Skip to content

GitLab

Commit 7f179833 authored by Bartlomiej Kocot's avatar Bartlomiej Kocot
Browse files

Fixes

parent f0bbc5db
Hide whitespace changes
Inline Side-by-side
Showing with 49 additions and 324 deletions
example/ck_tile/03_gemm/CMakeLists.txt
View file @ 7f179833
......@@ -3,4 +3,3 @@ add_executable(tile_example_gemm_universal EXCLUDE_FROM_ALL universal_gemm.cpp)
target_compile_options(tile_example_gemm_universal PRIVATE
-mllvm -enable-noalias-to-md-conversion=0
)
add_executable(tile_example_gemm_universal_pk_int4 EXCLUDE_FROM_ALL universal_gemm_pk_int4.cpp)
example/ck_tile/03_gemm/universal_gemm.cpp
View file @ 7f179833
......@@ -330,6 +330,13 @@ int run_gemm_example(int argc, char* argv[])
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "pk_int4_t")
{
// TODO: Add support for bhalf_t ADataType
return run_gemm_example_with_layouts<ck_tile::half_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
......@@ -353,6 +360,13 @@ int run_gemm_example(int argc, char* argv[])
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Col{}, Col{}, Row{});
}
else if(data_type == "pk_int4_t")
{
// TODO: Add support for bhalf_t ADataType
return run_gemm_example_with_layouts<ck_tile::half_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(argc, argv, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
......
example/ck_tile/03_gemm/universal_gemm_pk_int4.cpp deleted 100644 → 0
View file @ f0bbc5db
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <sstream>
#include <string>
#include <tuple>
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
// Memory friendly for Interwave scheduler
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 32;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t M_Warp = 4;
constexpr ck_tile::index_t N_Warp = 1;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
constexpr bool DoubleSmemBuffer = false;
#endif
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
// Compute friendly for Intrawave scheduler
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 16;
constexpr bool DoubleSmemBuffer = false;
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
// Compute friendly for Intrawave scheduler
// Using the ping pong reader in the lds level
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 16;
constexpr bool DoubleSmemBuffer = true;
#endif
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 1;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
// ===============================================
using GemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<kPadM,
kPadN,
kPadK,
DoubleSmemBuffer,
ALayout,
BLayout,
CLayout,
TransposeC>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
using BaseGemmPipeline = UNIVERSAL_GEMM_PIPELINE<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = GEMM_PIPELINE_SCHEDULER;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
UniversalGemmProblem::TransposeC>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
if(has_hot_loop)
{
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
else
{
std::ostringstream err;
err << "For compute pipeline tail number should always be Full, but have \"" << tail_num
<< "\" which is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
// Tail pipeline One to Seven
if(tail_num == ck_tile::TailNumber::One)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::One>{});
}
else if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
if constexpr(BaseGemmPipeline::PrefetchStages > 2)
{
if(tail_num == ck_tile::TailNumber::Two)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 3)
{
if(tail_num == ck_tile::TailNumber::Three)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 4)
{
if(tail_num == ck_tile::TailNumber::Four)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Four>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 5)
{
if(tail_num == ck_tile::TailNumber::Five)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Five>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 6)
{
if(tail_num == ck_tile::TailNumber::Six)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Six>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 7)
{
if(tail_num == ck_tile::TailNumber::Seven)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Seven>{});
}
}
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
if(tail_num == ck_tile::TailNumber::Three)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
}
else
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
}
#endif
}
else
{
// Tail number always Full - #PrefetchStages
if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<false>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
else
{
std::ostringstream err;
err << "When there's no hot loop, this tail number \"" << tail_num
<< "\" is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
}
return ave_time;
}
#include "run_gemm_example.inc"
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>(
argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "C")
{
return run_gemm_example_with_layouts<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>(
argc, argv, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4.hpp
View file @ 7f179833
......@@ -60,6 +60,11 @@ struct GemmPipelineAgBgCrCompV4 : public BaseGemmPipelineAgBgCrCompV4<Problem>
using CDataType = remove_cvref_t<typename Problem::CDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t APackedSize =
ck_tile::numeric_traits<remove_cvref_t<ADataType>>::PackedSize;
static constexpr index_t BPackedSize =
ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using CLayout = remove_cvref_t<typename Problem::CLayout>;
......@@ -139,12 +144,12 @@ struct GemmPipelineAgBgCrCompV4 : public BaseGemmPipelineAgBgCrCompV4<Problem>
(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_a =
A_LDS_Read_Width * sizeof(ADataType) / APackedSize == 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) / BPackedSize == 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;
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp
View file @ 7f179833
......@@ -21,6 +21,11 @@ struct BaseGemmPipelineAgBgCrMem
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t APackedSize =
ck_tile::numeric_traits<remove_cvref_t<ADataType>>::PackedSize;
static constexpr index_t BPackedSize =
ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
static constexpr index_t BlockSize = Problem::kBlockSize;
......@@ -33,9 +38,11 @@ struct BaseGemmPipelineAgBgCrMem
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 FullMemBandPrefetchStages =
integer_divide_ceil(MinMemInFlyBytes / WgpPerCU,
(MPerBlock * sizeof(ADataType) / APackedSize +
NPerBlock * sizeof(BDataType) / BPackedSize) *
KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp
View file @ 7f179833
......@@ -20,6 +20,11 @@ struct GemmPipelineAGmemBGmemCRegV2
using CDataType = remove_cvref_t<typename Problem::CDataType>;
using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
static constexpr index_t APackedSize =
ck_tile::numeric_traits<remove_cvref_t<ADataType>>::PackedSize;
static constexpr index_t BPackedSize =
ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kMPerBlock = BlockGemmShape::kM;
......@@ -37,13 +42,15 @@ struct GemmPipelineAGmemBGmemCRegV2
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{
return integer_divide_ceil(
sizeof(ADataType) *
Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(),
16) *
return integer_divide_ceil(sizeof(ADataType) *
Policy::template MakeALdsBlockDescriptor<Problem>()
.get_element_space_size() /
APackedSize,
16) *
16 +
sizeof(BDataType) *
Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size() /
BPackedSize;
}
template <typename ADramBlockWindowTmp,
......@@ -75,7 +82,8 @@ struct GemmPipelineAGmemBGmemCRegV2
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) *
integer_divide_ceil(
sizeof(ADataType) * a_lds_block_desc.get_element_space_size() / APackedSize, 16) *
16;
// B tile in LDS
......
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