CUTLASS 2.4 (Implicit GEMM convolution) (#147)
CUTLASS 2.4 (Implicit GEMM Convolution) Co-authored-by: Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
This commit is contained in:
Manish Gupta
co-authored by
Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
parent
c2b80ad4e4
commit
6615010cd0
@@ -21,3 +21,4 @@
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
add_subdirectory(unit)
|
||||
|
||||
|
||||
@@ -49,19 +49,14 @@ target_link_libraries(
|
||||
cutlass_test_unit_infra
|
||||
)
|
||||
|
||||
set(CUTLASS_INSTALL_TESTS ON CACHE BOOL "Install test executables")
|
||||
set(CUTLASS_TEST_EXECUTION_ENVIRONMENT "" CACHE BOOL "Environment in which to invoke unit test executables")
|
||||
|
||||
function(cutlass_test_unit_add_executable)
|
||||
function(cutlass_test_unit_add_executable NAME)
|
||||
|
||||
set(options)
|
||||
set(oneValueArgs)
|
||||
set(multiValueArgs)
|
||||
cmake_parse_arguments(_ "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
|
||||
|
||||
cutlass_add_executable(${__UNPARSED_ARGUMENTS})
|
||||
|
||||
list(GET __UNPARSED_ARGUMENTS 0 NAME)
|
||||
cutlass_add_executable(${NAME} ${__UNPARSED_ARGUMENTS})
|
||||
|
||||
target_link_libraries(
|
||||
${NAME}
|
||||
@@ -72,25 +67,13 @@ function(cutlass_test_unit_add_executable)
|
||||
|
||||
string(REGEX REPLACE cutlass_ "" NAME_STEM ${NAME})
|
||||
|
||||
add_test(c${NAME_STEM} ${NAME})
|
||||
set(CUTLASS_TEST_UNIT_TEST_COMMAND_OPTIONS --gtest_output=xml:${NAME_STEM}.gtest.xml)
|
||||
|
||||
add_custom_target(
|
||||
${NAME_STEM}
|
||||
COMMAND
|
||||
${CUTLASS_TEST_EXECUTION_ENVIRONMENT} $<TARGET_FILE:${NAME}>
|
||||
DEPENDS
|
||||
${NAME}
|
||||
cutlass_add_executable_tests(
|
||||
${NAME_STEM} ${NAME}
|
||||
TEST_COMMAND_OPTIONS CUTLASS_TEST_UNIT_TEST_COMMAND_OPTIONS
|
||||
)
|
||||
|
||||
if (CUTLASS_INSTALL_TESTS)
|
||||
|
||||
install(
|
||||
TARGETS ${NAME}
|
||||
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
|
||||
)
|
||||
|
||||
endif()
|
||||
|
||||
endfunction()
|
||||
|
||||
add_custom_target(cutlass_test_unit)
|
||||
@@ -99,6 +82,7 @@ add_custom_target(test_unit)
|
||||
set(SUBDIRS
|
||||
core
|
||||
gemm
|
||||
conv
|
||||
layout
|
||||
transform
|
||||
epilogue
|
||||
|
||||
@@ -0,0 +1,42 @@
|
||||
# Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
# provided that the following conditions are met:
|
||||
# * Redistributions of source code must retain the above copyright notice, this list of
|
||||
# conditions and the following disclaimer.
|
||||
# * 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.
|
||||
# * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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)
|
||||
add_custom_target(test_unit_conv)
|
||||
|
||||
set(CUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED ON CACHE BOOL
|
||||
"Enable/Disable convolution device reference for conv unit tests.")
|
||||
|
||||
if(CUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED)
|
||||
message(STATUS "Enable device reference verification in conv unit tests")
|
||||
list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED=1)
|
||||
endif()
|
||||
|
||||
foreach(SUBDIR
|
||||
device
|
||||
)
|
||||
|
||||
add_subdirectory(${SUBDIR})
|
||||
add_dependencies(cutlass_test_unit_conv cutlass_test_unit_conv_${SUBDIR})
|
||||
add_dependencies(test_unit_conv test_unit_conv_${SUBDIR})
|
||||
|
||||
endforeach()
|
||||
@@ -0,0 +1,148 @@
|
||||
# Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
# provided that the following conditions are met:
|
||||
# * Redistributions of source code must retain the above copyright notice, this list of
|
||||
# conditions and the following disclaimer.
|
||||
# * 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.
|
||||
# * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_device
|
||||
DEPENDS
|
||||
cutlass_test_unit_conv_device_simt
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm70
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm75
|
||||
cutlass_test_unit_conv_device_tensorop_f16_sm80
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm80
|
||||
cutlass_test_unit_conv_device_tensorop_f32_tf32_sm80
|
||||
cutlass_test_unit_conv_device_tensorop_s32
|
||||
cutlass_test_unit_conv_device_tensorop_s32_interleaved
|
||||
)
|
||||
|
||||
add_custom_target(
|
||||
test_unit_conv_device
|
||||
DEPENDS
|
||||
test_unit_conv_device_simt
|
||||
test_unit_conv_device_tensorop_f32_sm70
|
||||
test_unit_conv_device_tensorop_f32_sm75
|
||||
test_unit_conv_device_tensorop_f16_sm80
|
||||
test_unit_conv_device_tensorop_f32_sm80
|
||||
test_unit_conv_device_tensorop_f32_tf32_sm80
|
||||
test_unit_conv_device_tensorop_s32
|
||||
test_unit_conv_device_tensorop_s32_interleaved
|
||||
)
|
||||
|
||||
#
|
||||
# OpClassSimt (CUDA cores)
|
||||
#
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_simt
|
||||
|
||||
# F32
|
||||
conv2d_fprop_implicit_gemm_f32nhwc_f32nhwc_f32nhwc_simt_f32_sm50.cu
|
||||
|
||||
conv2d_fprop_implicit_gemm_f32nhwc_f32nhwc_f32nhwc_simt_f32_sm80.cu
|
||||
conv2d_dgrad_implicit_gemm_f32nhwc_f32nhwc_f32nhwc_simt_f32_sm80.cu
|
||||
conv2d_wgrad_implicit_gemm_f32nhwc_f32nhwc_f32nhwc_simt_f32_sm80.cu
|
||||
|
||||
# CF32
|
||||
conv2d_fprop_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm50.cu
|
||||
conv2d_dgrad_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm50.cu
|
||||
conv2d_wgrad_implicit_gemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32_sm50.cu
|
||||
|
||||
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
|
||||
)
|
||||
|
||||
#
|
||||
# OpClassTensorOp (Tensor cores)
|
||||
#
|
||||
|
||||
# Conv - F16 input, F32 output, F32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm70
|
||||
|
||||
conv2d_fprop_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm70.cu
|
||||
conv2d_dgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm70.cu
|
||||
conv2d_wgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm70.cu
|
||||
)
|
||||
|
||||
# Conv2d - F16 input, F32 output, F32 accumulation - SM75
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm75
|
||||
|
||||
conv2d_fprop_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm75.cu
|
||||
conv2d_dgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm75.cu
|
||||
conv2d_wgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm75.cu
|
||||
)
|
||||
|
||||
# Conv2d - F16 input, F16 output, F16 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f16_sm80
|
||||
|
||||
conv2d_fprop_implicit_gemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16_sm80.cu
|
||||
conv2d_dgrad_implicit_gemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16_sm80.cu
|
||||
conv2d_wgrad_implicit_gemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16_sm80.cu
|
||||
)
|
||||
|
||||
# Conv2d - F16 input, F32 output, F32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f32_sm80
|
||||
|
||||
|
||||
conv2d_fprop_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
conv2d_dgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
conv2d_wgrad_implicit_gemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
|
||||
conv3d_wgrad_implicit_gemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32_sm75.cu
|
||||
conv3d_wgrad_implicit_gemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32_sm80.cu
|
||||
)
|
||||
|
||||
# Conv2d - TF32 input, F32 output, F32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_f32_tf32_sm80
|
||||
|
||||
conv2d_fprop_implicit_gemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
conv2d_dgrad_implicit_gemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
conv2d_wgrad_implicit_gemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32_sm80.cu
|
||||
|
||||
conv3d_fprop_implicit_gemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32_sm80.cu
|
||||
conv3d_dgrad_implicit_gemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32_sm80.cu
|
||||
conv3d_wgrad_implicit_gemm_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32_sm80.cu
|
||||
)
|
||||
|
||||
# Conv2d - S8 input, S32 output, S32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_s32
|
||||
|
||||
conv2d_fprop_implicit_gemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32_sm75.cu
|
||||
conv2d_fprop_implicit_gemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32_sm75.cu
|
||||
conv2d_fprop_implicit_gemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32_sm80.cu
|
||||
conv2d_fprop_implicit_gemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32_sm80.cu
|
||||
)
|
||||
|
||||
# Conv2d - S8 interleaved input, S8 interleaved output, S32 accumulation
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_conv_device_tensorop_s32_interleaved
|
||||
|
||||
conv2d_fprop_implicit_gemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32_sm75.cu
|
||||
conv2d_fprop_implicit_gemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32_sm75.cu
|
||||
conv2d_fprop_implicit_gemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32_sm80.cu
|
||||
conv2d_fprop_implicit_gemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32_sm80.cu
|
||||
)
|
||||
+130
@@ -0,0 +1,130 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x2_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x2_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 64, 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<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
+314
@@ -0,0 +1,314 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+123
@@ -0,0 +1,123 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+118
@@ -0,0 +1,118 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM70_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM70_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm70,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM70_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm70,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM70_SUPPORTED
|
||||
+159
@@ -0,0 +1,159 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM75_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM75_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM75_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+286
@@ -0,0 +1,286 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic,
|
||||
cutlass::conv::StrideSupport::kStrided
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_32x4_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32_unity_stride,
|
||||
128x128_64x4_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+323
@@ -0,0 +1,323 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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 Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
64x64_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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 Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
test::conv::device::Conv2dProblemVector user_size;
|
||||
|
||||
user_size.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 8, 8, 4}, // input size (NHWC)
|
||||
{8, 1, 1, 4}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>(user_size));
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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<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::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
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<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 Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+124
@@ -0,0 +1,124 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_dgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Analytic_ImplicitGemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Dgrad_Optimized_ImplicitGemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dDgradKernel = typename cutlass::conv::kernel::DefaultConv2dDgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kUnity
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dDgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dDgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dDgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+222
@@ -0,0 +1,222 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x2_32x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<4>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x128_8x2_16x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 128, 8>,
|
||||
cutlass::gemm::GemmShape<16, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<4>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x2_32x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<4>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x2_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<4>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
|
||||
+397
@@ -0,0 +1,397 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x5_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
5,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x3_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+121
@@ -0,0 +1,121 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+124
@@ -0,0 +1,124 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
+81
@@ -0,0 +1,81 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM70_SUPPORTED)
|
||||
|
||||
TEST(SM70_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm70,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM70_SUPPORTED
|
||||
+121
@@ -0,0 +1,121 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+124
@@ -0,0 +1,124 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#if 0
|
||||
TEST(SM80_Device_Conv2d_Fprop_Precomputed_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
#endif
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+82
@@ -0,0 +1,82 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM50_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
128x128_8x2_64x64x8) {
|
||||
|
||||
/// 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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<128, 128, 8>,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
+321
@@ -0,0 +1,321 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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 Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
64x64_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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 Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
test::conv::device::Conv2dProblemVector user_size;
|
||||
|
||||
user_size.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 8, 8, 4}, // input size (NHWC)
|
||||
{8, 1, 1, 4}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>(user_size));
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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<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::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
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<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 Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+520
@@ -0,0 +1,520 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed_interleaved.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x128_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x128_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x256_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x64_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x256_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x128_128x2_32x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x128_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x128_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x256_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x64_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x256_128x2_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x128_128x2_32x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+521
@@ -0,0 +1,521 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed_interleaved.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x128_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x128_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x256_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x64_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x256_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x128_128x4_32x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x128_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x128_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
128x256_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
256x64_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x256_128x3_64x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4ncxhwx_s4cxrskx_s4ncxhwx_tensor_op_s32,
|
||||
64x128_128x4_32x64x128) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<64>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<64>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 64>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+119
@@ -0,0 +1,119 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<8, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+121
@@ -0,0 +1,121 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s4nhwc_s4nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::int4b_t;
|
||||
using ElementB = cutlass::int4b_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+679
@@ -0,0 +1,679 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed_interleaved.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x256_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x64_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x256_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x64_64x2_64x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x128_64x2_32x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x64_64x2_32x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x256_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x64_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x256_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x64_64x2_64x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x128_64x2_32x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x64_64x2_32x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+680
@@ -0,0 +1,680 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed_interleaved.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x256_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x64_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x256_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x64_64x4_64x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x128_64x4_32x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x64_64x6_32x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
6,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x256_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
256x64_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x256_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
128x64_64x4_64x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x128_64x4_32x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8ncxhwx_s8cxrskx_s8ncxhwx_tensor_op_s32,
|
||||
64x64_64x6_32x32x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementB, cutlass::layout::TensorCxRSKx<32>,
|
||||
ElementC, cutlass::layout::TensorNCxHWx<32>,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<1>,
|
||||
6,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE((test::conv::device::TestAllInterleavedConv2d<Conv2dFprop, 32>()));
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+119
@@ -0,0 +1,119 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x2_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
+120
@@ -0,0 +1,120 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Optimized_ImplicitGemm_s8nhwc_s8nhwc_s32nhwc_tensor_op_s32,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = int8_t;
|
||||
using ElementB = int8_t;
|
||||
using ElementC = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
64 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dFprop = cutlass::conv::device::ImplicitGemmConvolution<Conv2dFpropKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+81
@@ -0,0 +1,81 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_fprop.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Fprop_Analytic_ImplicitGemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dFpropKernel = typename cutlass::conv::kernel::DefaultConv2dFprop<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
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::TestAllConv2d<Conv2dFprop>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
@@ -0,0 +1,520 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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 sizes for Conv2d problem
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/layout/matrix.h"
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/conv/conv2d_problem_size.h"
|
||||
|
||||
#define CUTLASS_CONV_UNIT_TEST_RIGOROUS_SIZE_ENABLED 1
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
using Conv2dProblemVector = std::vector<cutlass::conv::Conv2dProblemSize>;
|
||||
|
||||
//
|
||||
// Structures to prune items from Conv2dProblemVector
|
||||
//
|
||||
// Specification template for pruning items for convolution problem lists
|
||||
template <typename T> struct Specification
|
||||
{
|
||||
virtual ~Specification() = default;
|
||||
virtual bool is_satisfied(T item) const = 0;
|
||||
};
|
||||
|
||||
// input size (NHWC) specification
|
||||
struct InputSizeSpecification : Specification<cutlass::conv::Conv2dProblemSize>
|
||||
{
|
||||
cutlass::Tensor4DCoord input_size;
|
||||
|
||||
InputSizeSpecification(cutlass::Tensor4DCoord input_size_) : input_size(input_size_) {}
|
||||
|
||||
bool is_satisfied(cutlass::conv::Conv2dProblemSize item) const override {
|
||||
return ((input_size.n() == item.N) && (input_size.h() == item.H) && (input_size.w() == item.W) && (input_size.c() == item.C));
|
||||
}
|
||||
};
|
||||
|
||||
// stride (stride_h, stride_w) specification
|
||||
struct StrideSpecification : Specification<cutlass::conv::Conv2dProblemSize>
|
||||
{
|
||||
cutlass::MatrixCoord stride;
|
||||
|
||||
StrideSpecification(cutlass::MatrixCoord stride_) : stride(stride_) {}
|
||||
|
||||
bool is_satisfied(cutlass::conv::Conv2dProblemSize item) const override {
|
||||
return ((stride.row() == item.stride_h) && (stride.column() == item.stride_h));
|
||||
}
|
||||
};
|
||||
|
||||
// channel (C,K) specification, must be multiple of minimum channel
|
||||
struct ChannelDivisibilitySpecification : Specification<cutlass::conv::Conv2dProblemSize>
|
||||
{
|
||||
int channel_multiple;
|
||||
|
||||
ChannelDivisibilitySpecification(int channel_multiple_) : channel_multiple(channel_multiple_) {}
|
||||
|
||||
bool is_satisfied(cutlass::conv::Conv2dProblemSize item) const override {
|
||||
return ((item.K % channel_multiple == 0) && (item.C % channel_multiple == 0));
|
||||
}
|
||||
};
|
||||
|
||||
//
|
||||
// Pruning function for items from Conv2dProblemVector based on a Specification
|
||||
//
|
||||
inline Conv2dProblemVector prune(Conv2dProblemVector const &items,
|
||||
Specification<cutlass::conv::Conv2dProblemSize> const &spec)
|
||||
{
|
||||
Conv2dProblemVector pruned_list;
|
||||
|
||||
for (auto& p : items)
|
||||
if (spec.is_satisfied(p))
|
||||
pruned_list.push_back(p);
|
||||
return pruned_list;
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
/// Structure TestbedConv2dProblemSizes initializes and holds conv default and
|
||||
/// important network sizes
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
struct TestbedConv2dProblemSizes {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
int minimum_channel_size;
|
||||
|
||||
Conv2dProblemVector conv2d_default_sizes;
|
||||
Conv2dProblemVector conv2d_rigorous_sizes;
|
||||
Conv2dProblemVector conv2d_resnet50_sizes;
|
||||
Conv2dProblemVector conv2d_resnet50_sizes_perf;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
/// Default ctor
|
||||
TestbedConv2dProblemSizes(int minimum_channel_size_ = 64): minimum_channel_size (minimum_channel_size_) {
|
||||
initialize_conv2d_default_sizes();
|
||||
initialize_conv2d_rigorous_sizes();
|
||||
initialize_conv2d_resnet50_sizes(conv2d_resnet50_sizes, 1 /*batch-size*/);
|
||||
|
||||
initialize_conv2d_resnet50_sizes(conv2d_resnet50_sizes_perf, 34 /*batch-size*/);
|
||||
filter_all();
|
||||
}
|
||||
|
||||
/// Eliminates some illegal cases
|
||||
void filter_all() {
|
||||
|
||||
Conv2dProblemVector *problems_vectors[] = {
|
||||
&conv2d_default_sizes,
|
||||
&conv2d_rigorous_sizes,
|
||||
&conv2d_resnet50_sizes,
|
||||
&conv2d_resnet50_sizes_perf
|
||||
};
|
||||
|
||||
for (Conv2dProblemVector *problems : problems_vectors) {
|
||||
Conv2dProblemVector filtered;
|
||||
|
||||
for (cutlass::conv::Conv2dProblemSize const & problem : *problems) {
|
||||
if (!(problem.C % minimum_channel_size)) {
|
||||
filtered.push_back(problem);
|
||||
}
|
||||
}
|
||||
|
||||
*problems = filtered;
|
||||
}
|
||||
}
|
||||
|
||||
// Add a few standard convolution problem sizes
|
||||
void initialize_conv2d_default_sizes() {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Very Small input size (1x8x8xminimum_channel_size), filter size (3x3 - 7x7), stride (1,1)
|
||||
// C < CTA::K and non-multiples of CTA::K. Typical CTA::K = {32, 64}
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 3, 3, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 4, 4, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 5, 5, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 6, 5, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 6, 6, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 8, 8, minimum_channel_size}, // input size (NHWC)
|
||||
{8, 7, 7, minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
// Medium input size (1x16x16x128), filter size (1x1, 2x2, 3x3, 5x5), stride (1, 1)
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 15, 19, 160}, // input size (NHWC)
|
||||
{224, 1, 1, 160}, // filter size (KRSC)
|
||||
{0, 0, 0, 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, 16, 16, 160}, // input size (NHWC)
|
||||
{224, 2, 3, 160}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 23, 21, 128}, // input size (NHWC)
|
||||
{224, 3, 3, 128}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 29, 37, 160}, // input size (NHWC)
|
||||
{224, 5, 5, 160}, // filter size (KRSC)
|
||||
{2, 2, 2, 2}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
// C > CTA::K and non-multiples of CTA::K. Typical CTA::K = {32, 64}
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 15, 19, 32 + minimum_channel_size}, // input size (NHWC)
|
||||
{96, 3, 3, 32 + minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // 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, 16, 16, 64 + minimum_channel_size}, // input size (NHWC)
|
||||
{96, 3, 3, 64 + minimum_channel_size}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
// Medium input size (1x16x16x128), filter size (1x1, 3,x3, 5x5), stride (2, 2)
|
||||
////////////////////////////////////////////////////////////////////////////////////
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 19, 37, 160}, // input size (NHWC)
|
||||
{224, 3, 3, 160}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 16, 16, 288}, // input size (NHWC)
|
||||
{160, 5, 5, 288}, // filter size (KRSC)
|
||||
{2, 2, 2, 2}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////
|
||||
// Additional input size
|
||||
/////////////////////////////////////////////////////////////////////////////
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{3, 28, 28, 256}, // input size (NHWC)
|
||||
{256, 2, 2, 256}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{32, 32, 32, 32}, // input size (NHWC)
|
||||
{32, 1, 1, 32}, // filter size (KRSC)
|
||||
{0, 0, 0, 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(
|
||||
{4, 3, 3, 128}, // input size (NHWC)
|
||||
{256, 3, 3, 128}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1}, // dilation (dilation_h, dilation_w)
|
||||
{4, 3, 3, 256} // output size (NPQK)
|
||||
));
|
||||
|
||||
conv2d_default_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{4, 1, 1, 256}, // input size (NHWC)
|
||||
{328, 3, 3, 256}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1}, // dilation (dilation_h, dilation_w)
|
||||
{4, 1, 1, 328} // output size (NPQK)
|
||||
));
|
||||
|
||||
}
|
||||
|
||||
|
||||
// Add a few large and rigorous convolution problem sizes
|
||||
void initialize_conv2d_rigorous_sizes() {
|
||||
|
||||
#if CUTLASS_CONV_UNIT_TEST_RIGOROUS_SIZE_ENABLED
|
||||
conv2d_rigorous_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 124, 224, 96}, // input size (NHWC)
|
||||
{24, 7, 7, 96}, // filter size (KRSC)
|
||||
{1, 229, 129, 32} // output size (NPQK)
|
||||
));
|
||||
|
||||
conv2d_rigorous_sizes.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 233, 35, 48}, // input size (NHWC)
|
||||
{24, 7, 5, 48}, // filter size (KRSC)
|
||||
{1, 233, 35, 24} // output size (NPQK)
|
||||
));
|
||||
|
||||
#endif
|
||||
|
||||
}
|
||||
|
||||
|
||||
// Add resent50 layers to unit testing sizes
|
||||
void initialize_conv2d_resnet50_sizes(Conv2dProblemVector &conv2d_problem_vector, int batch_size = 1){
|
||||
|
||||
#if 0 // Resnet50 first layer (layer_id = 0) with channel = 3 is not supported in cutlass
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
[1, 224, 224, 3], // input size (NHWC)
|
||||
[64, 7, 7, 3], // filter size (KRSC)
|
||||
[3, 3, 3, 3], // padding (pad_h, _, pad_w, _)
|
||||
[2, 2], // stride (stride_h, stride_w)
|
||||
[1, 1], // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
#endif
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 64}, // input size (NHWC)
|
||||
{256, 1, 1, 64}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 64}, // input size (NHWC)
|
||||
{64, 1, 1, 64}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 64}, // input size (NHWC)
|
||||
{64, 3, 3, 64}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 256}, // input size (NHWC)
|
||||
{64, 1, 1, 256}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 256}, // input size (NHWC)
|
||||
{512, 1, 1, 256}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 56, 56, 256}, // input size (NHWC)
|
||||
{128, 1, 1, 256}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 28, 28, 128}, // input size (NHWC)
|
||||
{128, 3, 3, 128}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 28, 28, 128}, // input size (NHWC)
|
||||
{512, 1, 1, 128}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 28, 28, 512}, // input size (NHWC)
|
||||
{128, 1, 1, 512}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 28, 28, 512}, // input size (NHWC)
|
||||
{1024, 1, 1, 512}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 28, 28, 512}, // input size (NHWC)
|
||||
{256, 1, 1, 512}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 14, 14, 256}, // input size (NHWC)
|
||||
{256, 3, 3, 256}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 14, 14, 256}, // input size (NHWC)
|
||||
{1024, 1, 1, 256}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 14, 14, 1024}, // input size (NHWC)
|
||||
{256, 1, 1, 1024}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 14, 14, 1024}, // input size (NHWC)
|
||||
{2048, 1, 1, 1024}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 14, 14, 1024}, // input size (NHWC)
|
||||
{512, 1, 1, 1024}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{2, 2}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 7, 7, 512}, // input size (NHWC)
|
||||
{512, 3, 3, 512}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 7, 7, 512}, // input size (NHWC)
|
||||
{2048, 1, 1, 512}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
conv2d_problem_vector.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{batch_size, 7, 7, 2048}, // input size (NHWC)
|
||||
{512, 1, 1, 2048}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
@@ -0,0 +1,558 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#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 "conv2d_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"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
template <typename Conv2d>
|
||||
class TestbedConv2d {
|
||||
public:
|
||||
|
||||
using ElementA = typename Conv2d::ElementA;
|
||||
using LayoutA = typename Conv2d::LayoutA;
|
||||
using ElementB = typename Conv2d::ElementB;
|
||||
using LayoutB = typename Conv2d::LayoutB;
|
||||
using ElementC = typename Conv2d::ElementC;
|
||||
using LayoutC = typename Conv2d::LayoutC;
|
||||
using ElementAccumulator = typename Conv2d::ElementAccumulator;
|
||||
using ElementCompute = typename Conv2d::ElementCompute;
|
||||
using EpilogueOutputOp = typename Conv2d::EpilogueOutputOp;
|
||||
|
||||
static cutlass::conv::Operator const kConvolutionalOperator = Conv2d::kConvolutionalOperator;
|
||||
|
||||
/// Reduction kernel
|
||||
using ReductionOp = cutlass::reduction::thread::ReduceAdd<
|
||||
ElementAccumulator,
|
||||
typename EpilogueOutputOp::ElementAccumulator,
|
||||
EpilogueOutputOp::kCount
|
||||
>;
|
||||
|
||||
using ReductionKernel = cutlass::reduction::kernel::ReduceSplitK<
|
||||
cutlass::MatrixShape<4, 32 * EpilogueOutputOp::kCount>,
|
||||
EpilogueOutputOp,
|
||||
ReductionOp
|
||||
>;
|
||||
|
||||
using ReductionDevice = cutlass::reduction::device::ReduceSplitK<ReductionKernel>;
|
||||
|
||||
|
||||
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<ElementC, LayoutC> tensor_D_computed;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_D_reference;
|
||||
|
||||
public:
|
||||
|
||||
TestbedConv2d(
|
||||
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) {
|
||||
scope = 3;
|
||||
}
|
||||
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::Conv2dProblemSize 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_D_computed.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_D_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
|
||||
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);
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_D_computed.sync_device();
|
||||
tensor_D_reference.sync_device();
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv2d::ImplicitGemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::conv::Conv2dProblemSize const &problem_size,
|
||||
cutlass::conv::SplitKMode const &split_k_mode = cutlass::conv::SplitKMode::kSerial,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waive test if CUDA device is insufficient
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
#if 0 //display conv2d problem size for debugging
|
||||
std::cout << problem_size << std::endl
|
||||
<< "alpha, beta: (" << float(alpha) << ", " << float(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
|
||||
Conv2d conv2d_op;
|
||||
|
||||
typename Conv2d::Arguments conv2d_args(
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D_computed.device_ref(),
|
||||
{alpha, beta},
|
||||
split_k_mode
|
||||
);
|
||||
|
||||
// find workspace requirement for parallel split-k reduction
|
||||
size_t workspace_size = Conv2d::get_workspace_size(conv2d_args);
|
||||
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
cutlass::Status status = conv2d_op.initialize(conv2d_args, workspace.get());
|
||||
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
std::cerr << "This test is not supported: " << cudaGetErrorString(error) << "\n";
|
||||
return true;
|
||||
}
|
||||
|
||||
// conv2d operation with parallel split-k-mode
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// conv2d output is written to workspace in global memory
|
||||
conv2d_args.ref_D.reset(reinterpret_cast<ElementC*>(workspace.get()));
|
||||
// accumulate mma for each cta in k-dimension (1.0 * A * B)
|
||||
conv2d_args.output_op = {ElementCompute(1), ElementCompute(0)};
|
||||
// update conv2d operator arguments
|
||||
status = conv2d_op.update(conv2d_args, workspace.get());
|
||||
}
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run conv2d operator
|
||||
status = conv2d_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// configure parallel reduction operator
|
||||
ReductionDevice reduction_op;
|
||||
|
||||
typename ReductionDevice::Arguments reduction_args(
|
||||
cutlass::conv::implicit_gemm_problem_size(kConvolutionalOperator, problem_size).mn(),
|
||||
problem_size.split_k_slices,
|
||||
cutlass::conv::implicit_gemm_tensor_c_size(kConvolutionalOperator, problem_size),
|
||||
{reinterpret_cast<ElementAccumulator*> (workspace.get()), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_D_computed.device_data(), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_C.device_data(), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{alpha, beta} // apply alpha, beta to obtain the following equation alpha * ReduceAdd(A * B) + beta * C
|
||||
);
|
||||
|
||||
status = reduction_op.initialize(reduction_args, nullptr);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run prallel reduction kernel
|
||||
status = reduction_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
bool passed = false;
|
||||
|
||||
tensor_D_computed.sync_host();
|
||||
|
||||
#if CUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED
|
||||
|
||||
cutlass::reference::device::Conv2d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementCompute,
|
||||
ElementAccumulator
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D_reference.device_ref(),
|
||||
alpha,
|
||||
beta);
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
EXPECT_EQ(result, cudaSuccess) << " device reference error: "
|
||||
<< cudaGetErrorString(result);
|
||||
|
||||
// sync host (copy device data to host) for dumping error output in case of mismatches
|
||||
tensor_D_reference.sync_host();
|
||||
|
||||
#else
|
||||
|
||||
cutlass::reference::host::Conv2d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementCompute,
|
||||
ElementAccumulator
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.host_ref(),
|
||||
tensor_B.host_ref(),
|
||||
tensor_C.host_ref(),
|
||||
tensor_D_reference.host_ref(),
|
||||
alpha,
|
||||
beta);
|
||||
|
||||
#endif
|
||||
passed = cutlass::reference::host::TensorEquals(
|
||||
tensor_D_computed.host_view(),
|
||||
tensor_D_reference.host_view());
|
||||
|
||||
EXPECT_TRUE(passed);
|
||||
|
||||
if (!passed) {
|
||||
std::stringstream fname;
|
||||
|
||||
fname << "error_Conv2d_ImplicitGemm_device_"
|
||||
<< (split_k_mode == cutlass::conv::SplitKMode::kSerial ? "serial_reduction_" : "parallel_reduction_")
|
||||
<< (Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kFprop ? "fprop_" :
|
||||
(Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ? "dgrad_" : "wgrad_"))
|
||||
<< "nhwc_"
|
||||
<< problem_size.N << "x"
|
||||
<< problem_size.H << "x"
|
||||
<< problem_size.W << "x"
|
||||
<< problem_size.C
|
||||
<< "_krsc_"
|
||||
<< problem_size.K << "x"
|
||||
<< problem_size.R << "x"
|
||||
<< problem_size.S << "x"
|
||||
<< problem_size.C
|
||||
<< "_padding_"
|
||||
<< problem_size.pad_h << "x"
|
||||
<< problem_size.pad_w
|
||||
<< "_stride_"
|
||||
<< problem_size.stride_h << "x"
|
||||
<< problem_size.stride_w
|
||||
<< "_dilation_"
|
||||
<< problem_size.dilation_h << "x"
|
||||
<< problem_size.dilation_w << "_"
|
||||
<< (problem_size.mode == cutlass::conv::Mode::kCrossCorrelation ? "xcorr_" : "conv_")
|
||||
<< Conv2d::ThreadblockShape::kM << "x"
|
||||
<< Conv2d::ThreadblockShape::kN << "x"
|
||||
<< Conv2d::ThreadblockShape::kK << "_"
|
||||
<< Conv2d::WarpShape::kM << "x"
|
||||
<< Conv2d::WarpShape::kN << "x"
|
||||
<< Conv2d::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"
|
||||
<< "\nD reference:\n" << tensor_D_reference.host_view() << "\n"
|
||||
<< "\nD computed:\n" << tensor_D_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::TestbedConv2dProblemSizes
|
||||
// Additionaly, each conv2d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
|
||||
// (conv_blacklist_sizes)
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
template <typename ImplicitGemm>
|
||||
bool TestAllConv2d(
|
||||
const Conv2dProblemVector & conv_test_sizes = Conv2dProblemVector(),
|
||||
const Conv2dProblemVector & conv_blacklist_sizes = Conv2dProblemVector()) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
TestbedConv2d<ImplicitGemm> testbed;
|
||||
|
||||
//
|
||||
// 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
|
||||
};
|
||||
|
||||
// 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) {
|
||||
|
||||
// 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::ImplicitGemmKernel::Mma::IteratorA::kStrideSupport ==
|
||||
cutlass::conv::StrideSupport::kUnity)) {
|
||||
if (!((conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
|
||||
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,
|
||||
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;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Sweep split-k-slice using serial and prallel reduction with non-unity alpha and non-zero beta for
|
||||
// a single conv2d problem size. Convolution unit tests take a long time to run so only sweep parameters
|
||||
// which are abolutely neccessary 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::Conv2dProblemSize conv2d_split_k_test_size (
|
||||
{1, 17, 11, 288}, // input size (NHWC)
|
||||
{160, 3, 3, 288}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
);
|
||||
|
||||
cutlass::conv::SplitKMode split_k_modes [] = {
|
||||
cutlass::conv::SplitKMode::kSerial,
|
||||
cutlass::conv::SplitKMode::kParallel,
|
||||
};
|
||||
|
||||
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(
|
||||
conv2d_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;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
@@ -0,0 +1,534 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#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 "conv2d_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/host_reorder.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"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
template <typename Conv2d, int InterleavedK>
|
||||
class InterleavedTestbedConv2d {
|
||||
public:
|
||||
|
||||
using ElementA = typename Conv2d::ElementA;
|
||||
using LayoutA = typename Conv2d::LayoutA;
|
||||
using ElementB = typename Conv2d::ElementB;
|
||||
using LayoutB = typename Conv2d::LayoutB;
|
||||
using ElementC = typename Conv2d::ElementC;
|
||||
using LayoutC = typename Conv2d::LayoutC;
|
||||
using ElementAccumulator = typename Conv2d::ElementAccumulator;
|
||||
using ElementCompute = typename Conv2d::ElementCompute;
|
||||
using EpilogueOutputOp = typename Conv2d::EpilogueOutputOp;
|
||||
|
||||
static cutlass::conv::Operator const kConvolutionalOperator = Conv2d::kConvolutionalOperator;
|
||||
|
||||
/// Reduction kernel
|
||||
using ReductionOp = cutlass::reduction::thread::ReduceAdd<
|
||||
ElementAccumulator,
|
||||
typename EpilogueOutputOp::ElementAccumulator,
|
||||
EpilogueOutputOp::kCount
|
||||
>;
|
||||
|
||||
using ReductionKernel = cutlass::reduction::kernel::ReduceSplitK<
|
||||
cutlass::MatrixShape<4, 32 * EpilogueOutputOp::kCount>,
|
||||
EpilogueOutputOp,
|
||||
ReductionOp
|
||||
>;
|
||||
|
||||
using ReductionDevice = cutlass::reduction::device::ReduceSplitK<ReductionKernel>;
|
||||
|
||||
|
||||
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<ElementB, LayoutB> tensor_B_reordered;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_C;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_D_computed;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_D_reference;
|
||||
|
||||
public:
|
||||
|
||||
InterleavedTestbedConv2d(
|
||||
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) {
|
||||
scope = 3;
|
||||
}
|
||||
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::Conv2dProblemSize 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_B_reordered.resize(implicit_gemm_tensor_b_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_C.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_D_computed.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_D_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
|
||||
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);
|
||||
|
||||
cutlass::reorder_convK<InterleavedK>(
|
||||
tensor_B_reordered.host_ref(), tensor_B.host_ref(), implicit_gemm_problem_size(kConvolutionalOperator, problem_size));
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_B_reordered.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_D_computed.sync_device();
|
||||
tensor_D_reference.sync_device();
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::conv::Conv2dProblemSize const &problem_size,
|
||||
cutlass::conv::SplitKMode const &split_k_mode = cutlass::conv::SplitKMode::kSerial,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
#if 0 //display conv2d problem size for debugging
|
||||
std::cout << problem_size << std::endl
|
||||
<< "alpha, beta: (" << float(alpha) << ", " << float(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
|
||||
Conv2d conv2d_op;
|
||||
|
||||
typename Conv2d::Arguments conv2d_args(
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B_reordered.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D_computed.device_ref(),
|
||||
{alpha, beta},
|
||||
split_k_mode
|
||||
);
|
||||
|
||||
// find workspace requirement for parallel split-k reduction
|
||||
size_t workspace_size = Conv2d::get_workspace_size(conv2d_args);
|
||||
|
||||
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
|
||||
|
||||
cutlass::Status status = conv2d_op.initialize(conv2d_args, workspace.get());
|
||||
|
||||
// conv2d operation with parallel split-k-mode
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// conv2d output is written to workspace in global memory
|
||||
conv2d_args.ref_D.reset(reinterpret_cast<ElementC*>(workspace.get()));
|
||||
// accumulate mma for each cta in k-dimension (1.0 * A * B)
|
||||
conv2d_args.output_op = {ElementCompute(1), ElementCompute(0)};
|
||||
// update conv2d operator arguments
|
||||
status = conv2d_op.update(conv2d_args, workspace.get());
|
||||
}
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run conv2d operator
|
||||
status = conv2d_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// configure parallel reduction operator
|
||||
ReductionDevice reduction_op;
|
||||
|
||||
typename ReductionDevice::Arguments reduction_args(
|
||||
cutlass::conv::implicit_gemm_problem_size(kConvolutionalOperator, problem_size).mn(),
|
||||
problem_size.split_k_slices,
|
||||
cutlass::conv::implicit_gemm_tensor_c_size(kConvolutionalOperator, problem_size),
|
||||
{reinterpret_cast<ElementAccumulator*> (workspace.get()), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_D_computed.device_data(), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_C.device_data(), tensor_C.stride(Conv2d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{alpha, beta} // apply alpha, beta to obtain the following equation alpha * ReduceAdd(A * B) + beta * C
|
||||
);
|
||||
|
||||
status = reduction_op.initialize(reduction_args, nullptr);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run prallel reduction kernel
|
||||
status = reduction_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
bool passed = false;
|
||||
|
||||
tensor_D_computed.sync_host();
|
||||
|
||||
#if CUTLASS_CONV_TEST_UNIT_REFERENCE_DEVICE_ENABLED
|
||||
|
||||
cutlass::reference::device::Conv2d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementCompute,
|
||||
ElementAccumulator,
|
||||
cutlass::NumericConverterClamp<ElementC, ElementCompute>
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D_reference.device_ref(),
|
||||
alpha,
|
||||
beta);
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
EXPECT_EQ(result, cudaSuccess) << " device reference error: "
|
||||
<< cudaGetErrorString(result);
|
||||
|
||||
// sync host (copy device data to host) for dumping error output in case of mismatches
|
||||
tensor_D_reference.sync_host();
|
||||
|
||||
#else
|
||||
|
||||
cutlass::reference::host::Conv2d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementCompute,
|
||||
ElementAccumulator,
|
||||
cutlass::NumericConverterClamp<ElementC, ElementCompute>
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.host_ref(),
|
||||
tensor_B.host_ref(),
|
||||
tensor_C.host_ref(),
|
||||
tensor_D_reference.host_ref(),
|
||||
alpha,
|
||||
beta);
|
||||
|
||||
#endif
|
||||
passed = cutlass::reference::host::TensorEquals(
|
||||
tensor_D_computed.host_view(),
|
||||
tensor_D_reference.host_view());
|
||||
|
||||
EXPECT_TRUE(passed);
|
||||
|
||||
if (!passed) {
|
||||
std::stringstream fname;
|
||||
|
||||
fname << "error_Conv2d_ImplicitGemm_device_"
|
||||
<< (split_k_mode == cutlass::conv::SplitKMode::kSerial ? "serial_reduction_" : "parallel_reduction_")
|
||||
<< (Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kFprop ? "fprop_" :
|
||||
(Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ? "dgrad_" : "wgrad_"))
|
||||
<< "nhwc_"
|
||||
<< problem_size.N << "x"
|
||||
<< problem_size.H << "x"
|
||||
<< problem_size.W << "x"
|
||||
<< problem_size.C
|
||||
<< "_krsc_"
|
||||
<< problem_size.K << "x"
|
||||
<< problem_size.R << "x"
|
||||
<< problem_size.S << "x"
|
||||
<< problem_size.C
|
||||
<< "_padding_"
|
||||
<< problem_size.pad_h << "x"
|
||||
<< problem_size.pad_w
|
||||
<< "_stride_"
|
||||
<< problem_size.stride_h << "x"
|
||||
<< problem_size.stride_w
|
||||
<< "_dilation_"
|
||||
<< problem_size.dilation_h << "x"
|
||||
<< problem_size.dilation_w << "_"
|
||||
<< (problem_size.mode == cutlass::conv::Mode::kCrossCorrelation ? "xcorr_" : "conv_")
|
||||
<< Conv2d::ThreadblockShape::kM << "x"
|
||||
<< Conv2d::ThreadblockShape::kN << "x"
|
||||
<< Conv2d::ThreadblockShape::kK << "_"
|
||||
<< Conv2d::WarpShape::kM << "x"
|
||||
<< Conv2d::WarpShape::kN << "x"
|
||||
<< Conv2d::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"
|
||||
<< "\nB_reordered =\n" << tensor_B_reordered.host_view() << "\n"
|
||||
<< "\nC:\n" << tensor_C.host_view() << "\n"
|
||||
<< "\nD reference:\n" << tensor_D_reference.host_view() << "\n"
|
||||
<< "\nD computed:\n" << tensor_D_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::TestbedConv2dProblemSizes
|
||||
// Additionaly, each conv2d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
|
||||
// (conv_blacklist_sizes)
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
template <typename ImplicitGemm, int InterleavedK>
|
||||
bool TestAllInterleavedConv2d(
|
||||
const Conv2dProblemVector & conv_test_sizes = Conv2dProblemVector(),
|
||||
const Conv2dProblemVector & conv_blacklist_sizes = Conv2dProblemVector()) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
InterleavedTestbedConv2d<ImplicitGemm, InterleavedK> testbed;
|
||||
|
||||
//
|
||||
// Get conv problem sizes to run conv operator
|
||||
//
|
||||
TestbedConv2dProblemSizes conv_problems(InterleavedK); // minimum channel size must be multiple of InterleavedK for interleaved layout
|
||||
|
||||
// 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
|
||||
};
|
||||
|
||||
// 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) {
|
||||
|
||||
ChannelDivisibilitySpecification channel_spec(InterleavedK); //input and output channels must be multiple of InterleavedK
|
||||
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;
|
||||
}
|
||||
|
||||
//
|
||||
// Procedurally disable certain cases
|
||||
//
|
||||
|
||||
// CUTLASS DGRAD's unity stride specialization only support stride {1, 1}
|
||||
if ((ImplicitGemm::kConvolutionalOperator ==
|
||||
cutlass::conv::Operator::kDgrad) &&
|
||||
(ImplicitGemm::ImplicitGemmKernel::Mma::IteratorA::kStrideSupport ==
|
||||
cutlass::conv::StrideSupport::kUnity)) {
|
||||
if (!((conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
|
||||
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,
|
||||
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 0
|
||||
// Sweep split-k-slice using serial and prallel reduction with non-unity alpha and non-zero beta for
|
||||
// a single conv2d problem size. Convolution unit tests take a long time to run so only sweep parameters
|
||||
// which are abolutely neccessary 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::Conv2dProblemSize conv2d_split_k_test_size (
|
||||
{1, 17, 11, 288}, // input size (NHWC)
|
||||
{160, 3, 3, 288}, // filter size (KRSC)
|
||||
{1, 1, 1, 1}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
);
|
||||
|
||||
cutlass::conv::SplitKMode split_k_modes [] = {
|
||||
cutlass::conv::SplitKMode::kSerial,
|
||||
cutlass::conv::SplitKMode::kParallel,
|
||||
};
|
||||
|
||||
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(
|
||||
conv2d_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;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
return passed;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
+172
@@ -0,0 +1,172 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x2_32x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x2_32x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM50_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x2_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm50,
|
||||
cutlass::gemm::GemmShape<32, 64, 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<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
+311
@@ -0,0 +1,311 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
64x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
32x64_8x4_32x64x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_cf32nhwc_cf32nhwc_cf32nhwc_simt_f32,
|
||||
128x128_8x4_64x32x8) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::complex<float>;
|
||||
using ElementB = cutlass::complex<float>;
|
||||
using ElementC = cutlass::complex<float>;
|
||||
using ElementAccumulator = cutlass::complex<float>;
|
||||
using ElementCompute = cutlass::complex<float>;
|
||||
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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<64, 32, 8>,
|
||||
cutlass::gemm::GemmShape<1, 1, 1>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAddComplex,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+122
@@ -0,0 +1,122 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f16nhwc_tensor_op_f16,
|
||||
128x128_64x3_64x64x64) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
using ElementCompute = cutlass::half_t;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+78
@@ -0,0 +1,78 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM70_SUPPORTED)
|
||||
|
||||
TEST(SM70_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm70,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM70_SUPPORTED
|
||||
|
||||
+78
@@ -0,0 +1,78 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM75_SUPPORTED
|
||||
|
||||
+161
@@ -0,0 +1,161 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_f16nhwc_f16nhwc_f32nhwc_tensor_op_f32,
|
||||
64x256_32x4_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32 >,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kOptimized,
|
||||
cutlass::conv::StrideSupport::kStrided
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
+321
@@ -0,0 +1,321 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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 Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
64x64_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 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 Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd,
|
||||
cutlass::conv::IteratorAlgorithm::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
test::conv::device::Conv2dProblemVector user_size;
|
||||
|
||||
user_size.push_back(cutlass::conv::Conv2dProblemSize(
|
||||
{1, 8, 8, 4}, // input size (NHWC)
|
||||
{8, 1, 1, 4}, // filter size (KRSC)
|
||||
{0, 0, 0, 0}, // padding (pad_h, _, pad_w, _)
|
||||
{1, 1}, // stride (stride_h, stride_w)
|
||||
{1, 1} // dilation (dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>(user_size));
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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<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::kAnalytic
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_simt_f32,
|
||||
32x64_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementB,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementC,
|
||||
cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassSimt,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 64, 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::kOptimized
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Optimized_ImplicitGemm_f32nhwc_f32nhwc_f32nhwc_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 Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
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<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 Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+81
@@ -0,0 +1,81 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_conv2d_wgrad.h"
|
||||
#include "cutlass/conv/device/implicit_gemm_convolution.h"
|
||||
|
||||
#include "conv2d_testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Conv2d_Wgrad_Analytic_ImplicitGemm_tf32nhwc_tf32nhwc_f32nhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv2dWgradKernel = typename cutlass::conv::kernel::DefaultConv2dWgrad<
|
||||
ElementA, cutlass::layout::TensorNHWC,
|
||||
ElementB, cutlass::layout::TensorNHWC,
|
||||
ElementC, cutlass::layout::TensorNHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::Kernel;
|
||||
|
||||
using Conv2dWgrad = cutlass::conv::device::ImplicitGemmConvolution<Conv2dWgradKernel>;
|
||||
|
||||
/// Run all unit test sizes with device-level Conv2d instance
|
||||
EXPECT_TRUE(test::conv::device::TestAllConv2d<Conv2dWgrad>());
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
+80
@@ -0,0 +1,80 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv3dDgradKernel = typename cutlass::conv::kernel::DefaultConv3dDgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::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
|
||||
+80
@@ -0,0 +1,80 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv3dFpropKernel = typename cutlass::conv::kernel::DefaultConv3dFprop<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::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 +1,248 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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 sizes for Conv2d problem
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/aligned_buffer.h"
|
||||
#include "cutlass/numeric_types.h"
|
||||
#include "cutlass/layout/matrix.h"
|
||||
#include "cutlass/layout/tensor.h"
|
||||
#include "cutlass/layout/pitch_linear.h"
|
||||
#include "cutlass/core_io.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/conv/conv2d_problem_size.h"
|
||||
#include "cutlass/conv/conv3d_problem_size.h"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
using Conv3dProblemVector = std::vector<cutlass::conv::Conv3dProblemSize>;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
/// Structure TestbedConv3dProblemSizes initializes and holds conv default and
|
||||
/// important network sizes
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
struct TestbedConv3dProblemSizes {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
int minimum_channel_size;
|
||||
Conv3dProblemVector conv3d_default_sizes;
|
||||
Conv3dProblemVector conv3d_vnet_medical_sizes;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
/// Default ctor
|
||||
TestbedConv3dProblemSizes(int minimum_channel_size_ = 64): minimum_channel_size (minimum_channel_size_) {
|
||||
|
||||
initialize_conv3d_default_sizes();
|
||||
initialize_conv3d_vnet_medical_sizes(conv3d_vnet_medical_sizes, 1 /*batch-size*/);
|
||||
|
||||
filter_all();
|
||||
}
|
||||
|
||||
/// Eliminates some illegal cases
|
||||
void filter_all() {
|
||||
|
||||
Conv3dProblemVector *problems_vectors[] = {
|
||||
&conv3d_default_sizes,
|
||||
&conv3d_vnet_medical_sizes
|
||||
};
|
||||
|
||||
for (Conv3dProblemVector *problems : problems_vectors) {
|
||||
Conv3dProblemVector filtered;
|
||||
|
||||
for (cutlass::conv::Conv3dProblemSize const & problem : *problems) {
|
||||
if (!(problem.C % minimum_channel_size)) {
|
||||
filtered.push_back(problem);
|
||||
}
|
||||
}
|
||||
|
||||
*problems = filtered;
|
||||
}
|
||||
}
|
||||
|
||||
// Add a few standard convolution problem sizes
|
||||
void initialize_conv3d_default_sizes() {
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 1, 3, 3, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 1, 1, 1, minimum_channel_size}, // 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)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 1, 16, 16, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 1, 3, 3, minimum_channel_size}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 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, 15, 19, 160}, // input size (NDHWC)
|
||||
{224, 1, 3, 6, 160}, // 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)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 2, 1, 1, minimum_channel_size}, // input size (NDHWC)
|
||||
{8, 2, 1, 1, minimum_channel_size}, // 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)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 1, 7, 7, minimum_channel_size}, // input size (NDHWC)
|
||||
{16, 1, 3, 3, minimum_channel_size}, // 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)
|
||||
));
|
||||
|
||||
conv3d_default_sizes.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{1, 11, 15, 19, 64}, // input size (NDHWC)
|
||||
{32, 4, 3, 6, 64}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({2, 1, 3}), // 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)
|
||||
));
|
||||
}
|
||||
|
||||
// Add vnet layers to unit testing sizes
|
||||
void initialize_conv3d_vnet_medical_sizes(Conv3dProblemVector &conv3d_problem_vector, int batch_size = 1) {
|
||||
|
||||
conv3d_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 32, 32, 32, 16}, // input size (NDHWC)
|
||||
{32, 2, 2, 2, 16}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({2, 2, 2}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
|
||||
conv3d_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 16, 16, 16, 32}, // input size (NDHWC)
|
||||
{32, 3, 3, 3, 32}, // 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_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 16, 16, 16, 32}, // input size (NDHWC)
|
||||
{64, 2, 2, 2, 32}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({2, 2, 2}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
|
||||
conv3d_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 8, 8, 8, 64}, // input size (NDHWC)
|
||||
{64, 3, 3, 3, 64}, // 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_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 8, 8, 8, 64}, // input size (NDHWC)
|
||||
{128, 2, 2, 2, 64}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({2, 2, 2}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
|
||||
conv3d_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 4, 4, 4, 128}, // input size (NDHWC)
|
||||
{128, 3, 3, 3, 128}, // 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_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 8, 8, 8, 128}, // input size (NDHWC)
|
||||
{128, 3, 3, 3, 128}, // 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_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 16, 16, 16, 64}, // input size (NDHWC)
|
||||
{64, 3, 3, 3, 64}, // 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_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 32, 32, 32, 16}, // input size (NDHWC)
|
||||
{64, 2, 2, 2, 16}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({2, 2, 2}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
|
||||
conv3d_problem_vector.push_back(cutlass::conv::Conv3dProblemSize(
|
||||
{batch_size, 16, 16, 16, 32}, // input size (NDHWC)
|
||||
{128, 2, 2, 2, 32}, // filter size (KTRSC)
|
||||
cutlass::Coord<3>({0, 0, 0}), // padding (pad_d, pad_h, pad_w)
|
||||
cutlass::Coord<3>({2, 2, 2}), // stride (stride_d, stride_h, stride_w)
|
||||
cutlass::Coord<3>({1, 1, 1}) // dilation (dilation_d, dilation_h, dilation_w)
|
||||
));
|
||||
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
@@ -0,0 +1,537 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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
|
||||
*/
|
||||
#pragma once
|
||||
|
||||
#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 "cutlass/util/reference/host/tensor_fill.h"
|
||||
|
||||
#include "cutlass/util/reference/host/convolution.h"
|
||||
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
|
||||
#include "cutlass/util/reference/device/convolution.h"
|
||||
#include "cutlass/util/reference/device/tensor_compare.h"
|
||||
|
||||
#include "conv3d_problems.h"
|
||||
#include "cutlass/core_io.h"
|
||||
|
||||
namespace test {
|
||||
namespace conv {
|
||||
namespace device {
|
||||
|
||||
template <typename Conv3d>
|
||||
class TestbedConv3d {
|
||||
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;
|
||||
|
||||
static cutlass::conv::Operator const kConvolutionalOperator = Conv3d::kConvolutionalOperator;
|
||||
|
||||
/// Reduction kernel
|
||||
using ReductionOp = cutlass::reduction::thread::ReduceAdd<
|
||||
ElementAccumulator,
|
||||
typename EpilogueOutputOp::ElementAccumulator,
|
||||
EpilogueOutputOp::kCount
|
||||
>;
|
||||
|
||||
using ReductionKernel = cutlass::reduction::kernel::ReduceSplitK<
|
||||
cutlass::MatrixShape<4, 32 * EpilogueOutputOp::kCount>,
|
||||
EpilogueOutputOp,
|
||||
ReductionOp
|
||||
>;
|
||||
|
||||
using ReductionDevice = cutlass::reduction::device::ReduceSplitK<ReductionKernel>;
|
||||
|
||||
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<ElementC, LayoutC> tensor_D_computed;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_D_reference;
|
||||
|
||||
public:
|
||||
|
||||
TestbedConv3d(
|
||||
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) {
|
||||
scope = 4;
|
||||
}
|
||||
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_D_computed.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
tensor_D_reference.resize(implicit_gemm_tensor_c_extent(kConvolutionalOperator, problem_size));
|
||||
|
||||
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);
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_C.sync_device();
|
||||
tensor_D_computed.sync_device();
|
||||
tensor_D_reference.sync_device();
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Conv3d::ImplicitGemmKernel::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.sharedMemPerMultiprocessor < 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()) {
|
||||
|
||||
// Waive test if CUDA device is insufficient.
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
#if 0 //display conv2d problem size for debugging
|
||||
std::cout << problem_size << std::endl
|
||||
<< "alpha, beta: (" << float(alpha) << ", " << float(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_D_computed.device_ref(),
|
||||
{alpha, beta},
|
||||
split_k_mode
|
||||
);
|
||||
|
||||
// find workspace requirement for parallel split-k reduction
|
||||
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;
|
||||
}
|
||||
|
||||
// conv3d operation with parallel split-k-mode
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// conv3d output is written to workspace in global memory
|
||||
conv3d_args.ref_D.reset(reinterpret_cast<ElementAccumulator*>(workspace.get()));
|
||||
// accumulate mma for each cta in k-dimension (1.0 * A * B)
|
||||
conv3d_args.output_op = {1.0, 0.0};
|
||||
// update conv3d operator arguments
|
||||
status = conv3d_op.update(conv3d_args, workspace.get());
|
||||
}
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run conv3d operator
|
||||
status = conv3d_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (split_k_mode == cutlass::conv::SplitKMode::kParallel) {
|
||||
|
||||
// configure parallel reduction operator
|
||||
ReductionDevice reduction_op;
|
||||
|
||||
typename ReductionDevice::Arguments reduction_args(
|
||||
cutlass::conv::implicit_gemm_problem_size(kConvolutionalOperator, problem_size).mn(),
|
||||
problem_size.split_k_slices,
|
||||
cutlass::conv::implicit_gemm_tensor_c_size(kConvolutionalOperator, problem_size),
|
||||
{reinterpret_cast<ElementAccumulator*> (workspace.get()), tensor_C.stride(Conv3d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_D_computed.device_data(), tensor_C.stride(Conv3d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{tensor_C.device_data(), tensor_C.stride(Conv3d::ImplicitGemmKernel::kTensorCStrideIdx)},
|
||||
{alpha, beta} // apply alpha, beta to obtain the following equation alpha * ReduceAdd(A * B) + beta * C
|
||||
);
|
||||
|
||||
status = reduction_op.initialize(reduction_args, nullptr);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// run prallel reduction kernel
|
||||
status = reduction_op();
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
bool passed = false;
|
||||
|
||||
cutlass::reference::host::Conv3d<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
LayoutC,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>(
|
||||
kConvolutionalOperator,
|
||||
problem_size,
|
||||
tensor_A.host_ref(),
|
||||
tensor_B.host_ref(),
|
||||
tensor_C.host_ref(),
|
||||
tensor_D_reference.host_ref(),
|
||||
alpha,
|
||||
beta
|
||||
);
|
||||
|
||||
tensor_D_computed.sync_host();
|
||||
|
||||
passed = cutlass::reference::host::TensorEquals(
|
||||
tensor_D_computed.host_view(),
|
||||
tensor_D_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_"))
|
||||
<< "ndhwc_"
|
||||
<< problem_size.N << "x"
|
||||
<< problem_size.D << "x"
|
||||
<< problem_size.H << "x"
|
||||
<< problem_size.W << "x"
|
||||
<< problem_size.C
|
||||
<< "_ktrsc_"
|
||||
<< 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"
|
||||
<< "\nD reference:\n" << tensor_D_reference.host_view() << "\n"
|
||||
<< "\nD computed:\n" << tensor_D_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::TestbedConv2dProblemSizes
|
||||
// Additionaly, each conv3d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
|
||||
// (conv_blacklist_sizes)
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <typename ImplicitGemm>
|
||||
bool TestAllConv3d(
|
||||
const Conv3dProblemVector & conv_test_sizes = Conv3dProblemVector(),
|
||||
const Conv3dProblemVector & conv_blacklist_sizes = Conv3dProblemVector()) {
|
||||
|
||||
bool passed = true;
|
||||
|
||||
//
|
||||
// Testbed object
|
||||
//
|
||||
|
||||
//TestbedConv3d<ImplicitGemm> testbed(cutlass::Distribution::Sequential, cutlass::Distribution::Sequential, cutlass::Distribution::Sequential);
|
||||
TestbedConv3d<ImplicitGemm> testbed;
|
||||
|
||||
//
|
||||
// Get conv problem sizes to run conv operator
|
||||
//
|
||||
TestbedConv3dProblemSizes conv3d_problems(128/cutlass::sizeof_bits<typename ImplicitGemm::ElementA>::value);
|
||||
|
||||
//
|
||||
// Get conv problem sizes to run conv operator
|
||||
//
|
||||
//TestbedConv3dProblemSizes conv_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::ImplicitGemmKernel::Mma::IteratorA::kStrideSupport ==
|
||||
cutlass::conv::StrideSupport::kUnity)) {
|
||||
if (!((conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
|
||||
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,
|
||||
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;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Sweep split-k-slice using serial reduction with non-unity alpha and non-zero beta for
|
||||
// a single conv2d problem size. Convolution unit tests take a long time to run so only sweep parameters
|
||||
// which are abolutely neccessary 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,
|
||||
cutlass::conv::SplitKMode::kParallel
|
||||
};
|
||||
|
||||
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;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace device
|
||||
} // namespace conv
|
||||
} // namespace test
|
||||
+78
@@ -0,0 +1,78 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_SM75_SUPPORTED)
|
||||
|
||||
TEST(SM75_Device_Conv3d_Wgrad_Analytic_ImplicitGemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x2_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::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_SM75_SUPPORTED
|
||||
|
||||
+159
@@ -0,0 +1,159 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x4_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::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_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x4_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
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>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Conv3d_Wgrad_Optimized_ImplicitGemm_f16ndhwc_f16ndhwc_f32ndhwc_tensor_op_f32,
|
||||
64x256_32x4_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
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_SM75_SUPPORTED
|
||||
|
||||
+120
@@ -0,0 +1,120 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * 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.
|
||||
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::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_tf32ndhwc_tf32ndhwc_f32ndhwc_tensor_op_f32,
|
||||
128x128_32x3_64x64x32) {
|
||||
|
||||
/// Conv operation element types for the Gemm equivalent (ImplicitGemm)
|
||||
using ElementA = cutlass::tfloat32_t;
|
||||
using ElementB = cutlass::tfloat32_t;
|
||||
using ElementC = float;
|
||||
using ElementAccumulator = float;
|
||||
using ElementCompute = float;
|
||||
|
||||
/// Device-level Conv2d instance
|
||||
using Conv3dWgradKernel = typename cutlass::conv::kernel::DefaultConv3dWgrad<
|
||||
ElementA, cutlass::layout::TensorNDHWC,
|
||||
ElementB, cutlass::layout::TensorNDHWC,
|
||||
ElementC, cutlass::layout::TensorNDHWC,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementC,
|
||||
128 / cutlass::sizeof_bits<ElementC>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
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
|
||||
@@ -804,7 +804,7 @@ TEST(SM61_Epilogue_threadblock_epilogue, simt_i8_i32_32x64_32x64x8) {
|
||||
// Output operator
|
||||
//
|
||||
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
kElementsPerAccess,
|
||||
ElementAccumulator,
|
||||
@@ -874,7 +874,7 @@ TEST(SM61_Epilogue_threadblock_epilogue, simt_i8_i32_32x128_32x64x8) {
|
||||
// Output operator
|
||||
//
|
||||
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
kElementsPerAccess,
|
||||
ElementAccumulator,
|
||||
@@ -944,7 +944,7 @@ TEST(SM61_Epilogue_threadblock_epilogue, simt_i8_i32_64x128_32x64x8) {
|
||||
// Output operator
|
||||
//
|
||||
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
kElementsPerAccess,
|
||||
ElementAccumulator,
|
||||
@@ -1014,7 +1014,7 @@ TEST(SM61_Epilogue_threadblock_epilogue, simt_i8_i32_128x128_32x64x8) {
|
||||
// Output operator
|
||||
//
|
||||
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
kElementsPerAccess,
|
||||
ElementAccumulator,
|
||||
@@ -1084,7 +1084,7 @@ TEST(SM61_Epilogue_threadblock_epilogue, simt_i8_i32_128x64_32x64x8) {
|
||||
// Output operator
|
||||
//
|
||||
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
kElementsPerAccess,
|
||||
ElementAccumulator,
|
||||
|
||||
@@ -20,120 +20,47 @@
|
||||
# STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
add_custom_target(
|
||||
cutlass_test_unit_gemm_device
|
||||
DEPENDS
|
||||
cutlass_test_unit_gemm_device_simt
|
||||
cutlass_test_unit_gemm_device_tensorop_sm70
|
||||
cutlass_test_unit_gemm_device_tensorop_sm75
|
||||
cutlass_test_unit_gemm_device_tensorop_f16_sm80
|
||||
cutlass_test_unit_gemm_device_tensorop_f32_sm80
|
||||
cutlass_test_unit_gemm_device_tensorop_f32_tf32_sm80
|
||||
cutlass_test_unit_gemm_device_tensorop_f64
|
||||
cutlass_test_unit_gemm_device_tensorop_s32_sm80
|
||||
cutlass_test_unit_gemm_device_wmma
|
||||
cutlass_test_unit_gemm_device_tensorop_planar_complex
|
||||
cutlass_test_unit_gemm_device_sparse_tensorop_sm80
|
||||
)
|
||||
|
||||
add_custom_target(
|
||||
test_unit_gemm_device
|
||||
DEPENDS
|
||||
test_unit_gemm_device_simt
|
||||
test_unit_gemm_device_tensorop_sm70
|
||||
test_unit_gemm_device_tensorop_sm75
|
||||
test_unit_gemm_device_tensorop_f16_sm80
|
||||
test_unit_gemm_device_tensorop_f32_sm80
|
||||
test_unit_gemm_device_tensorop_f32_tf32_sm80
|
||||
test_unit_gemm_device_tensorop_f64
|
||||
test_unit_gemm_device_tensorop_s32_sm80
|
||||
test_unit_gemm_device_wmma
|
||||
test_unit_gemm_device_tensorop_planar_complex
|
||||
test_unit_gemm_device_sparse_tensorop_sm80
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_simt
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm70.cu
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm75.cu
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm80.cu
|
||||
|
||||
gemm_universal_f16n_f16t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_universal_cf64n_cf64t_cf64t_tensor_op_f64_sm80.cu
|
||||
gemm_universal_cf64n_cf64t_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
gemm_universal_cf32n_cf32n_cf32n_tensor_op_f32_sm80.cu
|
||||
|
||||
gemm_cf64n_cf64t_cf64t_tensor_op_f64_sm80.cu
|
||||
gemm_cf64t_cf64n_cf64t_tensor_op_f64_sm80.cu
|
||||
|
||||
gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
|
||||
gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32_sm80.cu
|
||||
gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32_sm80.cu
|
||||
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sm80.cu
|
||||
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_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_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
|
||||
gemm_bf16n_bf16n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_bf16t_bf16t_bf16t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32t_tf32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32n_tf32t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32n_tf32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32t_tf32t_f32t_tensor_op_f32_sm80.cu
|
||||
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16n_f16n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_sparse_sm80.cu
|
||||
gemm_f16n_f16t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_sparse_sm80.cu
|
||||
gemm_f16t_f16n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16t_f16t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32t_f32n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32n_f32t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32t_f32t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32n_f32n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sparse_sm80.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sparse_sm80.cu
|
||||
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_slicedk_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_slicedk_sm80.cu
|
||||
|
||||
simt_sgemm_nt_sm80.cu
|
||||
simt_sgemm_tn_sm80.cu
|
||||
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s8n_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s8t_tensor_op_s32_sm80.cu
|
||||
gemm_s4t_s4n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sm80.cu
|
||||
gemm_b1t_b1n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_b1t_b1n_s32t_tensor_op_s32_sm80.cu
|
||||
|
||||
gemm_s8n_s8t_s8n_tensor_op_s32_sm80.cu
|
||||
gemm_s4n_s4t_s4n_tensor_op_s32_sm80.cu
|
||||
|
||||
gemm_f64n_f64t_f64t_tensor_op_f64_sm80.cu
|
||||
gemm_f64t_f64n_f64t_tensor_op_f64_sm80.cu
|
||||
|
||||
gemm_b1t_b1n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_b1t_b1n_s32n_tensor_op_s32_sm75.cu
|
||||
|
||||
gemm_f32n_f32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_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
|
||||
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_f16n_f16n_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16n_f16t_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16n_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16t_f32t_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_f16n_f16n_f32n_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16t_f32n_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_s8n_s8t_s8n_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s32n_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s8t_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s8n_tensor_op_s32_sm75.cu
|
||||
|
||||
gemm_s4n_s4t_s4n_tensor_op_s32_sm75.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_s4t_s4n_s32n_tensor_op_s32_sm75.cu
|
||||
|
||||
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
|
||||
gemm_f16t_f16t_f32t_volta_tensor_op_f32_sm70.cu
|
||||
|
||||
gemm_f16n_f16n_f16t_volta_tensor_op_f32_sm70.cu
|
||||
|
||||
gemm_f16n_f16t_f16t_volta_tensor_op_f16_sm70.cu
|
||||
gemm_f16t_f16n_f16t_volta_tensor_op_f16_sm70.cu
|
||||
|
||||
|
||||
simt_cgemm_nn_sm50.cu
|
||||
simt_cgemm_nt_sm50.cu
|
||||
simt_cgemm_tn_sm50.cu
|
||||
@@ -167,10 +94,158 @@ cutlass_test_unit_add_executable(
|
||||
simt_zgemm_tn_sm50.cu
|
||||
simt_zgemm_tt_sm50.cu
|
||||
|
||||
gemm_splitk_simt_sm50.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_sm70
|
||||
|
||||
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
|
||||
gemm_f16t_f16t_f32t_volta_tensor_op_f32_sm70.cu
|
||||
|
||||
gemm_f16n_f16n_f16t_volta_tensor_op_f32_sm70.cu
|
||||
|
||||
gemm_f16n_f16t_f16t_volta_tensor_op_f16_sm70.cu
|
||||
gemm_f16t_f16n_f16t_volta_tensor_op_f16_sm70.cu
|
||||
|
||||
gemm_splitk_tensor_op_sm70.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_sm75
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_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
|
||||
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_f16n_f16n_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16n_f16t_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16n_f32t_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16t_f32t_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_f16n_f16n_f32n_tensor_op_f32_sm75.cu
|
||||
gemm_f16t_f16t_f32n_tensor_op_f32_sm75.cu
|
||||
|
||||
gemm_s8n_s8t_s8n_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s32n_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s8t_tensor_op_s32_sm75.cu
|
||||
gemm_s8t_s8n_s8n_tensor_op_s32_sm75.cu
|
||||
|
||||
gemm_s4n_s4t_s4n_tensor_op_s32_sm75.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_s4t_s4n_s32n_tensor_op_s32_sm75.cu
|
||||
|
||||
gemm_b1t_b1n_s32t_tensor_op_s32_sm75.cu
|
||||
gemm_b1t_b1n_s32n_tensor_op_s32_sm75.cu
|
||||
|
||||
gemm_splitk_serial_tensor_op_sm75.cu
|
||||
gemm_splitk_tensor_op_sm75.cu
|
||||
gemm_splitk_tensor_op_sm70.cu
|
||||
gemm_splitk_simt_sm50.cu
|
||||
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_f16_sm80
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_slicedk_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_slicedk_sm80.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_f32_sm80
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_universal_f16n_f16t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sm80.cu
|
||||
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_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_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
|
||||
gemm_bf16n_bf16n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_bf16t_bf16t_bf16t_tensor_op_f32_sm80.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_f32_tf32_sm80
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_tf32t_tf32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32n_tf32t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32n_tf32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_tf32t_tf32t_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_universal_cf32n_cf32n_cf32n_tensor_op_f32_sm80.cu
|
||||
gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32_sm80.cu
|
||||
gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32_sm80.cu
|
||||
|
||||
gemm_f32n_f32n_f32t_tensor_op_f32_sm80.cu
|
||||
gemm_f32n_f32n_f32t_tensor_op_bf16_f32_sm80.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_f64
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_f64n_f64t_f64t_tensor_op_f64_sm80.cu
|
||||
gemm_f64t_f64n_f64t_tensor_op_f64_sm80.cu
|
||||
|
||||
gemm_universal_cf64n_cf64t_cf64t_tensor_op_f64_sm80.cu
|
||||
gemm_universal_cf64n_cf64t_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
gemm_cf64n_cf64t_cf64t_tensor_op_f64_sm80.cu
|
||||
gemm_cf64t_cf64n_cf64t_tensor_op_f64_sm80.cu
|
||||
gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian_sm80.cu
|
||||
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_s32_sm80
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s8n_tensor_op_s32_sm80.cu
|
||||
gemm_s8t_s8n_s8t_tensor_op_s32_sm80.cu
|
||||
gemm_s4t_s4n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sm80.cu
|
||||
gemm_b1t_b1n_s32n_tensor_op_s32_sm80.cu
|
||||
gemm_b1t_b1n_s32t_tensor_op_s32_sm80.cu
|
||||
|
||||
gemm_s8n_s8t_s8n_tensor_op_s32_sm80.cu
|
||||
gemm_s4n_s4t_s4n_tensor_op_s32_sm80.cu
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_wmma
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
# wmma floating point tests
|
||||
gemm_f16t_f16n_f16t_wmma_tensor_op_f16_sm70.cu
|
||||
@@ -222,5 +297,37 @@ cutlass_test_unit_add_executable(
|
||||
gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16_sm70.cu
|
||||
|
||||
gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32_sm70.cu
|
||||
|
||||
)
|
||||
|
||||
cutlass_test_unit_add_executable(
|
||||
cutlass_test_unit_gemm_device_tensorop_planar_complex
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_planar_complex_f16_f16_f32_tensor_op_sm70.cu
|
||||
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_sparse_tensorop_sm80
|
||||
|
||||
BATCH_SOURCES ON
|
||||
BATCH_SIZE 4
|
||||
|
||||
gemm_f16n_f16n_f16t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16n_f16n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16n_f16t_f16t_tensor_op_f16_sparse_sm80.cu
|
||||
gemm_f16n_f16t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16t_f16n_f16t_tensor_op_f16_sparse_sm80.cu
|
||||
gemm_f16t_f16n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f16t_f16t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32t_f32n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32n_f32t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32t_f32t_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_f32n_f32n_f32t_tensor_op_f32_sparse_sm80.cu
|
||||
gemm_s8t_s8n_s32t_tensor_op_s32_sparse_sm80.cu
|
||||
gemm_s4t_s4n_s32t_tensor_op_s32_sparse_sm80.cu
|
||||
)
|
||||
|
||||
|
||||
@@ -97,10 +97,45 @@ struct MultistageTestbed {
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Waives test if CUDA device is insufficient
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(cutlass::gemm::GemmCoord problem_size,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waives test if CUDA device is insufficient
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Allocate the GEMM workspace
|
||||
//
|
||||
@@ -144,7 +179,11 @@ struct MultistageTestbed {
|
||||
|
||||
cutlass::Status status = gemm_op.initialize(arguments);
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
std::cerr << "This test is not supported: " << cudaGetErrorString(error) << "\n";
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Run the GEMM
|
||||
|
||||
@@ -39,7 +39,8 @@
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f32n_f32t_f32t_simt_f32, 32x64x8_32x64x1) {
|
||||
@@ -246,4 +247,8 @@ TEST(SM80_Device_Gemm_f32n_f32t_f32t_simt_f32, 128x256x8_64x64x1) {
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -41,8 +41,10 @@
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
TEST(SM80_Device_Gemm_f32t_f32n_f32t_simt_f32, 32x64x8_32x64x1) {
|
||||
|
||||
using Element = float;
|
||||
@@ -246,4 +248,8 @@ TEST(SM80_Device_Gemm_f32t_f32n_f32t_simt_f32, 128x256x8_64x64x1) {
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -247,6 +247,36 @@ struct Testbed {
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
/// Determine if the CUDA device is sufficient to run the kernel
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::gemm::GemmCoord problem_size,
|
||||
@@ -254,6 +284,10 @@ struct Testbed {
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
this->initialize(problem_size);
|
||||
|
||||
@@ -279,7 +313,11 @@ struct Testbed {
|
||||
|
||||
cutlass::Status status = gemm_op.initialize(arguments, workspace.get());
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
cudaError_t error = cudaGetLastError();
|
||||
std::cerr << "This test is not supported: " << cudaGetErrorString(error) << "\n";
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Run the GEMM
|
||||
|
||||
@@ -31,6 +31,7 @@
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
#include <sstream>
|
||||
#include <stdexcept>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
@@ -100,6 +101,34 @@ struct TestbedComplex : public Testbed<Gemm> {
|
||||
return this->compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::gemm::GemmCoord problem_size,
|
||||
@@ -107,7 +136,17 @@ struct TestbedComplex : public Testbed<Gemm> {
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waive the test if device not sufficient
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Initialize workspace
|
||||
//
|
||||
|
||||
this->initialize(problem_size);
|
||||
|
||||
|
||||
//
|
||||
// Initialize the GEMM operator
|
||||
|
||||
@@ -99,6 +99,35 @@ struct InterleavedTestbed {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Waives test if CUDA device is insufficient
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
@@ -107,6 +136,10 @@ struct InterleavedTestbed {
|
||||
cutlass::gemm::GemmCoord problem_size,
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Allocate the GEMM workspace
|
||||
|
||||
@@ -295,6 +295,34 @@ struct SparseTestbed {
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::gemm::GemmCoord problem_size,
|
||||
@@ -302,6 +330,11 @@ struct SparseTestbed {
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
this->initialize(problem_size);
|
||||
|
||||
//
|
||||
@@ -327,7 +360,10 @@ struct SparseTestbed {
|
||||
|
||||
cutlass::Status status = gemm_op.initialize(arguments, workspace.get());
|
||||
|
||||
EXPECT_TRUE(status == cutlass::Status::kSuccess) << to_string(status);
|
||||
// This failure is likely due to insufficient device capabilities. Waive the test.
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return true;
|
||||
}
|
||||
|
||||
//
|
||||
// Run the GEMM
|
||||
|
||||
@@ -250,6 +250,34 @@ struct TestbedUniversal {
|
||||
return compare_reference(problem_size, alpha, beta);
|
||||
}
|
||||
|
||||
bool sufficient() const {
|
||||
//
|
||||
// Determine SMEM requirements and waive if not satisfied
|
||||
//
|
||||
|
||||
int smem_size = int(sizeof(typename Gemm::GemmKernel::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.sharedMemPerMultiprocessor < smem_size) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Executes one test
|
||||
bool run(
|
||||
cutlass::gemm::GemmUniversalMode mode,
|
||||
@@ -258,6 +286,11 @@ struct TestbedUniversal {
|
||||
ElementCompute alpha = ElementCompute(1),
|
||||
ElementCompute beta = ElementCompute(0)) {
|
||||
|
||||
// Waive test if insufficient CUDA device
|
||||
if (!sufficient()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
this->initialize(problem_size);
|
||||
|
||||
//
|
||||
|
||||
@@ -328,19 +328,17 @@ struct SparseTestbed {
|
||||
test::gemm::threadblock::kernel_multistage_mma_sparse<Mma>,
|
||||
cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
|
||||
|
||||
EXPECT_EQ(result, cudaSuccess)
|
||||
<< " cudaFuncSetAttribute "
|
||||
"cudaFuncAttributeMaxDynamicSharedMemorySize error: "
|
||||
<< cudaGetErrorString(result);
|
||||
if (result != cudaSuccess) {
|
||||
return true;
|
||||
}
|
||||
|
||||
result = cudaFuncSetAttribute(
|
||||
test::gemm::threadblock::kernel_multistage_mma_sparse<Mma>,
|
||||
cudaFuncAttributePreferredSharedMemoryCarveout, 100);
|
||||
|
||||
EXPECT_EQ(result, cudaSuccess)
|
||||
<< " cudaFuncSetAttribute "
|
||||
"cudaFuncAttributePreferredSharedMemoryCarveout error: "
|
||||
<< cudaGetErrorString(result);
|
||||
if (result != cudaSuccess) {
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
test::gemm::threadblock::kernel_multistage_mma_sparse<Mma>
|
||||
|
||||
@@ -266,19 +266,17 @@ struct Testbed {
|
||||
test::gemm::threadblock::kernel_multistage_mma<Mma>,
|
||||
cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
|
||||
|
||||
EXPECT_EQ(result, cudaSuccess)
|
||||
<< " cudaFuncSetAttribute "
|
||||
"cudaFuncAttributeMaxDynamicSharedMemorySize error: "
|
||||
<< cudaGetErrorString(result);
|
||||
if (result != cudaSuccess) {
|
||||
return true;
|
||||
}
|
||||
|
||||
result = cudaFuncSetAttribute(
|
||||
test::gemm::threadblock::kernel_multistage_mma<Mma>,
|
||||
cudaFuncAttributePreferredSharedMemoryCarveout, 100);
|
||||
|
||||
EXPECT_EQ(result, cudaSuccess)
|
||||
<< " cudaFuncSetAttribute "
|
||||
"cudaFuncAttributePreferredSharedMemoryCarveout error: "
|
||||
<< cudaGetErrorString(result);
|
||||
if (result != cudaSuccess) {
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
test::gemm::threadblock::kernel_multistage_mma<Mma>
|
||||
|
||||
@@ -199,6 +199,91 @@ TEST(SM70_warp_gemm_tensor_op_crosswise, 64x64x32_64x64x32_16x16x4) {
|
||||
|
||||
test::gemm::warp::Testbed<MmaTensorOp, cutlass::gemm::GemmShape<64, 64, 32> >().run();
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM70_warp_gemm_volta_tensor_op_canonical_f32_row_col, 64x64x16_64x64x4_8x8x4) {
|
||||
|
||||
using Shape = cutlass::gemm::GemmShape<64, 64, 4>;
|
||||
using InstructionShape = cutlass::gemm::GemmShape<8, 8, 4>;
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
|
||||
using Policy = cutlass::gemm::warp::MmaTensorOpPolicy<
|
||||
cutlass::arch::Mma<
|
||||
cutlass::gemm::GemmShape<16, 16, 4>,
|
||||
32,
|
||||
ElementA,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementB,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementC,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>,
|
||||
cutlass::MatrixShape<1, 1>
|
||||
>;
|
||||
|
||||
using MmaTensorOp = cutlass::gemm::warp::MmaVoltaTensorOp<
|
||||
Shape,
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
cutlass::layout::RowMajor,
|
||||
Policy
|
||||
>;
|
||||
|
||||
test::gemm::warp::Testbed<MmaTensorOp,
|
||||
cutlass::gemm::GemmShape<64, 64, 16> >()
|
||||
.run();
|
||||
}
|
||||
|
||||
TEST(SM70_warp_gemm_volta_tensor_op_canonical_f32_col_row, 64x64x16_64x64x4_8x8x4) {
|
||||
|
||||
using Shape = cutlass::gemm::GemmShape<64, 64, 4>;
|
||||
using InstructionShape = cutlass::gemm::GemmShape<8, 8, 4>;
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = float;
|
||||
using LayoutA = cutlass::layout::ColumnMajor;
|
||||
using LayoutB = cutlass::layout::RowMajor;
|
||||
|
||||
using Policy = cutlass::gemm::warp::MmaTensorOpPolicy<
|
||||
cutlass::arch::Mma<
|
||||
cutlass::gemm::GemmShape<16, 16, 4>,
|
||||
32,
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>,
|
||||
cutlass::MatrixShape<1, 1>
|
||||
>;
|
||||
|
||||
using MmaTensorOp = cutlass::gemm::warp::MmaVoltaTensorOp<
|
||||
Shape,
|
||||
ElementA,
|
||||
LayoutA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
ElementC,
|
||||
cutlass::layout::RowMajor,
|
||||
Policy
|
||||
>;
|
||||
|
||||
test::gemm::warp::Testbed<MmaTensorOp,
|
||||
cutlass::gemm::GemmShape<64, 64, 16> >()
|
||||
.run();
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM70_SUPPORTED
|
||||
|
||||
@@ -30,6 +30,7 @@
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/aligned_buffer.h"
|
||||
#include "cutlass/numeric_types.h"
|
||||
#include "cutlass/subbyte_reference.h"
|
||||
#include "cutlass/platform/platform.h"
|
||||
|
||||
@@ -1019,9 +1020,11 @@ __global__ void sparse_kernel(
|
||||
typename Mma::ElementB, ThreadblockShape::kN * ThreadblockShape::kK> smem_buffer_B;
|
||||
|
||||
__shared__ cutlass::AlignedBuffer<
|
||||
typename Mma::ElementE, ThreadblockShape::kM * ThreadblockShape::kK /
|
||||
typename Mma::ElementE, Mma::Shape::kM * Mma::Shape::kK /
|
||||
Mma::kSparse / Mma::kElementsPerElementE>
|
||||
smem_buffer_E;
|
||||
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
typename Mma::ElementA *smem_ptr_A = smem_buffer_A.data();
|
||||
@@ -1168,6 +1171,7 @@ struct SparseTestbed {
|
||||
|
||||
/// Allocates workspace in device memory
|
||||
SparseTestbed() {
|
||||
|
||||
tensor_A.reset(cutlass::make_Coord(ThreadblockShape::kM,
|
||||
ThreadblockShape::kK / Sparse));
|
||||
tensor_A_uncompressed.reset(
|
||||
|
||||
@@ -22,7 +22,6 @@
|
||||
|
||||
add_subdirectory(thread)
|
||||
add_subdirectory(kernel)
|
||||
|
||||
add_custom_target(
|
||||
cutlass_test_unit_reduction
|
||||
DEPENDS
|
||||
|
||||
@@ -81,7 +81,7 @@ __global__ void kernel_gemm_threadblock_tensor_op_multiplicand_store(
|
||||
}
|
||||
}
|
||||
|
||||
// Use iterator to scatter results
|
||||
// Use iterator to store results
|
||||
Iterator iter(ref_output, threadIdx.x);
|
||||
iter.store(frag);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user