AuxStoreDescriptor<
|
||||
EpilogueDescriptor, cutlass::layout::RowMajor, cutlass::half_t
|
||||
>;
|
||||
@@ -642,7 +642,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
typename AuxStoreDescriptor::CopyOpR2S>;
|
||||
|
||||
constexpr auto select_kernel = [](auto has_c, auto has_d) {
|
||||
using FusionCallbacks =
|
||||
using FusionCallbacks =
|
||||
cute::conditional_t<decltype(has_d){}, EVT_D, Sm90EVT<AuxStore, EVT_D>>;
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -660,7 +660,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -669,7 +669,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
CollectiveMainloop,
|
||||
CollectiveEpilogue>;
|
||||
|
||||
return *(GemmKernel *)(nullptr);
|
||||
return GemmKernel{};
|
||||
};
|
||||
|
||||
using GemmKernel = decltype(select_kernel(cute::C<has_c>{}, cute::C<true>{}));
|
||||
|
||||
+12
-12
@@ -86,7 +86,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -183,7 +183,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -271,7 +271,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -316,7 +316,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -361,7 +361,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -406,7 +406,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -451,7 +451,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -497,7 +497,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -542,7 +542,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -587,7 +587,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+2
-2
@@ -98,7 +98,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -151,7 +151,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 12
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+6
-6
@@ -70,7 +70,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperative;
|
||||
using FusionCallbacks = cutlass::epilogue::fusion::Sm90LinCombPerColumnReduce<
|
||||
cutlass::plus, cutlass::red, float, TileShape_MNK, cutlass::half_t, float, float>;
|
||||
cutlass::plus, cutlass::atomic_add, float, TileShape_MNK, cutlass::half_t, float, float>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -89,7 +89,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -116,7 +116,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperative;
|
||||
using FusionCallbacks = cutlass::epilogue::fusion::Sm90LinCombPerRowReduce<
|
||||
cutlass::plus, cutlass::red, float, TileShape_MNK, cutlass::half_t, float, float>;
|
||||
cutlass::plus, cutlass::atomic_add, float, TileShape_MNK, cutlass::half_t, float, float>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -135,7 +135,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -162,7 +162,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
|
||||
using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperative;
|
||||
using FusionCallbacks = cutlass::epilogue::fusion::Sm90LinCombScalarReduce<
|
||||
cutlass::plus, cutlass::red, float, cutlass::half_t, float, float>;
|
||||
cutlass::plus, cutlass::atomic_add, float, cutlass::half_t, float, float>;
|
||||
|
||||
using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
|
||||
cutlass::arch::Sm90, cutlass::arch::OpClassTensorOp,
|
||||
@@ -181,7 +181,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+2
-2
@@ -94,7 +94,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -144,7 +144,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_cooperative_epilogue, 25
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedCooperative
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+12
-12
@@ -762,7 +762,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f16_persistent_Epilogue, 64x
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -808,7 +808,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f16_persistent_Epilogue, 128
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -856,7 +856,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f16_persistent_Epilogue, 64x
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -902,7 +902,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f16_persistent_Epilogue, 128
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -950,7 +950,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_persistent_Epilogue, 64x
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -996,7 +996,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_persistent_Epilogue, 128
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1044,7 +1044,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_persistent_Epilogue, 64x
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1090,7 +1090,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_persistent_Epilogue, 128
|
||||
ElementB, LayoutB, 8,
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1133,7 +1133,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16n_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1176,7 +1176,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f16t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1219,7 +1219,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -1262,7 +1262,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
ElementB, LayoutB, 16 / sizeof(ElementB),
|
||||
ElementAccumulator,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
KernelSchedule
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+3
-3
@@ -99,7 +99,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -154,7 +154,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -209,7 +209,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
+10
-10
@@ -86,7 +86,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -129,7 +129,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -173,7 +173,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -218,7 +218,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -263,7 +263,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -308,7 +308,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32n_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -353,7 +353,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -398,7 +398,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -443,7 +443,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_persistent_epilogue, 128
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
@@ -488,7 +488,7 @@ TEST(SM90_Device_Gemm_f16t_f16n_f32t_tensor_op_gmma_f32_pingpong_epilogue, 128x1
|
||||
cutlass::half_t, LayoutB, 8,
|
||||
float,
|
||||
TileShape_MNK, ClusterShape_MNK,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<sizeof(typename CollectiveEpilogue::SharedStorage)>,
|
||||
cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
|
||||
cutlass::gemm::KernelTmaWarpSpecializedPingpong
|
||||
>::CollectiveOp;
|
||||
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user