CUTLASS 2.9 (#468)
This commit is contained in:
Andrew Kerr
@@ -0,0 +1,36 @@
|
||||
|
||||
# Copyright (c) 2017 - 2022 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.
|
||||
|
||||
|
||||
|
||||
cutlass_example_add_executable(
|
||||
35_gemm_softmax
|
||||
gemm_softmax.cu
|
||||
)
|
||||
|
||||
@@ -0,0 +1,407 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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 Generic epilogue for implementing certain kinds of fused epilogue behavior.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/fast_math.h"
|
||||
#include "cutlass/matrix_coord.h"
|
||||
#include "cutlass/semaphore.h"
|
||||
#include "cutlass/epilogue/threadblock/epilogue_base.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace epilogue {
|
||||
namespace threadblock {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
class EpilogueFusedVisitorConcept {
|
||||
public:
|
||||
|
||||
static int const kIterations = 1;
|
||||
static int const kElementsPerAccess = 4;
|
||||
using ElementOutput = float;
|
||||
using ElementAccumulator = float;
|
||||
using AccumulatorFragment = Array<ElementAccumulator, kElementsPerAccess>;
|
||||
|
||||
/// Arguments structure
|
||||
struct Arguments { };
|
||||
|
||||
/// Params structure
|
||||
struct Params {
|
||||
|
||||
Params() { }
|
||||
Params(Arguments const &args) { }
|
||||
};
|
||||
|
||||
/// Shared storage
|
||||
struct SharedStorage { };
|
||||
|
||||
public:
|
||||
|
||||
CUTLASS_DEVICE
|
||||
EpilogueFusedVisitorConcept(
|
||||
Params const ¶ms, ///< Parameters routed to the epilogue
|
||||
SharedStorage &shared_storage, ///< Shared storage needed by the functors here
|
||||
MatrixCoord const &problem_size, ///< Problem size of the output
|
||||
int thread_idx, ///< Thread index within the threadblock
|
||||
int warp_idx, ///< Warp index within the threadblock
|
||||
int lane_idx, ///< Lane index within the warp
|
||||
MatrixCoord const &threadblock_offset = MatrixCoord(0, 0)) { ///< Coordinate
|
||||
|
||||
}
|
||||
|
||||
/// Helper to indicate split-K behavior
|
||||
CUTLASS_DEVICE
|
||||
void set_k_partition(
|
||||
int split_k_index, ///< Index of this threadblock within split-K partitioned scheme
|
||||
int split_k_slices) { ///< Total number of split-K slices
|
||||
|
||||
}
|
||||
|
||||
/// Called to set the batch index
|
||||
CUTLASS_DEVICE
|
||||
void set_batch_index(int batch_idx) {
|
||||
|
||||
}
|
||||
|
||||
/// Called at the start of the epilogue just before iterating over accumulator slices
|
||||
CUTLASS_DEVICE
|
||||
void begin_epilogue() {
|
||||
|
||||
}
|
||||
|
||||
/// Called at the start of one step before starting accumulator exchange
|
||||
CUTLASS_DEVICE
|
||||
void begin_step(int step_idx) {
|
||||
|
||||
}
|
||||
|
||||
/// Called at the start of a row
|
||||
CUTLASS_DEVICE
|
||||
void begin_row(int row_idx) {
|
||||
|
||||
}
|
||||
|
||||
/// Called after accumulators have been exchanged for each accumulator vector
|
||||
CUTLASS_DEVICE
|
||||
void visit(
|
||||
int row_idx,
|
||||
int column_idx,
|
||||
int frag_idx,
|
||||
AccumulatorFragment const &accum) {
|
||||
|
||||
}
|
||||
|
||||
/// Called at the start of a row
|
||||
CUTLASS_DEVICE
|
||||
void end_row(int row_idx) {
|
||||
|
||||
}
|
||||
|
||||
/// Called after all accumulator elements have been visited
|
||||
CUTLASS_DEVICE
|
||||
void end_step(int step_idx) {
|
||||
|
||||
}
|
||||
|
||||
/// Called after all steps have been completed
|
||||
CUTLASS_DEVICE
|
||||
void end_epilogue() {
|
||||
|
||||
}
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Epilogue operator
|
||||
template <
|
||||
typename Visitor_, ///< Functor containing fused operations (satisfies EpilogueFusedVisitorConcept)
|
||||
typename Shape_, ///< Shape of threadblock tile (concept: GemmShape)
|
||||
typename WarpMmaOperator_, ///< Warp-level MMA operator (concept: gemm::warp::MmaTensorOp)
|
||||
int PartitionsK, ///< Number of partitions of the K dimension
|
||||
typename AccumulatorFragmentIterator_, ///< Fragment iterator selecting accumulators
|
||||
typename WarpTileIterator_, ///< Warp-scoped tile iterator writing accumulators to SMEM
|
||||
typename SharedLoadIterator_, ///< Threadblock-scoped tile iterator loading from SMEM
|
||||
typename Padding_, ///< Padding added to SMEM allocation to avoid bank conflicts (concept: MatrixShape)
|
||||
int FragmentsPerPartition = 1, ///< Used to coarsten the epilogue granularity
|
||||
int IterationsUnroll = ///< Used to reduce binary size when epilogue op is large
|
||||
(true || !IsEpilogueFunctorHeavy<Visitor_>::value)
|
||||
>
|
||||
class EpilogueWithVisitor :
|
||||
public EpilogueBase<
|
||||
Shape_,
|
||||
typename WarpMmaOperator_::Shape,
|
||||
PartitionsK,
|
||||
AccumulatorFragmentIterator_,
|
||||
WarpTileIterator_,
|
||||
Padding_,
|
||||
FragmentsPerPartition> {
|
||||
|
||||
public:
|
||||
|
||||
using Visitor = Visitor_;
|
||||
|
||||
using Base = EpilogueBase<
|
||||
Shape_,
|
||||
typename WarpMmaOperator_::Shape,
|
||||
PartitionsK,
|
||||
AccumulatorFragmentIterator_,
|
||||
WarpTileIterator_,
|
||||
Padding_,
|
||||
FragmentsPerPartition>;
|
||||
|
||||
using Shape = Shape_;
|
||||
using WarpMmaOperator = WarpMmaOperator_;
|
||||
static int const kPartitionsK = PartitionsK;
|
||||
|
||||
using AccumulatorFragmentIterator = AccumulatorFragmentIterator_;
|
||||
using WarpTileIterator = WarpTileIterator_;
|
||||
using SharedLoadIterator = SharedLoadIterator_;
|
||||
using Padding = Padding_;
|
||||
|
||||
using Layout = layout::RowMajor;
|
||||
using LongIndex = typename Layout::LongIndex;
|
||||
|
||||
/// The complete warp-level accumulator tile
|
||||
using AccumulatorTile = typename Base::AccumulatorTile;
|
||||
|
||||
/// Accumulator element
|
||||
using ElementAccumulator = typename WarpTileIterator::Element;
|
||||
|
||||
/// Output access size
|
||||
static int const kElementsPerAccess = Visitor::kElementsPerAccess;
|
||||
|
||||
/// Tensor reference to sync tensor
|
||||
using SyncTensorRef = typename cutlass::TensorRef<int, cutlass::layout::PackedVectorLayout>;
|
||||
|
||||
/// Array type used by output functor
|
||||
using AccumulatorAccessType = Array<
|
||||
typename WarpTileIterator::Element, kElementsPerAccess>;
|
||||
|
||||
/// Number of warps
|
||||
using WarpCount = typename Base::WarpCount;
|
||||
|
||||
static int constexpr kSmemTiles = Base::kFragmentsPerIteration > 1 ? Base::kFragmentsPerIteration : kPartitionsK;
|
||||
static int constexpr kSmemPointerOffset = Base::SharedStorage::StorageShape::kCount / kSmemTiles;
|
||||
|
||||
using SharedStorage = typename Base::SharedStorage;
|
||||
|
||||
private:
|
||||
|
||||
/// Loads fragment from shared memory aligned with output tensor
|
||||
SharedLoadIterator shared_load_iterator_;
|
||||
|
||||
public:
|
||||
|
||||
/// Constructor
|
||||
CUTLASS_DEVICE
|
||||
EpilogueWithVisitor(
|
||||
SharedStorage &shared_storage, ///< Shared storage object
|
||||
int thread_idx, ///< ID of a thread within the threadblock
|
||||
int warp_idx, ///< ID of warp within threadblock
|
||||
int lane_idx ///< Id of thread within warp
|
||||
):
|
||||
Base(shared_storage, thread_idx, warp_idx, lane_idx),
|
||||
shared_load_iterator_(shared_storage.reference(), thread_idx)
|
||||
{
|
||||
|
||||
}
|
||||
|
||||
/// Streams the result to global memory
|
||||
CUTLASS_DEVICE
|
||||
void operator()(
|
||||
Visitor & visitor,
|
||||
AccumulatorTile const &accumulators) { ///< Threadblock tile coordinate in GEMM (in units of threadblock tiles)
|
||||
|
||||
visitor.begin_epilogue();
|
||||
|
||||
//
|
||||
// Iterator over warp-level accumulator fragment
|
||||
//
|
||||
|
||||
AccumulatorFragmentIterator accum_fragment_iterator(accumulators);
|
||||
|
||||
//
|
||||
// Iterate over accumulator tile
|
||||
//
|
||||
|
||||
#pragma unroll(IterationsUnroll ? Visitor::kIterations : 1)
|
||||
for (int iter_idx = 0; iter_idx < Visitor::kIterations; ++iter_idx) {
|
||||
|
||||
//
|
||||
// Load the source
|
||||
//
|
||||
|
||||
visitor.begin_step(iter_idx);
|
||||
|
||||
//
|
||||
// Convert and store fragment
|
||||
//
|
||||
|
||||
__syncthreads();
|
||||
|
||||
acc2smem_source_needed<cutlass::make_index_sequence<Visitor::kIterations>>::push(
|
||||
iter_idx, accum_fragment_iterator, this->warp_tile_iterator_);
|
||||
|
||||
__syncthreads();
|
||||
|
||||
//
|
||||
// Load fragments from shared memory
|
||||
//
|
||||
|
||||
typename SharedLoadIterator::Fragment aligned_accum_fragment[kPartitionsK];
|
||||
|
||||
shared_load_iterator_.load(aligned_accum_fragment[0]);
|
||||
|
||||
// If the number of k-slices is > 1 - perform a reduction amongst the k-slices
|
||||
if (kPartitionsK > 1) {
|
||||
|
||||
plus <typename SharedLoadIterator::Fragment> add_fragments;
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for ( int i = 1; i < kPartitionsK; ++i) {
|
||||
shared_load_iterator_.add_pointer_offset(kSmemPointerOffset);
|
||||
shared_load_iterator_.load(aligned_accum_fragment[i]);
|
||||
aligned_accum_fragment[0] = add_fragments(aligned_accum_fragment[0], aligned_accum_fragment[i]);
|
||||
}
|
||||
|
||||
shared_load_iterator_.add_pointer_offset((1 - kPartitionsK) * kSmemPointerOffset);
|
||||
}
|
||||
|
||||
//
|
||||
// Iterate over output fragments
|
||||
//
|
||||
|
||||
AccumulatorAccessType const *accum_frag_ptr =
|
||||
reinterpret_cast<AccumulatorAccessType const *>(&aligned_accum_fragment[0]);
|
||||
|
||||
int const kAccumulatorFragmentCount = AccumulatorTile::kElements / (Visitor::kIterations * AccumulatorAccessType::kElements);
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int idx = 0; idx < kAccumulatorFragmentCount; ++idx) {
|
||||
|
||||
int row_idx = idx / SharedLoadIterator::ThreadMap::Iterations::kColumn;
|
||||
int col_idx = idx % SharedLoadIterator::ThreadMap::Iterations::kColumn;
|
||||
|
||||
// Start a new row of the output fragment
|
||||
if (!col_idx) {
|
||||
visitor.begin_row(row_idx);
|
||||
}
|
||||
|
||||
visitor.visit(
|
||||
row_idx,
|
||||
col_idx,
|
||||
idx,
|
||||
accum_frag_ptr[idx]
|
||||
);
|
||||
|
||||
// End the row of the output fragment
|
||||
if (col_idx + 1 == SharedLoadIterator::ThreadMap::Iterations::kColumn) {
|
||||
visitor.end_row(row_idx);
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Conclude the step
|
||||
//
|
||||
|
||||
visitor.end_step(iter_idx);
|
||||
}
|
||||
|
||||
visitor.end_epilogue();
|
||||
}
|
||||
|
||||
private:
|
||||
|
||||
|
||||
template<class Seq>
|
||||
struct acc2smem_source_needed;
|
||||
|
||||
template <size_t... Seq>
|
||||
struct acc2smem_source_needed<cutlass::index_sequence<Seq...>> {
|
||||
template<int Advance>
|
||||
CUTLASS_DEVICE
|
||||
static void helper(AccumulatorFragmentIterator accum_fragment_iterator,
|
||||
WarpTileIterator &warp_tile_iterator) {
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < Advance; i++) {
|
||||
++accum_fragment_iterator;
|
||||
}
|
||||
|
||||
typename AccumulatorFragmentIterator::Fragment accum_fragment;
|
||||
accum_fragment_iterator.load(accum_fragment);
|
||||
warp_tile_iterator.store(accum_fragment);
|
||||
}
|
||||
|
||||
CUTLASS_DEVICE
|
||||
static void push(size_t pos,
|
||||
AccumulatorFragmentIterator const &iterator_begin,
|
||||
WarpTileIterator &warp_tile_iterator) {
|
||||
int dummy[] = {(pos == Seq) && (helper<Seq>(iterator_begin, warp_tile_iterator), 0)...};
|
||||
}
|
||||
};
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Helper to create an EpilogueWithVisitor from an existing epilogue
|
||||
template <typename Visitor_, typename Existing_, bool IterationsUnroll = true>
|
||||
struct EpilogueWithVisitorFromExistingEpilogue {
|
||||
|
||||
using Epilogue = EpilogueWithVisitor<
|
||||
Visitor_,
|
||||
typename Existing_::Shape,
|
||||
typename Existing_::WarpMmaOperator,
|
||||
Existing_::kPartitionsK,
|
||||
typename Existing_::AccumulatorFragmentIterator,
|
||||
typename Existing_::WarpTileIterator,
|
||||
typename Existing_::SharedLoadIterator,
|
||||
typename Existing_::Padding,
|
||||
Existing_::kFragmentsPerIteration,
|
||||
IterationsUnroll
|
||||
>;
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
} // namespace epilogue
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -0,0 +1,681 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
/**
|
||||
|
||||
*/
|
||||
|
||||
#include <cmath>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
#include <limits>
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/arch/memory.h"
|
||||
#include "cutlass/arch/memory_sm75.h"
|
||||
#include "cutlass/gemm/device/gemm_complex.h"
|
||||
|
||||
#include "cutlass/util/command_line.h"
|
||||
#include "cutlass/util/host_tensor.h"
|
||||
|
||||
#include "cutlass/util/reference/host/gemm_complex.h"
|
||||
#include "cutlass/util/reference/host/tensor_reduce.h"
|
||||
#include "cutlass/util/reference/host/tensor_compare.h"
|
||||
#include "cutlass/util/reference/host/tensor_norm.h"
|
||||
#include "cutlass/util/reference/host/tensor_copy.h"
|
||||
#include "cutlass/util/reference/host/tensor_fill.h"
|
||||
#include "cutlass/util/reference/host/error_metrics.h"
|
||||
#include "cutlass/util/tensor_view_io.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#include "gemm_with_softmax.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
#define TRACE(x) { std::cout << "gemm_softmax.cu:" << __LINE__ << " " << x << std::endl; }
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
enum class Disposition {
|
||||
kPassed,
|
||||
kIncorrect,
|
||||
kNotVerified
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// Command line options parsing
|
||||
struct Options {
|
||||
|
||||
bool help;
|
||||
cutlass::gemm::GemmCoord problem_size;
|
||||
int batch_count;
|
||||
int iterations;
|
||||
unsigned seed;
|
||||
float alpha;
|
||||
float beta;
|
||||
bool verification_enabled;
|
||||
double tolerance;
|
||||
|
||||
Options():
|
||||
help(false),
|
||||
problem_size({16, 24, 64}),
|
||||
batch_count(1), // As a temporary limitation to the test bench, batch count must be 1. The kernels support arbitrary batching.
|
||||
iterations(20),
|
||||
seed(2022),
|
||||
alpha(1),
|
||||
beta(),
|
||||
verification_enabled(true),
|
||||
tolerance(0.01)
|
||||
{ }
|
||||
|
||||
bool valid() {
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// Parses the command line
|
||||
void parse(int argc, char const **args) {
|
||||
cutlass::CommandLine cmd(argc, args);
|
||||
|
||||
if (cmd.check_cmd_line_flag("help")) {
|
||||
help = true;
|
||||
}
|
||||
|
||||
cmd.get_cmd_line_argument("m", problem_size.m());
|
||||
cmd.get_cmd_line_argument("n", problem_size.n());
|
||||
cmd.get_cmd_line_argument("k", problem_size.k());
|
||||
|
||||
cmd.get_cmd_line_argument("alpha", alpha);
|
||||
cmd.get_cmd_line_argument("beta", beta);
|
||||
|
||||
cmd.get_cmd_line_argument("iterations", iterations);
|
||||
cmd.get_cmd_line_argument("verify", verification_enabled);
|
||||
cmd.get_cmd_line_argument("seed", seed);
|
||||
cmd.get_cmd_line_argument("tolerance", tolerance);
|
||||
}
|
||||
|
||||
/// Prints the usage statement.
|
||||
std::ostream & print_usage(std::ostream &out) const {
|
||||
|
||||
out << "35_gemm_softmax example\n\n"
|
||||
<< " This example uses the CUTLASS Library to compute GEMM + Softmax for arbitrary problem sizes.\n\n"
|
||||
<< "Options:\n\n"
|
||||
<< " --help If specified, displays this usage statement.\n\n"
|
||||
<< " --m=<int> GEMM M dimension\n"
|
||||
<< " --n=<int> GEMM N dimension\n"
|
||||
<< " --k=<int> GEMM K dimension\n"
|
||||
<< " --alpha=<f32> Epilogue scalar alpha\n"
|
||||
<< " --beta=<f32> Epilogue scalar beta\n\n"
|
||||
<< " --seed=<int> Random number seed (1*)\n\n"
|
||||
<< " --iterations=<int> Number of profiling iterations to perform (0 to disable profiling).\n\n"
|
||||
<< " --verify=<bool> If true, performs reference calculation.\n\n"
|
||||
<< " --tolerance <float> Error tolerance\n"
|
||||
;
|
||||
|
||||
out << "\n\nExamples:\n\n"
|
||||
<< "$ ./examples/35_gemm_softmax/35_gemm_softmax --m=1024 --n=512 \\\n"
|
||||
<< " --alpha=2 --beta=0.707 \n\n";
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
/// Returns true if the environment and Toolkit support this
|
||||
bool supported(bool verbose = true) const {
|
||||
|
||||
// Ampere Tensor Core operations exposed with mma.sync and ldmatrix are first available
|
||||
// in CUDA 11.0.
|
||||
//
|
||||
// CUTLASS must be compiled with CUDA 11.0 Toolkit to run these examples.
|
||||
if (!(__CUDACC_VER_MAJOR__ >= 11)) {
|
||||
if (verbose) {
|
||||
std::cerr << "Ampere Tensor Core operations must be compiled with CUDA 11.0 Toolkit or later." << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
cudaDeviceProp props;
|
||||
|
||||
cudaError_t error = cudaGetDeviceProperties(&props, 0);
|
||||
if (error != cudaSuccess) {
|
||||
if (verbose) {
|
||||
std::cerr << "cudaGetDeviceProperties() returned an error: " << cudaGetErrorString(error) << std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!((props.major * 10 + props.minor) >= 80)) {
|
||||
if (verbose) {
|
||||
std::cerr << "Ampere Tensor Core operations must be run on a machine with compute capability at least 80."
|
||||
<< std::endl;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
struct Testbed {
|
||||
|
||||
//
|
||||
// Type definitions
|
||||
//
|
||||
|
||||
|
||||
using ElementA = cutlass::half_t;
|
||||
using ElementB = cutlass::half_t;
|
||||
using ElementC = cutlass::half_t;
|
||||
using ElementD = cutlass::half_t;
|
||||
using ElementCompute = float;
|
||||
using ElementSoftmax = cutlass::half_t;
|
||||
|
||||
using LayoutA = cutlass::layout::RowMajor;
|
||||
using LayoutB = cutlass::layout::ColumnMajor;
|
||||
|
||||
using GemmSoftmax = cutlass::GemmSoftmax<
|
||||
ElementA, LayoutA,
|
||||
ElementB, LayoutB,
|
||||
ElementC,
|
||||
ElementCompute
|
||||
>;
|
||||
|
||||
using ElementN = typename GemmSoftmax::ElementN;
|
||||
using LayoutC = typename GemmSoftmax::LayoutC;
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
Options const &options;
|
||||
|
||||
cutlass::HostTensor<ElementA, LayoutA> tensor_A;
|
||||
cutlass::HostTensor<ElementB, LayoutB> tensor_B;
|
||||
cutlass::HostTensor<ElementC, LayoutC> tensor_C;
|
||||
cutlass::HostTensor<ElementD, LayoutC> tensor_D;
|
||||
cutlass::HostTensor<ElementN, LayoutC> tensor_N;
|
||||
cutlass::HostTensor<ElementSoftmax, LayoutC> tensor_Softmax;
|
||||
|
||||
cutlass::HostTensor<ElementD, LayoutC> reference_D;
|
||||
cutlass::HostTensor<ElementN, LayoutC> reference_N;
|
||||
cutlass::HostTensor<ElementSoftmax, LayoutC> reference_Softmax;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
Testbed(
|
||||
Options const &options_
|
||||
):
|
||||
options(options_)
|
||||
{
|
||||
|
||||
tensor_A.reset({options.problem_size.m(), options.problem_size.k()});
|
||||
tensor_B.reset({options.problem_size.k(), options.problem_size.n()});
|
||||
|
||||
tensor_C.reset({options.problem_size.m(), options.problem_size.n()});
|
||||
tensor_D.reset({options.problem_size.m(), options.problem_size.n()});
|
||||
|
||||
tensor_N.reset({options.problem_size.m(), 1});
|
||||
tensor_Softmax.reset({options.problem_size.m(), options.problem_size.n()});
|
||||
|
||||
reference_D.reset({options.problem_size.m(), options.problem_size.n()}, false);
|
||||
reference_N.reset({options.problem_size.m(), 1}, false);
|
||||
reference_Softmax.reset({options.problem_size.m(), options.problem_size.n()}, false);
|
||||
}
|
||||
|
||||
/// Run
|
||||
Disposition run() {
|
||||
|
||||
Disposition disposition = Disposition::kNotVerified;
|
||||
|
||||
//
|
||||
// Initialize the workspace
|
||||
//
|
||||
|
||||
initialize();
|
||||
|
||||
//
|
||||
// Launch device kernel
|
||||
//
|
||||
cutlass::Status status = cutlass::Status::kSuccess;
|
||||
|
||||
status = execute_device_kernel();
|
||||
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
std::cerr << "Device execution failed." << std::endl;
|
||||
return disposition;
|
||||
}
|
||||
|
||||
cudaError_t result = cudaDeviceSynchronize();
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "Device synchronize failed with error "
|
||||
<< cudaGetErrorString(result) << std::endl;
|
||||
return disposition;
|
||||
}
|
||||
|
||||
//
|
||||
// Compute the reference
|
||||
//
|
||||
compute_reference();
|
||||
|
||||
//
|
||||
// Verify
|
||||
//
|
||||
|
||||
if (options.verification_enabled) {
|
||||
|
||||
bool passed = verify();
|
||||
|
||||
if (passed) {
|
||||
disposition = Disposition::kPassed;
|
||||
}
|
||||
else {
|
||||
disposition = Disposition::kIncorrect;
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Profiling
|
||||
//
|
||||
if (options.iterations) {
|
||||
profile();
|
||||
}
|
||||
|
||||
return disposition;
|
||||
}
|
||||
|
||||
/// Random initialization
|
||||
void initialize() {
|
||||
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
tensor_A.host_view(),
|
||||
options.seed,
|
||||
ElementD(5),
|
||||
ElementD(-5),
|
||||
0
|
||||
);
|
||||
|
||||
cutlass::reference::host::TensorFillRandomUniform(
|
||||
tensor_B.host_view(),
|
||||
options.seed + 19,
|
||||
ElementD(5),
|
||||
ElementD(-5),
|
||||
0
|
||||
);
|
||||
|
||||
cutlass::reference::host::TensorFill(
|
||||
reference_D.host_view(),
|
||||
ElementD()
|
||||
);
|
||||
|
||||
cutlass::reference::host::TensorFill(
|
||||
reference_N.host_view(),
|
||||
ElementN()
|
||||
);
|
||||
|
||||
cutlass::reference::host::TensorFill(
|
||||
reference_Softmax.host_view(),
|
||||
ElementSoftmax()
|
||||
);
|
||||
|
||||
tensor_A.sync_device();
|
||||
tensor_B.sync_device();
|
||||
tensor_D.sync_device();
|
||||
tensor_N.sync_device();
|
||||
tensor_Softmax.sync_device();
|
||||
}
|
||||
|
||||
cutlass::Status execute_device_kernel() {
|
||||
|
||||
cutlass::Status status = cutlass::Status::kSuccess;
|
||||
|
||||
//
|
||||
// Setup arguments
|
||||
//
|
||||
|
||||
GemmSoftmax::Arguments args(
|
||||
options.problem_size,
|
||||
options.batch_count,
|
||||
tensor_A.device_ref(),
|
||||
tensor_B.device_ref(),
|
||||
tensor_C.device_ref(),
|
||||
tensor_D.device_ref(),
|
||||
{
|
||||
ElementCompute(options.alpha),
|
||||
ElementCompute(options.beta)
|
||||
},
|
||||
tensor_N.device_ref(),
|
||||
tensor_Softmax.device_ref()
|
||||
);
|
||||
|
||||
//
|
||||
// Launch
|
||||
//
|
||||
|
||||
GemmSoftmax gemm_softmax;
|
||||
|
||||
// Initialize
|
||||
status = gemm_softmax.initialize(args);
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
return status;
|
||||
}
|
||||
|
||||
// Run
|
||||
status = gemm_softmax();
|
||||
|
||||
return status;
|
||||
}
|
||||
|
||||
/// Reference calculation
|
||||
void compute_reference() {
|
||||
|
||||
// Compute GEMM
|
||||
|
||||
cutlass::reference::host::GemmComplex(
|
||||
options.problem_size,
|
||||
options.alpha,
|
||||
tensor_A.host_ref(),
|
||||
cutlass::ComplexTransform::kNone,
|
||||
tensor_B.host_ref(),
|
||||
cutlass::ComplexTransform::kNone,
|
||||
options.beta,
|
||||
tensor_C.host_ref(),
|
||||
reference_D.host_ref(),
|
||||
double()
|
||||
);
|
||||
|
||||
// Compute the norm
|
||||
for (int m = 0; m < options.problem_size.m(); ++m) {
|
||||
reference_N.at({m, 0}) = reference_D.at({m, 0});
|
||||
for (int n = 1; n < options.problem_size.n(); ++n) {
|
||||
reference_N.at({m, 0}) = std::max(reference_N.at({m, 0}), ElementN(reference_D.at({m, n})));
|
||||
}
|
||||
}
|
||||
|
||||
// Compute softmax
|
||||
for (int m = 0; m < options.problem_size.m(); ++m) {
|
||||
|
||||
float sum = float();
|
||||
|
||||
for (int n = 0; n < options.problem_size.n(); ++n) {
|
||||
sum += std::exp( float(reference_D.at({m, n})) - float(reference_N.at({m, 0})) );
|
||||
}
|
||||
|
||||
float inv_sum = float(1.0f / sum);
|
||||
|
||||
for (int n = 0; n < options.problem_size.n(); ++n) {
|
||||
|
||||
reference_Softmax.at({m, n}) = ElementSoftmax(
|
||||
std::exp( float(reference_D.at({m, n})) - float(reference_N.at({m, 0})) ) * inv_sum
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Emits all tensor values
|
||||
void emit_results() {
|
||||
std::cout << "D = \n" << tensor_D.host_view() << "\n\n";
|
||||
std::cout << "N = \n" << tensor_N.host_view() << "\n\n";
|
||||
std::cout << "Softmax = \n" << tensor_Softmax.host_view() << "\n\n";
|
||||
std::cout << "Reference N = \n" << reference_N.host_view() << "\n\n";
|
||||
std::cout << "Reference D = \n" << reference_D.host_view() << "\n\n";
|
||||
std::cout << "Reference Softmax = \n" << reference_Softmax.host_view() << "\n\n";
|
||||
}
|
||||
|
||||
/// Verifies the reference matches
|
||||
bool verify() {
|
||||
|
||||
tensor_D.sync_host();
|
||||
tensor_N.sync_host();
|
||||
tensor_Softmax.sync_host();
|
||||
|
||||
double const kThreshold = options.tolerance;
|
||||
|
||||
// Verification checks - set any of these to 'true' to override the verification checks.
|
||||
bool verified_D = false;
|
||||
bool verified_N = false;
|
||||
bool verified_Softmax = false;
|
||||
|
||||
// Verify softmax output
|
||||
if (!verified_D) {
|
||||
|
||||
double norm_diff = cutlass::reference::host::TensorNormDiff(
|
||||
tensor_D.host_view(),
|
||||
reference_D.host_view());
|
||||
|
||||
double norm_reference = cutlass::reference::host::TensorNorm(
|
||||
reference_D.host_view());
|
||||
|
||||
double rel_error = norm_diff / norm_reference;
|
||||
|
||||
if (rel_error > kThreshold) {
|
||||
std::cerr << "\n\nTensor D Relative error: " << rel_error << std::endl;
|
||||
}
|
||||
else {
|
||||
verified_D = true;
|
||||
}
|
||||
}
|
||||
|
||||
if (!verified_N) {
|
||||
|
||||
double norm_diff = cutlass::reference::host::TensorNormDiff(
|
||||
tensor_N.host_view(),
|
||||
reference_N.host_view());
|
||||
|
||||
double norm_reference = cutlass::reference::host::TensorNorm(
|
||||
reference_N.host_view());
|
||||
|
||||
double rel_error = norm_diff / norm_reference;
|
||||
|
||||
if (rel_error > kThreshold) {
|
||||
std::cerr << "\n\nTensor N Relative error: " << rel_error << std::endl;
|
||||
}
|
||||
else {
|
||||
verified_N = true;
|
||||
}
|
||||
}
|
||||
|
||||
if (!verified_Softmax) {
|
||||
|
||||
double norm_diff = cutlass::reference::host::TensorNormDiff(
|
||||
tensor_Softmax.host_view(),
|
||||
reference_Softmax.host_view());
|
||||
|
||||
double norm_reference = cutlass::reference::host::TensorNorm(
|
||||
reference_Softmax.host_view());
|
||||
|
||||
double rel_error = norm_diff / norm_reference;
|
||||
|
||||
if (rel_error > kThreshold) {
|
||||
std::cerr << "\n\nSoftmax Relative error: " << rel_error << std::endl;
|
||||
}
|
||||
else {
|
||||
verified_Softmax = true;
|
||||
}
|
||||
}
|
||||
|
||||
if (!verified_D || !verified_N || !verified_Softmax) {
|
||||
|
||||
std::cerr << "Verification check failed for tensor Softmax" << std::endl;
|
||||
|
||||
emit_results();
|
||||
|
||||
// Summarize which checks failed
|
||||
if (!verified_D) {
|
||||
std::cerr << "Verification of D tensor failed\n";
|
||||
}
|
||||
|
||||
if (!verified_N) {
|
||||
std::cerr << "Verification of N tensor failed\n";
|
||||
}
|
||||
|
||||
if (!verified_Softmax) {
|
||||
std::cerr << "Verification of Softmax tensor failed\n";
|
||||
}
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
/// Profiles
|
||||
bool profile() {
|
||||
|
||||
//
|
||||
// Profile
|
||||
//
|
||||
|
||||
cutlass::Status status = cutlass::Status::kSuccess;
|
||||
cudaError_t result;
|
||||
cudaEvent_t events[2];
|
||||
int const kIterations = options.iterations;
|
||||
|
||||
for (cudaEvent_t &evt : events) {
|
||||
result = cudaEventCreate(&evt);
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaEventCreate failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
result = cudaEventRecord(events[0]);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaEventRecord() failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
for (int iter = 0; iter < kIterations; ++iter) {
|
||||
|
||||
status = execute_device_kernel();
|
||||
|
||||
if (status != cutlass::Status::kSuccess) {
|
||||
std::cerr << "Device execution failed." << std::endl;
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
result = cudaEventRecord(events[1]);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaEventRecord() failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
result = cudaDeviceSynchronize();
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaDeviceSynchronize() failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
float elapsed_ms = 0;
|
||||
result = cudaEventElapsedTime(&elapsed_ms, events[0], events[1]);
|
||||
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaEventElapsedTime() failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
for (cudaEvent_t &evt : events) {
|
||||
result = cudaEventDestroy(evt);
|
||||
if (result != cudaSuccess) {
|
||||
std::cerr << "cudaEventDestroy() failed with error " << cudaGetErrorString(result) << std::endl;
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
int64_t flops = int64_t(options.problem_size.m()) * options.problem_size.n() * options.problem_size.k() * 2;
|
||||
int64_t bytes = (sizeof(ElementD) * 2 + sizeof(ElementSoftmax)) * options.problem_size.m() * options.problem_size.n();
|
||||
|
||||
double gflops_per_second = double(flops) * kIterations / double(elapsed_ms / 1000.0f) / double(1.0e9);
|
||||
double gbytes_per_second = double(bytes) * kIterations / double(elapsed_ms / 1000.0f) / double(1 << 30);
|
||||
|
||||
std::cout << " Problem: "
|
||||
<< options.problem_size.m() << "-by-" << options.problem_size.n() << "-by-" << options.problem_size.k()
|
||||
<< std::endl;
|
||||
|
||||
std::cout << " Runtime: " << elapsed_ms << " ms\n" << std::endl;
|
||||
|
||||
std::cout << " GFLOPs: " << gflops_per_second << " GFLOPs" << std::endl;
|
||||
std::cout << "Memory bandwidth: " << gbytes_per_second << " GiB/s" << std::endl;
|
||||
|
||||
return true;
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
int main(int argc, const char **argv) {
|
||||
|
||||
// Options parsing
|
||||
Options options;
|
||||
options.parse(argc, argv);
|
||||
|
||||
if (options.help) {
|
||||
options.print_usage(std::cout) << std::endl;
|
||||
return 0;
|
||||
}
|
||||
|
||||
if (!options.supported()) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
// Run
|
||||
Testbed testbed(options);
|
||||
|
||||
Disposition disposition = testbed.run();
|
||||
|
||||
std::cout << std::endl;
|
||||
|
||||
switch (disposition) {
|
||||
case Disposition::kPassed:
|
||||
std::cout << "Passed" << std::endl;
|
||||
break;
|
||||
case Disposition::kIncorrect:
|
||||
std::cout << "Incorrect" << std::endl;
|
||||
break;
|
||||
case Disposition::kNotVerified:
|
||||
std::cout << "Not verified" << std::endl;
|
||||
break;
|
||||
}
|
||||
|
||||
return (disposition == Disposition::kPassed ? 0 : -1);
|
||||
}
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -0,0 +1,498 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017 - 2022 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 GEMM kernel to support the 'epilogue visitor' model for fusion.
|
||||
|
||||
This source file will likely be moved to `include/cutlass/gemm/kernel/` in the future once
|
||||
its usage has been stabilized. For now, it is included in this example to demonstrate
|
||||
some basic output fusion options.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/fast_math.h"
|
||||
#include "cutlass/gemm/gemm.h"
|
||||
#include "cutlass/matrix_coord.h"
|
||||
#include "cutlass/complex.h"
|
||||
#include "cutlass/semaphore.h"
|
||||
|
||||
#include "cutlass/trace.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace gemm {
|
||||
namespace kernel {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template <
|
||||
typename Mma_, ///! Threadblock-scoped matrix multiply-accumulate
|
||||
typename Epilogue_, ///! Epilogue
|
||||
typename ThreadblockSwizzle_ ///! Threadblock swizzling function
|
||||
>
|
||||
struct GemmWithEpilogueVisitor {
|
||||
public:
|
||||
|
||||
using Mma = Mma_;
|
||||
using Epilogue = Epilogue_;
|
||||
using EpilogueVisitor = typename Epilogue::Visitor;
|
||||
using ThreadblockSwizzle = ThreadblockSwizzle_;
|
||||
|
||||
using ElementA = typename Mma::IteratorA::Element;
|
||||
using LayoutA = typename Mma::IteratorA::Layout;
|
||||
using TensorRefA = TensorRef<ElementA, LayoutA>;
|
||||
|
||||
using ElementB = typename Mma::IteratorB::Element;
|
||||
using LayoutB = typename Mma::IteratorB::Layout;
|
||||
using TensorRefB = TensorRef<ElementB, LayoutB>;
|
||||
|
||||
using ElementC = typename EpilogueVisitor::ElementOutput;
|
||||
using LayoutC = typename Epilogue::Layout;
|
||||
|
||||
static ComplexTransform const kTransformA = Mma::kTransformA;
|
||||
static ComplexTransform const kTransformB = Mma::kTransformB;
|
||||
using Operator = typename Mma::Operator;
|
||||
|
||||
using OperatorClass = typename Mma::Operator::OperatorClass;
|
||||
using ThreadblockShape = typename Mma::Shape;
|
||||
using WarpShape = typename Mma::Operator::Shape;
|
||||
using InstructionShape = typename Mma::Policy::Operator::InstructionShape;
|
||||
using ArchTag = typename Mma::ArchTag;
|
||||
|
||||
static int const kStages = Mma::kStages;
|
||||
static int const kAlignmentA = Mma::IteratorA::AccessType::kElements;
|
||||
static int const kAlignmentB = Mma::IteratorB::AccessType::kElements;
|
||||
static int const kAlignmentC = EpilogueVisitor::kElementsPerAccess;
|
||||
|
||||
/// Warp count (concept: GemmShape)
|
||||
using WarpCount = typename Mma::WarpCount;
|
||||
static int const kThreadCount = 32 * WarpCount::kCount;
|
||||
|
||||
/// Split-K preserves splits that are 128b aligned
|
||||
static int const kSplitKAlignment = const_max(
|
||||
128 / sizeof_bits<ElementA>::value,
|
||||
128 / sizeof_bits<ElementB>::value
|
||||
);
|
||||
|
||||
//
|
||||
// Structures
|
||||
//
|
||||
|
||||
/// Argument structure
|
||||
struct Arguments {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
GemmUniversalMode mode;
|
||||
GemmCoord problem_size;
|
||||
int batch_count;
|
||||
|
||||
TensorRefA ref_A;
|
||||
TensorRefB ref_B;
|
||||
|
||||
int64_t batch_stride_A;
|
||||
int64_t batch_stride_B;
|
||||
|
||||
typename EpilogueVisitor::Arguments epilogue_visitor;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
Arguments():
|
||||
mode(GemmUniversalMode::kGemm),
|
||||
batch_count(1)
|
||||
{ }
|
||||
|
||||
|
||||
/// constructs an arguments structure
|
||||
Arguments(
|
||||
GemmUniversalMode mode_,
|
||||
GemmCoord problem_size_,
|
||||
int batch_count_,
|
||||
TensorRefA ref_A_,
|
||||
TensorRefB ref_B_,
|
||||
int64_t batch_stride_A_,
|
||||
int64_t batch_stride_B_,
|
||||
typename EpilogueVisitor::Arguments epilogue_visitor_
|
||||
):
|
||||
mode(mode_),
|
||||
problem_size(problem_size_),
|
||||
batch_count(batch_count_),
|
||||
ref_A(ref_A_),
|
||||
ref_B(ref_B_),
|
||||
batch_stride_A(batch_stride_A_),
|
||||
batch_stride_B(batch_stride_B_),
|
||||
epilogue_visitor(epilogue_visitor_)
|
||||
{
|
||||
|
||||
}
|
||||
};
|
||||
|
||||
//
|
||||
// Structure for precomputing values in host memory and passing to kernels
|
||||
//
|
||||
|
||||
/// Parameters structure
|
||||
struct Params {
|
||||
|
||||
cutlass::gemm::GemmCoord problem_size;
|
||||
cutlass::gemm::GemmCoord grid_tiled_shape;
|
||||
int swizzle_log_tile;
|
||||
|
||||
typename Mma::IteratorA::Params params_A;
|
||||
typename Mma::IteratorB::Params params_B;
|
||||
|
||||
GemmUniversalMode mode;
|
||||
int batch_count;
|
||||
int gemm_k_size;
|
||||
|
||||
void * ptr_A;
|
||||
void * ptr_B;
|
||||
|
||||
int64_t batch_stride_A;
|
||||
int64_t batch_stride_B;
|
||||
|
||||
typename EpilogueVisitor::Params epilogue_visitor;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params():
|
||||
swizzle_log_tile(0),
|
||||
params_A(0),
|
||||
params_B(0),
|
||||
batch_count(0),
|
||||
gemm_k_size(0),
|
||||
mode(cutlass::gemm::GemmUniversalMode::kGemm),
|
||||
ptr_A(nullptr),
|
||||
ptr_B(nullptr),
|
||||
batch_stride_A(0),
|
||||
batch_stride_B(0)
|
||||
{ }
|
||||
|
||||
|
||||
Params(
|
||||
Arguments const &args
|
||||
):
|
||||
problem_size(args.problem_size),
|
||||
swizzle_log_tile(0),
|
||||
params_A(args.ref_A.layout()),
|
||||
params_B(args.ref_B.layout()),
|
||||
mode(args.mode),
|
||||
batch_count(args.batch_count),
|
||||
gemm_k_size(args.problem_size.k()),
|
||||
ptr_A(args.ref_A.data()),
|
||||
ptr_B(args.ref_B.data()),
|
||||
batch_stride_A(args.batch_stride_A),
|
||||
batch_stride_B(args.batch_stride_B),
|
||||
epilogue_visitor(args.epilogue_visitor)
|
||||
{
|
||||
|
||||
ThreadblockSwizzle threadblock_swizzle;
|
||||
|
||||
grid_tiled_shape = threadblock_swizzle.get_tiled_shape(
|
||||
args.problem_size,
|
||||
{ThreadblockShape::kM, ThreadblockShape::kN, ThreadblockShape::kK},
|
||||
args.batch_count);
|
||||
|
||||
if (args.mode == GemmUniversalMode::kGemm || args.mode == GemmUniversalMode::kGemmSplitKParallel) {
|
||||
|
||||
int const kAlignK = const_max(const_max(128 / sizeof_bits<ElementA>::value, 128 / sizeof_bits<ElementB>::value), 1);
|
||||
|
||||
gemm_k_size = round_up(ceil_div(args.problem_size.k(), args.batch_count), kAlignK);
|
||||
|
||||
if (gemm_k_size) {
|
||||
grid_tiled_shape.k() = ceil_div(args.problem_size.k(), gemm_k_size);
|
||||
}
|
||||
}
|
||||
|
||||
swizzle_log_tile = threadblock_swizzle.get_log_tile(grid_tiled_shape);
|
||||
}
|
||||
};
|
||||
|
||||
/// Shared memory storage structure
|
||||
union SharedStorage {
|
||||
|
||||
typename Mma::SharedStorage main_loop;
|
||||
|
||||
struct {
|
||||
typename Epilogue::SharedStorage epilogue;
|
||||
typename EpilogueVisitor::SharedStorage visitor;
|
||||
} epilogue;
|
||||
};
|
||||
|
||||
public:
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_DEVICE
|
||||
GemmWithEpilogueVisitor() { }
|
||||
|
||||
/// Determines whether kernel satisfies alignment
|
||||
static Status can_implement(
|
||||
cutlass::gemm::GemmCoord const & problem_size) {
|
||||
|
||||
CUTLASS_TRACE_HOST("GemmWithEpilogueVisitor::can_implement()");
|
||||
|
||||
static int const kAlignmentA = Mma::IteratorA::AccessType::kElements;
|
||||
static int const kAlignmentB = Mma::IteratorB::AccessType::kElements;
|
||||
static int const kAlignmentC = Epilogue::OutputTileIterator::kElementsPerAccess;
|
||||
|
||||
bool isAMisaligned = false;
|
||||
bool isBMisaligned = false;
|
||||
bool isCMisaligned = false;
|
||||
|
||||
if (platform::is_same<LayoutA, layout::RowMajor>::value) {
|
||||
isAMisaligned = problem_size.k() % kAlignmentA;
|
||||
} else if (platform::is_same<LayoutA, layout::ColumnMajor>::value) {
|
||||
isAMisaligned = problem_size.m() % kAlignmentA;
|
||||
} else if (platform::is_same<LayoutA, layout::ColumnMajorInterleaved<32>>::value
|
||||
|| platform::is_same<LayoutA, layout::ColumnMajorInterleaved<64>>::value) {
|
||||
isAMisaligned = problem_size.k() % kAlignmentA;
|
||||
}
|
||||
|
||||
if (platform::is_same<LayoutB, layout::RowMajor>::value) {
|
||||
isBMisaligned = problem_size.n() % kAlignmentB;
|
||||
} else if (platform::is_same<LayoutB, layout::ColumnMajor>::value) {
|
||||
isBMisaligned = problem_size.k() % kAlignmentB;
|
||||
} else if (platform::is_same<LayoutB, layout::RowMajorInterleaved<32>>::value
|
||||
|| platform::is_same<LayoutB, layout::RowMajorInterleaved<64>>::value) {
|
||||
isBMisaligned = problem_size.k() % kAlignmentB;
|
||||
}
|
||||
|
||||
if (platform::is_same<LayoutC, layout::RowMajor>::value) {
|
||||
isCMisaligned = problem_size.n() % kAlignmentC;
|
||||
} else if (platform::is_same<LayoutC, layout::ColumnMajor>::value) {
|
||||
isCMisaligned = problem_size.m() % kAlignmentC;
|
||||
} else if (platform::is_same<LayoutC, layout::ColumnMajorInterleaved<32>>::value
|
||||
|| platform::is_same<LayoutC, layout::ColumnMajorInterleaved<64>>::value) {
|
||||
isCMisaligned = problem_size.n() % kAlignmentC;
|
||||
}
|
||||
|
||||
if (isAMisaligned) {
|
||||
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for A operand");
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (isBMisaligned) {
|
||||
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for B operand");
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
if (isCMisaligned) {
|
||||
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for C operand");
|
||||
return Status::kErrorMisalignedOperand;
|
||||
}
|
||||
|
||||
CUTLASS_TRACE_HOST(" returning kSuccess");
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
|
||||
static Status can_implement(Arguments const &args) {
|
||||
return can_implement(args.problem_size);
|
||||
}
|
||||
|
||||
static size_t get_extra_workspace_size(Arguments const &args,
|
||||
cutlass::gemm::GemmCoord const &grid_tiled_shape) {
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
#define SPLIT_K_ENABLED 1
|
||||
|
||||
/// 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.swizzle_log_tile);
|
||||
|
||||
// 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;
|
||||
}
|
||||
|
||||
int offset_k = 0;
|
||||
int problem_size_k = params.problem_size.k();
|
||||
|
||||
ElementA *ptr_A = static_cast<ElementA *>(params.ptr_A);
|
||||
ElementB *ptr_B = static_cast<ElementB *>(params.ptr_B);
|
||||
|
||||
|
||||
#if SPLIT_K_ENABLED
|
||||
//
|
||||
// Fetch pointers based on mode.
|
||||
//
|
||||
if (params.mode == GemmUniversalMode::kGemm ||
|
||||
params.mode == GemmUniversalMode::kGemmSplitKParallel) {
|
||||
|
||||
if (threadblock_tile_offset.k() + 1 < params.grid_tiled_shape.k()) {
|
||||
|
||||
problem_size_k = (threadblock_tile_offset.k() + 1) * params.gemm_k_size;
|
||||
}
|
||||
|
||||
offset_k = threadblock_tile_offset.k() * params.gemm_k_size;
|
||||
}
|
||||
else if (params.mode == GemmUniversalMode::kBatched) {
|
||||
ptr_A += threadblock_tile_offset.k() * params.batch_stride_A;
|
||||
ptr_B += threadblock_tile_offset.k() * params.batch_stride_B;
|
||||
}
|
||||
else if (params.mode == GemmUniversalMode::kArray) {
|
||||
ptr_A = static_cast<ElementA * const *>(params.ptr_A)[threadblock_tile_offset.k()];
|
||||
ptr_B = static_cast<ElementB * const *>(params.ptr_B)[threadblock_tile_offset.k()];
|
||||
}
|
||||
#endif
|
||||
|
||||
// Compute initial location in logical coordinates
|
||||
cutlass::MatrixCoord tb_offset_A{
|
||||
threadblock_tile_offset.m() * Mma::Shape::kM,
|
||||
offset_k,
|
||||
};
|
||||
|
||||
cutlass::MatrixCoord tb_offset_B{
|
||||
offset_k,
|
||||
threadblock_tile_offset.n() * Mma::Shape::kN
|
||||
};
|
||||
|
||||
// Compute position within threadblock
|
||||
int thread_idx = threadIdx.x;
|
||||
|
||||
// Construct iterators to A and B operands
|
||||
typename Mma::IteratorA iterator_A(
|
||||
params.params_A,
|
||||
ptr_A,
|
||||
{params.problem_size.m(), problem_size_k},
|
||||
thread_idx,
|
||||
tb_offset_A);
|
||||
|
||||
typename Mma::IteratorB iterator_B(
|
||||
params.params_B,
|
||||
ptr_B,
|
||||
{problem_size_k, params.problem_size.n()},
|
||||
thread_idx,
|
||||
tb_offset_B);
|
||||
|
||||
// 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
|
||||
//
|
||||
|
||||
// Construct thread-scoped matrix multiply
|
||||
Mma mma(shared_storage.main_loop, thread_idx, warp_idx, lane_idx);
|
||||
|
||||
typename Mma::FragmentC accumulators;
|
||||
|
||||
accumulators.clear();
|
||||
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
int gemm_k_iterations = (problem_size_k - offset_k + Mma::Shape::kK - 1) / Mma::Shape::kK;
|
||||
|
||||
// Compute threadblock-scoped matrix multiply-add
|
||||
mma(
|
||||
gemm_k_iterations,
|
||||
accumulators,
|
||||
iterator_A,
|
||||
iterator_B,
|
||||
accumulators);
|
||||
|
||||
//
|
||||
// Masked tile iterators constructed from members
|
||||
//
|
||||
|
||||
threadblock_tile_offset = threadblock_swizzle.get_tile_offset(params.swizzle_log_tile);
|
||||
|
||||
//assume identity swizzle
|
||||
MatrixCoord threadblock_offset(
|
||||
threadblock_tile_offset.m() * Mma::Shape::kM,
|
||||
threadblock_tile_offset.n() * Mma::Shape::kN
|
||||
);
|
||||
|
||||
int block_idx = threadblock_tile_offset.m() + threadblock_tile_offset.n() * params.grid_tiled_shape.m();
|
||||
|
||||
//
|
||||
// Construct the epilogue visitor
|
||||
//
|
||||
|
||||
EpilogueVisitor epilogue_visitor(
|
||||
params.epilogue_visitor,
|
||||
shared_storage.epilogue.visitor,
|
||||
params.problem_size.mn(),
|
||||
thread_idx,
|
||||
warp_idx,
|
||||
lane_idx,
|
||||
threadblock_offset);
|
||||
|
||||
if (params.mode == GemmUniversalMode::kGemm) {
|
||||
// Indicate which position in a serial reduction the output operator is currently updating
|
||||
epilogue_visitor.set_k_partition(threadblock_tile_offset.k(), params.grid_tiled_shape.k());
|
||||
}
|
||||
else if (params.mode == GemmUniversalMode::kBatched || params.mode == GemmUniversalMode::kArray) {
|
||||
epilogue_visitor.set_batch_index(threadblock_tile_offset.k());
|
||||
}
|
||||
|
||||
// Construct the epilogue
|
||||
Epilogue epilogue(
|
||||
shared_storage.epilogue.epilogue,
|
||||
thread_idx,
|
||||
warp_idx,
|
||||
lane_idx);
|
||||
|
||||
// Execute the epilogue operator to update the destination tensor.
|
||||
epilogue(epilogue_visitor, accumulators);
|
||||
}
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace kernel
|
||||
} // namespace gemm
|
||||
} // namespace cutlass
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user