198 lines
7.1 KiB
C++
198 lines
7.1 KiB
C++
/***************************************************************************************************
|
|
* 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 Epilogue for threadblock scoped GEMMs.
|
|
|
|
This does not attempt to target any particular output layout. Instead, each threadblock
|
|
streams out its accumulator elements using 128b store operations. This assumes all threadblocks
|
|
have unique output tiles.
|
|
|
|
The target data layout is:
|
|
- threadblock indices mapped to linear offsets as (m, n, k), where m is fastest-changing
|
|
- threadblock output space partitioned into warps; each warp's region is contiguous
|
|
- per-thread accumulators partitioned into 128b accesses
|
|
- output memory striped across the threads of a warp
|
|
|
|
This enables very fast streaming of data, completely limited by the memory system. No predication
|
|
or data exchange is performed, and each threadblock is assumed to have a full region of memory
|
|
to write to.
|
|
|
|
This epilogue establishes an upper bound for epilogue performance and is suitable for
|
|
reductions across the GEMM K dimension which require a separate workspace.
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "cutlass/cutlass.h"
|
|
#include "cutlass/numeric_types.h"
|
|
#include "cutlass/array.h"
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
namespace epilogue {
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
template <
|
|
typename Shape_, ///< shape of accumulator tile (concept: MatrixShape)
|
|
int WarpCount, ///< number of warps
|
|
typename FragmentC_ ///< warp-level GEMM operator (concept: gemm::warp::Mma)
|
|
>
|
|
class EpilogueWorkspace {
|
|
public:
|
|
|
|
using Shape = Shape_;
|
|
using FragmentC = FragmentC_;
|
|
using ElementC = typename FragmentC::value_type;
|
|
|
|
static int const kWarpCount = WarpCount;
|
|
|
|
/// Optimize for 128b accesses
|
|
static int const kAccessSizeInBits = 128;
|
|
|
|
/// Warp size from the perspective of memory operations
|
|
static int const kWarpSize = 32;
|
|
|
|
/// Vector length of accesses
|
|
static int const kElementsPerAccess =
|
|
kAccessSizeInBits / sizeof_bits<ElementC>::value;
|
|
|
|
/// Number of stores per thread
|
|
static int const kIterations = FragmentC::kElements / kElementsPerAccess;
|
|
|
|
static_assert(
|
|
!(FragmentC::kElements % kElementsPerAccess),
|
|
"The number of accumulators must be divisible by the access size.");
|
|
|
|
/// Total number of vectorized accesses in warp (in units of vector)
|
|
static int const kWarpAccesses = kIterations * kWarpSize;
|
|
|
|
/// Total number of vectorized accesses in threadblock tile (in units of vector)
|
|
static int const kThreadblockAccesses = kWarpAccesses * kWarpCount;
|
|
|
|
/// Parameters structure
|
|
struct Params {
|
|
|
|
/// Pointer to C matrix
|
|
ElementC *ptr_C;
|
|
|
|
/// Stride between tiles along the GEMM N dimension (in units of vectors)
|
|
int stride_n;
|
|
|
|
/// Stride between tiles along the GEMM K dimension (in units of vectors)
|
|
int stride_k;
|
|
|
|
//
|
|
// Methods
|
|
//
|
|
|
|
CUTLASS_HOST_DEVICE
|
|
Params(
|
|
ElementC *ptr_C, ///< Pointer to C matrix
|
|
int stride_n_, ///< Stride between tiles along the GEMM N dimension (in units of ElementC)
|
|
int stride_k_ ///< Stride between tiles along the GEMM K dimension (in units of ElementC)
|
|
):
|
|
ptr_C(ptr_C), stride_n(stride_n_ / kElementsPerAccess), stride_k(stride_k_ / kElementsPerAccess) {
|
|
|
|
}
|
|
};
|
|
|
|
/// Shared storage allocation needed by the epilogue
|
|
struct SharedStorage {
|
|
// Intentionally empty
|
|
};
|
|
|
|
private:
|
|
|
|
struct alignas((kAccessSizeInBits / 8)) AccessType {
|
|
Array<ElementC, kElementsPerAccess> storage;
|
|
};
|
|
|
|
/// Constant reference to parameters object
|
|
AccessType *pointer_;
|
|
|
|
/// Stride between tiles along the n dimension (in vectors)
|
|
int stride_n_;
|
|
|
|
/// Stride between tiles along the k dimension (in vectors)
|
|
int stride_k_;
|
|
|
|
public:
|
|
|
|
/// Constructor
|
|
CUTLASS_DEVICE
|
|
EpilogueWorkspace(
|
|
Params const ¶ms, ///< Host-constructable params object
|
|
SharedStorage &, ///< Shared storage object
|
|
int warp_idx, ///< ID of warp within threadblock
|
|
int lane_idx ///< Id of thread within warp
|
|
|
|
):
|
|
pointer_(reinterpret_cast<AccessType *>(params.ptr_C)),
|
|
stride_n_(params.stride_n),
|
|
stride_k_(params.stride_k) {
|
|
|
|
// Add per-thread offset
|
|
pointer_ += lane_idx + warp_idx * kWarpAccesses;
|
|
}
|
|
|
|
/// Streams the result to global memory
|
|
CUTLASS_DEVICE
|
|
void operator()(
|
|
cutlass::gemm::GemmCoord problem_size, ///< Problem size of GEMM (units of ElementC)
|
|
cutlass::gemm::GemmCoord tb_tile_coord, ///< Threadblock tile coordinate in GEMM (in units of threadblock tiles)
|
|
FragmentC const &accum) { ///< Accumulator tile
|
|
|
|
// Compute offset for entire threadblock (note, per-thread offset has been folded in already)
|
|
AccessType *pointer = pointer_ +
|
|
tb_tile_coord.m() * kThreadblockAccesses +
|
|
tb_tile_coord.n() * stride_n_ +
|
|
tb_tile_coord.k() * stride_k_;
|
|
|
|
// Cast to vectorized view of accumulator fragments
|
|
AccessType const * src_pointer = reinterpret_cast<AccessType const *>(&accum);
|
|
|
|
// Write out accumulators at full speed
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for (int i = 0; i < kIterations; ++i) {
|
|
pointer[i * kWarpSize] = src_pointer[i];
|
|
}
|
|
}
|
|
};
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace epilogue
|
|
} // namespace cutlass
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|