66d9cddc83
* New updates. * Minor profiler updates Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
373 lines
14 KiB
C++
373 lines
14 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
* SPDX-License-Identifier: BSD-3-Clause
|
|
*
|
|
* Redistribution and use in source and binary forms, with or without
|
|
* modification, are permitted provided that the following conditions are met:
|
|
*
|
|
* 1. Redistributions of source code must retain the above copyright notice, this
|
|
* list of conditions and the following disclaimer.
|
|
*
|
|
* 2. 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.
|
|
*
|
|
* 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (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 testbed for kernel-level GEMM
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include <fstream>
|
|
|
|
#include "../../common/cutlass_unit_test.h"
|
|
|
|
#include "cutlass/aligned_buffer.h"
|
|
#include "cutlass/gemm/gemm.h"
|
|
#include "cutlass/layout/matrix.h"
|
|
#include "cutlass/layout/vector.h"
|
|
#include "cutlass/numeric_types.h"
|
|
|
|
#include "cutlass/core_io.h"
|
|
#include "cutlass/util/host_tensor.h"
|
|
#include "cutlass/util/tensor_view_io.h"
|
|
|
|
#include "cutlass/util/distribution.h"
|
|
#include "cutlass/util/reference/host/gemm.h"
|
|
#include "cutlass/util/reference/host/tensor_compare.h"
|
|
#include "cutlass/util/reference/host/tensor_fill.h"
|
|
|
|
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
|
|
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
|
|
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"
|
|
#include "cutlass/transform/threadblock/predicated_tile_iterator_2dthreadtile.h"
|
|
#include "cutlass/cutlass.h"
|
|
#include "cutlass/platform/platform.h"
|
|
|
|
namespace test {
|
|
namespace gemm {
|
|
namespace threadblock {
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <typename Mma>
|
|
__global__ void kernel_mma(cutlass::gemm::GemmCoord problem_size,
|
|
typename Mma::IteratorA::Params params_A,
|
|
typename Mma::IteratorA::TensorRef ref_A,
|
|
typename Mma::IteratorB::Params params_B,
|
|
typename Mma::IteratorB::TensorRef ref_B,
|
|
typename Mma::ElementC **ptr_C,
|
|
typename Mma::LayoutC::Stride::Index ldc) {
|
|
// Shared storage needed by threadblock-scoped matrix multiply-accumulate
|
|
__shared__ typename Mma::SharedStorage shared_storage;
|
|
|
|
// Compute threadblock location
|
|
cutlass::gemm::GemmCoord tb_tile_offset = {int(blockIdx.x), int(blockIdx.y),
|
|
0};
|
|
|
|
cutlass::MatrixCoord tb_offset_A{tb_tile_offset.m() * Mma::Shape::kM,
|
|
tb_tile_offset.k()};
|
|
|
|
cutlass::MatrixCoord tb_offset_B{tb_tile_offset.k(),
|
|
tb_tile_offset.n() * Mma::Shape::kN};
|
|
|
|
// Compute position within threadblock
|
|
int tb_thread_id = threadIdx.y * blockDim.x + threadIdx.x;
|
|
|
|
// Construct iterators to A and B operands
|
|
typename Mma::IteratorA iterator_A(params_A, ref_A.data(),
|
|
{problem_size.m(), problem_size.k()},
|
|
tb_thread_id, tb_offset_A);
|
|
|
|
typename Mma::IteratorB iterator_B(params_B, ref_B.data(),
|
|
{problem_size.k(), problem_size.n()},
|
|
tb_thread_id, tb_offset_B);
|
|
|
|
int warp_id = threadIdx.y;
|
|
int lane_id = threadIdx.x;
|
|
|
|
int partitionsK_idx = warp_id / (Mma::WarpCount::kM * Mma::WarpCount::kN);
|
|
|
|
// Construct thread-scoped matrix multiply
|
|
Mma mma(shared_storage, tb_thread_id, warp_id, threadIdx.x);
|
|
|
|
typename Mma::FragmentC accum;
|
|
|
|
accum.clear();
|
|
|
|
int gemm_k_iterations = (problem_size.k() + Mma::Shape::kK - 1) / Mma::Shape::kK;
|
|
|
|
// Compute threadblock-scoped matrix multiply-add
|
|
mma(gemm_k_iterations, accum, iterator_A, iterator_B, accum);
|
|
|
|
// Output results
|
|
typename Mma::Operator::IteratorC iterator_C({ptr_C[partitionsK_idx], ldc}, lane_id);
|
|
|
|
|
|
int warp_idx_mn = warp_id % (Mma::WarpCount::kM * Mma::WarpCount::kN);
|
|
iterator_C.add_tile_offset(
|
|
{(tb_tile_offset.m() * Mma::WarpCount::kM) +
|
|
(warp_idx_mn % Mma::WarpCount::kM),
|
|
(tb_tile_offset.n() * Mma::WarpCount::kN) +
|
|
(warp_idx_mn / Mma::WarpCount::kM)});
|
|
|
|
iterator_C.store(accum);
|
|
}
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Structure to compute the matrix product
|
|
template <
|
|
/// Threadblock-level matrix multiply-accumulate
|
|
typename MmaCore_>
|
|
struct Testbed {
|
|
/// Threadblock-level GEMM implementation
|
|
using MmaCore = MmaCore_;
|
|
using ThreadblockShape = typename MmaCore::Shape;
|
|
using WarpShape = typename MmaCore::WarpShape;
|
|
using InstructionShape = typename MmaCore::InstructionShape;
|
|
using ElementA = typename MmaCore::ElementA;
|
|
using LayoutA = typename MmaCore::LayoutA;
|
|
using ElementB = typename MmaCore::ElementB;
|
|
using LayoutB = typename MmaCore::LayoutB;
|
|
using ElementC = typename MmaCore::ElementC;
|
|
using LayoutC = typename MmaCore::LayoutC;
|
|
|
|
// Define iterators over tiles from the A operand
|
|
static const bool use_idp4a = cutlass::platform::is_same<ElementA, int8_t>::value &&
|
|
cutlass::platform::is_same<ElementB, int8_t>::value &&
|
|
cutlass::platform::is_same<typename MmaCore::OperatorClass, cutlass::arch::OpClassSimt>::value;
|
|
|
|
static const bool transposeA = cutlass::platform::is_same< LayoutA, cutlass::layout::ColumnMajor >::value;
|
|
static const bool transposeB = cutlass::platform::is_same< LayoutB, cutlass::layout::RowMajor >::value;
|
|
|
|
using IteratorA = typename cutlass::platform::conditional< use_idp4a,
|
|
cutlass::transform::threadblock::PredicatedTileIterator2dThreadTile<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA, LayoutA, 1, typename MmaCore::IteratorThreadMapA, transposeA> ,
|
|
|
|
cutlass::transform::threadblock::PredicatedTileIterator<
|
|
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
|
|
ElementA, LayoutA, 1, typename MmaCore::IteratorThreadMapA>
|
|
>::type;
|
|
|
|
// Define iterators over tiles from the B operand
|
|
using IteratorB = typename cutlass::platform::conditional< use_idp4a,
|
|
cutlass::transform::threadblock::PredicatedTileIterator2dThreadTile<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB, LayoutB, 0, typename MmaCore::IteratorThreadMapB, transposeB> ,
|
|
|
|
cutlass::transform::threadblock::PredicatedTileIterator<
|
|
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
|
|
ElementB, LayoutB, 0, typename MmaCore::IteratorThreadMapB>
|
|
>::type;
|
|
|
|
// Define the threadblock-scoped pipelined matrix multiply
|
|
using Mma = cutlass::gemm::threadblock::MmaPipelined<
|
|
typename MmaCore::Shape, IteratorA, typename MmaCore::SmemIteratorA,
|
|
IteratorB, typename MmaCore::SmemIteratorB, ElementC, LayoutC,
|
|
typename MmaCore::MmaPolicy>;
|
|
|
|
static int const kPartitionsK = MmaCore::MmaPolicy::kPartitionsK;
|
|
|
|
//
|
|
// Data members
|
|
//
|
|
|
|
cutlass::HostTensor<ElementA, LayoutA> matrix_A;
|
|
cutlass::HostTensor<ElementB, LayoutB> matrix_B;
|
|
cutlass::HostTensor<ElementC, LayoutC> matrix_C_computed[kPartitionsK];
|
|
cutlass::HostTensor<ElementC, LayoutC> matrix_C_reference;
|
|
cutlass::HostTensor<ElementC*, cutlass::layout::PackedVectorLayout> matrix_C_pointers;
|
|
|
|
cutlass::gemm::GemmCoord problem_size;
|
|
float alpha, beta;
|
|
|
|
//
|
|
// Methods
|
|
//
|
|
|
|
/// Allocates workspace in device memory
|
|
Testbed(int m, int n, int k, float alpha_, float beta_)
|
|
: problem_size(m, n, k), alpha(alpha_), beta(beta_) {
|
|
matrix_A.reset(cutlass::make_Coord(m, k));
|
|
matrix_B.reset(cutlass::make_Coord(k, n));
|
|
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 0; k < kPartitionsK; k++)
|
|
matrix_C_computed[k].reset(cutlass::make_Coord(m, n));
|
|
|
|
matrix_C_reference.reset(cutlass::make_Coord(m, n), false);
|
|
matrix_C_pointers.reset(cutlass::Coord<1>(kPartitionsK));
|
|
}
|
|
|
|
/// Runs the test
|
|
bool run(
|
|
dim3 grid, dim3 block,
|
|
cutlass::Distribution::Kind init_A = cutlass::Distribution::Uniform,
|
|
cutlass::Distribution::Kind init_B = cutlass::Distribution::Uniform) {
|
|
//
|
|
// initialize device memory
|
|
//
|
|
|
|
if (init_A == cutlass::Distribution::Uniform) {
|
|
|
|
int scope_max = 8;
|
|
int scope_min = -8;
|
|
|
|
if (cutlass::sizeof_bits<ElementA>::value == 4) {
|
|
scope_max = 2;
|
|
scope_min = -2;
|
|
} else if (cutlass::sizeof_bits<ElementA>::value == 1) {
|
|
scope_max = 2;
|
|
scope_min = 0;
|
|
}
|
|
|
|
uint64_t seed = 7;
|
|
cutlass::reference::host::TensorFillRandomUniform(
|
|
matrix_A.host_view(), seed, scope_max, scope_min, 0);
|
|
} else if (init_A == cutlass::Distribution::Sequential) {
|
|
cutlass::reference::host::BlockFillSequential(matrix_A.host_data(),
|
|
matrix_A.capacity());
|
|
} else if (init_A == cutlass::Distribution::Identity) {
|
|
cutlass::reference::host::TensorFillIdentity(matrix_A.host_view());
|
|
} else {
|
|
// TODO: Implement the rest
|
|
return false;
|
|
}
|
|
|
|
if (init_B == cutlass::Distribution::Uniform) {
|
|
|
|
int scope_max = 8;
|
|
int scope_min = -8;
|
|
|
|
if (cutlass::sizeof_bits<ElementB>::value == 4) {
|
|
scope_max = 2;
|
|
scope_min = -2;
|
|
} else if (cutlass::sizeof_bits<ElementB>::value == 1) {
|
|
scope_max = 2;
|
|
scope_min = 0;
|
|
}
|
|
|
|
uint64_t seed = 7;
|
|
cutlass::reference::host::TensorFillRandomUniform(
|
|
matrix_B.host_view(), seed + 16, scope_max, scope_min, 0);
|
|
} else if (init_B == cutlass::Distribution::Sequential) {
|
|
cutlass::reference::host::BlockFillSequential(matrix_B.host_data(),
|
|
matrix_B.capacity());
|
|
} else if (init_B == cutlass::Distribution::Identity) {
|
|
cutlass::reference::host::TensorFillIdentity(matrix_B.host_view());
|
|
} else {
|
|
// TODO: Implement the rest
|
|
return false;
|
|
}
|
|
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 0; k < kPartitionsK; k++)
|
|
cutlass::reference::host::TensorFill(matrix_C_computed[k].host_view());
|
|
|
|
cutlass::reference::host::TensorFill(matrix_C_reference.host_view());
|
|
|
|
matrix_A.sync_device();
|
|
matrix_B.sync_device();
|
|
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 0; k < kPartitionsK; k++)
|
|
matrix_C_computed[k].sync_device();
|
|
|
|
typename IteratorA::Params params_A(matrix_A.layout());
|
|
typename IteratorB::Params params_B(matrix_B.layout());
|
|
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 0; k < kPartitionsK; k++)
|
|
matrix_C_pointers.at(cutlass::Coord<1>(k)) = matrix_C_computed[k].device_data();
|
|
|
|
matrix_C_pointers.sync_device();
|
|
|
|
test::gemm::threadblock::kernel_mma<Mma><<<grid, block>>>(
|
|
problem_size, params_A, matrix_A.device_ref(), params_B,
|
|
matrix_B.device_ref(), matrix_C_pointers.device_data(),
|
|
matrix_C_computed[0].layout().stride(0));
|
|
|
|
//
|
|
// Check error code
|
|
//
|
|
|
|
cudaError_t result = cudaDeviceSynchronize();
|
|
EXPECT_EQ(result, cudaSuccess)
|
|
<< " kernel error: " << cudaGetErrorString(result);
|
|
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 0; k < kPartitionsK; k++)
|
|
matrix_C_computed[k].sync_host();
|
|
|
|
// TODO: this is temporary. it will be removed after slicing can de
|
|
// reduction
|
|
//
|
|
// Reduce matrix_C_computed
|
|
//
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int k = 1; k < kPartitionsK; k++) {
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int m = 0; m < matrix_C_computed[0].extent().row(); m++){
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for(int n = 0; n < matrix_C_computed[0].extent().column(); n++){
|
|
matrix_C_computed[0].at({m, n}) += matrix_C_computed[k].at({m, n});
|
|
}
|
|
}
|
|
}
|
|
|
|
cutlass::reference::host::Gemm<ElementA, LayoutA, ElementB, LayoutB,
|
|
ElementC, LayoutC, ElementC, ElementC,
|
|
typename MmaCore::Operator>
|
|
reference_gemm;
|
|
|
|
reference_gemm(
|
|
problem_size, ElementC(alpha), matrix_A.host_view(),
|
|
matrix_B.host_view(), ElementC(beta), matrix_C_reference.host_view());
|
|
|
|
bool passed = cutlass::reference::host::TensorEquals(
|
|
matrix_C_computed[0].host_view(), matrix_C_reference.host_view());
|
|
|
|
EXPECT_TRUE(passed);
|
|
|
|
if (!passed) {
|
|
std::ofstream output("mma_pipelined_testbed_errors.txt");
|
|
|
|
output
|
|
<< "A:\n" << matrix_A.host_view() << "\n"
|
|
<< "B:\n" << matrix_B.host_view() << "\n"
|
|
<< "Reference:\n"
|
|
<< matrix_C_reference.host_view() << "\n"
|
|
<< "Computed:\n"
|
|
<< matrix_C_computed[0].host_view() << "\n";
|
|
}
|
|
|
|
return passed;
|
|
}
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace threadblock
|
|
} // namespace gemm
|
|
} // namespace test
|