Skip to content

GitLab

Unverified Commit b8d11559 authored by amd-khushbu's avatar amd-khushbu Committed by GitHub
Browse files

Merge branch 'develop' into ck_profiler_m_instances

parents 7f3fe4e7 3b230208
Hide whitespace changes
Inline Side-by-side
Showing with 1009 additions and 151 deletions
include/ck_tile/ops/gemm/kernel/gemm_kernel.hpp
View file @ b8d11559
......@@ -8,6 +8,7 @@
#include "ck_tile/core.hpp"
#include "ck_tile/ops/common.hpp"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -75,6 +76,13 @@ struct GemmKernel
static constexpr auto I1 = number<1>();
static constexpr auto I2 = number<2>();
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "gemm", gemm_prec_str<ADataType, BDataType>, GemmPipeline::GetName());
// clang-format on
}
CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N, index_t KBatch)
{
return dim3(TilePartitioner::GridSize(M, N), 1, KBatch);
......@@ -455,7 +463,9 @@ struct GemmKernel
* @param a_ptr input A pointer
* @param b_ptr input B pointer
* @param c_ptr output C pointer
* @param smem_ptr_0 The start memory pointer of the shared memory block.
* @param kargs GEMM kernel arguments
* @param splitk_batch_offset splitk_batch_offset Utility structure used to calculate k batch.
* @param block_idx_m The GEMM's output M dimension tile index processed by this workgroup.
* @param block_idx_n The GEMM's output N dimension tile index processed by this workgroup.
*
......@@ -465,7 +475,7 @@ struct GemmKernel
CK_TILE_DEVICE static void RunGemm(const ADataType* a_ptr,
const BDataType* b_ptr,
CDataType* c_ptr,
void* smem_ptr,
void* smem_ptr_0,
const GemmKernelArgs& kargs,
const SplitKBatchOffset& splitk_batch_offset,
const index_t block_idx_m,
......@@ -483,15 +493,67 @@ struct GemmKernel
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
const auto& b_block_window = gemm_tile_windows.at(I1);
const auto& c_block_tile =
GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
const auto& c_block_tile = GemmPipeline{}.template operator()(
a_block_window, b_block_window, num_loop, smem_ptr_0);
// Run Epilogue Pipeline
auto& c_block_window = gemm_tile_windows.at(I2);
EpiloguePipeline{}
.template operator()<decltype(c_block_window), decltype(c_block_tile), DstInMemOp>(
c_block_window, c_block_tile, smem_ptr);
c_block_window, c_block_tile, smem_ptr_0);
}
/**
* @brief Runs single GEMM problem cooperatively by whole workgroup.
*
* @note RunGEMM2LDS in with two shared memory buffers using the ping pong buffer mechanism.
*
* @param a_ptr input A pointer
* @param b_ptr input B pointer
* @param c_ptr output C pointer
* @param smem_ptr_0 The starting pointer of 1st shared memory block.
* @param smem_ptr_1 The starting pointer of 2nd shared memory block.
* @param kargs GEMM kernel arguments
* @param splitk_batch_offset Utility structure used to calculate k batch.
* @param block_idx_m The GEMM's output M dimension tile index processed by this workgroup.
* @param block_idx_n The GEMM's output N dimension tile index processed by this workgroup.
*
* @tparam DstInMemOp Destination memory operation (default: set).
*/
template <memory_operation_enum DstInMemOp = memory_operation_enum::set>
CK_TILE_DEVICE static void RunGemm2LDS(const ADataType* a_ptr,
const BDataType* b_ptr,
CDataType* c_ptr,
void* __restrict__ smem_ptr_0,
void* __restrict__ smem_ptr_1,
const GemmKernelArgs& kargs,
const SplitKBatchOffset& splitk_batch_offset,
const index_t block_idx_m,
const index_t block_idx_n)
{
// Create Gemm tensor views, pad views and tile windows
const auto& gemm_tensor_views_tuple =
MakeGemmTensorViews<DstInMemOp>(a_ptr, b_ptr, c_ptr, kargs, splitk_batch_offset);
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k);
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
const auto& b_block_window = gemm_tile_windows.at(I1);
const auto& c_block_tile = GemmPipeline{}.template operator()(
a_block_window, b_block_window, num_loop, smem_ptr_0, smem_ptr_1);
// Run Epilogue Pipeline
auto& c_block_window = gemm_tile_windows.at(I2);
EpiloguePipeline{}
.template operator()<decltype(c_block_window), decltype(c_block_tile), DstInMemOp>(
c_block_window, c_block_tile, smem_ptr_0);
}
CK_TILE_DEVICE void operator()(GemmKernelArgs kargs) const
......@@ -509,11 +571,27 @@ struct GemmKernel
CDataType* c_ptr = static_cast<CDataType*>(kargs.c_ptr);
// allocate LDS
__shared__ char smem_ptr[GetSmemSize()];
__shared__ char smem_ptr_0[GetSmemSize()];
__shared__ char smem_ptr_1[GetSmemSize()];
if(kargs.k_batch == 1)
{
RunGemm(a_ptr, b_ptr, c_ptr, smem_ptr, kargs, splitk_batch_offset, i_m, i_n);
if constexpr(GemmPipeline::DoubleSmemBuffer == true)
{
RunGemm2LDS(a_ptr,
b_ptr,
c_ptr,
smem_ptr_0,
smem_ptr_1,
kargs,
splitk_batch_offset,
i_m,
i_n);
}
else
{
RunGemm(a_ptr, b_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
}
}
else
{
......@@ -522,8 +600,23 @@ struct GemmKernel
if constexpr(!(EpiloguePipeline::template GetVectorSizeC<CDataType>() % 2 != 0 &&
is_any_of<CDataType, fp16_t, bf16_t>::value))
{
RunGemm<memory_operation_enum::atomic_add>(
a_ptr, b_ptr, c_ptr, smem_ptr, kargs, splitk_batch_offset, i_m, i_n);
if constexpr(GemmPipeline::DoubleSmemBuffer == true)
{
RunGemm2LDS<memory_operation_enum::atomic_add>(a_ptr,
b_ptr,
c_ptr,
smem_ptr_0,
smem_ptr_1,
kargs,
splitk_batch_offset,
i_m,
i_n);
}
else
{
RunGemm<memory_operation_enum::atomic_add>(
a_ptr, b_ptr, c_ptr, smem_ptr_0, kargs, splitk_batch_offset, i_m, i_n);
}
}
}
}
......
include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp
View file @ b8d11559
......@@ -64,6 +64,18 @@ struct GroupedGemmKernel : public GemmKernel<TilePartitioner_, GemmPipeline_, Ep
}
};
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
using P_ = GemmPipeline;
return concat('_', "gemm_grouped", gemm_prec_str<ADataType, BDataType>,
concat('x', P_::kMPerBlock, P_::kNPerBlock, P_::kKPerBlock),
concat('x', P_::GetVectorSizeA(), P_::GetVectorSizeB(), P_::GetVectorSizeC()),
concat('x', P_::kPadM, P_::kPadN, P_::kPadK));
// clang-format on
}
__host__ static auto GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
-> std::size_t
{
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp
View file @ b8d11559
......@@ -41,20 +41,26 @@ struct GemmPipelineAgBgCrImplBase
store_tile(lds_tile_window, block_tile_tmp);
}
template <typename DstBlockTile, typename SrcTileWindow>
CK_TILE_DEVICE void LocalPrefetch(DstBlockTile& dst_block_tile,
const SrcTileWindow& lds_tile_window) const
{
load_tile(dst_block_tile, lds_tile_window);
}
CK_TILE_DEVICE auto GetABLdsTensorViews(void* p_smem) const
{
// A tile in LDS
ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
ADataType* __restrict__ 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;
constexpr index_t a_lds_block_space_size_aligned = integer_least_multiple(
sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16);
// B tile in LDS
BDataType* p_b_lds = static_cast<BDataType*>(
BDataType* __restrict__ 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);
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp
View file @ b8d11559
......@@ -10,6 +10,7 @@
#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_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"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -75,12 +76,23 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
static constexpr bool HasHotLoop = Problem::HasHotLoop;
static constexpr auto TailNum = Problem::TailNum;
static constexpr auto Scheduler = Problem::Scheduler;
using Base::PrefetchStages;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "pipeline_AgBgCrCompV3", BlockSize,
concat('x', GetVectorSizeA(), GetVectorSizeB(), GetVectorSizeC()),
concat('x', kPadM, kPadN, kPadK));
// clang-format on
}
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4.hpp 0 → 100644
View file @ b8d11559
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.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"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4_default_policy.hpp"
namespace ck_tile {
// A Tile Window: global memory
// B Tile Window: global memory
// C Distributed tensor: register
template <typename Problem>
struct BaseGemmPipelineAgBgCrCompV4
{
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)
{
if(num_loop % PrefetchStages == 1)
{
return TailNumber::Three;
}
else
{
return TailNumber::Two;
}
}
};
/**
* @brief Compute optimized pipeline version 4
*
* This version introduces a dual LDS window mechanism using a ping-pong buffer approach
* for more efficient data handling from global memory. Unlike compute version 3, this method
* allows one LDS to fetch data from global memory while the other LDS executes warps for MFMA
* matrix multiplication. This dual operation helps in keeping the Warp unit continuously busy,
* thereby significantly reducing memory load times and enhancing overall performance.
*
* @note This version shows improved performance over Compute Version 3 with the same block tile.
* It is particularly more efficient for large matrices where M, N, and K are greater than 8K,
* even when Compute Version 3's block size is twice that of Compute Version 4.
*/
template <typename Problem, typename Policy = GemmPipelineAgBgCrCompV4DefaultPolicy>
struct GemmPipelineAgBgCrCompV4 : public BaseGemmPipelineAgBgCrCompV4<Problem>
{
using Base = BaseGemmPipelineAgBgCrCompV4<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 GetVectorSizeA() { return Policy::template GetVectorSizeA<Problem>(); }
static constexpr index_t GetVectorSizeB() { return Policy::template GetVectorSizeB<Problem>(); }
static constexpr index_t GetVectorSizeC() { return Policy::template GetVectorSizeC<Problem>(); }
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
static constexpr bool HasHotLoop = Problem::HasHotLoop;
static constexpr auto TailNum = Problem::TailNum;
static constexpr auto Scheduler = Problem::Scheduler;
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
CK_TILE_HOST_DEVICE static constexpr auto IsTransposeC()
{
return Policy::template IsTransposeC<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 * GetVectorSizeA());
constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * GetVectorSizeB());
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);
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_read_inst = num_ds_read_inst_a + num_ds_read_inst_b;
constexpr auto num_ds_write_inst = A_LDS_Write_Inst_Num + B_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst = A_Buffer_Load_Inst_Num + B_Buffer_Load_Inst_Num;
constexpr auto num_issue = num_buffer_load_inst;
static_for<0, num_buffer_load_inst, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA : 1
__builtin_amdgcn_sched_group_barrier(
0x100, num_ds_read_inst / num_issue, 0); // DS read : 2
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA: 1
__builtin_amdgcn_sched_group_barrier(
0x200, num_ds_write_inst / num_issue, 0); // DS write : 1
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA : 1
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read :1
__builtin_amdgcn_sched_group_barrier(
0x008, C_MFMA_Inst_Num / num_issue - 3, 0); // MFMA : 5
});
__builtin_amdgcn_sched_barrier(0);
}
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* __restrict__ p_smem_0,
void* __restrict__ p_smem_1) 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>>,
"Data Type conflict on A and B matrix input data type.");
constexpr bool is_a_col_major =
std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
constexpr bool is_b_row_major = std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
static_assert(is_a_col_major
? (KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}])
: (MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}]),
"A block window has incorrect lengths for defined ALayout!");
static_assert(is_b_row_major
? (KPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I1{}])
: (NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
KPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I1{}]),
"B block window has incorrect lengths for defined BLayout!");
////////////// global window & register /////////////////
// A DRAM tile window for load
auto a_copy_dram_window =
make_tile_window_linear(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>());
// B DRAM tile window for load
auto b_copy_dram_window =
make_tile_window_linear(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>());
// A register tile for global load
constexpr auto ABlockTileDistr = a_copy_dram_window.get_tile_distribution();
constexpr auto BBlockTileDistr = 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_global_load_tile;
BBlockTile b_global_load_tile;
using ADramTileWindowStep = typename ADramBlockWindowTmp::BottomTensorIndex;
using BDramTileWindowStep = typename BDramBlockWindowTmp::BottomTensorIndex;
constexpr ADramTileWindowStep a_dram_tile_window_step =
is_a_col_major ? make_array(KPerBlock, 0) : make_array(0, KPerBlock);
constexpr BDramTileWindowStep b_dram_tile_window_step =
is_b_row_major ? make_array(KPerBlock, 0) : make_array(0, KPerBlock);
// global prefetch 0
// global read 0
Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
////////////// LDS desc, window & register /////////////////
auto&& [a_lds_block0, b_lds_block0] = Base::GetABLdsTensorViews(p_smem_0);
auto&& [a_lds_block1, b_lds_block1] = Base::GetABLdsTensorViews(p_smem_1);
auto a_copy_lds_window0 = make_tile_window(
a_lds_block0, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
auto a_copy_lds_window1 = make_tile_window(
a_lds_block1, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
auto b_copy_lds_window0 = make_tile_window(
b_lds_block0, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
auto b_copy_lds_window1 = make_tile_window(
b_lds_block1, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
// Block GEMM
auto block_gemm = BlockGemm();
auto c_block_tile = block_gemm.MakeCBlockTile();
// initialize C
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0
if constexpr(is_a_col_major)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegTileDistribution<Problem>());
transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
}
else
{
Base::LocalPrefill(a_copy_lds_window0, a_global_load_tile, a_element_func);
}
if constexpr(is_b_row_major)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegTileDistribution<Problem>());
transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
}
else
{
Base::LocalPrefill(b_copy_lds_window0, b_global_load_tile, b_element_func);
}
// global read 1
Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
block_sync_lds();
constexpr auto ALdsTileDistr = decltype(make_static_tile_distribution(
BlockGemm::MakeABlockDistributionEncode())){};
constexpr auto BLdsTileDistr = decltype(make_static_tile_distribution(
BlockGemm::MakeBBlockDistributionEncode())){};
using ALdsTile = decltype(make_static_distributed_tensor<ADataType>(ALdsTileDistr));
using BLdsTile = decltype(make_static_distributed_tensor<BDataType>(BLdsTileDistr));
ALdsTile a_block_tile0;
ALdsTile a_block_tile1;
BLdsTile b_block_tile0;
BLdsTile b_block_tile1;
auto a_lds_ld_window0 =
make_tile_window_linear(a_lds_block0,
make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
{0, 0},
ALdsTileDistr);
auto a_lds_ld_window1 =
make_tile_window_linear(a_lds_block1,
make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
{0, 0},
ALdsTileDistr);
auto b_lds_ld_window0 =
make_tile_window_linear(b_lds_block0,
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
{0, 0},
BLdsTileDistr);
auto b_lds_ld_window1 =
make_tile_window_linear(b_lds_block1,
make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
{0, 0},
BLdsTileDistr);
Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
Base::LocalPrefetch(b_block_tile0, b_lds_ld_window0);
if constexpr(is_a_col_major)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegTileDistribution<Problem>());
transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
Base::LocalPrefill(a_copy_lds_window1, a_shuffle_tmp, a_element_func);
}
else
{
Base::LocalPrefill(a_copy_lds_window1, a_global_load_tile, a_element_func);
}
if constexpr(is_b_row_major)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegTileDistribution<Problem>());
transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
Base::LocalPrefill(b_copy_lds_window1, b_shuffle_tmp, b_element_func);
}
else
{
Base::LocalPrefill(b_copy_lds_window1, b_global_load_tile, b_element_func);
}
Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
if(HasHotLoop)
{
// minus 2 because we have ping-pong double buffer.
index_t iCounter = __builtin_amdgcn_readfirstlane(num_loop - 2);
do
{
// ping
{
block_sync_lds();
Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
if constexpr(is_a_col_major)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegTileDistribution<Problem>());
transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
}
else
{
Base::LocalPrefill(
a_copy_lds_window0, a_global_load_tile, a_element_func);
}
if constexpr(is_b_row_major)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegTileDistribution<Problem>());
transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
}
else
{
Base::LocalPrefill(
b_copy_lds_window0, b_global_load_tile, b_element_func);
}
Base::GlobalPrefetch(
a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
Base::GlobalPrefetch(
b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
// gemm
block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
}
// pong
{
block_sync_lds();
Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
Base::LocalPrefetch(b_block_tile0, b_lds_ld_window0);
if constexpr(is_a_col_major)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegTileDistribution<Problem>());
transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
Base::LocalPrefill(a_copy_lds_window1, a_shuffle_tmp, a_element_func);
}
else
{
Base::LocalPrefill(
a_copy_lds_window1, a_global_load_tile, a_element_func);
}
if constexpr(is_b_row_major)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegTileDistribution<Problem>());
transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
Base::LocalPrefill(b_copy_lds_window1, b_shuffle_tmp, b_element_func);
}
else
{
Base::LocalPrefill(
b_copy_lds_window1, b_global_load_tile, b_element_func);
}
Base::GlobalPrefetch(
a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
Base::GlobalPrefetch(
b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
// gemm
block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
}
iCounter -= 2;
} while(iCounter > 1);
}
// tail 3
if(TailNum == TailNumber::Three)
{
// 3
{
block_sync_lds();
Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
if constexpr(is_a_col_major)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegTileDistribution<Problem>());
transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
}
else
{
Base::LocalPrefill(a_copy_lds_window0, a_global_load_tile, a_element_func);
}
if constexpr(is_b_row_major)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegTileDistribution<Problem>());
transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
}
else
{
Base::LocalPrefill(b_copy_lds_window0, b_global_load_tile, b_element_func);
}
block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
}
// 2
{
block_sync_lds();
Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
}
// 1
{
block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
__builtin_amdgcn_sched_barrier(0);
}
}
else
{
// 2
{
block_sync_lds();
Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
static_for<0, 8, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
});
__builtin_amdgcn_sched_barrier(0);
}
// 1
{
block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
__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_0,
void* p_smem_1) 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_0,
p_smem_1);
}
public:
template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
const BDramBlockWindowTmp& b_dram_block_window_tmp,
const index_t num_loop,
void* __restrict__ p_smem_0,
void* __restrict__ p_smem_1) 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_0,
p_smem_1);
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4_default_policy.hpp 0 → 100644
View file @ b8d11559
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
namespace ck_tile {
// Default policy for GemmPipelineAGmemBGmemCregComputeV4, except the block gemm method, it shares
// the same vector size implementation, SmemSize, Global memory tile distiribution as the
// UniversalGemm Pipeline Policy.
// Default policy class should not be templated, put template on
// member functions instead.
struct GemmPipelineAgBgCrCompV4DefaultPolicy
: public UniversalGemmBasePolicy<GemmPipelineAgBgCrCompV4DefaultPolicy>
{
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
{
using namespace ck_tile;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t KPack = GetSmemPackA<Problem>();
constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kKPerBlock / KPack>{}, number<kMPerBlock>{}, number<KPack>{}),
make_tuple(number<kMPerBlock * KPack>{}, number<KPack>{}, number<1>{}),
number<KPack>{},
number<1>{});
constexpr auto a_lds_block_desc = transform_tensor_descriptor(
a_lds_block_desc_0,
make_tuple(
make_pass_through_transform(number<kMPerBlock>{}),
make_merge_transform(make_tuple(number<kKPerBlock>{} / KPack, number<KPack>{}))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return a_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
{
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t KPack = GetSmemPackB<Problem>();
constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kKPerBlock / KPack>{}, number<kNPerBlock>{}, number<KPack>{}),
make_tuple(number<(kNPerBlock)*KPack>{}, number<KPack>{}, number<1>{}),
number<KPack>{},
number<1>{});
constexpr auto b_lds_block_desc = transform_tensor_descriptor(
b_lds_block_desc_0,
make_tuple(
make_pass_through_transform(number<kNPerBlock>{}),
make_merge_transform(make_tuple(number<kKPerBlock / KPack>{}, number<KPack>{}))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return b_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
{
using AccDataType = float;
using BlockWarps = typename Problem::BlockGemmShape::BlockWarps;
using WarpTile = typename Problem::BlockGemmShape::WarpTile;
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
AccDataType,
WarpTile::at(I0),
WarpTile::at(I1),
WarpTile::at(I2),
Problem::TransposeC>;
using BlockGemmPolicy = BlockGemmARegBRegCRegV1CustomPolicy<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
BlockWarps,
WarpGemm>;
return BlockGemmARegBRegCRegV1<Problem, BlockGemmPolicy>{};
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp
View file @ b8d11559
......@@ -7,6 +7,7 @@
#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"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -123,11 +124,23 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
// 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;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "pipeline_AgBgCrMe",
concat('x', MPerBlock, NPerBlock, KPerBlock),
concat('x', GetVectorSizeA(), GetVectorSizeB(), GetVectorSizeC()),
concat('x', kPadM, kPadN, kPadK));
// clang-format on
}
using Base::PrefetchStages;
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp
View file @ b8d11559
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <ostream>
#include <sstream>
#include "ck_tile/core.hpp"
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp
View file @ b8d11559
......@@ -5,6 +5,7 @@
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -39,6 +40,21 @@ struct GemmPipelineAGmemBGmemCRegV1
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
static constexpr index_t kLdsAlignmentInBytes = 16;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "pipeline_AGmemBGmemCRegV1",
concat('x', kMPerBlock, kNPerBlock, kKPerBlock, BlockSize),
concat('x', GetVectorSizeA(), GetVectorSizeB(), GetVectorSizeC()),
concat('x', kPadM, kPadN, kPadK));
// clang-format on
}
// For the basic gemm pipelien DoubleSmemBuffer set to be false naturally.
static constexpr bool DoubleSmemBuffer = false;
CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
......@@ -75,8 +91,9 @@ struct GemmPipelineAGmemBGmemCRegV1
auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
constexpr index_t a_lds_block_space_size_aligned =
integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
16;
integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(),
kLdsAlignmentInBytes) *
kLdsAlignmentInBytes;
// B tile in LDS
BDataType* p_b_lds = static_cast<BDataType*>(
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
View file @ b8d11559
......@@ -338,7 +338,7 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>);
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp
View file @ b8d11559
......@@ -5,6 +5,7 @@
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -25,6 +26,13 @@ struct GemmPipelineAGmemBGmemCRegV2
static constexpr index_t kNPerBlock = BlockGemmShape::kN;
static constexpr index_t kKPerBlock = BlockGemmShape::kK;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "pipeline_AGmemBGmemCRegV2",
concat('x', kMPerBlock, kNPerBlock, kKPerBlock, kBlockSize));
// clang-format on
}
CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp
View file @ b8d11559
......@@ -5,6 +5,7 @@
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -35,9 +36,21 @@ struct GemmPipelineProblemBase
static constexpr bool kPadN = Traits::kPadN;
static constexpr bool kPadK = Traits::kPadK;
static constexpr auto Scheduler = GemmPipelineScheduler::Default;
static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
static constexpr auto Scheduler = GemmPipelineScheduler::Default;
static constexpr index_t VectorLoadSize = Traits::_VectorSize;
[[nodiscard]] CK_TILE_HOST static const std::string GetName()
{
// clang-format off
return concat('_', "gemm_problem",
concat('x', VectorLoadSize, kBlockSize),
concat('x', kPadM, kPadN, kPadK),
Scheduler);
// clang-format on
}
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
{
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
......@@ -162,6 +175,8 @@ struct UniversalGemmPipelineProblem
static constexpr bool kPadN = Traits::kPadN;
static constexpr bool kPadK = Traits::kPadK;
static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
static constexpr auto Scheduler = Scheduler_;
static constexpr auto HasHotLoop = HasHotLoop_;
static constexpr auto TailNum = TailNum_;
......
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View file @ b8d11559
......@@ -9,8 +9,8 @@
namespace ck_tile {
// UniversalGemm Policy
struct UniversalGemmPipelineAgBgCrPolicy
template <typename Derived>
struct UniversalGemmBasePolicy
{
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
......@@ -113,7 +113,7 @@ struct UniversalGemmPipelineAgBgCrPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeC()
{
using BlockGemm = remove_cvref_t<decltype(GetBlockGemm<Problem>())>;
using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
using WG = typename BlockGemm::WarpGemm;
constexpr bool TransposeC = Problem::TransposeC;
......@@ -166,10 +166,116 @@ struct UniversalGemmPipelineAgBgCrPolicy
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto IsTransposeC()
{
return Problem::TransposeC;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
// Tile: MPerBlock X KPerBlock
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
MPerBlock,
KPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
// Tile: KPerBlock X MPerBlock
else
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
MPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
// Tile: KPerBlock X NPerBlock
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
NPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
// Tile: NPerBlock X KPerBlock
else
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
NPerBlock,
KPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegTileDistribution()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
MPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegTileDistribution()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
NPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
{
using BlockGemm = decltype(GetBlockGemm<Problem>());
using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
constexpr index_t KPack = BlockGemm::Traits::KPack;
return KPack;
}
......@@ -177,11 +283,43 @@ struct UniversalGemmPipelineAgBgCrPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackB()
{
using BlockGemm = decltype(GetBlockGemm<Problem>());
using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
constexpr index_t KPack = BlockGemm::Traits::KPack;
return KPack;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
{
constexpr auto a_lds_desc = Derived::template MakeALdsBlockDescriptor<Problem>();
constexpr index_t smem_size_a = integer_least_multiple(
sizeof(typename Problem::ADataType) * a_lds_desc.get_element_space_size(), 16);
return smem_size_a;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeB()
{
constexpr auto b_lds_desc = Derived::template MakeBLdsBlockDescriptor<Problem>();
constexpr index_t smem_size_b = integer_least_multiple(
sizeof(typename Problem::BDataType) * b_lds_desc.get_element_space_size(), 16);
return smem_size_b;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
constexpr index_t smem_size_b = GetSmemSizeB<Problem>();
return smem_size_a + smem_size_b;
}
};
// UniversalGemm Policy
struct UniversalGemmPipelineAgBgCrPolicy
: public UniversalGemmBasePolicy<UniversalGemmPipelineAgBgCrPolicy>
{
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
{
......@@ -421,133 +559,6 @@ struct UniversalGemmPipelineAgBgCrPolicy
#endif
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
{
constexpr index_t smem_size_a = sizeof(typename Problem::ADataType) *
MakeALdsBlockDescriptor<Problem>().get_element_space_size();
return smem_size_a;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeB()
{
constexpr index_t smem_size_b = sizeof(typename Problem::BDataType) *
MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
return smem_size_b;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
constexpr index_t smem_size_b = GetSmemSizeB<Problem>();
index_t smem_size = 0;
smem_size += smem_size_a + smem_size_b;
return smem_size;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
// Tile: MPerBlock X KPerBlock
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
MPerBlock,
KPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
// Tile: KPerBlock X MPerBlock
else
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
MPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
// Tile: KPerBlock X NPerBlock
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
NPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
// Tile: NPerBlock X KPerBlock
else
{
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
NPerBlock,
KPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::Make2DStaticTileDistribution();
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegTileDistribution()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
MPerBlock,
VecLoadSize,
ATileAccessPattern>;
return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegTileDistribution()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
KPerBlock,
NPerBlock,
VecLoadSize,
BTileAccessPattern>;
return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
{
......
include/ck_tile/ops/gemm/pipeline/tile_gemm_shape.hpp
View file @ b8d11559
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/concat.hpp"
namespace ck_tile {
......@@ -19,6 +20,16 @@ struct TileGemmShape
static constexpr index_t kM = BlockTile::at(number<0>{});
static constexpr index_t kN = BlockTile::at(number<1>{});
static constexpr index_t kK = BlockTile::at(number<2>{});
CK_TILE_HOST static std::string GetName()
{
// clang-format off
return concat('_', "tile_gemm_shape",
concat('x', kM, kN, kK, NumWarps),
concat('x', BlockWarps::at(number<0>{}), BlockWarps::at(number<1>{}), BlockWarps::at(number<2>{})),
concat('x', (WarpTile::at(number<0>{})), WarpTile::at(number<1>{}), WarpTile::at(number<2>{})));
// clang-format on
}
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp
View file @ b8d11559
......@@ -32,6 +32,7 @@ struct TileGemmTraits
template <bool kPadM_,
bool kPadN_,
bool kPadK_,
bool DoubleSmemBuffer_,
typename ALayout_,
typename BLayout_,
typename CLayout_,
......@@ -42,6 +43,8 @@ struct TileGemmUniversalTraits
static constexpr bool kPadN = kPadN_;
static constexpr bool kPadK = kPadK_;
static constexpr bool DoubleSmemBuffer = DoubleSmemBuffer_;
using ALayout = ALayout_;
using BLayout = BLayout_;
using CLayout = CLayout_;
......
include/ck_tile/ops/image_to_column.hpp
View file @ b8d11559
......@@ -8,3 +8,4 @@
#include "ck_tile/ops/image_to_column/pipeline/tile_image_to_column_shape.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/layernorm2d.hpp
View file @ b8d11559
......@@ -11,3 +11,4 @@
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/norm_reduce.hpp
View file @ b8d11559
......@@ -8,3 +8,4 @@
#include "ck_tile/ops/norm_reduce/thread/thread_welford.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/permute.hpp
View file @ b8d11559
......@@ -7,3 +7,4 @@
#include "ck_tile/ops/permute/pipeline/generic_petmute_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
include/ck_tile/ops/reduce.hpp
View file @ b8d11559
......@@ -9,3 +9,4 @@
#include "ck_tile/ops/reduce/block/block_reduce2d_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"
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