ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 13:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[ckProfiler] Add infrastructure and instances to profile gemm_universal with B preshuffle (#2427)

Browse Source 
* works on mi300

* fix(profiler): add error message for unsupported type/layout

* refactor(preshuffle.inc): add type aliases for code readability
This commit is contained in:
Aviral Goel
2025-07-01 21:34:52 -04:00
committed by GitHub
parent a03682cb80
commit 36df1cbd0a
42 changed files with 2632 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

180
profiler/src/profile_gemm_universal_preshuffle.cpp Normal file
View File

@@ -0,0 +1,180 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include "profiler/profile_gemm_universal_preshuffle_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
// MK_KN_MN, // 0
MK_NK_MN = 1, // 1
// KM_KN_MN, // 2
// KM_NK_MN, // 3
};
enum struct GemmDataType
{
F8_F8_F16 = 0, // 0
F8_F8_BF16 = 1, // 1
// F32_F32_F32, // 0
// F16_F16_F16, // 1
// BF16_BF16_BF16, // 2
// INT8_INT8_INT8, // 3
// F8_F16_F16, // 4
// F16_F8_F16, // 5
// F16_F16_F16_F8, // 6
// F8_F8_BF16, // 7
// F16_I4_F16, // 8
// BF16_I4_BF16, // 9
};
#define OP_NAME "gemm_universal_preshuffle"
#define OP_DESC "Universal GEMM Preshuffle"
int profile_gemm_universal_preshuffle(int argc, char* argv[])
{
if(argc != 15 && argc != 18)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f8->bf16, 1: f8->f16)\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: split k into mulitiple batch\n");
printf("optional:\n");
printf("arg15: number of warm-up cycles (default 1)\n");
printf("arg16: number of iterations (default 10)\n");
printf("arg17: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
int M;
int N;
int StrideA;
int StrideB;
// Analyze the unsupported matrix shapes, switch the M and N number
if(std::stoi(argv[9]) % 8 != 0 && std::stoi(argv[8]) % 8 == 0)
{
M = std::stoi(argv[9]);
StrideA = std::stoi(argv[12]);
N = std::stoi(argv[8]);
StrideB = std::stoi(argv[11]);
}
else
{
M = std::stoi(argv[8]);
StrideA = std::stoi(argv[11]);
N = std::stoi(argv[9]);
StrideB = std::stoi(argv[12]);
}
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 K = std::stoi(argv[10]);
const int StrideC = std::stoi(argv[13]);
const int KBatch = std::stoi(argv[14]);
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 18)
{
n_warmup = std::stoi(argv[15]);
n_iter = std::stoi(argv[16]);
rotating = std::stoull(argv[17]) * 1024 * 1024;
}
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) || defined(CK_USE_WMMA_FP8)
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto comp_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 ComputeDataType = decltype(comp_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_universal_preshuffle_impl<ADataType,
BDataType,
ComputeDataType,
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,
KBatch,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F8_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94) || defined(CK_USE_WMMA_FP8)
if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_universal_preshuffle);