Skip to content

GitLab

Unverified Commit f83e9701 authored by Haocong WANG's avatar Haocong WANG Committed by GitHub
Browse files

[GEMM] Gemm universal device operation (#1154) 



* Optimize GEMM on MI200/300:
1. Add new blockwise gemm pipeline
2. Add irregular splitk intances

* clang format + typo fix

* Fix a bug

* initial commit

* Add more instances to irregular splitk

* blkgemm pipeline v1~4 prototype

* Sanity Checked. Known issue:
1. Poor performance of splitk
2. Register spill on blkgemmpipeline v3

* Sanity and Performance fix:
1. fix a bug related to sanity in grouped b2c mapping
2. fix a bug related to sanity and performance in splitk offset

* Sanity and API update:
1. Remove prefetch stage
2. Fix valid check bug
3, Add first gemm_universal instance into ckProfiler

* Add NN instances for gemm universal

* 1. Add NT instances for gemm_universal
2. Fix a bug about Kpadding in gemm_universal

* Fix a bug regarding padding Odd K number

* remove kernel print

* Fix KPadding bug...

* Update safety check

* another try to fix kpadding..

* Sanity checked

* new instances..

* clang format+typo fix

* remove clang format script's change

* Add non-hotloop compile option

* 1. Add fp16xfp8 example
2. pull packed convert f8 from pr1150

* Some miscs.. opt and fix

* Add pipeline description docs

* Split universal gemm instance library to cut profiler compiling time

* uncomment cmakefile

* Fix a bug caused by blockwise_gemm_pipe_v2

* reduce default splitk to 1

* Add 224x256x64 tile size

* update, including:
1. Experiment pipeline 5~7
2. Optimization for pipeline 4
3. Organized instance library

* temp save

* temp save

* Permuted lds layout, sanity and function checked

* clang format

* Move OOB check from RunRead to RunWrite, for better software pipeline.
TODO: agpr spill when NN layout

* clangformat

* A/B splitpipe scheduler for v3

* Fix two bugs

* bug fix

* fix a bug in oob check

* Example for mixed fp16_fp8 gemm

* Clean experimental code blocks

* Add mixed precision gemm into profiler

* tempsave

* optimize m/n major lds layout

* Add RRR GEMM  mixed precision instances

* Optimize f8 matrix transpose

* Add test_gemm_universal

* A/B spilt schedule for blkpip v5

* Take ds_read2 into iglp scheduling scheme

* format

* fixed cmake

* Add llvm-option into CI cmake flag

---------
Co-authored-by: default avatarJing Zhang <jizhan@amd.com>
parent 7cdf5a96
Hide whitespace changes
Inline Side-by-side
Showing with 3097 additions and 113 deletions
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ f83e9701
......@@ -259,46 +259,20 @@ struct BlockToCTileMap_M00_N0_M01Adapt : BlockToCTileMap_M00_N0_M01Adapt<MPerBlo
BlockToCTileMap_M00_N0_M01Adapt;
};
// Rows of column-vectors
// This C-tile map dynamically adjusts M01 when C-tile index is out of range
template <index_t GroupNum, index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N = void>
struct BlockToCTileMap_Grouped_M00_N0_M01Adapt;
// Grouped Rows of column-vectors WGP mapping
// Optimized for MI300-like multipe-die chip
template <index_t GroupNum, index_t MPerBlock, index_t NPerBlock>
struct BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, void>
struct BlockToCTileMap_Grouped_M00_N0_M01Adapt
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
__host__ __device__ BlockToCTileMap_Grouped_M00_N0_M01Adapt() = default;
__host__ __device__ BlockToCTileMap_Grouped_M00_N0_M01Adapt(
const BlockToCTileMap_Grouped_M00_N0_M01Adapt&) = default;
__host__ __device__
BlockToCTileMap_Grouped_M00_N0_M01Adapt(BlockToCTileMap_Grouped_M00_N0_M01Adapt&&) = default;
__host__ __device__ BlockToCTileMap_Grouped_M00_N0_M01Adapt&
operator=(const BlockToCTileMap_Grouped_M00_N0_M01Adapt&) = default;
__host__ __device__ BlockToCTileMap_Grouped_M00_N0_M01Adapt&
operator=(BlockToCTileMap_Grouped_M00_N0_M01Adapt&&) = default;
__host__ __device__ BlockToCTileMap_Grouped_M00_N0_M01Adapt(index_t M,
index_t N,
index_t M01 = 8)
: M_(M), N_(N), M01_(M01)
{
#if 0
if(get_thread_global_1d_id()==0){
printf("Ctor called, M= %d, N= %d, M01 = %d\n", M_, N_, M01_);
}
#endif
}
template <typename CGridDesc_M_N>
__host__ __device__
BlockToCTileMap_Grouped_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01 = 8)
: BlockToCTileMap_Grouped_M00_N0_M01Adapt(
c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), M01)
{
}
__host__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
......@@ -309,12 +283,6 @@ struct BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, v
return M0 * N0;
}
template <typename CGridDesc_M_N>
__host__ static constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
{
return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
......@@ -329,67 +297,82 @@ struct BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, v
const auto M0 = math::integer_divide_ceil(M_, MPerBlock);
const auto N0 = math::integer_divide_ceil(N_, NPerBlock);
block_1d_id = block_1d_id % (M0 * N0); // swallow batch index
const auto group_size = math::integer_divide_ceil(M0 * N0, GroupNum);
auto group_id = block_1d_id % GroupNum;
auto remap_block_1d_id = group_id * group_size + block_1d_id / GroupNum;
index_t idx_N0 = remap_block_1d_id % N0;
index_t idx_M0 = remap_block_1d_id / N0;
const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
index_t idx_M00 = idx_M0 / M01_;
index_t idx_M01 = idx_M0 % M01_;
index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
/**
* idxN0
*
* |< mtx N >|
*
* NPerBlock NPerBlock NPerBlock NPerBlock
* N_0 N_1 N_2 N_3
* - |-----------|-----------|-----------|-----|-----|-
* ^ | - - 0 |/----> 2 | | | |
* | | | / | | | | | M_0 MPerBlock
* | M | /| | | | | |
* |-0---|---/-|-----|-----|-----------|-----|-----|-
* | 1 | / | | | blockid | | |
* idxM0 | | | / | V | 5 | | | M_1 MPerBlock
* | - V 1 | - 3 | | | |
* |-----------|-----------|-----------|-----|-----|-
* mtx M | | | | | |
* | | | | | | M_2 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* | | | | | |
* | | | | | | M_3 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* V | | | | | |
* - |-----------|-----------|-----------|-----|-----|- M_4 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* Example:
* assume:
* M0 = 5
* N0 = 4
* block_1d_id = 5
* M01 = 2
*
* idx_N0 = 1
* idx_M0 = 1
* M01_adapt = 2
* idx_M00 = 0
* idx_M01 = 1
* idx_N0_M01_local = 5
* output {1, 2}
*/
return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
idx_N0_M01_local / M01_adapt);
if(M0 == 1)
{
return make_tuple(0, block_1d_id);
}
else if(N0 == 1)
{
return make_tuple(block_1d_id, 0);
}
// block_1d_id = block_1d_id % (M0 * N0); // swallow batch index
else
{
const auto group_size = math::integer_divide_ceil(M0 * N0, GroupNum);
const auto big_group_num = GroupNum - (group_size * GroupNum - M0 * N0);
auto group_id_x = block_1d_id % GroupNum;
auto group_id_y = block_1d_id / GroupNum;
auto remap_block_1d_id =
group_id_x <= big_group_num
? group_id_x * group_size + group_id_y
: group_id_x * group_size + big_group_num - group_id_x + group_id_y;
index_t idx_N0 = remap_block_1d_id % N0;
index_t idx_M0 = remap_block_1d_id / N0;
const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
index_t idx_M00 = idx_M0 / M01_;
index_t idx_M01 = idx_M0 % M01_;
index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
/**
* idxN0
*
* |< mtx N >|
*
* NPerBlock NPerBlock NPerBlock NPerBlock
* N_0 N_1 N_2 N_3
* - |-----------|-----------|-----------|-----|-----|-
* ^ | - - 0 |/----> 2 | | | |
* | | | / | | | | | M_0 MPerBlock
* | M | /| | | | | |
* |-0---|---/-|-----|-----|-----------|-----|-----|-
* | 1 | / | | | blockid | | |
* idxM0 | | | / | V | 5 | | | M_1 MPerBlock
* | - V 1 | - 3 | | | |
* |-----------|-----------|-----------|-----|-----|-
* mtx M | | | | | |
* | | | | | | M_2 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* | | | | | |
* | | | | | | M_3 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* V | | | | | |
* - |-----------|-----------|-----------|-----|-----|- M_4 MPerBlock
* | | | | | |
* |-----------|-----------|-----------|-----|-----|-
* Example:
* assume:
* M0 = 5
* N0 = 4
* block_1d_id = 5
* M01 = 2
*
* idx_N0 = 1
* idx_M0 = 1
* M01_adapt = 2
* idx_M00 = 0
* idx_M01 = 1
* idx_N0_M01_local = 5
* output {1, 2}
*/
return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
idx_N0_M01_local / M01_adapt);
}
}
template <typename CTileIdx, typename CTileDim>
......@@ -405,15 +388,6 @@ struct BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, v
index_t M01_;
};
// keep the redundant type argument for backward compatibility
template <index_t GroupNum, index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
struct BlockToCTileMap_Grouped_M00_N0_M01Adapt
: BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, void>
{
using BlockToCTileMap_Grouped_M00_N0_M01Adapt<GroupNum, MPerBlock, NPerBlock, void>::
BlockToCTileMap_Grouped_M00_N0_M01Adapt;
};
// columns of row-vectors
// This C-tile map dynamically adjusts N01 when C-tile index is out of range
template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N = void>
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_selector.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
// kernel function Blockers:
// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
// two lds chunks.
// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
// buffer when we declare __shared__ inside blkgemmpipe
template <typename GridwiseGemm,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid,
p_shared,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
template <typename GridwiseGemm,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_2lds(typename GridwiseGemm::Argument karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
__shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid,
p_shared_0,
p_shared_1,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename AccDataType,
typename CShuffleDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA>
struct GridwiseGemm_xdl_cshuffle_v3
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1Number = Number<AK1Value>{};
static constexpr auto BK1Number = Number<BK1Value>{};
static constexpr index_t KPack =
math::max(math::lcm(AK1Number, BK1Number),
MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
__host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
{
return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
}
__host__ static auto CalculateMPadded(index_t M)
{
return math::integer_least_multiple(M, MPerBlock);
}
__host__ static auto CalculateNPadded(index_t N)
{
return math::integer_least_multiple(N, NPerBlock);
}
__host__ static auto CalculateKPadded(index_t K)
{
return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
}
__host__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
}
__host__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
}
__host__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{
auto K_t = K_Batch * KPerBlock;
return (K + K_t - 1) / K_t * KPerBlock;
}
__host__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
{
constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
auto K_t = K_Batch * KReadVec;
return (K + K_t - 1) / K_t * KReadVec;
}
__host__ static auto CalculateMBlock(index_t M)
{
return math::integer_divide_ceil(M, MPerBlock);
}
__host__ static auto CalculateNBlock(index_t N)
{
return math::integer_divide_ceil(N, NPerBlock);
}
template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
__host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
{
constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
return transform_tensor_descriptor(
TileDesc_K0_MN_K1{},
make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
make_unmerge_transform(make_tuple(
Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
}
__device__ static auto MakeAGridDescriptor_AK0_M_AK1(
index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
__device__ static auto MakeBGridDescriptor_BK0_N_BK1(
index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
}
}();
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(N, NPad - N),
make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
}
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
{
constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
}
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
}
__host__ __device__ static auto
MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
}
struct Problem
{
__host__ Problem(index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t KBatch_)
: M{M_},
N{N_},
K{K_},
StrideA{StrideA_},
StrideB{StrideB_},
StrideC{StrideC_},
KBatch{KBatch_},
MPadded{CalculateMPadded(M_)},
NPadded{CalculateNPadded(N_)},
KRead{CalculateKRead(K_, KBatch_)},
KPadded{CalculateKPadded(K_, KBatch_)},
AK0{CalculateAK0Padded(K_, KBatch_)},
BK0{CalculateBK0Padded(K_, KBatch_)},
MBlock{CalculateMBlock(M_)},
NBlock{CalculateNBlock(N_)}
{
}
__host__ void Print() const
{
std::cout << "problem {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SC:" << StrideC << ", "
<< "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", "
<< "KRead:" << KRead << ", "
<< "KP:" << KPadded << ", "
<< "AK0:" << AK0 << ", "
<< "BK0:" << BK0 << ", "
<< "MBlock: " << MBlock << ", "
<< "NBlock: " << NBlock << "}" << std::endl;
}
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
index_t StrideC;
index_t KBatch;
index_t MPadded;
index_t NPadded;
index_t KRead;
index_t KPadded;
index_t AK0;
index_t BK0;
index_t MBlock;
index_t NBlock;
};
// Argument
struct Argument : public tensor_operation::device::BaseArgument, public Problem
{
__host__ Argument(const ADataType* p_a_grid_,
const BDataType* p_b_grid_,
CDataType* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t k_batch_)
: Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_, k_batch_},
p_a_grid{p_a_grid_},
p_b_grid{p_b_grid_},
p_c_grid{p_c_grid_}
{
}
const ADataType* p_a_grid;
const BDataType* p_b_grid;
CDataType* p_c_grid;
};
struct SplitKBatchOffset
{
__device__ SplitKBatchOffset(Argument& karg)
{
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
a_k_split_offset = blockIdx.z * karg.KRead;
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
a_k_split_offset = blockIdx.z * karg.KRead * karg.M;
}
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
{
b_k_split_offset = blockIdx.z * karg.KRead * karg.N;
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
{
b_k_split_offset = blockIdx.z * karg.KRead;
}
if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
{
karg.K = karg.KRead;
}
else
{
karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
}
}
index_t a_k_split_offset;
index_t b_k_split_offset;
};
__device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
}
// xor tensor transformation request more unnecessary vgpr usage, would cause register spill
// in some cases.
else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(ADataType) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(ADataType);
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(make_xor_transform(make_tuple(Number<MPerBlock / MLdsLayer>{},
Number<AK0Number * MLdsLayer>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_ak0_mldslayer_m_ak1,
make_tuple(make_pass_through_transform(AK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return a_lds_block_desc_ak0_m_ak1;
}
else // ColumnMajor A
{
// kfold and mpair dimension is not always required.
// more dimension in merge_transform increase the difficulty of generating immarg offset
// for compiler.
constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
constexpr auto M1 = MPerBlock / M0;
constexpr auto KThreadWrite = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
constexpr auto KThreadRead = 64 / MPerXdl;
constexpr auto K0PerThreadRead = AK0Number / KThreadRead;
constexpr auto kfold = (AK1Number * M0 * sizeof(ADataType) > 128)
? 1
: 128 / (AK1Number * M0 * sizeof(ADataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=mpair<=n0
constexpr auto mpair = (AK1Number * MPerXdl * sizeof(ADataType) > 128)
? 1
: ((128 / (AK1Number * MPerXdl * sizeof(ADataType))) > M0
? M0
: 128 / (AK1Number * MPerXdl * sizeof(ADataType)));
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<K0PerThreadWrite>{},
Number<KThreadReadPerm * M1>{},
Number<kfold * M0 / mpair>{},
Number<mpair>{},
AK1Number));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_xor_transform(
make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
make_pass_through_transform(Number<mpair>{}),
make_pass_through_transform(AK1Number)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
make_pass_through_transform(Number<mpair>{}),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<1>{},
Sequence<2>{},
Sequence<0, 3>{},
Sequence<4, 5>{},
Sequence<6>{},
Sequence<7>{}));
constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(Number<KThreadReadPerm>{},
Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<kfold>{},
Number<K0PerThreadWrite>{})),
make_merge_transform_v3_division_mod(
make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
make_pass_through_transform(AK1Number)),
make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return a_lds_block_desc_ak0_m_ak1;
}
}
__device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
// NLdsLayer * K0 as logical Bank
constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(BDataType) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(BDataType);
;
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(
BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(make_xor_transform(make_tuple(Number<NPerBlock / NLdsLayer>{},
Number<BK0Number * NLdsLayer>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_bk0_nldslayer_n_bk1,
make_tuple(make_pass_through_transform(BK0Number),
make_merge_transform_v3_division_mod(
make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_lds_block_desc_bk0_n_bk1;
}
else // RowMajor B
{
constexpr auto N0 = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I1);
constexpr auto N1 = NPerBlock / N0;
constexpr auto KThreadWrite = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I0);
constexpr auto K0PerThreadWrite = BK0Number / KThreadWrite;
constexpr auto KThreadRead = 64 / NPerXdl;
constexpr auto K0PerThreadRead = BK0Number / KThreadRead;
constexpr auto kfold = (BK1Number * N0 * sizeof(BDataType) > 128)
? 1
: 128 / (BK1Number * N0 * sizeof(BDataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=npair<=n0
constexpr auto npair = (BK1Number * NPerXdl * sizeof(BDataType) > 128)
? 1
: ((128 / (BK1Number * NPerXdl * sizeof(BDataType))) > N0
? N0
: 128 / (BK1Number * NPerXdl * sizeof(BDataType)));
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<K0PerThreadWrite>{},
Number<KThreadReadPerm * N1>{},
Number<kfold * N0 / npair>{},
Number<npair>{},
BK1Number));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_xor_transform(
make_tuple(Number<KThreadReadPerm * N1>{}, Number<kfold * N0 / npair>{})),
make_pass_through_transform(Number<npair>{}),
make_pass_through_transform(BK1Number)),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
make_tuple(
Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(Number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<N1>{})),
make_unmerge_transform(make_tuple(Number<kfold>{}, Number<N0 / npair>{})),
make_pass_through_transform(Number<npair>{}),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
Sequence<3>{},
Sequence<4>{},
Sequence<5>{}),
make_tuple(Sequence<1>{},
Sequence<2>{},
Sequence<0, 3>{},
Sequence<4, 5>{},
Sequence<6>{},
Sequence<7>{}));
constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(Number<KThreadReadPerm>{},
Number<KThreadWrite / kfold / KThreadReadPerm>{},
Number<kfold>{},
Number<K0PerThreadWrite>{})),
make_merge_transform_v3_division_mod(
make_tuple(Number<N0 / npair>{}, Number<npair>{}, Number<N1>{})),
make_pass_through_transform(BK1Number)),
make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return b_lds_block_desc_bk0_n_bk1;
}
}
__device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
using BlockwiseGemmPipe =
remove_cvref_t<decltype(BlockGemmPipeline_Selector<
BlkGemmPipelineVer,
BlkGemmPipeSched,
BlockSize,
ADataType,
BDataType,
ComputeTypeA,
AccDataType,
decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack>())>;
__device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned * sizeof(ADataType) +
b_block_space_size_aligned * sizeof(BDataType)),
c_block_size * sizeof(CShuffleDataType));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
__host__ static constexpr bool CheckValidity(const Argument& karg)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
if(!(karg.M % MPerBlock == 0))
{
#if DEBUG_LOG
std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
#endif // DEBUG_LOG
return false;
}
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
if(!(karg.N % NPerBlock == 0))
{
#if DEBUG_LOG
std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
<< __FILE__ << ":" << __LINE__ << ", in function: " << __func__
<< std::endl;
#endif // DEBUG_LOG
return false;
}
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
auto K_t = karg.KBatch * KPerBlock;
if(!(karg.K % K_t == 0))
{
#if DEBUG_LOG
std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
<< karg.K << " " << __FILE__ << ":" << __LINE__
<< ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
else
{
constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
auto K_t = karg.KBatch * KReadVec;
auto KReadPadSplited = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
{
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
if(karg.K % ABlockTransferSrcScalarPerVector != 0)
{
#if DEBUG_LOG
std::cout << "Arg K (" << karg.K
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
else
{
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
{
#if DEBUG_LOG
std::cout << "Arg M (" << karg.M
<< ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
<< ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
if(karg.N % BBlockTransferSrcScalarPerVector != 0)
{
#if DEBUG_LOG
std::cout << "Arg N (" << karg.N
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
else
{
if(karg.K % BBlockTransferSrcScalarPerVector != 0)
{
#if DEBUG_LOG
std::cout << "Arg K (" << karg.K
<< ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
<< BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
<< __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
#if DEBUG_LOG
std::cout << "Arg N (" << karg.N
<< ") value is not a multiple of "
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
else
{
if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
#if DEBUG_LOG
std::cout << "Arg M (" << karg.M
<< ") value is not a multiple of "
"CShuffleBlockTransferScalarPerVector_NPerBlock ("
<< CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
}
// check gridwise gemm pipeline
const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1)
{
if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
{
return false;
}
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
}
__host__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
{
const index_t num_loop = K / KPerBlock;
return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
}
template <typename CGridDesc>
__device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
{
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
// return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942
using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
// using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
void* p_shared,
const Problem& problem)
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const AElementwiseOperation a_element_op{};
const BElementwiseOperation b_element_op{};
const CElementwiseOperation c_element_op{};
// divide block work by [M, N]
const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0Number, MPerBlock, AK1Number>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ADataType,
ADataType,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0Number, NPerBlock, BK1Number>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
// Cast after lds
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<BDataType*>(p_shared) +
a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
// Blockwise GEMM pipeline
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
void* p_shared_0,
void* p_shared_1,
const Problem& problem)
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const AElementwiseOperation a_element_op{};
const BElementwiseOperation b_element_op{};
const CElementwiseOperation c_element_op{};
// divide block work by [M, N]
const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0Number, MPerBlock, AK1Number>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ADataType,
ADataType,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0Number, NPerBlock, BK1Number>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BDataType,
BDataType,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<BDataType*>(p_shared_0) +
a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ADataType*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<BDataType*>(p_shared_1) +
a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
// Blockwise GEMM pipeline
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_bufs,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_bufs,
b_block_slice_copy_step,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared_0),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ f83e9701
......@@ -202,15 +202,17 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto src_data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
// maintain a container record is_src_valid, waiting for RunWrite use.
const bool is_src_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
src_oob_thread_scratch_tuple_(thread_scratch_id)
.template SetAsType<bool>(src_data_idx_seq, is_src_valid);
using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
using src_vector_t = typename src_vector_type::type;
// copy data from src_buf into src_vector_container
auto src_vector_container = src_vector_type{
src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
auto src_vector_container =
src_vector_type{src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), true)};
using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
using dst_vector_t = typename dst_vector_type::type;
......@@ -305,12 +307,78 @@ struct ThreadwiseTensorSliceTransfer_v3r1
dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
});
#else
// OOB Check
constexpr auto src_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
constexpr auto src_dim_access_order = SrcDimAccessOrder{};
constexpr auto ordered_src_access_lengths =
container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
// loop over tensor and copy
static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_src_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate src data index
constexpr auto src_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
: ordered_src_access_lengths[i] - 1 -
ordered_src_access_idx[i];
});
return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
src_scalar_per_access;
}();
constexpr auto src_data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
using vector_t = typename vector_type_maker<DstData, SrcScalarPerVector>::type::type;
auto op_r = src_thread_scratch_tuple_(thread_scratch_id)
.template GetAsType<vector_t>(src_data_idx_seq);
const bool is_src_valid = src_oob_thread_scratch_tuple_(thread_scratch_id)
.template GetAsType<bool>(src_data_idx_seq);
auto op_r_v = is_src_valid ? op_r : vector_t(0);
src_thread_scratch_tuple_(thread_scratch_id)
.template SetAsType<vector_t>(src_data_idx_seq, op_r_v);
});
// sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
// TODO make this logic more generic for more sub-dword datatype
if constexpr(SrcVectorDim != DstVectorDim &&
((is_same<half_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
(is_same<int8_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
(is_same<f8_t, remove_cvref_t<DstData>>::value &&
SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
{
// each transpose does
......@@ -386,6 +454,7 @@ struct ThreadwiseTensorSliceTransfer_v3r1
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
// if there is transpose, it's done here
// if there is oob check, it's done here
// TODO move this elsewhere
TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
......@@ -738,6 +807,16 @@ struct ThreadwiseTensorSliceTransfer_v3r1
return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
}
__device__ static constexpr auto GetSrcOOBThreadScratchDescriptor()
{
constexpr auto src_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
return make_naive_tensor_descriptor_packed(src_access_lengths);
}
__device__ static constexpr auto GetDstThreadScratchDescriptor()
{
// 1st stage of transforms
......@@ -789,6 +868,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
private:
static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
static constexpr auto src_oob_thread_scratch_desc_ =
decltype(GetSrcThreadScratchDescriptor()){};
static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
using SrcThreadScratch =
......@@ -798,6 +879,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1
decltype(src_thread_scratch_desc_),
true>;
using SrcOOBThreadScratch =
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
bool, // apply data_convert with SrcThreadScratch
1,
decltype(src_oob_thread_scratch_desc_),
true>;
using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
DstData,
DstScalarPerVector,
......@@ -805,6 +893,7 @@ struct ThreadwiseTensorSliceTransfer_v3r1
true>;
StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
StaticallyIndexedArray<SrcOOBThreadScratch, NumThreadScratch> src_oob_thread_scratch_tuple_;
DstThreadScratch dst_thread_scratch_;
......
include/ck/utility/blkgemmpipe_scheduler.hpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
namespace ck {
enum struct BlockGemmPipelineScheduler
{
Intrawave,
Interwave,
};
enum struct TailNumber
{
// Single / Double buffer pipeline
Odd,
Even,
// Long prefetch pipeline, up to 8
One,
Two,
Three,
Four,
Five,
Six,
Seven,
// Unroll stages > Prefetch stages, number of loop is multiple of unroll stages
Empty,
// Unroll stages <= Prefetch stages, number of loop is multiple of unroll stages add
// prefetchstages
Full,
};
template <index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t ABufferLoadWidth,
index_t BBufferLoadWidth,
index_t ALDSWriteWidth,
index_t BLDSWriteWidth,
index_t ALDSReadWidth,
index_t BLDSReadWidth,
index_t MRepeat,
index_t NRepeat,
index_t MPerXDL,
index_t NPerXDL,
index_t KPerXDL>
struct BlockwiseGemmXdlops_pipeline_hotloop_inst
{
static constexpr index_t WaveSize = 64;
static constexpr index_t WaveNumM = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t WaveNumN = NPerBlock / (NRepeat * NPerXDL);
static constexpr index_t A_LDS_Read_Width = ALDSReadWidth;
static constexpr index_t B_LDS_Read_Width = BLDSReadWidth;
static constexpr index_t A_Buffer_Load_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * ABufferLoadWidth);
static constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * BBufferLoadWidth);
static constexpr index_t A_LDS_Write_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * ALDSWriteWidth);
static constexpr index_t B_LDS_Write_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * BLDSWriteWidth);
static constexpr index_t A_LDS_Read_Inst_Num =
WaveNumN * MPerBlock * KPerBlock / (BlockSize * ALDSReadWidth);
static constexpr index_t B_LDS_Read_Inst_Num =
WaveNumM * MPerBlock * KPerBlock / (BlockSize * BLDSReadWidth);
static constexpr index_t C_MFMA_Inst_Num =
MPerBlock * NPerBlock * KPerBlock / (BlockSize / WaveSize) / (MPerXDL * NPerXDL * KPerXDL);
static constexpr auto Print()
{
printf(" Blk/Wave Size: %d, %d, M/N/K PerBlk: %d, %d, %d, M/N/K PerXdl: %d, %d, %d\n",
BlockSize,
WaveSize,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
KPerXDL);
printf(" A/B buffer load inst: %d, %d\n A/B LDS write inst: %d, %d\n A/B LDS read inst: "
"%d, %d\n C MFMA inst: %d\n",
A_Buffer_Load_Inst_Num,
B_Buffer_Load_Inst_Num,
A_LDS_Write_Inst_Num,
B_LDS_Write_Inst_Num,
A_LDS_Read_Inst_Num,
B_LDS_Read_Inst_Num,
C_MFMA_Inst_Num);
}
};
} // namespace ck
include/ck/utility/data_type.hpp
View file @ f83e9701
......@@ -163,6 +163,13 @@ struct scalar_type<bf8_t>
static constexpr index_t vector_size = 1;
};
template <>
struct scalar_type<bool>
{
using type = bool;
static constexpr index_t vector_size = 1;
};
template <typename T>
struct vector_type<T, 1>
{
......
include/ck/utility/synchronization.hpp
View file @ f83e9701
......@@ -10,10 +10,12 @@ namespace ck {
__device__ void block_sync_lds()
{
#if CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
asm volatile("\
s_waitcnt lgkmcnt(0) \n \
s_barrier \
" ::);
// asm volatile("\
// s_waitcnt lgkmcnt(0) \n \
// s_barrier \
// " ::);
__builtin_amdgcn_s_waitcnt(0xc07f);
__builtin_amdgcn_s_barrier();
#else
__syncthreads();
#endif
......
include/ck/utility/transpose_vectors.hpp
View file @ f83e9701
......@@ -162,4 +162,83 @@ struct transpose_vectors<int8_t, NX, NY>
}
};
// transpose f8 4x4
__device__ void transpose_f8_4x4(const f8x4_t& x0,
const f8x4_t& x1,
const f8x4_t& x2,
const f8x4_t& x3,
f8x4_t& y0,
f8x4_t& y1,
f8x4_t& y2,
f8x4_t& y3)
{
int32_t t0, t1;
int32_t z0, z1, z2, z3;
constexpr int32_t m0 = 0x05010400;
constexpr int32_t m1 = 0x05040100;
constexpr int32_t m2 = 0x07060302;
constexpr int32_t m3 = 0x07030602;
// ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
// -- -- -- -- -- -- -- -- - - - -
// index 7 6 5 4 3 2 1 0 33 77 44 88
// index is reversed because of little endianness (least significant bits first)
t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m0);
t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m0);
z0 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
z1 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m3);
t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m3);
z2 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
z3 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
y0 = bit_cast<f8x4_t>(z0);
y1 = bit_cast<f8x4_t>(z1);
y2 = bit_cast<f8x4_t>(z2);
y3 = bit_cast<f8x4_t>(z3);
}
template <index_t NX, index_t NY>
struct transpose_vectors<f8_t, NX, NY>
{
// we got [NY * NX] amount of S data to be transposed
static constexpr index_t s_per_x = NY;
static constexpr index_t s_per_y = NX;
using S = f8_t;
using VX = vector_type<f8_t, s_per_x>;
using VY = vector_type<f8_t, s_per_y>;
__device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
StaticallyIndexedArray<VY&, NY>& vy_tuple)
{
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static_assert((NX % 4 == 0 && NY % 4 == 0), "wrong!");
// loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
static_for<0, NY, 4>{}([&](auto iy) {
static_for<0, NX, 4>{}([&](auto ix) {
// reference to 4 f8 data from vx_tuple
const auto& x_s4_0 = vx_tuple[ix].template AsType<f8x4_t>()[iy / I4];
const auto& x_s4_1 = vx_tuple[ix + I1].template AsType<f8x4_t>()[iy / I4];
const auto& x_s4_2 = vx_tuple[ix + I2].template AsType<f8x4_t>()[iy / I4];
const auto& x_s4_3 = vx_tuple[ix + I3].template AsType<f8x4_t>()[iy / I4];
// reference to 4 f8 data from vy_tuple
auto& y_s4_0 = vy_tuple(iy).template AsType<f8x4_t>()(ix / I4);
auto& y_s4_1 = vy_tuple(iy + I1).template AsType<f8x4_t>()(ix / I4);
auto& y_s4_2 = vy_tuple(iy + I2).template AsType<f8x4_t>()(ix / I4);
auto& y_s4_3 = vy_tuple(iy + I3).template AsType<f8x4_t>()(ix / I4);
// transpose
transpose_f8_4x4(x_s4_0, x_s4_1, x_s4_2, x_s4_3, y_s4_0, y_s4_1, y_s4_2, y_s4_3);
});
});
}
};
} // namespace ck
include/ck/utility/type.hpp
View file @ f83e9701
......@@ -43,6 +43,8 @@ __host__ __device__ constexpr Y bit_cast(const X& x)
#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST
Y y;
// auto t = reinterpret_cast<const Y*>(&x);
// y = *t;
__builtin_memcpy(&y, &x, sizeof(X));
return y;
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ f83e9701
......@@ -21,7 +21,7 @@ template <typename ADataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename ComputeTypeA = ADataType,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA>
struct ReferenceGemm : public device::BaseOperator
{
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmV2<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceGemmV2<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/utility/host_tensor_generator.hpp
View file @ f83e9701
......@@ -31,6 +31,18 @@ struct GeneratorTensor_1
}
};
template <>
struct GeneratorTensor_1<ck::half_t>
{
float value = 1.0;
template <typename... Is>
ck::bhalf_t operator()(Is...)
{
return ck::type_convert<ck::half_t>(value);
}
};
template <>
struct GeneratorTensor_1<ck::bhalf_t>
{
......@@ -43,6 +55,20 @@ struct GeneratorTensor_1<ck::bhalf_t>
}
};
#if defined CK_ENABLE_FP8
template <>
struct GeneratorTensor_1<ck::f8_t>
{
float value = 1.0;
template <typename... Is>
ck::bhalf_t operator()(Is...)
{
return ck::type_convert<ck::f8_t>(value);
}
};
#endif
template <>
struct GeneratorTensor_1<int8_t>
{
......
library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt 0 → 100644
View file @ f83e9701
# ONLY XDL_KERNELS
set(GEMM_UNIVERSAL_INSTANCES)
list(APPEND GEMM_UNIVERSAL_INSTANCES
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp
)
add_instance_library(device_gemm_universal_instance ${GEMM_UNIVERSAL_INSTANCES})
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = half_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 8, 4, 32, 32, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 32, 8, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 128, 32, 8, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 4, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
// Memory friendly
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 32, 64, 8, 2, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 32, 64, 8, 4, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 32, 64, 8, 4, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 8, 4, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 4, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 8, 4, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 8, 4, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 8, 4, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 64, 64, 8, 4, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 8, 4, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 128, 64, 8, 4, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 8, 4, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
DeviceGemm_Xdl_CShuffleV3< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 64, 8, 4, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances<GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances<GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances<GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances<GemmMNPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave, GemmKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp 0 → 100644
View file @ f83e9701
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave, GemmMNKPadding>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment