Skip to content

GitLab

Commit 896f8b4c authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

add gemm_api and instances

parent 73a076ee
Hide whitespace changes
Inline Side-by-side
Showing with 249 additions and 217 deletions
example/ck_tile/03_gemm/instances/gemm_universal_mem_f16_f16_f16_mk_nk_mn.cpp 0 → 100644
View file @ 896f8b4c
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_universal_mem_instance_common.hpp"
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
template float gemm_<trait_<ck_tile::half_t,
ck_tile::half_t,
float,
ck_tile::half_t,
Row,
Col,
Row,
128,
32,
64,
4,
1,
1,
32,
32,
8,
false,
false,
false>>(const A&, const S&);
example/ck_tile/03_gemm/instances/gemm_universal_mem_instance_common.hpp 0 → 100644
View file @ 896f8b4c
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include <iostream>
#include "gemm_basic.hpp"
using A = ck_tile::GemmHostArgs;
using S = ck_tile::stream_config;
template <typename ADataType_,
typename BDataType_,
typename AccDataType_,
typename CDataType_,
typename ALayout_,
typename BLayout_,
typename CLayout_,
ck_tile::index_t M_Tile_,
ck_tile::index_t N_Tile_,
ck_tile::index_t K_Tile_,
ck_tile::index_t M_Warp_,
ck_tile::index_t N_Warp_,
ck_tile::index_t K_Warp_,
ck_tile::index_t M_Warp_Tile_,
ck_tile::index_t N_Warp_Tile_,
ck_tile::index_t K_Warp_Tile_,
bool kPadM_,
bool kPadN_,
bool kPadK_>
using trait_ = gemm_traits_<ADataType_,
BDataType_,
AccDataType_,
CDataType_,
ALayout_,
BLayout_,
CLayout_,
M_Tile_,
N_Tile_,
K_Tile_,
M_Warp_,
N_Warp_,
K_Warp_,
M_Warp_Tile_,
N_Warp_Tile_,
K_Warp_Tile_,
kPadM_,
kPadN_,
kPadK_>;
template <typename Traits_>
float gemm_(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
using GemmShape = ck_tile::TileGemmShape<
ck_tile::sequence<Traits_::M_Tile, Traits_::N_Tile, Traits_::K_Tile>,
ck_tile::sequence<Traits_::M_Warp, Traits_::N_Warp, Traits_::K_Warp>,
ck_tile::sequence<Traits_::M_Warp_Tile, Traits_::N_Warp_Tile, Traits_::K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
using GemmEpilogue =
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename Traits_::AccDataType,
typename Traits_::CDataType,
Traits_::kPadM,
Traits_::kPadN>>;
using GemmTraits = ck_tile::TileGemmTraits<Traits_::kPadM,
Traits_::kPadN,
Traits_::kPadK,
typename Traits_::ALayout,
typename Traits_::BLayout,
typename Traits_::CLayout>;
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<typename Traits_::ADataType,
typename Traits_::BDataType,
typename Traits_::AccDataType,
GemmShape,
GemmTraits>>;
constexpr int kBlockPerCu = 1;
const ck_tile::index_t k_grain = args.k_batch * Traits_::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * Traits_::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;
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
ck_tile::UniversalGemmPipelineProblem<typename Traits_::ADataType,
typename Traits_::BDataType,
typename Traits_::AccDataType,
GemmShape,
GemmTraits,
ck_tile::GemmPipelineScheduler::Interwave,
has_hot_loop_v,
tail_number_v>>;
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)
{
// 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)
{
static_assert(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>{});
}
}
}
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;
}
example/ck_tile/03_gemm/run_gemm_example.inc
View file @ 896f8b4c
...@@ -28,8 +28,13 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf, ...@@ -28,8 +28,13 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
args.stride_B = stride_B; args.stride_B = stride_B;
args.stride_C = stride_C; args.stride_C = stride_C;
float ave_time = gemm_calc<ALayout, BLayout, CLayout>( gemm_traits traits{DataTypeTraits<ADataType>{}.name,
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat}); std::is_same_v<ALayout, Row>,
std::is_same_v<BLayout, Row>,
std::is_same_v<CLayout, Row>};
float ave_time =
gemm(traits, args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte = std::size_t num_byte =
...@@ -210,9 +215,6 @@ int run_gemm_example(int argc, char* argv[]) ...@@ -210,9 +215,6 @@ int run_gemm_example(int argc, char* argv[])
if(!result) if(!result)
return -1; return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout"); std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout"); std::string b_layout = arg_parser.get_str("b_layout");
...@@ -224,16 +226,14 @@ int run_gemm_example(int argc, char* argv[]) ...@@ -224,16 +226,14 @@ int run_gemm_example(int argc, char* argv[])
{ {
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{}); return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
} }
// TODO: Fixme: with latest changes to GemmPipelineAGmemBGmemCRegV1DefaultPolicy below do not else if(a_layout == "C" && b_layout == "C")
// work. {
// else if(a_layout == "C" && b_layout == "C") return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
// { }
// return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{}); else if(a_layout == "C" && b_layout == "R")
// } {
// else if(a_layout == "C" && b_layout == "R") return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
// { }
// return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
// }
else else
{ {
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!"); throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
......
example/ck_tile/03_gemm/universal_gemm.cpp
View file @ 896f8b4c
// SPDX-License-Identifier: MIT // 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.
#include <hip/hip_runtime.h> #include <hip/hip_runtime.h>
...@@ -14,207 +14,6 @@ ...@@ -14,207 +14,6 @@
#include "ck_tile/host.hpp" #include "ck_tile/host.hpp"
#include "gemm_basic.hpp" #include "gemm_basic.hpp"
#define CK_TILE_PIPELINE_COMPUTE 1
#define CK_TILE_PIPELINE_MEMORY 2
#ifndef CK_TILE_PIPELINE_DEFAULT
#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE
#endif
template <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;
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
// 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 = 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;
#endif
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
// ===============================================
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::GemmTilePartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<
#endif
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
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;
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<
#endif
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
ck_tile::GemmPipelineScheduler::Interwave,
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
ck_tile::GemmPipelineScheduler::Intrawave,
#endif
has_hot_loop_v,
tail_number_v>>;
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)
{
// 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>{});
}
}
}
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" #include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
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