CUTLASS 2.2 (#96)
Adds support for NVIDIA Ampere Architecture features. CUDA 11 Toolkit recommended.
This commit is contained in:
Andrew Kerr
@@ -1,4 +1,4 @@
|
||||
# Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
# 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:
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
# Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
# 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:
|
||||
@@ -26,6 +26,64 @@ cutlass_test_unit_add_executable(
|
||||
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_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_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
|
||||
@@ -149,4 +207,5 @@ 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
|
||||
|
||||
)
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -62,7 +62,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 128x256x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -84,7 +84,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 256x128x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -106,7 +106,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 128x128x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -128,7 +128,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 64x128x512_32x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -150,7 +150,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 128x64x512_64x32x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -172,7 +172,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_tensor_op_s32, 64x64x512_32x32x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
|
||||
@@ -0,0 +1,373 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x256x1024_64x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 256x128x1024_64x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x128x1024_64x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 256x64x1024_64x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x256x1024_64x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x128x1024_32x64x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<32, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x64x1024_64x32x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 32, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x64x1024_32x32x1024) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<32, 32, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x256x512_64x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 256x128x512_64x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x128x512_64x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 256x64x512_64x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x256x512_64x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x128x512_32x64x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 512>,
|
||||
cutlass::gemm::GemmShape<32, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 128x64x512_64x32x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 512>,
|
||||
cutlass::gemm::GemmShape<64, 32, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_XOR_b1t_b1n_s32n_tensor_op_s32, 64x64x512_32x32x512) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>,
|
||||
cutlass::gemm::GemmShape<32, 32, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 128x256x512_64x64x512_8x8
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc
|
||||
>;
|
||||
@@ -104,7 +104,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 256x128x512_64x64x512_8x8
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 128x128x512_64x64x512_8x8
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -166,7 +166,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 64x128x512_32x64x512_8x8x
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -197,7 +197,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 128x64x512_64x32x512_8x8x
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -228,7 +228,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32n_wmma_tensor_op_s32, 64x64x512_32x32x512_8x8x1
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -62,7 +62,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 128x256x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -84,7 +84,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 256x128x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -106,7 +106,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 128x128x512_64x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -128,7 +128,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 64x128x512_32x64x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -150,7 +150,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 128x64x512_64x32x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
@@ -172,7 +172,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_tensor_op_s32, 64x64x512_32x32x512) {
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2, 128, 128,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
|
||||
@@ -0,0 +1,374 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x256x1024_64x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 256x128x1024_64x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x128x1024_64x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 256x64x1024_64x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x256x1024_64x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x128x1024_32x64x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 1024>,
|
||||
cutlass::gemm::GemmShape<32, 64, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x64x1024_64x32x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<64, 32, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x64x1024_32x32x1024, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 1024>,
|
||||
cutlass::gemm::GemmShape<32, 32, 1024>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x256x512_64x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 256x128x512_64x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x128x512_64x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 256x64x512_64x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x256x512_64x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 512>,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x128x512_32x64x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 512>,
|
||||
cutlass::gemm::GemmShape<32, 64, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 128x64x512_64x32x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 512>,
|
||||
cutlass::gemm::GemmShape<64, 32, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_XOR_b1t_b1n_s32t_tensor_op_s32, 64x64x512_32x32x512, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::uint1b_t, cutlass::layout::RowMajor, cutlass::uint1b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 512>,
|
||||
cutlass::gemm::GemmShape<32, 32, 512>, cutlass::gemm::GemmShape<16, 8, 256>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6, 128, 128,
|
||||
false, cutlass::arch::OpXorPopc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 128x256x512_64x64x512_8x8
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc
|
||||
>;
|
||||
@@ -104,7 +104,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 256x128x512_64x64x512_8x8
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 128x128x512_64x64x512_8x8
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -166,7 +166,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 64x128x512_32x64x512_8x8x
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -197,7 +197,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 128x64x512_64x32x512_8x8x
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
@@ -228,7 +228,7 @@ TEST(SM75_Device_Gemm_b1t_b1n_s32t_wmma_tensor_op_s32, 64x64x512_32x32x512_8x8x1
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2, 128, 128, false,
|
||||
cutlass::arch::OpXorPopc>;
|
||||
|
||||
|
||||
@@ -0,0 +1,353 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, 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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, 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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, 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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16n_bf16n_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor,
|
||||
cutlass::bfloat16_t, cutlass::layout::ColumnMajor, ElementOutput,
|
||||
cutlass::layout::RowMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -0,0 +1,337 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_bf16t_bf16t_bf16t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = cutlass::bfloat16_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::bfloat16_t, cutlass::layout::RowMajor, cutlass::bfloat16_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -0,0 +1,253 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Operands data type: complex<float>
|
||||
// Rounding: float -> tfloat32_t (half_ulp_truncate)
|
||||
// Instruction operand data type: tfloat32_t (real part) and tfloat32_t (imaginary part)
|
||||
// Math instruction: MMA.1688.F32.TF32
|
||||
// Instruction output/accumulation data type: f32 (real part) and f32 (imaginary part)
|
||||
// Output data type: complex<float>
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 64x64x16_16x32x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 64x64x16_32x32x16) {
|
||||
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 128x64x16_64x32x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 16>,
|
||||
cutlass::gemm::GemmShape<64, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 64x128x16_32x64x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32n_cf32t_cf32t_tensor_op_tf32_f32, 128x128x16_32x64x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,252 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Operands data type: complex<float>
|
||||
// Rounding: float -> tfloat32_t (round to nearest)
|
||||
// Instruction operand data type: tfloat32_t (real part) and tfloat32_t (imaginary part)
|
||||
// Math instruction: MMA.1688.F32.TF32
|
||||
// Instruction output/accumulation data type: f32 (real part) and f32 (imaginary part)
|
||||
// Output data type: complex<float>
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 64x64x16_16x32x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 64x64x16_32x32x16) {
|
||||
|
||||
|
||||
using Element = cutlass::complex<float>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 128x64x16_64x32x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 16>,
|
||||
cutlass::gemm::GemmShape<64, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 64x128x16_32x64x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf32t_cf32n_cf32t_tensor_op_tf32_f32, 128x128x16_32x64x16) {
|
||||
|
||||
using Element = cutlass::complex<float>;;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,192 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian, 32x32x8_16x16x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<16, 16, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian, 64x64x16_16x32x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64_gaussian, 64x64x8_16x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<16, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,246 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 32x32x8_16x16x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<16, 16, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 64x64x16_16x32x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<16, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 64x64x8_16x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<16, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 64x64x16_32x32x16) {
|
||||
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64n_cf64t_cf64t_tensor_op_f64, 64x64x8_32x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,191 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian, 32x32x8_16x16x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<16, 16, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian, 64x64x8_32x16x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 16, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64_gaussian, 64x64x16_32x16x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
cutlass::arch::OpMultiplyAddGaussianComplex
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -0,0 +1,299 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed_complex.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 32x32x8_16x16x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<16, 16, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 64x64x8_32x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 64x128x8_32x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 128x64x8_32x32x8) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 8>,
|
||||
cutlass::gemm::GemmShape<32, 32, 8>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 32x32x16_16x16x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 64x64x16_32x32x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 64x128x16_32x32x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_cf64t_cf64n_cf64t_tensor_op_f64, 128x64x16_32x32x16) {
|
||||
|
||||
using Element = cutlass::complex<double>;
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmComplex<
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::layout::ColumnMajor,
|
||||
Element,
|
||||
cutlass::layout::RowMajor,
|
||||
Element,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
Element,
|
||||
1,
|
||||
Element,
|
||||
Element
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmComplex<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -141,7 +141,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -105,7 +105,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -139,7 +139,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -122,7 +122,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -153,7 +153,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -205,7 +205,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f16t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -256,7 +256,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_volta_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -321,7 +321,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -355,7 +355,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -389,7 +389,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -133,7 +133,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -164,7 +164,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -195,7 +195,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -226,7 +226,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -257,7 +257,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -288,7 +288,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -320,7 +320,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -354,7 +354,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -388,7 +388,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -122,7 +122,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -153,7 +153,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -205,7 +205,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,337 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16n_f32n_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -73,7 +73,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32n_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -108,7 +108,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -142,7 +142,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -122,7 +122,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -153,7 +153,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -205,7 +205,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM75_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,340 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x256x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 256x128x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x128x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 256x64x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x256x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x128x64_32x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x64x64_64x32x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x64x64_32x32x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x256x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 256x128x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x128x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 256x64x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x256x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x128x32_32x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 128x64x32_64x32x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16n_f32t_tensor_op_f32, 64x64x32_32x32x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -256,7 +256,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_volta_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 64x128x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -293,7 +293,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -327,7 +327,7 @@ TEST(SM70_Device_Gemm_f16n_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -141,7 +141,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -105,7 +105,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -139,7 +139,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16_sliced_k, 64x64x64_64x32x32)
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16_sliced_k, 128x64x64_64x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x64x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64> ,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x128x64_32x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x64x64_64x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x64x64_32x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 256x64x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16n_f16t_f16t_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -0,0 +1,77 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/gemm/device/gemm_universal.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmUniversal_f16n_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
|
||||
/*
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
*/
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -63,7 +63,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -94,7 +94,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -125,7 +125,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -156,7 +156,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -187,7 +187,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -218,7 +218,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -249,7 +249,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_volta_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -321,7 +321,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -355,7 +355,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -389,7 +389,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -73,7 +73,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -108,7 +108,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -142,7 +142,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,339 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x256x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 256x128x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x128x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 256x64x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x256x64_64x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x128x64_32x64x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x64x64_64x32x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x64x64_32x32x64, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x256x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 256x128x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x128x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 256x64x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x256x32_64x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x128x32_32x64x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 128x64x32_64x32x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_f16n_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32, {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -63,7 +63,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -94,7 +94,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -125,7 +125,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -156,7 +156,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -187,7 +187,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -218,7 +218,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -249,7 +249,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_volta_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 64x128x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -293,7 +293,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -327,7 +327,7 @@ TEST(SM70_Device_Gemm_f16n_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -73,7 +73,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 128x256x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -105,7 +105,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 128x64x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -137,7 +137,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 64x128x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -170,7 +170,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 64x64x32_32
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -202,7 +202,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 64x64x64_32
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -234,7 +234,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 128x128x64_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -270,7 +270,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -305,7 +305,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_singlestage_wmma_tensor_op_f16, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -141,7 +141,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f32, 128x128x32_64x64x16_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -106,7 +106,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f32, 64x64x32_64x64x16_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -140,7 +140,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16n_wmma_tensor_op_f32, 64x64x32_64x64x16_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -73,7 +73,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 128x256x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -105,7 +105,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 128x64x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -137,7 +137,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 64x128x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -170,7 +170,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 64x64x32_32
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -202,7 +202,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 64x64x64_32
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -234,7 +234,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 128x128x64_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -270,7 +270,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -305,7 +305,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_singlestage_wmma_tensor_op_f16, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16_sliced_k, 64x64x64_64x32x32)
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,83 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16_sliced_k, 128x64x64_64x64x32) {
|
||||
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,339 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 256x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x128x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 256x64x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x256x64_64x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x128x64_32x64x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x64x64_64x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x64x64_32x32x64) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 256x64x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x256x32_64x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f16t_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
using ElementOutput = cutlass::half_t;
|
||||
using ElementAccumulator = cutlass::half_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -256,7 +256,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_volta_tensor_op_f16, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -321,7 +321,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -355,7 +355,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -389,7 +389,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -133,7 +133,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -164,7 +164,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -195,7 +195,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -226,7 +226,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -257,7 +257,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -288,7 +288,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -319,7 +319,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -353,7 +353,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -387,7 +387,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -141,7 +141,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -74,7 +74,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32, 128x64x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -106,7 +106,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32, 64x128x32_6
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -138,7 +138,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32, 64x64x32_32
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -174,7 +174,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
@@ -209,7 +209,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_singlestage_wmma_tensor_op_f32, 128x128x32_
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
kStages
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16n_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // CUTLASS_ARCH_MMA_SM80_SUPPORTED
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 64x64x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -256,7 +256,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_volta_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 64x128x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -293,7 +293,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -327,7 +327,7 @@ TEST(SM70_Device_Gemm_f16t_f16n_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -141,7 +141,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -106,7 +106,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -140,7 +140,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16n_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -289,7 +289,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -321,7 +321,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -355,7 +355,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -389,7 +389,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f16, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x128x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -133,7 +133,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 128x64x32_64x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -164,7 +164,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -195,7 +195,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -226,7 +226,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -257,7 +257,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -288,7 +288,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x128x32_32x64x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -320,7 +320,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -354,7 +354,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_32x8x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -388,7 +388,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f16t_wmma_tensor_op_f32, 64x64x32_64x64x32_8x32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32n_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -105,7 +105,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -139,7 +139,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32n_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM75_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 256x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x128x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 256x64x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x256x64_64x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x128x64_32x64x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x64x64_64x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x64x64_32x32x64) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 256x64x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x256x32_64x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 32>,
|
||||
cutlass::gemm::GemmShape<32, 64, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 32>,
|
||||
cutlass::gemm::GemmShape<64, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_f16t_f16t_f32t_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::half_t, cutlass::layout::RowMajor, cutlass::half_t,
|
||||
cutlass::layout::RowMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>, cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementAccumulator>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 10>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -70,7 +70,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 128x256x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -101,7 +101,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 256x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -132,7 +132,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -163,7 +163,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 64x128x32_32x64x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -194,7 +194,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 128x64x32_64x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -225,7 +225,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_volta_tensor_op_f32, 64x64x32_32x32x32) {
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 64x64x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -134,7 +134,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 128x256x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -165,7 +165,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 256x128x32_64x64x32_16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -196,7 +196,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 128x64x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -227,7 +227,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 64x128x32_64x32x32_16x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -258,7 +258,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 64x64x32_32x32x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -293,7 +293,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_32x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -327,7 +327,7 @@ TEST(SM70_Device_Gemm_f16t_f16t_f32t_wmma_tensor_op_f32, 128x128x32_64x64x32_8x3
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,87 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface using BF16.
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/arch/mma.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f32t_f32n_f32t_tensor_op_bf16_f32, 128x128x32_64x64x32) {
|
||||
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::layout::ColumnMajor,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
4,
|
||||
4,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddFastBF16
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f32n_f32n_f32t_tensor_op_f32, 128x128x32_64x64x32) {
|
||||
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
float,
|
||||
cutlass::layout::ColumnMajor,
|
||||
float,
|
||||
cutlass::layout::ColumnMajor,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 32>,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -0,0 +1,212 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64, 32x32x16_16x16x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64, 64x64x16_32x32x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64, 128x64x16_64x32x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 16>,
|
||||
cutlass::gemm::GemmShape<64, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64, 64x128x16_32x64x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64n_f64t_f64t_tensor_op_f64, 128x128x16_32x64x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -0,0 +1,212 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64, 32x32x16_16x16x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<16, 16, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64, 64x64x16_32x32x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 16>,
|
||||
cutlass::gemm::GemmShape<32, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64, 64x128x16_32x64x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64, 128x64x16_64x32x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 16>,
|
||||
cutlass::gemm::GemmShape<64, 32, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_f64t_f64n_f64t_tensor_op_f64, 128x128x16_32x64x16) {
|
||||
|
||||
using ElementOutput = double;
|
||||
using ElementAccumulator = double;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
double,
|
||||
cutlass::layout::RowMajor,
|
||||
double,
|
||||
cutlass::layout::ColumnMajor,
|
||||
ElementOutput,
|
||||
cutlass::layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 16>,
|
||||
cutlass::gemm::GemmShape<32, 64, 16>,
|
||||
cutlass::gemm::GemmShape<8, 8, 4>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
ElementOutput,
|
||||
1,
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3
|
||||
>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -63,7 +63,7 @@ using gemm_planar_complex_s884_tn_base = typename cutlass::gemm::kernel::Default
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
@@ -107,7 +107,7 @@ using gemm_planar_complex_s884_nt_base = typename cutlass::gemm::kernel::Default
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
@@ -0,0 +1,217 @@
|
||||
/***************************************************************************************************
|
||||
* 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-level GEMM API for Planar Complex.
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/gemm/kernel/default_gemm_planar_complex_universal.h"
|
||||
#include "cutlass/gemm/device/gemm_universal_base.h"
|
||||
#include "cutlass/gemm/device/gemm_universal_adapter.h"
|
||||
|
||||
#include "testbed_planar_complex.h"
|
||||
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s1688_tn_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s1688_tn : gemm_planar_complex_s1688_tn_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_GemmPlanarComplex_f16t_f16n_f32n_tensor_op_f32_1688, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s1688_tn>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s1688_hc_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s1688_hc : gemm_planar_complex_s1688_hc_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_GemmPlanarComplex_f16h_f16c_f32n_tensor_op_f32_1688, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s1688_hc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s1688_nt_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s1688_nt : gemm_planar_complex_s1688_nt_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_GemmPlanarComplex_f16n_f16t_f32n_tensor_op_f32_1688, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s1688_nt>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s1688_ch_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm75,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 8>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s1688_ch : gemm_planar_complex_s1688_ch_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM75_Device_GemmPlanarComplex_f16c_f16h_f32n_tensor_op_f32_1688, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s1688_ch>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM75_SUPPORTED)
|
||||
@@ -0,0 +1,216 @@
|
||||
/***************************************************************************************************
|
||||
* 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-level GEMM API for Planar Complex.
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/gemm/kernel/default_gemm_planar_complex_universal.h"
|
||||
#include "cutlass/gemm/device/gemm_universal_adapter.h"
|
||||
|
||||
#include "testbed_planar_complex.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s16816_tn_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s16816_tn : gemm_planar_complex_s16816_tn_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmPlanarComplex_f16t_f16n_f32n_tensor_op_f32_16816, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s16816_tn>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s16816_hc_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s16816_hc : gemm_planar_complex_s16816_hc_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmPlanarComplex_f16h_f16c_f32n_tensor_op_f32_16816, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s16816_hc>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s16816_nt_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kNone,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s16816_nt : gemm_planar_complex_s16816_nt_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmPlanarComplex_f16n_f16t_f32n_tensor_op_f32_16816, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s16816_nt>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
using gemm_planar_complex_s16816_ch_base = typename cutlass::gemm::kernel::DefaultGemmPlanarComplexUniversal<
|
||||
cutlass::half_t,
|
||||
cutlass::layout::ColumnMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
cutlass::half_t,
|
||||
cutlass::layout::RowMajor,
|
||||
cutlass::ComplexTransform::kConjugate,
|
||||
8,
|
||||
float,
|
||||
cutlass::layout::RowMajor,
|
||||
float,
|
||||
cutlass::arch::OpClassTensorOp,
|
||||
cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 32>,
|
||||
cutlass::gemm::GemmShape<32, 32, 32>,
|
||||
cutlass::gemm::GemmShape<16, 8, 16>,
|
||||
cutlass::epilogue::thread::LinearCombinationPlanarComplex<
|
||||
float,
|
||||
4,
|
||||
float,
|
||||
float
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
cutlass::arch::OpMultiplyAdd
|
||||
>::GemmKernel;
|
||||
|
||||
struct gemm_planar_complex_s16816_ch : gemm_planar_complex_s16816_ch_base {
|
||||
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_GemmPlanarComplex_f16c_f16h_f32n_tensor_op_f32_16816, 64x64x32_32x32x32) {
|
||||
|
||||
using Gemm = cutlass::gemm::device::GemmUniversalAdapter<gemm_planar_complex_s16816_ch>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemmPlanarComplex<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -107,7 +107,7 @@ TEST(SM75_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -143,7 +143,7 @@ TEST(SM75_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -179,7 +179,7 @@ TEST(SM75_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,213 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "multistage_testbed_interleaved.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<64>,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajorInterleaved<64>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
32,
|
||||
32,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 64> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<64>,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajorInterleaved<64>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
32,
|
||||
32,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 64> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<64>,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajorInterleaved<64>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
32,
|
||||
32,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 64> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s4n_s4t_s4n_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
|
||||
using ElementOutput = cutlass::int4b_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<64>,
|
||||
cutlass::int4b_t,
|
||||
cutlass::layout::RowMajorInterleaved<64>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
32,
|
||||
32,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 64> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -167,7 +167,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -199,7 +199,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x64x128_64x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -231,7 +231,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x64x128_32x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,354 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x256x256_64x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 256x128x256_64x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x128x256_64x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 256x64x256_64x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x256x256_64x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x128x256_32x64x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 256>,
|
||||
cutlass::gemm::GemmShape<32, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x64x256_64x32x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 256>,
|
||||
cutlass::gemm::GemmShape<64, 32, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x64x256_32x32x256) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>,
|
||||
cutlass::gemm::GemmShape<32, 32, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 256x64x128_64x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x256x128_64x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
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<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 128x64x128_64x32x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
TEST(SM80_Device_Gemm_s4t_s4n_s32n_tensor_op_s32, 64x64x128_32x32x128) {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::ColumnMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif //#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 128x256x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -104,7 +104,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 256x128x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -136,7 +136,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 128x128x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -168,7 +168,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 64x128x128_32x64x128_8x8x
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -200,7 +200,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 128x64x128_64x32x128_8x8x
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -232,7 +232,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32n_wmma_tensor_op_s32, 64x64x128_32x32x128_8x8x3
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -167,7 +167,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -199,7 +199,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x64x128_64x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -231,7 +231,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x64x128_32x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,357 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x256x256_64x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 256x128x256_64x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x128x256_64x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>,
|
||||
cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 256x64x256_64x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x256x256_64x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 256>,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x128x256_32x64x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 256>,
|
||||
cutlass::gemm::GemmShape<32, 64, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x64x256_64x32x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 256>,
|
||||
cutlass::gemm::GemmShape<64, 32, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x64x256_32x32x256, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 256>,
|
||||
cutlass::gemm::GemmShape<32, 32, 256>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x256x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 256x128x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x128x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 256x64x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x256x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x256x128_32x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 128x64x128_64x32x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s4t_s4n_s32t_tensor_op_s32, 64x64x128_32x32x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
cutlass::int4b_t, cutlass::layout::RowMajor, cutlass::int4b_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>, cutlass::gemm::GemmShape<16, 8, 64>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -72,7 +72,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 128x256x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -104,7 +104,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 256x128x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -136,7 +136,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 128x128x128_64x64x128_8x8
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -168,7 +168,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 64x128x128_32x64x128_8x8x
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -200,7 +200,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 128x64x128_64x32x128_8x8x
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -232,7 +232,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s32t_wmma_tensor_op_s32, 64x64x128_32x32x128_8x8x3
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 128x256x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 256x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 128x128x128_64x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -167,7 +167,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 64x128x128_32x64x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -199,7 +199,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 128x64x128_64x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -231,7 +231,7 @@ TEST(SM75_Device_Gemm_s4t_s4n_s4n_tensor_op_s32, 64x64x128_32x32x128) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -69,7 +69,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 32x64x64_16x32x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 64x64x64_32x32x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -137,7 +137,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x64x64_64x32x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -171,7 +171,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 64x128x64_32x64x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -205,7 +205,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x128x64_64x64x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -239,7 +239,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 256x128x64_64x64x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -273,7 +273,7 @@ TEST(SM75_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x256x64_64x64x64) {
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,361 @@
|
||||
/***************************************************************************************************
|
||||
* 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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
|
||||
#include "multistage_testbed_interleaved.h"
|
||||
|
||||
#if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 64x64x64_32x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
6,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x64x64_64x32x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 64x128x64_32x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
4,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x128x64_64x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 256x128x64_64x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 64x256x64_64x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 256x64x64_64x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
TEST(SM80_Device_Gemm_s8n_s8t_s8n_tensor_op_s32, 128x256x64_64x64x64) {
|
||||
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t,
|
||||
cutlass::layout::ColumnMajorInterleaved<32>,
|
||||
int8_t,
|
||||
cutlass::layout::RowMajorInterleaved<32>,
|
||||
ElementOutput,
|
||||
cutlass::layout::ColumnMajorInterleaved<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::FastLinearCombinationClamp<
|
||||
ElementOutput,
|
||||
64 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
3,
|
||||
16,
|
||||
16,
|
||||
false,
|
||||
cutlass::arch::OpMultiplyAddSaturate,
|
||||
true
|
||||
>;
|
||||
|
||||
test::gemm::device::MultistageInterleavedTestbed<Gemm, 32> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 128x256x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 256x128x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 128x128x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -167,7 +167,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 64x128x64_32x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -199,7 +199,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 128x64x64_64x32x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -231,7 +231,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_tensor_op_s32, 64x64x64_32x32x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_wmma_tensor_op_s32, 128x128x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -102,7 +102,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_wmma_tensor_op_s32, 64x128x64_32x32x64_16x16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -136,7 +136,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32n_wmma_tensor_op_s32, 64x128x64_32x64x64_8x32x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x256x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 256x128x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -135,7 +135,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x128x64_64x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -167,7 +167,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x128x64_32x64x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -199,7 +199,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x64x64_64x32x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -231,7 +231,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x64x64_32x32x64) {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,355 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#include "testbed.h"
|
||||
|
||||
#if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x256x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 256x128x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x128x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 256x64x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x256x128_64x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x128x128_32x64x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x64x128_64x32x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x64x128_32x32x128, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x256x64_64x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 256x128x64_64x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x128x64_64x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 256x64x64_64x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x256x64_64x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 64>,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x128x64_32x64x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 64>,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 128x64x64_64x32x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 64>,
|
||||
cutlass::gemm::GemmShape<64, 32, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L1(SM80_Device_Gemm_s8t_s8n_s32t_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
using ElementOutput = int32_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = int32_t;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t,
|
||||
cutlass::layout::ColumnMajor, ElementOutput, cutlass::layout::RowMajor,
|
||||
ElementAccumulator, cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>,
|
||||
cutlass::gemm::GemmShape<32, 32, 64>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::LinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value,
|
||||
ElementAccumulator, ElementCompute>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#endif // #if defined(CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -71,7 +71,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_wmma_tensor_op_s32, 128x128x32_64x64x32_16x16
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -103,7 +103,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_wmma_tensor_op_s32, 64x128x64_32x32x64_16x16x
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -137,7 +137,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_wmma_tensor_op_s32, 64x128x64_32x64x64_32x8x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -171,7 +171,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s32t_wmma_tensor_op_s32, 64x128x64_32x64x64_8x32x1
|
||||
ElementAccumulator,
|
||||
ElementAccumulator
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -59,7 +59,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x256x64_64x64x64,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2>;
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
|
||||
@@ -78,7 +78,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x128x64_64x64x64,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2>;
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
} )
|
||||
@@ -96,7 +96,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x128x64_64x64x64,
|
||||
cutlass::gemm::GemmShape<64, 64, 64>, cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2>;
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
|
||||
@@ -115,7 +115,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x128x64_32x64x64,
|
||||
cutlass::gemm::GemmShape<32, 64, 64>, cutlass::gemm::GemmShape<8, 8, 16>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle, 2>;
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 2>;
|
||||
|
||||
EXPECT_TRUE(test::gemm::device::TestAllGemm<Gemm>());
|
||||
|
||||
@@ -146,7 +146,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x64x64_64x32x64,
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -179,7 +179,7 @@ CUTLASS_TEST_L0(SM75_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
ElementAccumulator,
|
||||
ElementCompute
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
@@ -0,0 +1,368 @@
|
||||
/**************************************************************************************************
|
||||
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 GEMM interface
|
||||
*/
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "../../common/cutlass_unit_test.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/gemm/device/gemm.h"
|
||||
#include "multistage_testbed.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
#include "cutlass/util/reference/host/gemm.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
#if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x256x128_64x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x128x128_64x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x128x128_64x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 128, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x64x128_64x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<256, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x256x128_64x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 256, 128>,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x128x128_32x64x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 128, 128>,
|
||||
cutlass::gemm::GemmShape<32, 64, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x64x128_64x32x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<128, 64, 128>,
|
||||
cutlass::gemm::GemmShape<64, 32, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x64x128_32x32x128, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, ElementAccumulator,
|
||||
cutlass::arch::OpClassTensorOp, cutlass::arch::Sm80,
|
||||
cutlass::gemm::GemmShape<64, 64, 128>,
|
||||
cutlass::gemm::GemmShape<32, 32, 128>, cutlass::gemm::GemmShape<16, 8, 32>,
|
||||
cutlass::epilogue::thread::FastLinearCombinationClamp<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x256x64_64x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x128x64_64x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x128x64_64x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 256x64x64_64x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x256x64_64x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 3>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x128x64_32x64x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 128 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 128x64x64_64x32x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 4>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
CUTLASS_TEST_L0(SM80_Device_Gemm_s8t_s8n_s8n_tensor_op_s32, 64x64x64_32x32x64, {
|
||||
using ElementOutput = int8_t;
|
||||
using ElementAccumulator = int32_t;
|
||||
using ElementCompute = float;
|
||||
|
||||
using Gemm = cutlass::gemm::device::Gemm<
|
||||
int8_t, cutlass::layout::RowMajor, int8_t, cutlass::layout::ColumnMajor,
|
||||
ElementOutput, cutlass::layout::ColumnMajor, 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::FastLinearCombinationClamp<
|
||||
ElementOutput, 64 / cutlass::sizeof_bits<ElementOutput>::value>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>, 6>;
|
||||
|
||||
test::gemm::device::MultistageTestbed<Gemm> testbed;
|
||||
|
||||
EXPECT_TRUE(testbed.run_all());
|
||||
} )
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
#endif // if (CUTLASS_ARCH_MMA_SM80_SUPPORTED)
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
* 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:
|
||||
@@ -69,7 +69,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s8n_wmma_tensor_op_s32, 128x128x32_64x64x32_16x16x
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -98,7 +98,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s8n_wmma_tensor_op_s32, 64x128x64_32x32x64_16x16x1
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -130,7 +130,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s8n_wmma_tensor_op_s32, 64x128x64_32x64x64_32x8x16
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
@@ -162,7 +162,7 @@ TEST(SM75_Device_Gemm_s8t_s8n_s8n_wmma_tensor_op_s32, 64x128x64_32x64x64_8x32x16
|
||||
ElementOutput,
|
||||
128 / cutlass::sizeof_bits<ElementOutput>::value
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,
|
||||
2
|
||||
>;
|
||||
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user