mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 09:16:52 +00:00
initial stream-k implementation with example (#699)
* initial stream-k implementation with example * fix unexpected change in err * improve a little bit performance by reorganize pipeline. * improve perf a little bit by swizzle block idx * add profiler * update example * fix spelling * shrink karg for streamk * support dynamic buffer using memory coherence glc_slc bit from template * control memory coherence while construct dynamic buffer * update reduction for streamk(not ready yet) * Add template parameter to make_dynamic_buffer to support amd_buffer coherence setting * fix build issue * fix several bug * now result is correct, everything works (but has scratch) * remove scratch by manually reset coordinate * update device code * fix a bug in final reduce * fix something in example * update async memset * fix enum as camel case * modify coherence enum name * clean code and use atomic streamk by default * remove unused var * throw exception if have empty pointer * fix format * fix CI warning * fix type in init * modify CI error * filter out on gfx10+ * restore changed example code --------- Co-authored-by: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
This commit is contained in:
carlushuang
@@ -3,6 +3,7 @@ set(PROFILER_SOURCES
|
||||
profiler.cpp
|
||||
profile_gemm.cpp
|
||||
profile_gemm_splitk.cpp
|
||||
profile_gemm_streamk.cpp
|
||||
profile_gemm_bilinear.cpp
|
||||
profile_gemm_bias_add_reduce.cpp
|
||||
profile_gemm_add_add_fastgelu.cpp
|
||||
@@ -48,6 +49,7 @@ target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
|
||||
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
|
||||
|
||||
155
profiler/src/profile_gemm_streamk.cpp
Normal file
155
profiler/src/profile_gemm_streamk.cpp
Normal file
@@ -0,0 +1,155 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "profiler/profile_gemm_streamk_impl.hpp"
|
||||
#include "profiler_operation_registry.hpp"
|
||||
|
||||
enum struct GemmMatrixLayout
|
||||
{
|
||||
MK_KN_MN, // 0
|
||||
MK_NK_MN, // 1
|
||||
KM_KN_MN, // 2
|
||||
KM_NK_MN, // 3
|
||||
};
|
||||
|
||||
enum struct GemmDataType
|
||||
{
|
||||
F32_F32_F32, // 0
|
||||
F16_F16_F16, // 1
|
||||
BF16_BF16_BF16, // 2
|
||||
INT8_INT8_INT8, // 3
|
||||
};
|
||||
|
||||
#define OP_NAME "gemm_streamk"
|
||||
#define OP_DESC "StreamK GEMM"
|
||||
|
||||
int profile_gemm_streamk(int argc, char* argv[])
|
||||
{
|
||||
if(argc < 14)
|
||||
{
|
||||
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
|
||||
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
|
||||
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
|
||||
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
|
||||
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
|
||||
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
|
||||
printf("arg4: verification (0: no; 1: yes)\n");
|
||||
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
|
||||
printf("arg6: print tensor value (0: no; 1: yes)\n");
|
||||
printf("arg7: time kernel (0=no, 1=yes)\n");
|
||||
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
|
||||
printf("arg14: num_sk_blocks (optional)\n");
|
||||
exit(1);
|
||||
}
|
||||
|
||||
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
|
||||
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
|
||||
const bool do_verification = std::stoi(argv[4]);
|
||||
const int init_method = std::stoi(argv[5]);
|
||||
const bool do_log = std::stoi(argv[6]);
|
||||
const bool time_kernel = std::stoi(argv[7]);
|
||||
|
||||
const int M = std::stoi(argv[8]);
|
||||
const int N = std::stoi(argv[9]);
|
||||
const int K = std::stoi(argv[10]);
|
||||
|
||||
const int StrideA = std::stoi(argv[11]);
|
||||
const int StrideB = std::stoi(argv[12]);
|
||||
const int StrideC = std::stoi(argv[13]);
|
||||
const uint32_t NumSKBlocks =
|
||||
argc >= 15 ? static_cast<uint32_t>(std::stoul(std::string(argv[14]))) : 0xffffffff;
|
||||
|
||||
using F32 = float;
|
||||
using F16 = ck::half_t;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
auto profile = [&](auto a_type,
|
||||
auto b_type,
|
||||
auto acc_type,
|
||||
auto c_type,
|
||||
auto a_layout,
|
||||
auto b_layout,
|
||||
auto c_layout) {
|
||||
using ADataType = decltype(a_type);
|
||||
using BDataType = decltype(b_type);
|
||||
using AccDataType = decltype(acc_type);
|
||||
using CDataType = decltype(c_type);
|
||||
|
||||
using ALayout = decltype(a_layout);
|
||||
using BLayout = decltype(b_layout);
|
||||
using CLayout = decltype(c_layout);
|
||||
|
||||
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
|
||||
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
|
||||
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
|
||||
|
||||
bool pass = ck::profiler::profile_gemm_streamk_impl<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
do_verification,
|
||||
init_method,
|
||||
do_log,
|
||||
time_kernel,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
(StrideA <= 0) ? DefaultStrideA : StrideA,
|
||||
(StrideB <= 0) ? DefaultStrideB : StrideB,
|
||||
(StrideC <= 0) ? DefaultStrideC : StrideC,
|
||||
NumSKBlocks);
|
||||
|
||||
return pass ? 0 : 1;
|
||||
};
|
||||
|
||||
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
|
||||
{
|
||||
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
|
||||
{
|
||||
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
|
||||
{
|
||||
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
|
||||
{
|
||||
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << "this data_type & layout is not implemented" << std::endl;
|
||||
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_streamk);
|
||||
Reference in New Issue
Block a user