630 lines
24 KiB
Plaintext
630 lines
24 KiB
Plaintext
/***************************************************************************************************
|
|
* Copyright (c) 2017-2018, 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 "cutlass/wmma_matrix.h"
|
|
|
|
#ifdef CUTLASS_USE_WMMA_API
|
|
|
|
#include "cutlass_unit_tests.h"
|
|
#include "tools/test/unit/gemm/gemm_testbed.h"
|
|
#include "tools/util/half.h"
|
|
|
|
#include "cutlass/gemm/gemm_global_stream.h"
|
|
#include "cutlass/gemm/gemm_shared_stream.h"
|
|
#include "cutlass/gemm/wmma_gemm_multiply_add.h"
|
|
#include "cutlass/gemm/wmma_gemm_global_tile.h"
|
|
#include "cutlass/gemm/wmma_gemm_shared_tile.h"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
struct ProblemDesc {
|
|
int m, n, k;
|
|
inline __device__ ProblemDesc(int m_, int n_, int k_) : m(m_), n(n_), k(k_) {}
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <typename StoreIterator_, typename LoadIterator_>
|
|
union SharedStorage {
|
|
// Storage to store the data.
|
|
typename StoreIterator_::SharedStorage store;
|
|
// Storage to load the data.
|
|
typename LoadIterator_::SharedStorage load;
|
|
};
|
|
|
|
template <class> struct Debug {};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <typename Threads_, int kW_, bool = (Threads_::kW > kW_)>
|
|
struct ReshapeThreadsA {
|
|
typedef cutlass::Shape<Threads_::kD, Threads_::kH, Threads_::kW> Threads;
|
|
};
|
|
|
|
template <typename Threads_, int kW_>
|
|
struct ReshapeThreadsA<Threads_, kW_, true> {
|
|
typedef cutlass::Shape<Threads_::kD, Threads_::kH * Threads_::kW / kW_, kW_> Threads;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <typename Threads_, int kH_, bool = (Threads_::kW > kH_)>
|
|
struct ReshapeThreadsB {
|
|
typedef cutlass::Shape<Threads_::kD, Threads_::kH, Threads_::kW> Threads;
|
|
};
|
|
|
|
template <typename Threads_, int kH_>
|
|
struct ReshapeThreadsB<Threads_, kH_, true> {
|
|
typedef cutlass::Shape<Threads_::kD, Threads_::kH * Threads_::kW / kH_, kH_> Threads;
|
|
};
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
#if 1
|
|
template <typename Traits_>
|
|
static __global__ void kernel_nt(half const *d_a, int lda, half const *d_b, int ldb, float *d_c,
|
|
int ldc) {
|
|
#if 0
|
|
// The default configuration of threads.
|
|
typedef cutlass::Shape<1, Warps_::kCount, 32> Threads_;
|
|
// The threads.
|
|
typedef typename ReshapeThreadsA<Threads_, OutputTile_::kW>::Threads ThreadsA;
|
|
// The threads.
|
|
typedef typename ReshapeThreadsB<Threads_, OutputTile_::kH>::Threads ThreadsB;
|
|
// The number of elements loaded per LDG.
|
|
int const kScalarsPerLdg = 1;
|
|
// The tile for A.
|
|
typedef cutlass::Shape<1, OutputTile_::kD, OutputTile_::kW> TileA;
|
|
// The tile for B.
|
|
typedef cutlass::Shape<1, OutputTile_::kD, OutputTile_::kH> TileB;
|
|
// The tile for C.
|
|
typedef cutlass::Shape<1, Warps_::kH*WmmaShape_::kH, OutputTile_::kW> TileC;
|
|
#endif
|
|
|
|
// The problem descriptor.
|
|
ProblemDesc desc(Traits_::OutputTile::kW, Traits_::OutputTile::kH, Traits::OutputTile::kD);
|
|
|
|
// The elements computed by a single warp.
|
|
typedef typename cutlass::ShapeDiv<OutputTile_, Warps_>::Shape AccumulatorsPerWarp;
|
|
|
|
// Global memory load for A.
|
|
typedef cutlass::gemm::GemmGlobalIteratorAb<
|
|
cutlass::gemm::GemmGlobalIteratorTraits<
|
|
cutlass::GemmOperand::kA, cutlass::MatrixLayout::kColumnMajor, half const, TileA, ThreadsA, kScalarsPerLdg>
|
|
>
|
|
GlobalLoadIteratorA;
|
|
|
|
// Shared store iterator for A.
|
|
typedef cutlass::gemm::GemmSharedStoreIteratorAb<
|
|
cutlass::gemm::GemmSharedStoreIteratorAbTraits<
|
|
half, TileA, ThreadsA, kScalarsPerLdg>
|
|
>
|
|
SharedStoreIteratorA;
|
|
|
|
// The global stream for A.
|
|
typedef cutlass::gemm::GlobalLoadStream<
|
|
GlobalLoadIteratorA,
|
|
cutlass::Copy<typename GlobalLoadIteratorA::Fragment>,
|
|
SharedStoreIteratorA>
|
|
GlobalLoadStreamA;
|
|
|
|
// Shared load iterator for A.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorA<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorAbTraits<
|
|
cutlass::GemmOperand::kA,
|
|
cutlass::MatrixLayout::kColumnMajor, half, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedLoadIteratorA;
|
|
|
|
// Global memory load for B.
|
|
typedef cutlass::gemm::GemmGlobalIteratorAb<
|
|
cutlass::gemm::GemmGlobalIteratorTraits<
|
|
cutlass::GemmOperand::kB, cutlass::MatrixLayout::kRowMajor, half const, TileB, ThreadsB, kScalarsPerLdg> >
|
|
GlobalLoadIteratorB;
|
|
|
|
// Shared store iterator for B.
|
|
typedef cutlass::gemm::GemmSharedStoreIteratorAb<
|
|
cutlass::gemm::GemmSharedStoreIteratorAbTraits<
|
|
half, TileB, ThreadsB, kScalarsPerLdg> >
|
|
SharedStoreIteratorB;
|
|
|
|
// The global stream for B.
|
|
typedef cutlass::gemm::GlobalLoadStream<GlobalLoadIteratorB,
|
|
cutlass::Copy<typename GlobalLoadIteratorB::Fragment>,
|
|
SharedStoreIteratorB>
|
|
GlobalLoadStreamB;
|
|
|
|
// Shared load iterator for B.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorB<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorAbTraits<
|
|
cutlass::GemmOperand::kB,
|
|
cutlass::MatrixLayout::kRowMajor, half, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedLoadIteratorB;
|
|
|
|
// Share memory to exchange data for A.
|
|
__shared__ SharedStorage<GlobalLoadStreamA, SharedLoadIteratorA> shared_storage_a;
|
|
|
|
// Share memory to exchange data for B.
|
|
__shared__ SharedStorage<GlobalLoadStreamB, SharedLoadIteratorB> shared_storage_b;
|
|
|
|
// Iterator to load A.
|
|
typename GlobalLoadStreamA::Params global_params_a;
|
|
global_params_a.initialize(desc, d_a, lda);
|
|
GlobalLoadStreamA global_load_a(global_params_a, shared_storage_a.store, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Iterator to load B.
|
|
typename GlobalLoadStreamB::Params global_params_b;
|
|
global_params_b.initialize(desc, d_b, ldb);
|
|
GlobalLoadStreamB global_load_b(global_params_b, shared_storage_b.store, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Load A/B.
|
|
global_load_a.copy();
|
|
global_load_b.copy();
|
|
|
|
// Copy to shared memory.
|
|
global_load_a.commit();
|
|
global_load_b.commit();
|
|
|
|
// Make sure the data is in shared memory.
|
|
__syncthreads();
|
|
|
|
// Load iterator A.
|
|
typename SharedLoadIteratorA::Params shared_params_a;
|
|
shared_params_a.initialize(desc);
|
|
SharedLoadIteratorA shared_load_a(shared_params_a, shared_storage_a.load);
|
|
|
|
// Load iterator B.
|
|
typename SharedLoadIteratorB::Params shared_params_b;
|
|
shared_params_b.initialize(desc);
|
|
SharedLoadIteratorB shared_load_b(shared_params_b, shared_storage_b.load);
|
|
|
|
// Copy A from shared memory.
|
|
typename SharedLoadIteratorA::Fragment fragment_a;
|
|
cutlass::gemm::load_shared(shared_load_a, fragment_a);
|
|
|
|
// Copy B from shared memory.
|
|
typename SharedLoadIteratorB::Fragment fragment_b;
|
|
cutlass::gemm::load_shared(shared_load_b, fragment_b);
|
|
|
|
// The functor to do WMMA.
|
|
typedef cutlass::gemm::WmmaGemmMultiplyAdd<
|
|
cutlass::MatrixLayout::kColumnMajor,
|
|
cutlass::MatrixLayout::kRowMajor,
|
|
cutlass::MatrixLayout::kColumnMajor,
|
|
float,
|
|
AccumulatorsPerWarp,
|
|
WmmaShape_> WmmaGemmMultiplyAdd;
|
|
|
|
// The output fragment.
|
|
typename WmmaGemmMultiplyAdd::Accumulators fragment_c;
|
|
fragment_c.clear();
|
|
|
|
// Do the WMMA.
|
|
WmmaGemmMultiplyAdd multiply_add;
|
|
multiply_add.multiply_add(fragment_a, fragment_b, fragment_c, fragment_c);
|
|
|
|
// Global memory stream to store D.
|
|
typedef cutlass::gemm::WmmaGemmGlobalIteratorCd<
|
|
cutlass::gemm::WmmaGemmGlobalIteratorCdTraits<
|
|
float, TileC, ThreadsA, 1>
|
|
>
|
|
GlobalStoreIteratorD;
|
|
typedef cutlass::gemm::GlobalStoreStream<GlobalStoreIteratorD> GlobalStoreStreamD;
|
|
|
|
// The shared memory to store D.
|
|
__shared__ typename GlobalStoreStreamD::SharedStorage shared_storage_stream_d;
|
|
|
|
// Iterator to store C.
|
|
typename GlobalStoreStreamD::Params global_params_d;
|
|
global_params_d.initialize(desc, d_c, ldc);
|
|
GlobalStoreStreamD global_store_d(global_params_d, shared_storage_stream_d, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Shared store iterator/stream for C.
|
|
typedef cutlass::gemm::WmmaGemmSharedStoreIteratorD<
|
|
cutlass::gemm::WmmaGemmSharedStoreIteratorDTraits<
|
|
cutlass::MatrixLayout::kColumnMajor, float, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedStoreIteratorD;
|
|
typedef cutlass::gemm::SharedStoreStream<SharedStoreIteratorD> SharedStoreStreamD;
|
|
|
|
// Shared load iterator/stream for D.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorD<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorDTraits<
|
|
float, typename SharedStoreIteratorD::Tile, ThreadsA, 1> >
|
|
SharedLoadIteratorD;
|
|
typedef cutlass::gemm::SharedLoadStream<SharedLoadIteratorD> SharedLoadStreamD;
|
|
|
|
// The shared memory structure to swizzle D.
|
|
union SharedStorageD {
|
|
typename SharedStoreStreamD::SharedStorage store;
|
|
typename SharedLoadStreamD::SharedStorage load;
|
|
};
|
|
|
|
// The shared memory for D.
|
|
__shared__ SharedStorageD shared_storage_d;
|
|
|
|
// Store iterator D.
|
|
typename SharedStoreStreamD::Params shared_store_params_d;
|
|
shared_store_params_d.initialize();
|
|
|
|
// Store iterator D.
|
|
typename SharedLoadStreamD::Params shared_load_params_d;
|
|
shared_load_params_d.initialize();
|
|
|
|
// The number of WMMA in the tile H/W dimension (N/M in GEMM).
|
|
int const kWmmaPerH = OutputTile_::kH / Warps_::kH / WmmaShape_::kH;
|
|
int const kWmmaPerW = OutputTile_::kW / Warps_::kW / WmmaShape_::kW;
|
|
|
|
// Iterate over the data.
|
|
for (int i = 0; i < kWmmaPerH; ++i) {
|
|
// Make sure the shared memory can be written to.
|
|
__syncthreads();
|
|
|
|
// Create the iterator to store to SMEM.
|
|
SharedStoreStreamD shared_store_d(shared_store_params_d,
|
|
shared_storage_d.store,
|
|
fragment_c,
|
|
i*kWmmaPerW);
|
|
shared_store_d.copy();
|
|
shared_store_d.commit();
|
|
|
|
// Make sure the shared memory was written.
|
|
__syncthreads();
|
|
|
|
// Create the iterator to load from SMEM.
|
|
SharedLoadStreamD shared_load_d(shared_load_params_d, shared_storage_d.load);
|
|
shared_load_d.copy();
|
|
shared_load_d.commit();
|
|
|
|
// Copy the data.
|
|
cutlass::Copy<typename SharedLoadStreamD::Fragment> copy;
|
|
copy.transform(shared_load_d.fragment(), global_store_d.fragment());
|
|
|
|
// Copy the data to global memory.
|
|
global_store_d.copy();
|
|
global_store_d.commit();
|
|
}
|
|
}
|
|
#else
|
|
template <typename OutputTile_, typename Warps_, typename WmmaShape_>
|
|
static __global__ void kernel_nt(half const *d_a, int lda, half const *d_b, int ldb, float *d_c,
|
|
int ldc) {
|
|
// The default configuration of threads.
|
|
typedef cutlass::Shape<1, Warps_::kCount, 32> Threads_;
|
|
// The threads.
|
|
typedef typename ReshapeThreadsA<Threads_, OutputTile_::kW>::Threads ThreadsA;
|
|
// The threads.
|
|
typedef typename ReshapeThreadsB<Threads_, OutputTile_::kH>::Threads ThreadsB;
|
|
// The number of elements loaded per LDG.
|
|
int const kScalarsPerLdg = 1;
|
|
// The tile for A.
|
|
typedef cutlass::Shape<1, OutputTile_::kD, OutputTile_::kW> TileA;
|
|
// The tile for B.
|
|
typedef cutlass::Shape<1, OutputTile_::kD, OutputTile_::kH> TileB;
|
|
// The tile for C.
|
|
typedef cutlass::Shape<1, Warps_::kH*WmmaShape_::kH, OutputTile_::kW> TileC;
|
|
|
|
// The problem descriptor.
|
|
ProblemDesc desc(OutputTile_::kW, OutputTile_::kH, OutputTile_::kD);
|
|
|
|
// The elements computed by a single warp.
|
|
typedef typename cutlass::ShapeDiv<OutputTile_, Warps_>::Shape AccumulatorsPerWarp;
|
|
|
|
// Global memory load for A.
|
|
typedef cutlass::gemm::GemmGlobalIteratorAb<
|
|
cutlass::gemm::GemmGlobalIteratorTraits<
|
|
cutlass::GemmOperand::kA, cutlass::MatrixLayout::kColumnMajor, half const, TileA, ThreadsA, kScalarsPerLdg>
|
|
>
|
|
GlobalLoadIteratorA;
|
|
|
|
// Shared store iterator for A.
|
|
typedef cutlass::gemm::GemmSharedStoreIteratorAb<
|
|
cutlass::gemm::GemmSharedStoreIteratorAbTraits<
|
|
half, TileA, ThreadsA, kScalarsPerLdg>
|
|
>
|
|
SharedStoreIteratorA;
|
|
|
|
// The global stream for A.
|
|
typedef cutlass::gemm::GlobalLoadStream<
|
|
GlobalLoadIteratorA,
|
|
cutlass::Copy<typename GlobalLoadIteratorA::Fragment>,
|
|
SharedStoreIteratorA>
|
|
GlobalLoadStreamA;
|
|
|
|
// Shared load iterator for A.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorA<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorAbTraits<
|
|
cutlass::GemmOperand::kA,
|
|
cutlass::MatrixLayout::kColumnMajor, half, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedLoadIteratorA;
|
|
|
|
// Global memory load for B.
|
|
typedef cutlass::gemm::GemmGlobalIteratorAb<
|
|
cutlass::gemm::GemmGlobalIteratorTraits<
|
|
cutlass::GemmOperand::kB, cutlass::MatrixLayout::kRowMajor, half const, TileB, ThreadsB, kScalarsPerLdg> >
|
|
GlobalLoadIteratorB;
|
|
|
|
// Shared store iterator for B.
|
|
typedef cutlass::gemm::GemmSharedStoreIteratorAb<
|
|
cutlass::gemm::GemmSharedStoreIteratorAbTraits<
|
|
half, TileB, ThreadsB, kScalarsPerLdg> >
|
|
SharedStoreIteratorB;
|
|
|
|
// The global stream for B.
|
|
typedef cutlass::gemm::GlobalLoadStream<GlobalLoadIteratorB,
|
|
cutlass::Copy<typename GlobalLoadIteratorB::Fragment>,
|
|
SharedStoreIteratorB>
|
|
GlobalLoadStreamB;
|
|
|
|
// Shared load iterator for B.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorB<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorAbTraits<
|
|
cutlass::GemmOperand::kB,
|
|
cutlass::MatrixLayout::kRowMajor, half, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedLoadIteratorB;
|
|
|
|
// Share memory to exchange data for A.
|
|
__shared__ SharedStorage<GlobalLoadStreamA, SharedLoadIteratorA> shared_storage_a;
|
|
|
|
// Share memory to exchange data for B.
|
|
__shared__ SharedStorage<GlobalLoadStreamB, SharedLoadIteratorB> shared_storage_b;
|
|
|
|
// Iterator to load A.
|
|
typename GlobalLoadStreamA::Params global_params_a;
|
|
global_params_a.initialize(desc, d_a, lda);
|
|
GlobalLoadStreamA global_load_a(global_params_a, shared_storage_a.store, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Iterator to load B.
|
|
typename GlobalLoadStreamB::Params global_params_b;
|
|
global_params_b.initialize(desc, d_b, ldb);
|
|
GlobalLoadStreamB global_load_b(global_params_b, shared_storage_b.store, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Load A/B.
|
|
global_load_a.copy();
|
|
global_load_b.copy();
|
|
|
|
// Copy to shared memory.
|
|
global_load_a.commit();
|
|
global_load_b.commit();
|
|
|
|
// Make sure the data is in shared memory.
|
|
__syncthreads();
|
|
|
|
// Load iterator A.
|
|
typename SharedLoadIteratorA::Params shared_params_a;
|
|
shared_params_a.initialize(desc);
|
|
SharedLoadIteratorA shared_load_a(shared_params_a, shared_storage_a.load);
|
|
|
|
// Load iterator B.
|
|
typename SharedLoadIteratorB::Params shared_params_b;
|
|
shared_params_b.initialize(desc);
|
|
SharedLoadIteratorB shared_load_b(shared_params_b, shared_storage_b.load);
|
|
|
|
// Copy A from shared memory.
|
|
typename SharedLoadIteratorA::Fragment fragment_a;
|
|
cutlass::gemm::load_shared(shared_load_a, fragment_a);
|
|
|
|
// Copy B from shared memory.
|
|
typename SharedLoadIteratorB::Fragment fragment_b;
|
|
cutlass::gemm::load_shared(shared_load_b, fragment_b);
|
|
|
|
// The functor to do WMMA.
|
|
typedef cutlass::gemm::WmmaGemmMultiplyAdd<
|
|
cutlass::MatrixLayout::kColumnMajor,
|
|
cutlass::MatrixLayout::kRowMajor,
|
|
cutlass::MatrixLayout::kColumnMajor,
|
|
float,
|
|
AccumulatorsPerWarp,
|
|
WmmaShape_> WmmaGemmMultiplyAdd;
|
|
|
|
// The output fragment.
|
|
typename WmmaGemmMultiplyAdd::Accumulators fragment_c;
|
|
fragment_c.clear();
|
|
|
|
// Do the WMMA.
|
|
WmmaGemmMultiplyAdd multiply_add;
|
|
multiply_add.multiply_add(fragment_a, fragment_b, fragment_c, fragment_c);
|
|
|
|
// Global memory stream to store D.
|
|
typedef cutlass::gemm::WmmaGemmGlobalIteratorCd<
|
|
cutlass::gemm::WmmaGemmGlobalIteratorCdTraits<
|
|
float, TileC, ThreadsA, 1>
|
|
>
|
|
GlobalStoreIteratorD;
|
|
typedef cutlass::gemm::GlobalStoreStream<GlobalStoreIteratorD> GlobalStoreStreamD;
|
|
|
|
// The shared memory to store D.
|
|
__shared__ typename GlobalStoreStreamD::SharedStorage shared_storage_stream_d;
|
|
|
|
// Iterator to store C.
|
|
typename GlobalStoreStreamD::Params global_params_d;
|
|
global_params_d.initialize(desc, d_c, ldc);
|
|
GlobalStoreStreamD global_store_d(global_params_d, shared_storage_stream_d, desc.m, desc.n, desc.k,
|
|
cutlass::make_Coord(0, 0, 0));
|
|
|
|
// Shared store iterator/stream for C.
|
|
typedef cutlass::gemm::WmmaGemmSharedStoreIteratorD<
|
|
cutlass::gemm::WmmaGemmSharedStoreIteratorDTraits<
|
|
cutlass::MatrixLayout::kColumnMajor, float, OutputTile_, Warps_, WmmaShape_> >
|
|
SharedStoreIteratorD;
|
|
typedef cutlass::gemm::SharedStoreStream<SharedStoreIteratorD> SharedStoreStreamD;
|
|
|
|
// Shared load iterator/stream for D.
|
|
typedef cutlass::gemm::WmmaGemmSharedLoadIteratorD<
|
|
cutlass::gemm::WmmaGemmSharedLoadIteratorDTraits<
|
|
float, typename SharedStoreIteratorD::Tile, ThreadsA, 1> >
|
|
SharedLoadIteratorD;
|
|
typedef cutlass::gemm::SharedLoadStream<SharedLoadIteratorD> SharedLoadStreamD;
|
|
|
|
// The shared memory structure to swizzle D.
|
|
union SharedStorageD {
|
|
typename SharedStoreStreamD::SharedStorage store;
|
|
typename SharedLoadStreamD::SharedStorage load;
|
|
};
|
|
|
|
// The shared memory for D.
|
|
__shared__ SharedStorageD shared_storage_d;
|
|
|
|
// Store iterator D.
|
|
typename SharedStoreStreamD::Params shared_store_params_d;
|
|
shared_store_params_d.initialize();
|
|
|
|
// Store iterator D.
|
|
typename SharedLoadStreamD::Params shared_load_params_d;
|
|
shared_load_params_d.initialize();
|
|
|
|
// The number of WMMA in the tile H/W dimension (N/M in GEMM).
|
|
int const kWmmaPerH = OutputTile_::kH / Warps_::kH / WmmaShape_::kH;
|
|
int const kWmmaPerW = OutputTile_::kW / Warps_::kW / WmmaShape_::kW;
|
|
|
|
// Iterate over the data.
|
|
for (int i = 0; i < kWmmaPerH; ++i) {
|
|
// Make sure the shared memory can be written to.
|
|
__syncthreads();
|
|
|
|
// Create the iterator to store to SMEM.
|
|
SharedStoreStreamD shared_store_d(shared_store_params_d,
|
|
shared_storage_d.store,
|
|
fragment_c,
|
|
i*kWmmaPerW);
|
|
shared_store_d.copy();
|
|
shared_store_d.commit();
|
|
|
|
// Make sure the shared memory was written.
|
|
__syncthreads();
|
|
|
|
// Create the iterator to load from SMEM.
|
|
SharedLoadStreamD shared_load_d(shared_load_params_d, shared_storage_d.load);
|
|
shared_load_d.copy();
|
|
shared_load_d.commit();
|
|
|
|
// Copy the data.
|
|
cutlass::Copy<typename SharedLoadStreamD::Fragment> copy;
|
|
copy.transform(shared_load_d.fragment(), global_store_d.fragment());
|
|
|
|
// Copy the data to global memory.
|
|
global_store_d.copy();
|
|
global_store_d.commit();
|
|
}
|
|
}
|
|
#endif
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <typename OutputTile_, typename Warps_, typename WmmaShape_>
|
|
void run() {
|
|
/// Testbed type.
|
|
typedef test::GemmTestbed<cutlass::half_t, cutlass::half_t, float, float, float> GemmTestbed;
|
|
|
|
// Create the testbed.
|
|
GemmTestbed testbed(OutputTile_::kW, // M
|
|
OutputTile_::kH, // N
|
|
OutputTile_::kD, // K
|
|
cutlass::convert(cutlass::MatrixLayout::kColumnMajor),
|
|
cutlass::convert(cutlass::MatrixLayout::kRowMajor), 1, 0,
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP,
|
|
cutlass::convert(cutlass::MatrixLayout::kColumnMajor));
|
|
|
|
// Initialize.
|
|
testbed.initialize();
|
|
|
|
// Launch the kernel.
|
|
kernel_nt<OutputTile_, Warps_, WmmaShape_><<<1, 32*Warps_::kCount>>>(
|
|
testbed.ptr_A(), testbed.lda(),
|
|
testbed.ptr_B(), testbed.ldb(),
|
|
testbed.ptr_computed(), testbed.ldc());
|
|
ASSERT_EQ(cudaSuccess, cudaGetLastError());
|
|
|
|
// Make sure it worked as expected.
|
|
ASSERT_TRUE(testbed.verify_with_host());
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_16x16x16_16x16x16) {
|
|
run<cutlass::Shape<16, 16, 16>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_16x32x16_16x16x16) {
|
|
run<cutlass::Shape<16, 32, 16>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_32x16x16_16x16x16) {
|
|
run<cutlass::Shape<16, 16, 32>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_64x16x16_16x16x16) {
|
|
run<cutlass::Shape<16, 16, 64>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_64x64x16_16x16x16) {
|
|
run<cutlass::Shape<16, 64, 64>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_128x128x16_16x16x16) {
|
|
run<cutlass::Shape<16, 128, 128>, cutlass::Shape<1, 2, 2>, cutlass::Shape<16, 16, 16> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_32x8x16_32x8x16) {
|
|
run<cutlass::Shape<16, 8, 32>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 8, 32> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_128x128x16_32x8x16) {
|
|
run<cutlass::Shape<16, 128, 128>, cutlass::Shape<1, 2, 2>, cutlass::Shape<16, 8, 32> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_8x32x16_8x32x16) {
|
|
run<cutlass::Shape<16, 32, 8>, cutlass::Shape<1, 1, 1>, cutlass::Shape<16, 32, 8> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
TEST(WmmaGemm, multiply_add_f32_128x128x16_8x32x16) {
|
|
run<cutlass::Shape<16, 128, 128>, cutlass::Shape<1, 2, 2>, cutlass::Shape<16, 32, 8> >();
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
#endif // defined CUTLASS_USE_WMMA_API
|