fb335f6a5f
CUTLASS 2.0 Substantially refactored for - Better performance, particularly for native Turing Tensor Cores - Robust and durable templates spanning the design space - Encapsulated functionality embodying modern C++11 programming techniques - Optimized containers and data types for efficient, generic, portable device code Updates to: - Quick start guide - Documentation - Utilities - CUTLASS Profiler Native Turing Tensor Cores - Efficient GEMM kernels targeting Turing Tensor Cores - Mixed-precision floating point, 8-bit integer, 4-bit integer, and binarized operands Coverage of existing CUTLASS functionality: - GEMM kernels targeting CUDA and Tensor Cores in NVIDIA GPUs - Volta Tensor Cores through native mma.sync and through WMMA API - Optimizations such as parallel reductions, threadblock rasterization, and intra-threadblock reductions - Batched GEMM operations - Complex-valued GEMMs Note: this commit and all that follow require a host compiler supporting C++11 or greater.
486 lines
12 KiB
Plaintext
486 lines
12 KiB
Plaintext
/***************************************************************************************************
|
|
* Copyright (c) 2017-2019, 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 Unit tests for thread-level GEMM
|
|
*/
|
|
|
|
#include <fstream>
|
|
|
|
#include "../../common/cutlass_unit_test.h"
|
|
|
|
#include "cutlass/aligned_buffer.h"
|
|
|
|
#include "cutlass/gemm/warp/mma_simt.h"
|
|
#include "cutlass/gemm/warp/mma_simt_policy.h"
|
|
|
|
#include "cutlass/epilogue/thread/linear_combination.h"
|
|
#include "cutlass/epilogue/threadblock/default_epilogue_simt.h"
|
|
|
|
#include "cutlass/util/host_tensor.h"
|
|
#include "cutlass/util/tensor_view_io.h"
|
|
#include "cutlass/util/reference/host/tensor_fill.h"
|
|
|
|
#include "testbed.h"
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
//
|
|
// Real-valued half precision tests
|
|
//
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_32x64_32x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<32, 64, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<32, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<4, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_64x64_64x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<8, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_64x128_64x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<64, 128, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<8, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_128x128_64x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<128, 128, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<8, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_128x256_64x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<128, 256, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<8, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
TEST(SM60_Epilogue_threadblock_epilogue, simt_f16_256x128_64x64x8) {
|
|
|
|
//
|
|
// Define the warp-level matrix multiply
|
|
//
|
|
|
|
using Element = cutlass::half_t;
|
|
using ElementOutput = cutlass::half_t;
|
|
using ElementAccumulator = cutlass::half_t;
|
|
using ElementCompute = cutlass::half_t;
|
|
|
|
int const kElementsPerAccess = 1;
|
|
|
|
using Shape = cutlass::gemm::GemmShape<256, 128, 8>;
|
|
using WarpShape = cutlass::gemm::GemmShape<64, 64, 8>;
|
|
using ElementC = ElementAccumulator;
|
|
using LayoutA = cutlass::layout::ColumnMajor;
|
|
using LayoutB = cutlass::layout::RowMajor;
|
|
using LayoutC = cutlass::layout::RowMajor;
|
|
|
|
using ElementOutput = Element;
|
|
using ElementAccumulator = Element;
|
|
using ElementCompute = Element;
|
|
|
|
using WarpMmaSimt = cutlass::gemm::warp::MmaSimt<
|
|
WarpShape,
|
|
Element,
|
|
LayoutA,
|
|
Element,
|
|
LayoutB,
|
|
Element,
|
|
LayoutC,
|
|
cutlass::gemm::warp::MmaSimtPolicy<
|
|
cutlass::MatrixShape<4, 8>,
|
|
cutlass::layout::RowMajorInterleaved<2>,
|
|
cutlass::gemm::GemmShape<8, 4, 1>
|
|
>
|
|
>;
|
|
|
|
//
|
|
// Output operator
|
|
//
|
|
|
|
using OutputOp = cutlass::epilogue::thread::LinearCombination<
|
|
ElementOutput,
|
|
kElementsPerAccess,
|
|
ElementAccumulator,
|
|
ElementCompute
|
|
>;
|
|
|
|
//
|
|
// Define the epilogue
|
|
//
|
|
|
|
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueSimt<
|
|
Shape,
|
|
WarpMmaSimt,
|
|
OutputOp,
|
|
kElementsPerAccess
|
|
>::Epilogue;
|
|
|
|
//
|
|
// Instantiate epilogue
|
|
//
|
|
|
|
EpilogueTestbed<Epilogue> testbed;
|
|
|
|
bool passed = testbed.run_all();
|
|
|
|
EXPECT_TRUE(passed);
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////////////////////////////////////
|