Skip to content

GitLab

Commit 6e3c786e authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents 1bb510cb 261f1759
Hide whitespace changes
Inline Side-by-side
Showing with 1170 additions and 31 deletions
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp 0 → 100644
View file @ 6e3c786e
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
template <typename Problem, typename Policy>
struct GemmPipelineAgBgCrImplBase
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t MPerBlock = BlockGemmShape::kM;
static constexpr index_t NPerBlock = BlockGemmShape::kN;
static constexpr index_t KPerBlock = BlockGemmShape::kK;
template <typename DstBlockTile, typename SrcTileWindow>
CK_TILE_DEVICE void GlobalPrefetch(DstBlockTile& dst_block_tile,
SrcTileWindow& dram_tile_window) const
{
load_tile(dst_block_tile, dram_tile_window);
move_tile_window(dram_tile_window, {0, KPerBlock});
}
template <typename DstTileWindow, typename SrcBlockTile, typename ElementFunction>
CK_TILE_DEVICE void LocalPrefill(DstTileWindow& lds_tile_window,
const SrcBlockTile& src_block_tile,
const ElementFunction& element_func) const
{
const auto block_tile_tmp = tile_elementwise_in(element_func, src_block_tile);
store_tile(lds_tile_window, block_tile_tmp);
}
CK_TILE_DEVICE auto GetABLdsTensorViews(void* p_smem) const
{
// A tile in LDS
ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
// TODO: LDS alignment should come from Policy!
constexpr index_t a_lds_block_space_size_aligned =
integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
16;
// B tile in LDS
BDataType* p_b_lds = static_cast<BDataType*>(
static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);
return make_tuple(std::move(a_lds_block), std::move(b_lds_block));
}
template <typename ADramBlockWindowTmp, typename ALdsTensorView>
CK_TILE_DEVICE auto GetAWindows(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const ALdsTensorView& a_lds_block_view) const
{
// A DRAM tile window for load
auto a_copy_dram_window =
make_tile_window(a_dram_block_window_tmp.get_bottom_tensor_view(),
make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
a_dram_block_window_tmp.get_window_origin(),
Policy::template MakeADramTileDistribution<Problem>());
// A LDS tile window for store
auto a_copy_lds_window =
make_tile_window(a_lds_block_view,
make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
{0, 0},
a_copy_dram_window.get_tile_distribution());
auto a_lds_gemm_window = make_tile_window(
a_lds_block_view, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
return make_tuple(std::move(a_copy_dram_window),
std::move(a_copy_lds_window),
std::move(a_lds_gemm_window));
}
template <typename BDramBlockWindowTmp, typename BLdsTensorView>
CK_TILE_DEVICE auto GetBWindows(const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BLdsTensorView& b_lds_block_view) const
{
auto b_copy_dram_window =
make_tile_window(b_dram_block_window_tmp.get_bottom_tensor_view(),
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
b_dram_block_window_tmp.get_window_origin(),
Policy::template MakeBDramTileDistribution<Problem>());
// B LDS tile window for store
auto b_copy_lds_window =
make_tile_window(b_lds_block_view,
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
{0, 0},
b_copy_dram_window.get_tile_distribution());
auto b_lds_gemm_window = make_tile_window(
b_lds_block_view, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
return make_tuple(std::move(b_copy_dram_window),
std::move(b_copy_lds_window),
std::move(b_lds_gemm_window));
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp 0 → 100644
View file @ 6e3c786e
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
namespace ck_tile {
// A Tile Window: global memory
// B Tile Window: global memory
// C Distributed tensor: register
template <typename Problem>
struct BaseGemmPipelineAgBgCrCompV3
{
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
CK_TILE_HOST static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
CK_TILE_HOST static constexpr TailNumber GetBlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
}
};
// Compute optimized pipeline
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy>
struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
{
using Base = BaseGemmPipelineAgBgCrCompV3<Problem>;
using PipelineImplBase = GemmPipelineAgBgCrImplBase<Problem, Policy>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using CDataType = remove_cvref_t<typename Problem::CDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using CLayout = remove_cvref_t<typename Problem::CLayout>;
using BlockGemm = remove_cvref_t<decltype(Policy::template GetBlockGemm<Problem>())>;
using I0 = number<0>;
using I1 = number<1>;
using I2 = number<2>;
static constexpr index_t BlockSize = Problem::kBlockSize;
static constexpr index_t MPerBlock = BlockGemmShape::kM;
static constexpr index_t NPerBlock = BlockGemmShape::kN;
static constexpr index_t KPerBlock = BlockGemmShape::kK;
static constexpr index_t VectorSizeA = Problem::VectorSizeA;
static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
// Where is the right place for HasHotLoop and TailNum ???
static constexpr bool HasHotLoop = Problem::HasHotLoop;
static constexpr auto TailNum = Problem::TailNum;
static constexpr auto Scheduler = Problem::Scheduler;
using Base::PrefetchStages;
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <GemmPipelineScheduler Scheduler>
struct PipelineImpl : public PipelineImplBase
{
};
template <>
struct PipelineImpl<GemmPipelineScheduler::Intrawave> : public PipelineImplBase
{
using Base = PipelineImplBase;
CK_TILE_DEVICE static constexpr auto HotLoopScheduler()
{
constexpr index_t MPerXDL = BlockGemmShape::WarpTile::at(I0{});
constexpr index_t NPerXDL = BlockGemmShape::WarpTile::at(I1{});
constexpr index_t KPerXDL = BlockGemmShape::WarpTile::at(I2{});
constexpr index_t WaveSize = 64;
constexpr index_t WaveNumM = BlockGemmShape::BlockWarps::at(I0{});
constexpr index_t WaveNumN = BlockGemmShape::BlockWarps::at(I1{});
constexpr index_t A_LDS_Read_Width = KPerXDL;
constexpr index_t B_LDS_Read_Width = KPerXDL;
constexpr index_t A_Buffer_Load_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * VectorSizeA);
constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * VectorSizeB);
constexpr index_t A_LDS_Write_Inst_Num = MPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t B_LDS_Write_Inst_Num = NPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t A_LDS_Read_Inst_Num =
WaveNumN * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t B_LDS_Read_Inst_Num =
WaveNumM * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock /
(BlockSize / WaveSize) /
(MPerXDL * NPerXDL * KPerXDL);
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a = A_LDS_Read_Width * sizeof(ADataType) == 16
? A_LDS_Read_Inst_Num
: A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_read_inst_b = B_LDS_Read_Width * sizeof(BDataType) == 16
? B_LDS_Read_Inst_Num
: B_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_write_inst_a = A_LDS_Write_Inst_Num;
constexpr auto num_ds_write_inst_b = B_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = B_Buffer_Load_Inst_Num;
constexpr auto num_mfma_inst = C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle =
A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_b_issue_cycle =
B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto ds_read_b_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
constexpr auto num_dsread_a_mfma =
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
constexpr auto num_dsread_b_mfma =
(num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
// stage 1
// Separate this part?
// constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
// sizeof(ComputeDataType) /
// sizeof(BDataType)
// ? sizeof(ComputeDataType) /
// sizeof(ADataType) : sizeof(ComputeDataType)
// / sizeof(BDataType);
constexpr auto num_mfma_stage1 =
num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
constexpr auto num_mfma_per_issue =
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
});
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
});
// stage 2
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
ds_read_a_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x100,
num_ds_read_inst_a - (num_dsread_a_mfma - 1) * ds_read_a_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
ds_read_b_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x100,
num_ds_read_inst_b - (num_dsread_b_mfma - 1) * ds_read_b_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
}
template <bool HasHotLoop,
TailNumber TailNum,
typename ADramBlockWindowTmp,
typename BDramBlockWindowTmp,
typename AElementFunction,
typename BElementFunction>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const AElementFunction& a_element_func,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BElementFunction& b_element_func,
index_t num_loop,
void* p_smem) const
{
static_assert(
std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
std::is_same_v<BDataType,
remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
"A/B Dram block window should have the same data type as appropriate "
"([A|B]DataType) defined in Problem definition!");
static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
"A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
" or KPerBlock!");
// ------------------------------------------------------------------------------------
// Definitions of all needed tiles
// A/B tiles in LDS
auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
// A DRAM tile window for load
// A LDS tile window for store
// A LDS tile for block GEMM
auto&& [a_copy_dram_window, a_copy_lds_window, a_lds_gemm_window] =
Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
// B DRAM tile window for load
// B LDS tile window for store
// B LDS tile for block GEMM
auto&& [b_copy_dram_window, b_copy_lds_window, b_lds_gemm_window] =
Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
// Block GEMM
auto block_gemm = BlockGemm();
auto c_block_tile = block_gemm.MakeCBlockTile();
using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
using BBlockTileDistr = decltype(b_copy_dram_window.get_tile_distribution());
using ABlockTile =
decltype(make_static_distributed_tensor<ADataType>(ABlockTileDistr{}));
using BBlockTile =
decltype(make_static_distributed_tensor<BDataType>(BBlockTileDistr{}));
ABlockTile a_block_tile;
BBlockTile b_block_tile;
// -----------------------------------------------------------------------------------------
// Gemm pipeline start
// prefetch
// global read 0
Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
// initialize C
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0
Base::LocalPrefill(a_copy_lds_window, a_block_tile, a_element_func);
Base::LocalPrefill(b_copy_lds_window, b_block_tile, b_element_func);
Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasHotLoop)
{
index_t i = 0;
do
{
block_sync_lds();
Base::LocalPrefill(a_copy_lds_window, a_block_tile, a_element_func);
Base::LocalPrefill(b_copy_lds_window, b_block_tile, b_element_func);
Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
}
// Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
// latency
// __builtin_amdgcn_sched_barrier(0);
return c_block_tile;
}
};
template <typename ADramBlockWindowTmp,
typename BDramBlockWindowTmp,
typename AElementFunction,
typename BElementFunction>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const AElementFunction& a_element_func,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BElementFunction& b_element_func,
index_t num_loop,
void* p_smem) const
{
return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
a_dram_block_window_tmp,
a_element_func,
b_dram_block_window_tmp,
b_element_func,
num_loop,
p_smem);
}
template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
index_t num_loop,
void* p_smem) const
{
return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
a_dram_block_window_tmp,
[](const ADataType& a) { return a; },
b_dram_block_window_tmp,
[](const BDataType& b) { return b; },
num_loop,
p_smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp 0 → 100644
View file @ 6e3c786e
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
namespace ck_tile {
// A Tile Window: global memory
// B Tile Window: global memory
// C Distributed tensor: register
template <typename Problem>
struct BaseGemmPipelineAgBgCrMem
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t BlockSize = Problem::kBlockSize;
static constexpr index_t MPerBlock = BlockGemmShape::kM;
static constexpr index_t NPerBlock = BlockGemmShape::kN;
static constexpr index_t KPerBlock = BlockGemmShape::kK;
// TODO: Is this 32K value gfx9 arch specific?
static constexpr index_t MinMemInFlyBytes = 32768;
static constexpr index_t WgpPerCU =
(4 * get_warp_size() / BlockSize) >= 1 ? 4 * get_warp_size() / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = integer_divide_ceil(
MinMemInFlyBytes / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t LocalPrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
CK_TILE_HOST static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
CK_TILE_HOST static constexpr TailNumber GetBlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
return TailNumber::One;
}
else if(num_loop % PrefetchStages == 2)
{
return TailNumber::Two;
}
else if(num_loop % PrefetchStages == 3)
{
return TailNumber::Three;
}
else if(num_loop % PrefetchStages == 4)
{
return TailNumber::Four;
}
else if(num_loop % PrefetchStages == 5)
{
return TailNumber::Five;
}
else if(num_loop % PrefetchStages == 6)
{
return TailNumber::Six;
}
else if(num_loop % PrefetchStages == 7)
{
return TailNumber::Seven;
}
else
{
return TailNumber::Full;
}
}
};
// Maximum Global Memory throughput pipeline with >=32KB data in fly
// GlobalPrefetchStages: >=2
// LocalPreFillStages: 1
// LocalPreFetchStages: 0
// LocalSharedMemoryBuffer: 1
template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy>
struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
{
using Base = BaseGemmPipelineAgBgCrMem<Problem>;
using PipelineImplBase = GemmPipelineAgBgCrImplBase<Problem, Policy>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using CDataType = remove_cvref_t<typename Problem::CDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using CLayout = remove_cvref_t<typename Problem::CLayout>;
using BlockGemm = remove_cvref_t<decltype(Policy::template GetBlockGemm<Problem>())>;
using I0 = number<0>;
using I1 = number<1>;
using I2 = number<2>;
static constexpr index_t MPerBlock = BlockGemmShape::kM;
static constexpr index_t NPerBlock = BlockGemmShape::kN;
static constexpr index_t KPerBlock = BlockGemmShape::kK;
static constexpr index_t VectorSizeA = Problem::VectorSizeA;
static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
// Where is the right place for HasHotLoop and TailNum ???
static constexpr bool HasHotLoop = Problem::HasHotLoop;
static constexpr auto TailNum = Problem::TailNum;
static constexpr auto Scheduler = Problem::Scheduler;
using Base::PrefetchStages;
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <GemmPipelineScheduler Scheduler>
struct PipelineImpl : public PipelineImplBase
{
};
template <>
struct PipelineImpl<GemmPipelineScheduler::Intrawave> : public PipelineImplBase
{
using Base = PipelineImplBase;
template <bool HasHotLoop,
TailNumber TailNum,
typename ADramBlockWindowTmp,
typename BDramBlockWindowTmp,
typename AElementFunction,
typename BElementFunction>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const AElementFunction& a_element_func,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BElementFunction& b_element_func,
index_t num_loop,
void* p_smem) const
{
static_assert(
std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
std::is_same_v<BDataType,
remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
"A/B Dram block window should have the same data type as appropriate "
"([A|B]DataType) defined in Problem definition!");
static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
"A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
" or KPerBlock!");
// ------------------------------------------------------------------------------------
// Definitions of all needed tiles
// A/B tiles in LDS
// With c++20 could simplify to below line.
// Currently get error: captured structured bindings are a C++20 extension
// auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
auto ab_lds_blocks = Base::GetABLdsTensorViews(p_smem);
auto& a_lds_block = ab_lds_blocks.at(I0{});
auto& b_lds_block = ab_lds_blocks.at(I1{});
// A DRAM tile window for load
// A LDS tile window for store
// A LDS tile for block GEMM
auto a_windows = Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
auto& a_copy_dram_window = a_windows.at(I0{});
auto& a_copy_lds_window = a_windows.at(I1{});
auto& a_lds_gemm_window = a_windows.at(I2{});
// B DRAM tile window for load
// B LDS tile window for store
// B LDS tile for block GEMM
auto b_windows = Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
auto& b_copy_dram_window = b_windows.at(I0{});
auto& b_copy_lds_window = b_windows.at(I1{});
auto& b_lds_gemm_window = b_windows.at(I2{});
// Block GEMM
auto block_gemm = BlockGemm();
auto c_block_tile = block_gemm.MakeCBlockTile();
using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
using BBlockTileDistr = decltype(b_copy_dram_window.get_tile_distribution());
using ABlockTile =
decltype(make_static_distributed_tensor<ADataType>(ABlockTileDistr{}));
using BBlockTile =
decltype(make_static_distributed_tensor<BDataType>(BBlockTileDistr{}));
tuple_array<ABlockTile, PrefetchStages> a_block_tiles;
tuple_array<BBlockTile, PrefetchStages> b_block_tiles;
// -----------------------------------------------------------------------------------------
// Gemm pipeline start
// prefetch
// global read 0
Base::GlobalPrefetch(a_block_tiles.get(I0{}), a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(I0{}), b_copy_dram_window);
// initialize C
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0
Base::LocalPrefill(a_copy_lds_window, a_block_tiles.get(I0{}), a_element_func);
Base::LocalPrefill(b_copy_lds_window, b_block_tiles.get(I0{}), b_element_func);
// Global prefetch [1, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto prefetch_idx) {
Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}), a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}), b_copy_dram_window);
});
// main body
if constexpr(HasHotLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto prefetch_idx) {
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_sync_lds();
Base::LocalPrefill(
a_copy_lds_window,
a_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
a_element_func);
Base::LocalPrefill(
b_copy_lds_window,
b_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
b_element_func);
Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}),
a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}),
b_copy_dram_window);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
auto HotLoopTail = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto prefetch_idx) {
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
block_sync_lds();
Base::LocalPrefill(a_copy_lds_window,
a_block_tiles.get(number<prefetch_idx>{}),
a_element_func);
Base::LocalPrefill(b_copy_lds_window,
b_block_tiles.get(number<prefetch_idx>{}),
b_element_func);
});
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
};
if constexpr(TailNum == TailNumber::One)
{
block_sync_lds();
block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
}
else if constexpr(TailNum == TailNumber::Two)
{
HotLoopTail(number<2>{});
}
else if constexpr(TailNum == TailNumber::Three)
{
HotLoopTail(number<3>{});
}
else if constexpr(TailNum == TailNumber::Four)
{
HotLoopTail(number<4>{});
}
else if constexpr(TailNum == TailNumber::Five)
{
HotLoopTail(number<5>{});
}
else if constexpr(TailNum == TailNumber::Six)
{
HotLoopTail(number<6>{});
}
else if constexpr(TailNum == TailNumber::Seven)
{
HotLoopTail(number<7>{});
}
else if constexpr(TailNum == TailNumber::Full)
{
HotLoopTail(number<PrefetchStages>{});
}
return c_block_tile;
}
};
template <>
struct PipelineImpl<GemmPipelineScheduler::Interwave> : public PipelineImplBase
{
using Base = PipelineImplBase;
template <bool HasHotLoop,
TailNumber TailNum,
typename ADramBlockWindowTmp,
typename BDramBlockWindowTmp,
typename AElementFunction,
typename BElementFunction>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const AElementFunction& a_element_func,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BElementFunction& b_element_func,
index_t num_loop,
void* p_smem) const
{
static_assert(
std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
std::is_same_v<BDataType,
remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
"A/B Dram block window should have the same data type as appropriate "
"([A|B]DataType) defined in Problem definition!");
static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
"A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
" or KPerBlock!");
// ------------------------------------------------------------------------------------
// Definitions of all needed tiles
// A/B tiles in LDS
// With c++20 could simplify to below line.
// Currently get error: captured structured bindings are a C++20 extension
// auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
auto ab_lds_blocks = Base::GetABLdsTensorViews(p_smem);
auto& a_lds_block = ab_lds_blocks.at(I0{});
auto& b_lds_block = ab_lds_blocks.at(I1{});
// A DRAM tile window for load
// A LDS tile window for store
// A LDS tile for block GEMM
auto a_windows = Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
auto& a_copy_dram_window = a_windows.at(I0{});
auto& a_copy_lds_window = a_windows.at(I1{});
auto& a_lds_gemm_window = a_windows.at(I2{});
// B DRAM tile window for load
// B LDS tile window for store
// B LDS tile for block GEMM
auto b_windows = Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
auto& b_copy_dram_window = b_windows.at(I0{});
auto& b_copy_lds_window = b_windows.at(I1{});
auto& b_lds_gemm_window = b_windows.at(I2{});
// Block GEMM
auto block_gemm = BlockGemm();
auto c_block_tile = block_gemm.MakeCBlockTile();
using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
using BBlockTileDistr = decltype(b_copy_dram_window.get_tile_distribution());
using ABlockTile =
decltype(make_static_distributed_tensor<ADataType>(ABlockTileDistr{}));
using BBlockTile =
decltype(make_static_distributed_tensor<BDataType>(BBlockTileDistr{}));
tuple_array<ABlockTile, PrefetchStages> a_block_tiles;
tuple_array<BBlockTile, PrefetchStages> b_block_tiles;
// -----------------------------------------------------------------------------------------
// Gemm pipeline start
// prefetch
// global read 0
Base::GlobalPrefetch(a_block_tiles.get(I0{}), a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(I0{}), b_copy_dram_window);
// initialize C
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0
Base::LocalPrefill(a_copy_lds_window, a_block_tiles.get(I0{}), a_element_func);
Base::LocalPrefill(b_copy_lds_window, b_block_tiles.get(I0{}), b_element_func);
// Global prefetch [1, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto prefetch_idx) {
Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}), a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}), b_copy_dram_window);
});
// main body
if constexpr(HasHotLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto prefetch_idx) {
block_sync_lds();
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
// no second block_sync_lds because it's interwave
Base::LocalPrefill(
a_copy_lds_window,
a_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
a_element_func);
Base::LocalPrefill(
b_copy_lds_window,
b_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
b_element_func);
Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}),
a_copy_dram_window);
Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}),
b_copy_dram_window);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
auto HotLoopTail = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto prefetch_idx) {
block_sync_lds();
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
// no second block_sync_lds because it's interwave
Base::LocalPrefill(a_copy_lds_window,
a_block_tiles.get(number<prefetch_idx>{}),
a_element_func);
Base::LocalPrefill(b_copy_lds_window,
b_block_tiles.get(number<prefetch_idx>{}),
b_element_func);
});
block_sync_lds();
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
};
if constexpr(TailNum == TailNumber::One)
{
block_sync_lds();
block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
}
else if constexpr(TailNum == TailNumber::Two)
{
HotLoopTail(number<2>{});
}
else if constexpr(TailNum == TailNumber::Three)
{
HotLoopTail(number<3>{});
}
else if constexpr(TailNum == TailNumber::Four)
{
HotLoopTail(number<4>{});
}
else if constexpr(TailNum == TailNumber::Five)
{
HotLoopTail(number<5>{});
}
else if constexpr(TailNum == TailNumber::Six)
{
HotLoopTail(number<6>{});
}
else if constexpr(TailNum == TailNumber::Seven)
{
HotLoopTail(number<7>{});
}
else if constexpr(TailNum == TailNumber::Full)
{
HotLoopTail(number<PrefetchStages>{});
}
return c_block_tile;
}
};
template <typename ADramBlockWindowTmp,
typename BDramBlockWindowTmp,
typename AElementFunction,
typename BElementFunction>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const AElementFunction& a_element_func,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const BElementFunction& b_element_func,
index_t num_loop,
void* p_smem) const
{
return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
a_dram_block_window_tmp,
a_element_func,
b_dram_block_window_tmp,
b_element_func,
num_loop,
p_smem);
}
template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
index_t num_loop,
void* p_smem) const
{
return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
a_dram_block_window_tmp,
[](const ADataType& a) { return a; },
b_dram_block_window_tmp,
[](const BDataType& b) { return b; },
num_loop,
p_smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp 0 → 100644
View file @ 6e3c786e
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <ostream>
#include "ck_tile/core.hpp"
namespace ck_tile {
enum struct GemmPipelineScheduler
{
Default,
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,
};
} // namespace ck_tile
inline std::ostream& operator<<(std::ostream& os, const ck_tile::GemmPipelineScheduler& s)
{
switch(s)
{
case ck_tile::GemmPipelineScheduler::Default: os << "Default"; break;
case ck_tile::GemmPipelineScheduler::Intrawave: os << "Intrawave"; break;
case ck_tile::GemmPipelineScheduler::Interwave: os << "Interwave"; break;
default: os << "";
}
return os;
}
inline std::ostream& operator<<(std::ostream& os, const ck_tile::TailNumber& s)
{
switch(s)
{
case ck_tile::TailNumber::Odd: os << "Odd"; break;
case ck_tile::TailNumber::Even: os << "Even"; break;
case ck_tile::TailNumber::One: os << "One"; break;
case ck_tile::TailNumber::Two: os << "Two"; break;
case ck_tile::TailNumber::Three: os << "Three"; break;
case ck_tile::TailNumber::Four: os << "Four"; break;
case ck_tile::TailNumber::Five: os << "Five"; break;
case ck_tile::TailNumber::Six: os << "Six"; break;
case ck_tile::TailNumber::Seven: os << "Seven"; break;
case ck_tile::TailNumber::Empty: os << "Empty"; break;
case ck_tile::TailNumber::Full: os << "Full"; break;
default: os << "";
}
return os;
}
include/ck_tile/ops/gemm/pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp → include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp
View file @ 6e3c786e
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include "ck_tile/core.hpp" #include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
namespace ck_tile { namespace ck_tile {
// A Tile Window: global memory // A Tile Window: global memory
// B Tile Window: global memory // B Tile Window: global memory
// C Distributed tensor: register // C Distributed tensor: register
template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy> template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy>
struct BlockGemmPipelineAGmemBGmemCRegV1 struct GemmPipelineAGmemBGmemCRegV1
{ {
using ADataType = remove_cvref_t<typename Problem::ADataType>; using ADataType = remove_cvref_t<typename Problem::ADataType>;
using BDataType = remove_cvref_t<typename Problem::BDataType>; using BDataType = remove_cvref_t<typename Problem::BDataType>;
using CDataType = remove_cvref_t<typename Problem::CDataType>; using CDataType = remove_cvref_t<typename Problem::CDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>; using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t kBlockSize = Problem::kBlockSize; using ALayout = remove_cvref_t<typename Problem::ALayout>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using CLayout = remove_cvref_t<typename Problem::CLayout>;
static constexpr index_t BlockSize = Problem::kBlockSize;
static constexpr index_t kMPerBlock = BlockGemmShape::kM; static constexpr index_t kMPerBlock = BlockGemmShape::kM;
static constexpr index_t kNPerBlock = BlockGemmShape::kN; static constexpr index_t kNPerBlock = BlockGemmShape::kN;
static constexpr index_t kKPerBlock = BlockGemmShape::kK; static constexpr index_t kKPerBlock = BlockGemmShape::kK;
static constexpr index_t AlignmentA = Problem::AlignmentA; static constexpr index_t VectorSizeA = Problem::VectorSizeA;
static constexpr index_t AlignmentB = Problem::AlignmentB; static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t AlignmentC = Problem::AlignmentC; static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadA = Problem::kPadA; static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadB = Problem::kPadB; static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadC = Problem::kPadC; static constexpr bool kPadK = Problem::kPadK;
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetStaticLdsSize() CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{ {
return ck_tile::integer_divide_ceil( return integer_divide_ceil(
sizeof(ADataType) * sizeof(ADataType) *
Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(), Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(),
16) * 16) *
...@@ -44,7 +48,7 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -44,7 +48,7 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size(); Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
} }
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{ {
return Policy::template GetSmemSize<Problem>(); return Policy::template GetSmemSize<Problem>();
} }
...@@ -97,11 +101,8 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -97,11 +101,8 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
Policy::template MakeADramTileDistribution<Problem>()); Policy::template MakeADramTileDistribution<Problem>());
// A LDS tile window for store // A LDS tile window for store
auto a_copy_lds_window = auto a_copy_lds_window = make_tile_window(
make_tile_window(a_lds_block, a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
a_copy_dram_window.get_tile_distribution());
// B DRAM tile window for load // B DRAM tile window for load
auto b_copy_dram_window = auto b_copy_dram_window =
...@@ -111,11 +112,8 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -111,11 +112,8 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
Policy::template MakeBDramTileDistribution<Problem>()); Policy::template MakeBDramTileDistribution<Problem>());
// B LDS tile window for store // B LDS tile window for store
auto b_copy_lds_window = auto b_copy_lds_window = make_tile_window(
make_tile_window(b_lds_block, b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
b_copy_dram_window.get_tile_distribution());
// A LDS tile for block GEMM // A LDS tile for block GEMM
auto a_lds_gemm_window = make_tile_window( auto a_lds_gemm_window = make_tile_window(
...@@ -126,7 +124,7 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -126,7 +124,7 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0}); b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
// Block GEMM // Block GEMM
constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>(); auto block_gemm = Policy::template GetBlockGemm<Problem>();
// Acc register tile // Acc register tile
auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){}; auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
...@@ -145,12 +143,32 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -145,12 +143,32 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile); tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0 // LDS write 0
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile); if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>)
store_tile(a_copy_lds_window, a_block_tile_tmp); {
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegBlockDescriptor<Problem>());
shuffle_tile(a_shuffle_tmp, a_block_tile);
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_shuffle_tmp);
store_tile(a_copy_lds_window, a_block_tile_tmp);
}
else
{
store_tile(a_copy_lds_window, tile_elementwise_in(a_element_func, a_block_tile));
}
// LDS write 0 // LDS write 0
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile); if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
store_tile(b_copy_lds_window, b_block_tile_tmp); {
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp, b_block_tile);
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_shuffle_tmp);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
else
{
store_tile(b_copy_lds_window, tile_elementwise_in(b_element_func, b_block_tile));
}
} }
index_t iCounter = num_loop - 1; index_t iCounter = num_loop - 1;
...@@ -176,8 +194,19 @@ struct BlockGemmPipelineAGmemBGmemCRegV1 ...@@ -176,8 +194,19 @@ struct BlockGemmPipelineAGmemBGmemCRegV1
store_tile(a_copy_lds_window, a_block_tile_tmp); store_tile(a_copy_lds_window, a_block_tile_tmp);
// LDS write i + 1 // LDS write i + 1
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile); if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
store_tile(b_copy_lds_window, b_block_tile_tmp); {
auto b_shuffle_tmp_loop = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp_loop, b_block_tile);
store_tile(b_copy_lds_window,
tile_elementwise_in(b_element_func, b_shuffle_tmp_loop));
}
else
{
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
iCounter--; iCounter--;
} }
......
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