877bdcace6
CUTLASS 1.3 Release - Efficient GEMM kernel targeting Volta Tensor Cores via mma.sync instruction added in CUDA 10.1.
343 lines
12 KiB
Plaintext
343 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 Defines unit tests for HostTensor and HostMatrix.
|
|
|
|
HostTensor is a utility class for allocating memory on the host and on the selected CUDA device
|
|
and presenting a TensorView of this memory.
|
|
|
|
HostMatrix is new in CUTLASS 1.1 that offers a matrix-like interface to a HostTensor with rank 2.
|
|
Several examples are shown in this source file.
|
|
*/
|
|
|
|
#include "cutlass_unit_test.h"
|
|
|
|
#include "cutlass/matrix_traits.h"
|
|
|
|
#include "tools/util/tensor_view_io.h"
|
|
#include "tools/util/host_tensor.h"
|
|
#include "tools/util/host_matrix.h"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace test {
|
|
|
|
/// Kernel to compute a thread's unique coordinate within a CUDA kernel grid and write a value
|
|
/// using a CUTLASS TensorView.
|
|
template <typename TensorView>
|
|
__global__ void fill_sequential(TensorView view) {
|
|
|
|
// Compute the thread's coordinate in the 2D CUDA kernel grid
|
|
cutlass::Coord<2> coord = cutlass::make_Coord(
|
|
blockIdx.x * blockDim.x + threadIdx.x,
|
|
blockIdx.y * blockDim.y + threadIdx.y
|
|
);
|
|
|
|
// Write a value into the view
|
|
if (view.contains(coord)) {
|
|
view.at(coord) = coord[0] + view.size(0) * coord[1];
|
|
}
|
|
}
|
|
|
|
} // namespace test
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
// This test constructs a CUTLASS HostTensor with column-major layout.
|
|
TEST(HostTensor, fill_sequential_column_major) {
|
|
|
|
int const M = 16;
|
|
int const N = 32;
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M, N);
|
|
|
|
// Construct a rank=2 host tensor of size M-by-N with leading dimension M
|
|
cutlass::HostTensor<
|
|
int,
|
|
2,
|
|
cutlass::MatrixLayout::ColumnMajor> host_tensor(cutlass::make_Coord(M, 1), bounds);
|
|
|
|
// Fill it with zeros and synchronize device
|
|
host_tensor.fill(0);
|
|
host_tensor.sync_device();
|
|
|
|
// Launch a CUDA kernel by obtaining a TensorView of the device memory
|
|
dim3 block(16, 16);
|
|
dim3 grid((M + block.x - 1) / block.x, (N + block.y - 1) / block.y);
|
|
|
|
test::fill_sequential<<< grid, block >>>(host_tensor.device_view());
|
|
|
|
ASSERT_EQ(cudaDeviceSynchronize(), cudaSuccess);
|
|
|
|
// Synchronize the host data
|
|
host_tensor.sync_host();
|
|
|
|
// Verify host_tensor contains sequential elements
|
|
int errors = 0;
|
|
for (int n = 0; n < N; ++n) {
|
|
for (int m = 0; m < M; ++m) {
|
|
int expected = m + n * M;
|
|
int got = host_tensor.at(cutlass::make_Coord(m, n));
|
|
if (expected != got) {
|
|
++errors;
|
|
}
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(errors, 0) << std::setw(4) << host_tensor << std::endl;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
// This test constructs a CUTLASS HostTensor with column-major interleaved layout
|
|
TEST(HostTensor, fill_sequential_column_major_interleaved) {
|
|
|
|
int const M = 16;
|
|
int const N = 16;
|
|
int const kInterleave = 4;
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M, N);
|
|
|
|
// Define a mapping function for column-major interleaved layout
|
|
typedef cutlass::MatrixLayout::ColumnMajorInterleaved<kInterleave> TensorRefMapFunc;
|
|
|
|
// Construct a rank=2 host tensor of size M-by-N
|
|
cutlass::HostTensor<
|
|
int,
|
|
2,
|
|
TensorRefMapFunc > host_tensor(TensorRefMapFunc::stride(M), bounds);
|
|
|
|
// Fill it with zeros and synchronize device
|
|
host_tensor.fill(0);
|
|
host_tensor.sync_device();
|
|
|
|
// Launch a CUDA kernel by obtaining a TensorView of the device memory
|
|
dim3 block(16, 16);
|
|
dim3 grid((M + block.x - 1) / block.x, (N + block.y - 1) / block.y);
|
|
|
|
test::fill_sequential<<< grid, block >>>(host_tensor.device_view());
|
|
|
|
ASSERT_EQ(cudaDeviceSynchronize(), cudaSuccess);
|
|
|
|
// Synchronize the host data
|
|
host_tensor.sync_host();
|
|
|
|
// Verify host_tensor contains sequential elements
|
|
int errors = 0;
|
|
for (int n = 0; n < N; ++n) {
|
|
for (int m = 0; m < M; ++m) {
|
|
int expected = m + n * M;
|
|
int got = host_tensor.at(cutlass::make_Coord(m, n));
|
|
if (got != expected) {
|
|
++errors;
|
|
}
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(errors, 0) << std::setw(4) << host_tensor << std::endl;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
//
|
|
// cutlass::HostMatrix extends cutlass::HostTensor of rank=2 to facilitate allocate and operating
|
|
// on matrices in device memory.
|
|
//
|
|
// cutlass::HostMatrix<T> accommodates both row-major and column-major matrices with a single
|
|
// leading dimension.
|
|
//
|
|
// The first test demonstrates use of HostMatrix<> in the same circumstances as HostTensor but with
|
|
// simplifcations to the calling interface.
|
|
//
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
// This test constructs a CUTLASS cutlass::HostMatrix with column-major layout.
|
|
TEST(HostMatrix, fill_sequential_column_major) {
|
|
|
|
int const M = 16;
|
|
int const N = 32;
|
|
int const ldm = M + 2; // define leading dimension with padding
|
|
|
|
cutlass::Coord<2> bounds = cutlass::make_Coord(M, N);
|
|
|
|
// Construct a HostMatrix of size M-by-N with leading dimension ldm
|
|
cutlass::HostMatrix<int> host_matrix(bounds, cutlass::MatrixLayout::kColumnMajor, ldm);
|
|
|
|
// Fill it with zeros and synchronize device
|
|
host_matrix.fill(0);
|
|
host_matrix.sync_device();
|
|
|
|
// Launch a CUDA kernel by obtaining a TensorView of the device memory
|
|
dim3 block(16, 16);
|
|
dim3 grid((M + block.x - 1) / block.x, (N + block.y - 1) / block.y);
|
|
|
|
test::fill_sequential<<< grid, block >>>(host_matrix.device_view());
|
|
|
|
ASSERT_EQ(cudaDeviceSynchronize(), cudaSuccess);
|
|
|
|
// Synchronize the host data
|
|
host_matrix.sync_host();
|
|
|
|
// Verify host_matrix contains sequential elements
|
|
int errors = 0;
|
|
for (int n = 0; n < N; ++n) {
|
|
for (int m = 0; m < M; ++m) {
|
|
int expected = m + n * M;
|
|
int got = host_matrix.at(cutlass::make_Coord(m, n));
|
|
if (expected != got) {
|
|
++errors;
|
|
}
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(errors, 0) << std::setw(4) << host_matrix << std::endl;
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
//
|
|
// Previously, cutlass::HostTensorView<> offered a gemm() method defined for the H and W dimensions.
|
|
// The other dimensions were ignored.
|
|
//
|
|
// To improve the interface, we We have moved this into the HostMatrixView<> and HostMatrix<>
|
|
// classes which require rank=2. To accommodate matrix operands of differing layout, we have extracted
|
|
// the host-side GEMM implementation into cutlass::reference::host::Gemm() which can compute the
|
|
// general matrix product of matrices with arbitrary layout.
|
|
//
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
// This test constructs a CUTLASS cutlass::HostMatrix with column-major layout.
|
|
TEST(HostMatrix, gemm) {
|
|
|
|
// Problem size intentionally small, as reference check has complexity O(MNK).
|
|
int const M = 32;
|
|
int const N = 16;
|
|
int const K = 4;
|
|
|
|
int const lda = M;
|
|
int const ldb = N;
|
|
int const ldc = M;
|
|
|
|
// Construct matrix operands
|
|
cutlass::HostMatrix<int> A(cutlass::make_Coord(M, K), cutlass::MatrixLayout::kColumnMajor, lda);
|
|
cutlass::HostMatrix<int> B(cutlass::make_Coord(K, N), cutlass::MatrixLayout::kRowMajor, ldb);
|
|
cutlass::HostMatrix<int> C(cutlass::make_Coord(M, N), cutlass::MatrixLayout::kColumnMajor, ldc);
|
|
|
|
A.fill_sequential();
|
|
B.fill_sequential();
|
|
C.fill(0);
|
|
|
|
int alpha = 1;
|
|
|
|
// Compute host-side GEMM reference
|
|
cutlass::reference::host::Gemm(
|
|
cutlass::gemm::GemmCoord(K, N, M),
|
|
alpha,
|
|
A.host_ref(),
|
|
B.host_ref(),
|
|
int(0), // beta
|
|
C.host_ref());
|
|
|
|
// Verify result
|
|
int errors = 0;
|
|
|
|
// Primitive reference implementation for matrix product
|
|
for (int i = 0; i < M; ++i) {
|
|
for (int j = 0; j < N; ++j) {
|
|
int result = 0;
|
|
for (int k = 0; k < K; ++k) {
|
|
result += A.at(cutlass::make_Coord(i, k)) * B.at(cutlass::make_Coord(k, j));
|
|
}
|
|
if (C.at(cutlass::make_Coord(i, j)) != alpha * result) {
|
|
++errors;
|
|
}
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(errors, 0) << "GEMM error\n"
|
|
<< "A =\n" << A << "\nB = \n" << B << "\nC =\n" << C << "\n";
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
// When layout is known at compile time, we may be use the corresponding helper classes to smplify
|
|
// matrix instantiation. The matrix layout becomes part of the type which reduces the StorageRank
|
|
// of the internal stride vector.
|
|
//
|
|
// Apart from specifying the matrix layout at compile time, this test is functionally identical to
|
|
// HostMatrix.gemm.
|
|
//
|
|
TEST(HostMatrix, gemm_compile_time_layout) {
|
|
|
|
// Problem size intentionally small, as reference check has complexity O(MNK).
|
|
int const M = 32;
|
|
int const N = 16;
|
|
int const K = 4;
|
|
|
|
int const lda = M;
|
|
int const ldb = N;
|
|
int const ldc = M;
|
|
|
|
// Construct matrix operands
|
|
cutlass::HostMatrixColumnMajor<int> A(cutlass::make_Coord(M, K), lda);
|
|
cutlass::HostMatrixRowMajor<int> B(cutlass::make_Coord(K, N), ldb);
|
|
cutlass::HostMatrixColumnMajor<int> C(cutlass::make_Coord(M, N), ldc);
|
|
|
|
A.fill_sequential();
|
|
B.fill_sequential();
|
|
C.fill(0);
|
|
|
|
int alpha = 1;
|
|
|
|
// Compute host-side GEMM reference
|
|
cutlass::reference::host::Gemm(
|
|
cutlass::gemm::GemmCoord(K, N, M),
|
|
alpha,
|
|
A.host_ref(),
|
|
B.host_ref(),
|
|
int(0), // beta
|
|
C.host_ref());
|
|
|
|
// Verify result
|
|
int errors = 0;
|
|
|
|
// Primitive reference implementation for matrix product
|
|
for (int i = 0; i < M; ++i) {
|
|
for (int j = 0; j < N; ++j) {
|
|
int result = 0;
|
|
for (int k = 0; k < K; ++k) {
|
|
result += A.at(cutlass::make_Coord(i, k)) * B.at(cutlass::make_Coord(k, j));
|
|
}
|
|
if (C.at(cutlass::make_Coord(i, j)) != alpha * result) {
|
|
++errors;
|
|
}
|
|
}
|
|
}
|
|
|
|
EXPECT_EQ(errors, 0) << "GEMM error\n"
|
|
<< "A =\n" << A << "\nB = \n" << B << "\nC =\n" << C << "\n";
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|