ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Flatmm merge (#2168)

Browse Source 
* sync with function interface of cshuffleepiloge,fix flatmm build fail

* move code from solin/flatmm which add mfma16*16*32fp8 and optimize flatmm

---------

Co-authored-by: solin <bingzhou@amd.com>

[ROCm/composable_kernel commit: 6a3960c1e1]
This commit is contained in:
BingYuan.Zhou
2025-05-08 12:59:57 +08:00
committed by GitHub
parent a86620111c
commit 2428249bc7
11 changed files with 552 additions and 192 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/ck_tile/18_flatmm/CMakeLists.txt
View File

@@ -3,5 +3,6 @@ add_executable(tile_example_flatmm_basic EXCLUDE_FROM_ALL flatmm_basic.cpp)
set(EXAMPLE_FLATMM_COMPILE_OPTIONS)
# list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
# list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-unused-variable -Wno-unused-parameter)
# list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-unused-local-typedef)
list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -DUSING_MFMA_16x16x32=1 -DENABLE_FP8=1 -Wno-unused-local-typedef)
#list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -DUSING_MFMA_32x32x16=1 -DENABLE_FP8=1 -Wno-unused-local-typedef)
target_compile_options(tile_example_flatmm_basic PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})

162
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -12,7 +12,13 @@
#include "ck_tile/host.hpp"
#include "flatmm_basic.hpp"
template <typename ALayout, typename BLayout, typename CLayout>
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_config& s)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
@@ -23,18 +29,32 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
constexpr int kBlockPerCu = 2;
// This part comes from the Codegen
#if defined(USING_MFMA_16x16x32) || defined(ENABLE_FP16)
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t K_Tile = 128;
constexpr ck_tile::index_t M_Warp = 1;
constexpr ck_tile::index_t N_Warp = 4;
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 ck_tile::index_t M_Warp_Tile = is_8bit_type<ADataType>::value ? 16 : 32;
constexpr ck_tile::index_t N_Warp_Tile = is_8bit_type<ADataType>::value ? 16 : 32;
constexpr ck_tile::index_t K_Warp_Tile = is_8bit_type<ADataType>::value ? 64 : 16;
#elif defined(USING_MFMA_32x32x16) && defined(ENABLE_FP8)
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 128;
constexpr ck_tile::index_t M_Warp = 1;
constexpr ck_tile::index_t N_Warp = 8;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = is_8bit_type<ADataType>::value ? 32 : 32;
constexpr ck_tile::index_t N_Warp_Tile = is_8bit_type<ADataType>::value ? 32 : 32;
constexpr ck_tile::index_t K_Warp_Tile = is_8bit_type<ADataType>::value ? 32 : 16;
#endif
using CodegenFlatmmShape =
ck_tile::TileFlatmmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
@@ -49,54 +69,112 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
AccDataType,
CodegenFlatmmShape,
CodegenGemmTraits>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
const auto Run = [&](const auto memory_operation_) {
constexpr auto memory_operation = memory_operation_.value;
using CodegenFlatmmPolicy = ck_tile::UniversalFlatmmPipelineAgBgCrPolicy;
using CodegenFlatmmPipeline =
ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenFlatmmPolicy>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::FlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue>;
using CodegenFlatmmPolicy = ck_tile::UniversalFlatmmPipelineAgBgCrPolicy;
using CodegenFlatmmPipeline =
ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenFlatmmPolicy>;
auto kargs = Kernel::MakeKernelArgs(args);
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::FlatmmKernel<TilePartitioner, CodegenFlatmmPipeline, GemmEpilogue>;
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
auto kargs = Kernel::MakeKernelArgs(args);
if(!Kernel::IsSupportedArgument(kargs))
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;
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
if(args.k_batch == 1)
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
if(s.log_level_ > 0)
else
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
#include "run_flatmm_example.inc"
int run_flatmm_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")
{
if(data_type == "fp16")
{
run_flatmm_example_with_layouts<ck_tile::half_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
run_flatmm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
run_flatmm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
run_flatmm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
return -1;
}
int main(int argc, char* argv[]) { return !run_flatmm_example(argc, argv); }

52
example/ck_tile/18_flatmm/flatmm_basic.hpp
View File

@@ -31,7 +31,7 @@
#error "unsupported CK_TILE_PIPELINE_DEFAULT value"
#endif
template <typename DataType>
template <typename ADataType, typename BDataType = ADataType, typename CDataType = ADataType>
struct GemmBasicTypeConfig;
template <>
@@ -44,9 +44,47 @@ struct GemmBasicTypeConfig<ck_tile::half_t>
// ToDo: Add more bias config to support different categories of GEMM.
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf16_t>
{
using ADataType = ck_tile::bf16_t;
using BDataType = ck_tile::bf16_t;
using AccDataType = float;
using CDataType = ck_tile::bf16_t;
};
template <>
struct GemmBasicTypeConfig<ck_tile::fp8_t>
{
using ADataType = ck_tile::fp8_t;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
// ToDo: Add more bias config to support different categories of GEMM.
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf8_t>
{
using ADataType = ck_tile::bf8_t;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
template <>
struct DataTypeTraits<float>
{
@@ -65,13 +103,11 @@ struct DataTypeTraits<ck_tile::half_t>
static constexpr const char* name = "fp16";
};
using Types = GemmBasicTypeConfig<ck_tile::half_t>;
// Specific type aliases for easy access
using ADataType = Types::ADataType;
using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType;
template <typename T>
struct is_8bit_type
: std::bool_constant<std::is_same_v<T, ck_tile::fp8_t> || std::is_same_v<T, ck_tile::bf8_t>>
{
};
auto create_args(int argc, char* argv[])
{

79
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -1,6 +1,20 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <type_traits>
template <typename T>
constexpr const char* DataTypeToString() {
if constexpr (std::is_same_v<T, ck_tile::half_t>) {
return "fp16";
} else if constexpr (std::is_same_v<T, ck_tile::fp8_t>) {
return "fp8";
} else if constexpr (std::is_same_v<T, ck_tile::bf8_t>) {
return "bf8";
} else {
return "unknown";
}
}
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
@@ -11,7 +25,7 @@ static constexpr inline auto is_row_major(Layout layout_)
// mfma_type, 0:32x32, 1:16x16
template <typename T>
auto shuffle_b(const ck_tile::HostTensor<T>& t, std::string mfma_dtype, int mfma_type = 0)
auto shuffle_b(const ck_tile::HostTensor<T>& t, std::string mfma_dtype, int mfma_type)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
@@ -29,13 +43,13 @@ auto shuffle_b(const ck_tile::HostTensor<T>& t, std::string mfma_dtype, int mfma
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 0)
else if((mfma_dtype == "int8" || mfma_dtype == "fp8" || mfma_dtype == "bf8") && mfma_type == 0)
{
ck_tile::HostTensor<T> t_view({n_ / 32, 32, k_ / 32, 2, 16});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {0, 2, 3, 1, 4});
}
else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 1)
else if((mfma_dtype == "int8" || mfma_dtype == "fp8" || mfma_dtype == "bf8") && mfma_type == 1)
{
ck_tile::HostTensor<T> t_view({n_ / 16, 16, k_ / 64, 4, 16});
std::copy(t.begin(), t.end(), t_view.begin());
@@ -44,6 +58,7 @@ auto shuffle_b(const ck_tile::HostTensor<T>& t, std::string mfma_dtype, int mfma
return t;
}
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
@@ -64,7 +79,13 @@ auto calculate_rtol_atol(const ck_tile::index_t K,
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
}
template <typename ALayout, typename BLayout, typename CLayout>
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
ck_tile::DeviceMem& b_shuffle_dev_buf,
ck_tile::DeviceMem& c_dev_buf,
@@ -91,7 +112,7 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
args.stride_B = stride_B;
args.stride_C = stride_C;
float ave_time = flatmm_calc<ALayout, BLayout, CLayout>(
float ave_time = flatmm_calc<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::size_t flop = std::size_t(2) * M * N * K;
@@ -100,7 +121,7 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Run Flatmm kernel with M =" << M << " N =" << N << " K =" << K
std::cout << "Run Flatmm kernel with DataType = " << DataTypeToString<ADataType>() << " M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
@@ -108,7 +129,10 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
return ave_time;
}
template <typename ALayout, typename BLayout, typename CLayout>
template <typename PrecType,
typename ALayout,
typename BLayout,
typename CLayout>
int run_flatmm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
@@ -119,6 +143,11 @@ int run_flatmm_example_with_layouts(int argc,
if(!result)
return -1;
using ADataType = typename GemmBasicTypeConfig<PrecType>::ADataType;
using BDataType = typename GemmBasicTypeConfig<PrecType>::BDataType;
using CDataType = typename GemmBasicTypeConfig<PrecType>::CDataType;
using AccDataType = typename GemmBasicTypeConfig<PrecType>::AccDataType;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
@@ -154,11 +183,17 @@ int run_flatmm_example_with_layouts(int argc,
// do pre-shuffle
std::string mfma = arg_parser.get_str("prec");
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b(b_origin_host, mfma, 0);
#if defined(USING_MFMA_16x16x32) && defined(ENABLE_FP8)
ck_tile::index_t mfma_type = 1;
#else
ck_tile::index_t mfma_type = 0;
#endif
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b(b_origin_host, mfma, mfma_type);
ck_tile::DeviceMem b_shuffle_dev_buf(b_shuffle_host.get_element_space_size_in_bytes());
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
invoke_flatmm<ALayout, BLayout, CLayout>(a_dev_buf,
invoke_flatmm<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
@@ -184,7 +219,7 @@ int run_flatmm_example_with_layouts(int argc,
a_host, b_origin_host, c_ref_host);
const float max_accumulated_value =
*std::max_element(c_ref_host.mData.begin(), c_ref_host.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_rslt_host,
c_ref_host,
"Error: Incorrect results!",
@@ -242,7 +277,7 @@ int run_flatmm_example_with_layouts(int argc,
c_gpu_ref_dev_buf.FromDevice(c_gpu_ref_host.data());
const float max_accumulated_value =
*std::max_element(c_gpu_ref_host.mData.begin(), c_gpu_ref_host.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_rslt_host,
c_gpu_ref_host,
"Error: Incorrect results!",
@@ -257,25 +292,3 @@ int run_flatmm_example_with_layouts(int argc,
return pass;
}
int run_flatmm_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 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_flatmm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}