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.
284 lines
8.6 KiB
Plaintext
284 lines
8.6 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.
|
|
*
|
|
**************************************************************************************************/
|
|
#include "../common/cutlass_unit_test.h"
|
|
|
|
#include "cutlass/tensor_view.h"
|
|
#include "cutlass/layout/matrix.h"
|
|
|
|
#include "cutlass/util/tensor_view_io.h"
|
|
#include "cutlass/util/host_tensor.h"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(TensorView, rank2_contiguous_dynamic) {
|
|
int const M = 8;
|
|
int const N = 16;
|
|
|
|
typedef cutlass::TensorView<int, cutlass::layout::ContiguousMatrix> ContiguousTensorView;
|
|
|
|
cutlass::layout::Matrix layouts[] = {
|
|
cutlass::layout::Matrix::kColumnMajor,
|
|
cutlass::layout::Matrix::kRowMajor
|
|
};
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M - 2, N - 2);
|
|
|
|
for (int i = 0; i < 2; ++i) {
|
|
|
|
int matrix_data[M * N] = { 0 };
|
|
|
|
int row_stride;
|
|
int col_stride;
|
|
|
|
if (layouts[i] == cutlass::layout::Matrix::kColumnMajor) {
|
|
row_stride = 1;
|
|
col_stride = M;
|
|
}
|
|
else {
|
|
row_stride = N;
|
|
col_stride = 1;
|
|
}
|
|
|
|
// Use helper to determine stride vector from leading dimension
|
|
ContiguousTensorView view(
|
|
matrix_data,
|
|
cutlass::layout::ContiguousMatrix::packed(cutlass::make_Coord(M, N), layouts[i]),
|
|
bounds);
|
|
|
|
ASSERT_TRUE(view.good());
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
cutlass::Coord<2> coord = cutlass::make_Coord(m, n);
|
|
if (view.contains(coord)) {
|
|
view.at(coord) = m * N + n;
|
|
}
|
|
}
|
|
}
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
int expected = 0;
|
|
if (m < bounds[0] && n < bounds[1]) {
|
|
expected = int(m * N + n);
|
|
}
|
|
EXPECT_EQ(matrix_data[m * row_stride + n * col_stride], expected);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
//
|
|
// Uncomment the following line to observe output from printing TensorView objects
|
|
//
|
|
|
|
// #define OBSERVE_TENSORVIEW_IO // uncomment to enable printing
|
|
|
|
#ifdef OBSERVE_TENSORVIEW_IO
|
|
|
|
// This test construct a TensorView of rank=2 with matrix layouts known at runtime. This
|
|
// uses TensorRefMapFunc classes defined in cutlass/matrix_traits.h to define the mapping
|
|
// from logical tensor indices to storage in memory.
|
|
//
|
|
// Helpers in tools/util/tensor_view_io.h print both the logical TensorView and the
|
|
// linear memory of the tensor.
|
|
TEST(TensorView, contiguous) {
|
|
|
|
int const M = 8;
|
|
int const N = 16;
|
|
|
|
typedef cutlass::TensorView<
|
|
int32_t,
|
|
cutlass::layout::ContiguousLayout> ContiguousTensorView;
|
|
|
|
cutlass::MatrixLayout layouts[] = {
|
|
cutlass::MatrixLayout::kColumnMajor,
|
|
cutlass::MatrixLayout::kRowMajor
|
|
};
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M, N);
|
|
|
|
for (int i = 0; i < 2; ++i) {
|
|
|
|
int matrix_data[M * N] = { 0 };
|
|
|
|
int ldm;
|
|
int row_stride;
|
|
int col_stride;
|
|
|
|
if (layouts[i] == cutlass::MatrixLayout::kColumnMajor) {
|
|
row_stride = 1;
|
|
col_stride = M;
|
|
ldm = col_stride;
|
|
}
|
|
else {
|
|
row_stride = N;
|
|
col_stride = 1;
|
|
ldm = row_stride;
|
|
}
|
|
|
|
// Use helper to determine stride vector from leading dimension
|
|
ContiguousTensorView view(
|
|
matrix_data,
|
|
cutlass::layout::ContiguousLayout::stride(layouts[i], ldm),
|
|
bounds);
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
cutlass::Coord<2> coord = cutlass::make_Coord(m, n);
|
|
if (view.contains(coord)) {
|
|
view.at(coord) = m * N + n;
|
|
}
|
|
}
|
|
}
|
|
|
|
std::cout << "---------\n";
|
|
std::cout << (layouts[i] == cutlass::MatrixLayout::kColumnMajor ?
|
|
"Column-major:" : "Row-major:") << "\n\n";
|
|
|
|
std::cout << "Logical view:\n";
|
|
std::cout.width(4);
|
|
std::cout << view << "\n" << std::endl; // Print TensorView object.
|
|
|
|
std::cout << "Linear memory:";
|
|
for (int idx = 0; idx < view.capacity(); ++idx) {
|
|
if (!(idx % (layouts[i] == cutlass::MatrixLayout::kColumnMajor ? M : N))) {
|
|
std::cout << std::endl;
|
|
}
|
|
std::cout << std::setw(4) << view.at(idx) << " ";
|
|
}
|
|
|
|
std::cout << "\n" << std::endl;
|
|
}
|
|
}
|
|
|
|
// This test is similar to the previous except it uses a column-major, interleaved data
|
|
// layout. The test prints both the logical representation (a typical column-major matrix)
|
|
// and a representation of linear memory.
|
|
//
|
|
// Note, the interleave=4 structure implies that every four consecutive elements in the
|
|
// same row shall be adjacent in memory followed by the next row.
|
|
TEST(TensorView, rank2_column_major_interleaved) {
|
|
int const M = 16;
|
|
int const N = 16;
|
|
int const kInterleave = 4;
|
|
|
|
int matrix_data[M * N] = {0};
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M, N);
|
|
|
|
// Define the TensorRefMapFunc for a column-major interleaved matrix format
|
|
typedef cutlass::layout::ColumnMajorInterleaved<kInterleave> TensorRefMapFunc;
|
|
|
|
// Define a TensorView of rank=2 using the column-major interleaved mapping function
|
|
typedef cutlass::TensorView<
|
|
int,
|
|
TensorRefMapFunc> InterleavedTensorView;
|
|
|
|
InterleavedTensorView view(
|
|
matrix_data,
|
|
TensorRefMapFunc::stride(M),
|
|
bounds);
|
|
|
|
// Initialize
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
view.at(cutlass::make_Coord(m, n)) = m + n * M;
|
|
}
|
|
}
|
|
|
|
// Print logical view
|
|
std::cout << "Column-major, interleave=" << kInterleave << " (logical view):\n";
|
|
|
|
std::cout << std::setw(4) << view << "\n" << std::endl;
|
|
|
|
// Now define a linear view of the same data in memory
|
|
typedef cutlass::TensorView<int, 2, cutlass::layout::RowMajor> LinearTensorView;
|
|
|
|
LinearTensorView linear_view(matrix_data, cutlass::make_Coord(N), bounds);
|
|
|
|
std::cout << "Linear view in memory:\n";
|
|
std::cout << std::setw(4) << linear_view << std::endl;
|
|
}
|
|
|
|
#endif
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(TensorView, int4) {
|
|
|
|
int const M = 4;
|
|
int const N = 8;
|
|
|
|
using T = cutlass::int4b_t;
|
|
|
|
cutlass::HostTensor<T, cutlass::layout::RowMajor> tensor({M, N});
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
T x = T(n ^ m); // some simple hash
|
|
tensor.host_view().at({m, n}) = x;
|
|
}
|
|
}
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
int x = (n ^ m); // some simple hash
|
|
EXPECT_TRUE(int(tensor.host_view().at({m, n})) == x);
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(tensor.size(), M * N);
|
|
}
|
|
|
|
TEST(TensorView, uint4) {
|
|
|
|
int const M = 4;
|
|
int const N = 8;
|
|
|
|
using T = cutlass::uint4b_t;
|
|
|
|
cutlass::HostTensor<T, cutlass::layout::RowMajor> tensor({M, N});
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
T x = T(n ^ m); // some simple hash
|
|
tensor.host_view().at({m, n}) = x;
|
|
}
|
|
}
|
|
|
|
for (int m = 0; m < M; ++m) {
|
|
for (int n = 0; n < N; ++n) {
|
|
int x = (n ^ m); // some simple hash
|
|
EXPECT_TRUE(int(tensor.host_view().at({m, n})) == x);
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(tensor.size(), M * N);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|