CUTLASS 2.5
This commit is contained in:
Andrew Kerr
@@ -0,0 +1,409 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, 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 Template for a pipelined GEMM kernel. Does not compute batching or support split-K.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/gemm/gemm.h"
|
||||
#include "cutlass/matrix_coord.h"
|
||||
#include "cutlass/semaphore.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace gemm {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
typename B2bMma_, ///! Threadblock-scoped matrix multiply-accumulate
|
||||
typename Epilogue_, ///! Epilogue
|
||||
typename ThreadblockSwizzle_, ///! Threadblock swizzling function
|
||||
bool SplitKSerial ///! If true, code supporting split-K via serial reduction is enabled.
|
||||
>
|
||||
struct B2bGemm {
|
||||
|
||||
using B2bMma = B2bMma_;
|
||||
using Epilogue = Epilogue_;
|
||||
using OutputOp0 = typename B2bMma::OutputOp;
|
||||
using OutputOp1 = typename Epilogue::OutputOp;
|
||||
using ThreadblockSwizzle = ThreadblockSwizzle_;
|
||||
static bool const kSplitKSerial = SplitKSerial;
|
||||
|
||||
/// Warp count (concept: GemmShape)
|
||||
using WarpCount0 = typename B2bMma::WarpCount0;
|
||||
static int const kThreadCount = 32 * WarpCount0::kCount;
|
||||
|
||||
/// Parameters structure
|
||||
struct Params {
|
||||
cutlass::gemm::GemmCoord problem_size_0;
|
||||
cutlass::gemm::GemmCoord problem_size_1;
|
||||
cutlass::gemm::GemmCoord grid_tiled_shape;
|
||||
typename B2bMma::IteratorA0::Params params_A0;
|
||||
typename B2bMma::IteratorA0::TensorRef ref_A0;
|
||||
typename B2bMma::IteratorB0::Params params_B0;
|
||||
typename B2bMma::IteratorB0::TensorRef ref_B0;
|
||||
typename Epilogue::OutputTileIterator::Params params_C0;
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C0;
|
||||
typename B2bMma::IteratorB1::Params params_B1;
|
||||
typename B2bMma::IteratorB1::TensorRef ref_B1;
|
||||
typename Epilogue::OutputTileIterator::Params params_C1;
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C1;
|
||||
typename Epilogue::OutputTileIterator::Params params_D1;
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_D1;
|
||||
typename OutputOp0::Params output_op_0;
|
||||
typename OutputOp1::Params output_op_1;
|
||||
int *semaphore;
|
||||
int gemm_k_iterations_0;
|
||||
int gemm_k_size_0;
|
||||
int gemm_k_iterations_1;
|
||||
int gemm_k_size_1;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(): semaphore(0), gemm_k_iterations_0(0), gemm_k_size_0(0),
|
||||
gemm_k_iterations_1(0), gemm_k_size_1(0) { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(
|
||||
cutlass::gemm::GemmCoord const & problem_size_0,
|
||||
cutlass::gemm::GemmCoord const & problem_size_1,
|
||||
cutlass::gemm::GemmCoord const & grid_tiled_shape,
|
||||
typename B2bMma::IteratorA0::TensorRef ref_A0,
|
||||
typename B2bMma::IteratorB0::TensorRef ref_B0,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C0,
|
||||
typename B2bMma::IteratorB1::TensorRef ref_B1,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C1,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_D1,
|
||||
typename OutputOp0::Params output_op_0 = typename OutputOp0::Params(),
|
||||
typename OutputOp1::Params output_op_1 = typename OutputOp1::Params(),
|
||||
int *workspace = nullptr
|
||||
):
|
||||
problem_size_0(problem_size_0),
|
||||
problem_size_1(problem_size_1),
|
||||
grid_tiled_shape(grid_tiled_shape),
|
||||
params_A0(ref_A0.layout()),
|
||||
ref_A0(ref_A0),
|
||||
params_B0(ref_B0.layout()),
|
||||
ref_B0(ref_B0),
|
||||
params_C0(ref_C0.layout()),
|
||||
ref_C0(ref_C0),
|
||||
params_B1(ref_B1.layout()),
|
||||
ref_B1(ref_B1),
|
||||
params_C1(ref_C1.layout()),
|
||||
ref_C1(ref_C1),
|
||||
params_D1(ref_D1.layout()),
|
||||
ref_D1(ref_D1),
|
||||
output_op_0(output_op_0),
|
||||
output_op_1(output_op_1) {
|
||||
|
||||
int total_gemm_k_iterations_0 = (problem_size_0.k() + B2bMma::Shape0::kK - 1) / B2bMma::Shape0::kK;
|
||||
int gemm_k_iterations_0 = (total_gemm_k_iterations_0 + grid_tiled_shape.k() - 1) / grid_tiled_shape.k();
|
||||
gemm_k_size_0 = gemm_k_iterations_0 * B2bMma::Shape0::kK;
|
||||
int total_gemm_k_iterations_1 = (problem_size_1.k() + B2bMma::Shape1::kK - 1) / B2bMma::Shape1::kK;
|
||||
int gemm_k_iterations_1 = (total_gemm_k_iterations_1 + grid_tiled_shape.k() - 1) / grid_tiled_shape.k();
|
||||
gemm_k_size_1 = gemm_k_iterations_1 * B2bMma::Shape1::kK;
|
||||
|
||||
semaphore = workspace;
|
||||
}
|
||||
};
|
||||
|
||||
/// Shared memory storage structure
|
||||
union SharedStorage {
|
||||
typename B2bMma::B2bMmaSharedStorage main_loop;
|
||||
typename Epilogue::SharedStorage epilogue;
|
||||
};
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
B2bGemm() { }
|
||||
|
||||
/// Determines whether kernel satisfies alignment
|
||||
static Status can_implement(
|
||||
cutlass::gemm::GemmCoord const & problem_size_0,
|
||||
cutlass::gemm::GemmCoord const & problem_size_1,
|
||||
typename B2bMma::IteratorA0::TensorRef ref_A0,
|
||||
typename B2bMma::IteratorB0::TensorRef ref_B0,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C0,
|
||||
typename B2bMma::IteratorB1::TensorRef ref_B1,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_C1,
|
||||
typename Epilogue::OutputTileIterator::TensorRef ref_D1) {
|
||||
|
||||
static int const kAlignmentA = B2bMma::IteratorA0::AccessType::kElements;
|
||||
static int const kAlignmentB = B2bMma::IteratorB0::AccessType::kElements;
|
||||
static int const kAlignmentC = Epilogue::OutputTileIterator::kElementsPerAccess;
|
||||
|
||||
if (!TensorRef_aligned(ref_A0, kAlignmentA)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (!TensorRef_aligned(ref_B0, kAlignmentB)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (!TensorRef_aligned(ref_C0, kAlignmentC)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (!TensorRef_aligned(ref_B1, kAlignmentB)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (!TensorRef_aligned(ref_C1, kAlignmentC)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (!TensorRef_aligned(ref_D1, kAlignmentC)) {
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if ((problem_size_0.m() % kAlignmentA) || (problem_size_0.k() % kAlignmentA) ||
|
||||
(problem_size_0.n() % kAlignmentB) || (problem_size_0.k() % kAlignmentB) ||
|
||||
(problem_size_0.m() % kAlignmentC) || (problem_size_0.n() % kAlignmentC) ||
|
||||
(problem_size_1.m() % kAlignmentA) || (problem_size_1.k() % kAlignmentA) ||
|
||||
(problem_size_1.n() % kAlignmentB) || (problem_size_1.k() % kAlignmentB) ||
|
||||
(problem_size_1.m() % kAlignmentC) || (problem_size_1.n() % kAlignmentC)) {
|
||||
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
|
||||
/// Executes one GEMM
|
||||
CUTLASS_DEVICE
|
||||
void operator()(Params const ¶ms, SharedStorage &shared_storage) {
|
||||
|
||||
// Compute threadblock location
|
||||
ThreadblockSwizzle threadblock_swizzle;
|
||||
|
||||
cutlass::gemm::GemmCoord threadblock_tile_offset =
|
||||
threadblock_swizzle.get_tile_offset(params.grid_tiled_shape);
|
||||
|
||||
// Early exit if CTA is out of range
|
||||
if (params.grid_tiled_shape.m() <= threadblock_tile_offset.m() ||
|
||||
params.grid_tiled_shape.n() <= threadblock_tile_offset.n()) {
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
// Compute initial location in logical coordinates
|
||||
cutlass::MatrixCoord tb_offset_A0{
|
||||
threadblock_tile_offset.m() * B2bMma::Shape0::kM,
|
||||
threadblock_tile_offset.k() * params.gemm_k_size_0,
|
||||
};
|
||||
|
||||
cutlass::MatrixCoord tb_offset_B0{
|
||||
threadblock_tile_offset.k() * params.gemm_k_size_0,
|
||||
threadblock_tile_offset.n() * B2bMma::Shape0::kN
|
||||
};
|
||||
|
||||
cutlass::MatrixCoord tb_offset_B1{
|
||||
threadblock_tile_offset.k() * params.gemm_k_size_1,
|
||||
threadblock_tile_offset.n() * B2bMma::Shape1::kN
|
||||
};
|
||||
|
||||
// Problem size is a function of threadblock index in the K dimension
|
||||
int problem_size_k_0 = min(
|
||||
params.problem_size_0.k(),
|
||||
(threadblock_tile_offset.k() + 1) * params.gemm_k_size_0);
|
||||
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
int gemm_k_iterations_0 = (problem_size_k_0 - tb_offset_A0.column() + B2bMma::Shape0::kK - 1) / B2bMma::Shape0::kK;
|
||||
|
||||
// Problem size is a function of threadblock index in the K dimension
|
||||
int problem_size_k_1 = min(
|
||||
params.problem_size_1.k(),
|
||||
(threadblock_tile_offset.k() + 1) * params.gemm_k_size_1);
|
||||
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
// int gemm_k_iterations_1 = (problem_size_k_1 - tb_offset_B1.row() + B2bMma::Shape1::kK - 1) / B2bMma::Shape1::kK;
|
||||
|
||||
|
||||
// Compute position within threadblock
|
||||
int thread_idx = threadIdx.x;
|
||||
|
||||
// Construct iterators to A and B operands
|
||||
typename B2bMma::IteratorA0 iterator_A0(
|
||||
params.params_A0,
|
||||
params.ref_A0.data(),
|
||||
{params.problem_size_0.m(), problem_size_k_0},
|
||||
thread_idx,
|
||||
tb_offset_A0);
|
||||
|
||||
typename B2bMma::IteratorB0 iterator_B0(
|
||||
params.params_B0,
|
||||
params.ref_B0.data(),
|
||||
{problem_size_k_0, params.problem_size_0.n()},
|
||||
thread_idx,
|
||||
tb_offset_B0);
|
||||
|
||||
typename B2bMma::IteratorB1 iterator_B1(
|
||||
params.params_B1,
|
||||
params.ref_B1.data(),
|
||||
{problem_size_k_1, params.problem_size_1.n()},
|
||||
thread_idx,
|
||||
tb_offset_B1);
|
||||
|
||||
|
||||
// Broadcast the warp_id computed by lane 0 to ensure dependent code
|
||||
// is compiled as warp-uniform.
|
||||
int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
|
||||
int lane_idx = threadIdx.x % 32;
|
||||
|
||||
//
|
||||
// Main loop
|
||||
//
|
||||
|
||||
OutputOp0 output_op_0(params.output_op_0);
|
||||
|
||||
// Construct thread-scoped matrix multiply
|
||||
B2bMma b2bMma(shared_storage.main_loop, thread_idx, warp_idx, lane_idx);
|
||||
|
||||
typename B2bMma::FragmentC0 src_accum;
|
||||
typename B2bMma::FragmentC1 accumulators;
|
||||
|
||||
src_accum.clear();
|
||||
accumulators.clear();
|
||||
|
||||
if (!kSplitKSerial || gemm_k_iterations_0 > 0) {
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
b2bMma(gemm_k_iterations_0, accumulators, iterator_A0, iterator_B0, iterator_B1, src_accum, output_op_0);
|
||||
}
|
||||
|
||||
//
|
||||
// Epilogue
|
||||
//
|
||||
|
||||
OutputOp1 output_op_1(params.output_op_1);
|
||||
|
||||
//
|
||||
// Masked tile iterators constructed from members
|
||||
//
|
||||
|
||||
threadblock_tile_offset =
|
||||
threadblock_swizzle.get_tile_offset(params.grid_tiled_shape);
|
||||
|
||||
//assume identity swizzle
|
||||
MatrixCoord threadblock_offset(
|
||||
threadblock_tile_offset.m() * B2bMma::Shape1::kM,
|
||||
threadblock_tile_offset.n() * B2bMma::Shape1::kN
|
||||
);
|
||||
|
||||
int block_idx = threadblock_tile_offset.m() + threadblock_tile_offset.n() * params.grid_tiled_shape.m();
|
||||
|
||||
// Construct the semaphore.
|
||||
Semaphore semaphore(params.semaphore + block_idx, thread_idx);
|
||||
|
||||
// If performing a reduction via split-K, fetch the initial synchronization
|
||||
if (kSplitKSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
// Fetch the synchronization lock initially but do not block.
|
||||
semaphore.fetch();
|
||||
|
||||
// Indicate which position in a serial reduction the output operator is currently updating
|
||||
output_op_1.set_k_partition(threadblock_tile_offset.k(), params.grid_tiled_shape.k());
|
||||
}
|
||||
|
||||
// Tile iterator loading from source tensor.
|
||||
typename Epilogue::OutputTileIterator iterator_C1(
|
||||
params.params_C1,
|
||||
params.ref_C1.data(),
|
||||
params.problem_size_1.mn(),
|
||||
thread_idx,
|
||||
threadblock_offset
|
||||
);
|
||||
|
||||
// Tile iterator writing to destination tensor.
|
||||
typename Epilogue::OutputTileIterator iterator_D1(
|
||||
params.params_D1,
|
||||
params.ref_D1.data(),
|
||||
params.problem_size_1.mn(),
|
||||
thread_idx,
|
||||
threadblock_offset
|
||||
);
|
||||
|
||||
Epilogue epilogue(
|
||||
shared_storage.epilogue,
|
||||
thread_idx,
|
||||
warp_idx,
|
||||
lane_idx);
|
||||
|
||||
// Wait on the semaphore - this latency may have been covered by iterator construction
|
||||
if (kSplitKSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
// For subsequent threadblocks, the source matrix is held in the 'D' tensor.
|
||||
if (threadblock_tile_offset.k()) {
|
||||
iterator_C1 = iterator_D1;
|
||||
}
|
||||
|
||||
semaphore.wait(threadblock_tile_offset.k());
|
||||
|
||||
__threadfence();
|
||||
}
|
||||
|
||||
// Execute the epilogue operator to update the destination tensor.
|
||||
epilogue(output_op_1, iterator_D1, accumulators, iterator_C1);
|
||||
|
||||
//
|
||||
// Release the semaphore
|
||||
//
|
||||
|
||||
if (kSplitKSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
int lock = 0;
|
||||
if (params.grid_tiled_shape.k() == threadblock_tile_offset.k() + 1) {
|
||||
|
||||
// The final threadblock resets the semaphore for subsequent grids.
|
||||
lock = 0;
|
||||
}
|
||||
else {
|
||||
// Otherwise, the semaphore is incremented
|
||||
lock = threadblock_tile_offset.k() + 1;
|
||||
}
|
||||
|
||||
__threadfence();
|
||||
semaphore.release(lock);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace gemm
|
||||
} // namespace cutlass
|
||||
|
||||
@@ -0,0 +1,475 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, 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 Template for a pipelined Implicit GEMM kernel.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/aligned_buffer.h"
|
||||
#include "cutlass/array.h"
|
||||
#include "cutlass/numeric_types.h"
|
||||
#include "cutlass/matrix_shape.h"
|
||||
#include "cutlass/semaphore.h"
|
||||
#include "cutlass/tensor_ref.h"
|
||||
#include "cutlass/layout/tensor.h"
|
||||
#include "cutlass/gemm/gemm.h"
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/conv/conv2d_problem_size.h"
|
||||
#include "cutlass/conv/conv3d_problem_size.h"
|
||||
#include "cutlass/epilogue/threadblock/output_iterator_parameter.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace conv {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
typename B2bMma_, ///! Threadblock-scoped matrix multiply-accumulate
|
||||
typename Epilogue_, ///! Epilogue
|
||||
typename ThreadblockSwizzle_, ///! Threadblock swizzling function
|
||||
conv::Operator ConvOperator, ///! Convolutional operator (Fprop, Dgrad, Wgrad)
|
||||
typename ConvProblemSize_ = Conv2dProblemSize ///! Convolutional operator on 2D or 3D problem
|
||||
>
|
||||
struct B2bImplicitGemmConvolution {
|
||||
|
||||
using B2bMma = B2bMma_;
|
||||
using Epilogue = Epilogue_;
|
||||
using EpilogueOutputOp0 = typename B2bMma::OutputOp;
|
||||
using EpilogueOutputOp1 = typename Epilogue::OutputOp;
|
||||
using ThreadblockSwizzle = ThreadblockSwizzle_;
|
||||
static Operator const kConvolutionalOperator = ConvOperator;
|
||||
|
||||
using ElementA = typename B2bMma::IteratorA0::Element;
|
||||
using LayoutA = typename B2bMma::IteratorA0::Layout;
|
||||
using ElementB = typename B2bMma::IteratorB0::Element;
|
||||
using LayoutB = typename B2bMma::IteratorB0::Layout;
|
||||
using ElementC = typename EpilogueOutputOp1::ElementOutput;
|
||||
|
||||
/// Set output tensor C layout
|
||||
using LayoutC = LayoutA;
|
||||
|
||||
using ElementAccumulator = typename EpilogueOutputOp0::ElementAccumulator;
|
||||
using ElementCompute = typename EpilogueOutputOp0::ElementCompute;
|
||||
|
||||
using WarpMmaOperator0 = typename B2bMma::Policy0::Operator;
|
||||
using WarpMmaOperator1 = typename B2bMma::Policy1::Operator;
|
||||
|
||||
using ArchMmaOperator = typename WarpMmaOperator0::ArchMmaOperator;
|
||||
using MathOperator = typename ArchMmaOperator::Operator;
|
||||
|
||||
using OperatorClass = typename WarpMmaOperator0::OperatorClass;
|
||||
using ArchTag = typename WarpMmaOperator0::ArchTag;
|
||||
|
||||
using ThreadblockShape0 = typename B2bMma::Shape0;
|
||||
using ThreadblockShape1 = typename B2bMma::Shape1;
|
||||
using WarpShape0 = typename WarpMmaOperator0::Shape;
|
||||
using WarpShape1 = typename WarpMmaOperator1::Shape;
|
||||
using InstructionShape = typename ArchMmaOperator::Shape;
|
||||
|
||||
static int const kStages = B2bMma::kStages;
|
||||
static IteratorAlgorithm const kIteratorAlgorithm = B2bMma::IteratorA0::kIteratorAlgorithm;
|
||||
|
||||
/// Warp count (concept: GemmShape)
|
||||
using WarpCount0 = typename B2bMma::WarpCount0;
|
||||
static int const kThreadCount = 32 * WarpCount0::kCount;
|
||||
|
||||
using TensorRefA0 = typename B2bMma::IteratorA0::TensorRef;
|
||||
using TensorRefB0 = typename B2bMma::IteratorB0::TensorRef;
|
||||
using TensorRefB1 = typename B2bMma::IteratorB1::TensorRef;
|
||||
using TensorRefC = cutlass::TensorRef<ElementC, LayoutC>;
|
||||
|
||||
/// Check iterator A and B convolution dimension are the same and
|
||||
// set device::B2bImplicitGemmConvolution::kConvDim
|
||||
static_assert(B2bMma::IteratorA0::kConvDim == B2bMma::IteratorB0::kConvDim,
|
||||
"Convolution on different different dimensions is not supported");
|
||||
static int const kConvDim = B2bMma::IteratorA0::kConvDim;
|
||||
|
||||
/// Conv dimension and problem size structure (Conv2d or Conv3d)
|
||||
using ConvProblemSize = ConvProblemSize_;
|
||||
|
||||
/// Wgrad C stride idx for implicit gemm algorithm
|
||||
// Conv2d row-major matrix C (KxRSC)
|
||||
// Conv3d row-major matrix C (KxTRSC)
|
||||
static int const kWgradCStrideIdx =
|
||||
cutlass::platform::is_same<LayoutC, cutlass::layout::TensorNHWC>::value ? 2 : 3;
|
||||
|
||||
/// This chooses the appropriate stride element of the C tensor.
|
||||
static int const kTensorCStrideIdx =
|
||||
(kConvolutionalOperator == conv::Operator::kWgrad ? kWgradCStrideIdx : 0);
|
||||
|
||||
//
|
||||
//
|
||||
//
|
||||
using ConvOutputIteratorParameter = epilogue::threadblock::ConvOutputIteratorParameter<
|
||||
LayoutC,
|
||||
typename Epilogue::OutputTileIterator::Layout,
|
||||
TensorRefC,
|
||||
ConvOperator,
|
||||
ConvProblemSize
|
||||
>;
|
||||
|
||||
/// Argument structure
|
||||
struct Arguments {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
ConvProblemSize problem_size_0;
|
||||
ConvProblemSize problem_size_1;
|
||||
TensorRefA0 ref_A0;
|
||||
TensorRefB0 ref_B0;
|
||||
TensorRefC ref_C0;
|
||||
TensorRefB1 ref_B1;
|
||||
TensorRefC ref_C1;
|
||||
TensorRefC ref_D1;
|
||||
typename EpilogueOutputOp0::Params output_op_0;
|
||||
typename EpilogueOutputOp1::Params output_op_1;
|
||||
SplitKMode split_k_mode;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
/// Default ctor
|
||||
CUTLASS_HOST_DEVICE
|
||||
Arguments() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Arguments(
|
||||
ConvProblemSize const & problem_size_0,
|
||||
ConvProblemSize const & problem_size_1
|
||||
):
|
||||
problem_size_0(problem_size_0),
|
||||
problem_size_1(problem_size_1) { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Arguments(
|
||||
ConvProblemSize const & problem_size_0,
|
||||
ConvProblemSize const & problem_size_1,
|
||||
TensorRefA0 const & ref_A0,
|
||||
TensorRefB0 const & ref_B0,
|
||||
TensorRefC const & ref_C0,
|
||||
TensorRefB1 const & ref_B1,
|
||||
TensorRefC const & ref_C1,
|
||||
TensorRefC const & ref_D1,
|
||||
typename EpilogueOutputOp0::Params const & output_op_0,
|
||||
typename EpilogueOutputOp1::Params const & output_op_1,
|
||||
SplitKMode const & split_k_mode = SplitKMode::kSerial
|
||||
):
|
||||
problem_size_0(problem_size_0),
|
||||
problem_size_1(problem_size_1),
|
||||
ref_A0(ref_A0),
|
||||
ref_B0(ref_B0),
|
||||
ref_C0(ref_C0),
|
||||
ref_B1(ref_B1),
|
||||
ref_C1(ref_C1),
|
||||
ref_D1(ref_D1),
|
||||
output_op_0(output_op_0),
|
||||
output_op_1(output_op_1),
|
||||
split_k_mode(split_k_mode)
|
||||
{
|
||||
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
/// Parameters structure
|
||||
struct Params {
|
||||
ConvProblemSize problem_size_0;
|
||||
ConvProblemSize problem_size_1;
|
||||
cutlass::gemm::GemmCoord grid_tiled_shape;
|
||||
gemm::GemmCoord implicit_gemm_problem_size_0;
|
||||
gemm::GemmCoord implicit_gemm_problem_size_1;
|
||||
int gemm_k_iterations_0;
|
||||
int gemm_k_iterations_1;
|
||||
typename B2bMma::IteratorA0::Params iterator_A0;
|
||||
typename B2bMma::IteratorA0::Element const *ptr_A0;
|
||||
typename B2bMma::IteratorB0::Params iterator_B0;
|
||||
typename B2bMma::IteratorB0::Element const *ptr_B0;
|
||||
typename Epilogue::OutputTileIterator::Params iterator_C0;
|
||||
typename Epilogue::OutputTileIterator::Element *ptr_C0;
|
||||
typename B2bMma::IteratorB1::Params iterator_B1;
|
||||
typename B2bMma::IteratorB1::Element const *ptr_B1;
|
||||
typename Epilogue::OutputTileIterator::Params iterator_C1;
|
||||
typename Epilogue::OutputTileIterator::Element *ptr_C1;
|
||||
typename Epilogue::OutputTileIterator::Params iterator_D1;
|
||||
typename Epilogue::OutputTileIterator::Element *ptr_D1;
|
||||
typename EpilogueOutputOp0::Params output_op_0;
|
||||
typename EpilogueOutputOp1::Params output_op_1;
|
||||
int *semaphore;
|
||||
SplitKMode split_k_mode;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(): gemm_k_iterations_0(0), gemm_k_iterations_1(0) { }
|
||||
|
||||
///
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(
|
||||
Arguments const &args,
|
||||
int *semaphore = nullptr
|
||||
):
|
||||
problem_size_0(args.problem_size_0),
|
||||
problem_size_1(args.problem_size_1),
|
||||
implicit_gemm_problem_size_0(cutlass::conv::implicit_gemm_problem_size(kConvolutionalOperator, args.problem_size_0)),
|
||||
implicit_gemm_problem_size_1(cutlass::conv::implicit_gemm_problem_size(kConvolutionalOperator, args.problem_size_1)),
|
||||
grid_tiled_shape(grid_tiled_shape),
|
||||
iterator_A0(B2bMma::IteratorA0::getParams(args.problem_size_0, args.ref_A0.layout())),
|
||||
ptr_A0(args.ref_A0.data()),
|
||||
iterator_B0(args.problem_size_0, args.ref_B0.layout()),
|
||||
ptr_B0(args.ref_B0.data()),
|
||||
iterator_C0(ConvOutputIteratorParameter::layout(args.ref_C0)),
|
||||
ptr_C0(args.ref_C0.data()),
|
||||
iterator_B1(args.problem_size_1, args.ref_B1.layout()),
|
||||
ptr_B1(args.ref_B1.data()),
|
||||
iterator_C1(ConvOutputIteratorParameter::layout(args.ref_C1)),
|
||||
ptr_C1(args.ref_C1.data()),
|
||||
iterator_D1(ConvOutputIteratorParameter::layout(args.ref_D1)),
|
||||
ptr_D1(args.ref_D1.data()),
|
||||
output_op_0(args.output_op_0),
|
||||
output_op_1(args.output_op_1),
|
||||
semaphore(semaphore),
|
||||
split_k_mode(args.split_k_mode)
|
||||
{
|
||||
gemm_k_iterations_0 = implicit_gemm_k_iterations(kConvolutionalOperator, ThreadblockShape0::kK, args.problem_size_0);
|
||||
gemm_k_iterations_1 = implicit_gemm_k_iterations(kConvolutionalOperator, ThreadblockShape1::kK, args.problem_size_1);
|
||||
|
||||
ThreadblockSwizzle threadblock_swizzle;
|
||||
|
||||
grid_tiled_shape = threadblock_swizzle.get_tiled_shape(
|
||||
implicit_gemm_problem_size_0,
|
||||
{ThreadblockShape0::kM, ThreadblockShape0::kN, ThreadblockShape0::kK},
|
||||
args.problem_size_0.split_k_slices);
|
||||
}
|
||||
};
|
||||
|
||||
/// Shared memory storage structure
|
||||
union SharedStorage {
|
||||
typename B2bMma::B2bMmaSharedStorage main_loop;
|
||||
typename Epilogue::SharedStorage epilogue;
|
||||
};
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
B2bImplicitGemmConvolution() { }
|
||||
|
||||
/// Executes one ImplicitGEMM
|
||||
CUTLASS_DEVICE
|
||||
void operator()(Params const ¶ms, SharedStorage &shared_storage) {
|
||||
|
||||
// Compute threadblock location
|
||||
ThreadblockSwizzle threadblock_swizzle;
|
||||
|
||||
cutlass::gemm::GemmCoord threadblock_tile_idx =
|
||||
threadblock_swizzle.get_tile_offset(params.grid_tiled_shape);
|
||||
|
||||
// Early exit if CTA is out of range
|
||||
if (params.grid_tiled_shape.m() <= threadblock_tile_idx.m() ||
|
||||
params.grid_tiled_shape.n() <= threadblock_tile_idx.n()) {
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
// Compute position within threadblock
|
||||
int thread_idx = threadIdx.x;
|
||||
|
||||
// Construct iterators to A and B operands
|
||||
typename B2bMma::IteratorA0 iterator_A0(
|
||||
params.iterator_A0,
|
||||
params.problem_size_0,
|
||||
params.ptr_A0,
|
||||
thread_idx,
|
||||
MatrixCoord(
|
||||
threadblock_tile_idx.m() * B2bMma::Shape0::kM,
|
||||
threadblock_tile_idx.k() * B2bMma::Shape0::kK
|
||||
)
|
||||
);
|
||||
|
||||
typename B2bMma::IteratorB0 iterator_B0(
|
||||
params.iterator_B0,
|
||||
params.problem_size_0,
|
||||
params.ptr_B0,
|
||||
thread_idx,
|
||||
MatrixCoord(
|
||||
threadblock_tile_idx.k() * B2bMma::Shape0::kK,
|
||||
threadblock_tile_idx.n() * B2bMma::Shape0::kN
|
||||
)
|
||||
);
|
||||
|
||||
typename B2bMma::IteratorB1 iterator_B1(
|
||||
params.iterator_B1,
|
||||
params.problem_size_1,
|
||||
params.ptr_B1,
|
||||
thread_idx,
|
||||
MatrixCoord(
|
||||
threadblock_tile_idx.k() * B2bMma::Shape1::kK,
|
||||
threadblock_tile_idx.n() * B2bMma::Shape1::kN
|
||||
)
|
||||
);
|
||||
|
||||
|
||||
// Broadcast the warp_id computed by lane 0 to ensure dependent code
|
||||
// is compiled as warp-uniform.
|
||||
int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
|
||||
int lane_idx = threadIdx.x % 32;
|
||||
|
||||
//
|
||||
// Main loop
|
||||
//
|
||||
|
||||
EpilogueOutputOp0 output_op_0(params.output_op_0);
|
||||
|
||||
// Construct thread-scoped matrix multiply
|
||||
B2bMma b2bMma(shared_storage.main_loop, thread_idx, warp_idx, lane_idx);
|
||||
|
||||
typename B2bMma::FragmentC0 src_accum;
|
||||
typename B2bMma::FragmentC1 accumulators;
|
||||
|
||||
src_accum.clear();
|
||||
accumulators.clear();
|
||||
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
b2bMma(params.gemm_k_iterations_0, accumulators, iterator_A0, iterator_B0, iterator_B1, src_accum, output_op_0);
|
||||
|
||||
//
|
||||
// Epilogue
|
||||
//
|
||||
|
||||
EpilogueOutputOp1 output_op_1(params.output_op_1);
|
||||
|
||||
// Construct the semaphore.
|
||||
int block_idx = threadblock_tile_idx.m() + threadblock_tile_idx.n() * params.grid_tiled_shape.m();
|
||||
|
||||
Semaphore semaphore(params.semaphore + block_idx, thread_idx);
|
||||
|
||||
// Compute logical position within grid
|
||||
threadblock_tile_idx =
|
||||
threadblock_swizzle.get_tile_offset(params.grid_tiled_shape);
|
||||
|
||||
// If performing a reduction via split-K, fetch the initial synchronization
|
||||
if (params.split_k_mode == SplitKMode::kSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
// Fetch the synchronization lock initially but do not block.
|
||||
semaphore.fetch();
|
||||
|
||||
// Indicate which position in a serial reduction the output operator is currently updating
|
||||
output_op_1.set_k_partition(threadblock_tile_idx.k(), params.grid_tiled_shape.k());
|
||||
}
|
||||
|
||||
MatrixCoord threadblock_offset(
|
||||
threadblock_tile_idx.m() * B2bMma::Shape1::kM,
|
||||
threadblock_tile_idx.n() * B2bMma::Shape1::kN
|
||||
);
|
||||
|
||||
// Tile iterator writing to destination tensor
|
||||
typename Epilogue::OutputTileIterator iterator_D1(
|
||||
params.iterator_D1,
|
||||
params.ptr_D1,
|
||||
ConvOutputIteratorParameter::extent(params.problem_size_1),
|
||||
thread_idx,
|
||||
threadblock_offset
|
||||
);
|
||||
|
||||
// Tile iterator reading from source accumulator tensor
|
||||
typename Epilogue::OutputTileIterator iterator_C1(
|
||||
params.iterator_C1,
|
||||
params.ptr_C1,
|
||||
ConvOutputIteratorParameter::extent(params.problem_size_1),
|
||||
thread_idx,
|
||||
threadblock_offset
|
||||
);
|
||||
|
||||
|
||||
// Construct the epilogue
|
||||
Epilogue epilogue(
|
||||
shared_storage.epilogue,
|
||||
thread_idx,
|
||||
warp_idx,
|
||||
lane_idx);
|
||||
|
||||
// Wait on the semaphore - this latency may have been covered by iterator construction
|
||||
if (params.split_k_mode == SplitKMode::kSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
// For subsequent threadblocks, the source matrix is held in the 'D' tensor.
|
||||
if (threadblock_tile_idx.k()) {
|
||||
iterator_C1 = iterator_D1;
|
||||
}
|
||||
|
||||
semaphore.wait(threadblock_tile_idx.k());
|
||||
|
||||
__threadfence();
|
||||
}
|
||||
// Each split-k-slice writes to a unique tensor location
|
||||
else if (params.split_k_mode == SplitKMode::kParallel) {
|
||||
iterator_D1.add_pointer_offset(threadblock_tile_idx.k() *
|
||||
cutlass::conv::implicit_gemm_tensor_c_size(ConvOperator, params.problem_size_1));
|
||||
}
|
||||
|
||||
// Run efficient epilogue
|
||||
epilogue(output_op_1, iterator_D1, accumulators, iterator_C1);
|
||||
|
||||
//
|
||||
// Release the semaphore
|
||||
//
|
||||
|
||||
if (params.split_k_mode == SplitKMode::kSerial && params.grid_tiled_shape.k() > 1) {
|
||||
|
||||
int lock = 0;
|
||||
if (params.grid_tiled_shape.k() == threadblock_tile_idx.k() + 1) {
|
||||
|
||||
// The final threadblock resets the semaphore for subsequent grids.
|
||||
lock = 0;
|
||||
}
|
||||
else {
|
||||
// Otherwise, the semaphore is incremented
|
||||
lock = threadblock_tile_idx.k() + 1;
|
||||
}
|
||||
|
||||
semaphore.release(lock);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace conv
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,442 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2021, 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
|
||||
Default kernel-level GEMM definitions combine threadblock-scoped matrix multiply-add with
|
||||
the appropriate threadblock-scoped epilogue.
|
||||
|
||||
Note, CUTLASS epilogues universally target row-major outputs. Column-major outputs are
|
||||
accommodated by exchanging A and B operands and assuming transposed layouts. Partial
|
||||
specializations here choose 'device::GemmTransposed' to implement this functionality.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/layout/matrix.h"
|
||||
#include "cutlass/numeric_types.h"
|
||||
|
||||
#include "cutlass/epilogue/threadblock/epilogue.h"
|
||||
#include "cutlass/epilogue/thread/linear_combination.h"
|
||||
|
||||
#include "cutlass/gemm/gemm.h"
|
||||
#include "cutlass/gemm/kernel/gemm_pipelined.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_sm80.h"
|
||||
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
|
||||
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_tensor_op.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h"
|
||||
#include "cutlass/epilogue/threadblock/default_epilogue_simt.h"
|
||||
|
||||
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"
|
||||
|
||||
#include "kernel/b2b_gemm.h"
|
||||
#include "threadblock/default_b2b_mma.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace gemm {
|
||||
namespace kernel {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA_,
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA_,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB_,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB_,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC_,
|
||||
/// Layout type for C and D matrix operands
|
||||
typename LayoutC_,
|
||||
/// Element type for internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Operator class tag
|
||||
typename OperatorClass,
|
||||
/// Tag indicating architecture to tune for
|
||||
typename ArchTag,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of stages used in the pipelined mainloop
|
||||
int Stages,
|
||||
/// If true, kernel is configured to support serial reduction in the epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator,
|
||||
/// Beta is zero or not
|
||||
bool IsBetaZero = false
|
||||
>
|
||||
struct DefaultB2bGemm;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Partial specialization for Ampere Architecture
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Element type for internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of stages used in the pipelined mainloop
|
||||
int Stages,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator>
|
||||
struct DefaultB2bGemm<ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementC,
|
||||
layout::RowMajor, ElementAccumulator, arch::OpClassTensorOp,
|
||||
arch::Sm80, ThreadblockShape0, ThreadblockShape1,
|
||||
WarpShape0, WarpShape1, InstructionShape,
|
||||
EpilogueOutputOp0, EpilogueOutputOp1, ThreadblockSwizzle, Stages, SplitKSerial,
|
||||
Operator> {
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
|
||||
ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, Stages, Operator, EpilogueOutputOp0>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue =
|
||||
typename cutlass::epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Partial specialization for Turing Architecture
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Layout type for A matrix operand
|
||||
typename LayoutA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Layout type for B matrix operand
|
||||
typename LayoutB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Element type for internal accumulation
|
||||
typename ElementAccumulator,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// If true, kernel is configured to support serial reduction in the epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator
|
||||
>
|
||||
struct DefaultB2bGemm<
|
||||
ElementA, LayoutA, kAlignmentA,
|
||||
ElementB, LayoutB, kAlignmentB,
|
||||
ElementC, layout::RowMajor,
|
||||
ElementAccumulator,
|
||||
arch::OpClassTensorOp,
|
||||
arch::Sm75,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
EpilogueOutputOp0,
|
||||
EpilogueOutputOp1,
|
||||
ThreadblockSwizzle,
|
||||
2,
|
||||
SplitKSerial,
|
||||
Operator
|
||||
> {
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA,
|
||||
LayoutA,
|
||||
kAlignmentA,
|
||||
ElementB,
|
||||
LayoutB,
|
||||
kAlignmentB,
|
||||
ElementAccumulator,
|
||||
layout::RowMajor,
|
||||
arch::OpClassTensorOp,
|
||||
arch::Sm75,
|
||||
ThreadblockShape0,
|
||||
ThreadblockShape1,
|
||||
WarpShape0,
|
||||
WarpShape1,
|
||||
InstructionShape,
|
||||
2,
|
||||
Operator,
|
||||
EpilogueOutputOp0
|
||||
>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::DefaultEpilogueTensorOp<
|
||||
ThreadblockShape1,
|
||||
typename B2bMma::Operator1,
|
||||
kPartitionsK1,
|
||||
EpilogueOutputOp1,
|
||||
EpilogueOutputOp1::kCount
|
||||
>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
|
||||
/// Partial specialization for Ampere Integer Matrix Multiply Interleaved layout
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of stages used in the pipelined mainloop
|
||||
int Stages,
|
||||
/// Number of Interleaved k
|
||||
int InterleavedK,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator,
|
||||
/// Is Beta zero or not
|
||||
bool IsBetaZero>
|
||||
struct DefaultB2bGemm<
|
||||
ElementA, layout::ColumnMajorInterleaved<InterleavedK>, kAlignmentA,
|
||||
ElementB, layout::RowMajorInterleaved<InterleavedK>, kAlignmentB,
|
||||
ElementC, layout::ColumnMajorInterleaved<InterleavedK>, int32_t,
|
||||
arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, EpilogueOutputOp0, EpilogueOutputOp1,
|
||||
ThreadblockSwizzle, Stages,
|
||||
SplitKSerial, Operator, IsBetaZero> {
|
||||
using LayoutA = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
using LayoutB = layout::RowMajorInterleaved<InterleavedK>;
|
||||
using LayoutC = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
|
||||
using ElementAccumulator = int32_t;
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
|
||||
ElementAccumulator, LayoutC, arch::OpClassTensorOp, arch::Sm80,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, Stages, Operator, EpilogueOutputOp0,
|
||||
true>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::
|
||||
DefaultInterleavedEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
64 / sizeof_bits<ElementC>::value, InterleavedK,
|
||||
IsBetaZero>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
|
||||
/// Partial specialization for Turing Integer Tensor Core Interleaved layout
|
||||
template <
|
||||
/// Element type for A matrix operand
|
||||
typename ElementA,
|
||||
/// Access granularity of A matrix in units of elements
|
||||
int kAlignmentA,
|
||||
/// Element type for B matrix operand
|
||||
typename ElementB,
|
||||
/// Access granularity of B matrix in units of elements
|
||||
int kAlignmentB,
|
||||
/// Element type for C and D matrix operands
|
||||
typename ElementC,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape0,
|
||||
/// Threadblock-level tile size (concept: GemmShape)
|
||||
typename ThreadblockShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape0,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename WarpShape1,
|
||||
/// Warp-level tile size (concept: GemmShape)
|
||||
typename InstructionShape,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp0,
|
||||
/// Epilogue output operator
|
||||
typename EpilogueOutputOp1,
|
||||
/// Threadblock-level swizzling operator
|
||||
typename ThreadblockSwizzle,
|
||||
/// Number of Interleaved k
|
||||
int InterleavedK,
|
||||
/// If true, kernel is configured to support serial reduction in the
|
||||
/// epilogue
|
||||
bool SplitKSerial,
|
||||
/// Operation performed by GEMM
|
||||
typename Operator,
|
||||
/// Is Beta zero or not
|
||||
bool IsBetaZero>
|
||||
struct DefaultB2bGemm<ElementA, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
kAlignmentA, ElementB,
|
||||
layout::RowMajorInterleaved<InterleavedK>, kAlignmentB,
|
||||
ElementC, layout::ColumnMajorInterleaved<InterleavedK>,
|
||||
int32_t, arch::OpClassTensorOp, arch::Sm75,
|
||||
ThreadblockShape0, ThreadblockShape1, WarpShape0, WarpShape1,
|
||||
InstructionShape, EpilogueOutputOp0, EpilogueOutputOp1,
|
||||
ThreadblockSwizzle, 2, SplitKSerial, Operator, IsBetaZero> {
|
||||
using LayoutA = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
using LayoutB = layout::RowMajorInterleaved<InterleavedK>;
|
||||
using LayoutC = layout::ColumnMajorInterleaved<InterleavedK>;
|
||||
|
||||
using ElementAccumulator = int32_t;
|
||||
|
||||
/// Define the threadblock-scoped matrix multiply-accumulate
|
||||
using B2bMma = typename cutlass::gemm::threadblock::DefaultB2bMma<
|
||||
ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementAccumulator, LayoutC,
|
||||
arch::OpClassTensorOp, arch::Sm75, ThreadblockShape0, ThreadblockShape1,
|
||||
WarpShape0, WarpShape1, InstructionShape, 2, Operator, EpilogueOutputOp0, true>::ThreadblockB2bMma;
|
||||
|
||||
static const int kPartitionsK1 = ThreadblockShape1::kK / WarpShape1::kK;
|
||||
|
||||
/// Define the epilogue for the 2nd Gemm
|
||||
using Epilogue = typename cutlass::epilogue::threadblock::
|
||||
DefaultInterleavedEpilogueTensorOp<
|
||||
ThreadblockShape1, typename B2bMma::Operator1, kPartitionsK1, EpilogueOutputOp1,
|
||||
64 / sizeof_bits<ElementC>::value, InterleavedK,
|
||||
IsBetaZero>::Epilogue;
|
||||
|
||||
/// Define the kernel-level GEMM operator.
|
||||
using B2bGemmKernel = kernel::B2bGemm<B2bMma, Epilogue, ThreadblockSwizzle, SplitKSerial>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace gemm
|
||||
} // namespace cutlass
|
||||
Reference in New Issue
Block a user