mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 17:26:00 +00:00
[GEMM] Gemm universal device operation (#1154)
* Optimize GEMM on MI200/300: 1. Add new blockwise gemm pipeline 2. Add irregular splitk intances * clang format + typo fix * Fix a bug * initial commit * Add more instances to irregular splitk * blkgemm pipeline v1~4 prototype * Sanity Checked. Known issue: 1. Poor performance of splitk 2. Register spill on blkgemmpipeline v3 * Sanity and Performance fix: 1. fix a bug related to sanity in grouped b2c mapping 2. fix a bug related to sanity and performance in splitk offset * Sanity and API update: 1. Remove prefetch stage 2. Fix valid check bug 3, Add first gemm_universal instance into ckProfiler * Add NN instances for gemm universal * 1. Add NT instances for gemm_universal 2. Fix a bug about Kpadding in gemm_universal * Fix a bug regarding padding Odd K number * remove kernel print * Fix KPadding bug... * Update safety check * another try to fix kpadding.. * Sanity checked * new instances.. * clang format+typo fix * remove clang format script's change * Add non-hotloop compile option * 1. Add fp16xfp8 example 2. pull packed convert f8 from pr1150 * Some miscs.. opt and fix * Add pipeline description docs * Split universal gemm instance library to cut profiler compiling time * uncomment cmakefile * Fix a bug caused by blockwise_gemm_pipe_v2 * reduce default splitk to 1 * Add 224x256x64 tile size * update, including: 1. Experiment pipeline 5~7 2. Optimization for pipeline 4 3. Organized instance library * temp save * temp save * Permuted lds layout, sanity and function checked * clang format * Move OOB check from RunRead to RunWrite, for better software pipeline. TODO: agpr spill when NN layout * clangformat * A/B splitpipe scheduler for v3 * Fix two bugs * bug fix * fix a bug in oob check * Example for mixed fp16_fp8 gemm * Clean experimental code blocks * Add mixed precision gemm into profiler * tempsave * optimize m/n major lds layout * Add RRR GEMM mixed precision instances * Optimize f8 matrix transpose * Add test_gemm_universal * A/B spilt schedule for blkpip v5 * Take ds_read2 into iglp scheduling scheme * format * fixed cmake * Add llvm-option into CI cmake flag --------- Co-authored-by: Jing Zhang <jizhan@amd.com>
This commit is contained in:
Haocong WANG
164
profiler/src/profile_gemm_universal.cpp
Normal file
164
profiler/src/profile_gemm_universal.cpp
Normal file
@@ -0,0 +1,164 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
|
||||
#include "profiler/profile_gemm_universal_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
|
||||
F8_F16_F16, // 4
|
||||
F16_F8_F16, // 5
|
||||
F16_F16_F16_F8, // 6
|
||||
};
|
||||
|
||||
#define OP_NAME "gemm_universal"
|
||||
#define OP_DESC "Universal GEMM"
|
||||
|
||||
int profile_gemm_universal(int argc, char* argv[])
|
||||
{
|
||||
if(argc != 15 && argc != 17)
|
||||
{
|
||||
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
|
||||
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, "
|
||||
"comp f8)\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");
|
||||
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 int KBatch = std::stoi(argv[14]);
|
||||
|
||||
int n_warmup = 1;
|
||||
int n_iter = 10;
|
||||
if(argc == 17)
|
||||
{
|
||||
n_warmup = std::stoi(argv[15]);
|
||||
n_iter = std::stoi(argv[16]);
|
||||
}
|
||||
|
||||
using F32 = float;
|
||||
using F16 = ck::half_t;
|
||||
using F8 = ck::f8_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_universal_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,
|
||||
KBatch,
|
||||
n_warmup,
|
||||
n_iter);
|
||||
|
||||
return pass ? 0 : 1;
|
||||
};
|
||||
|
||||
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_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
|
||||
{
|
||||
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
|
||||
{
|
||||
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
|
||||
}
|
||||
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
|
||||
{
|
||||
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, 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_universal);
|
||||
Reference in New Issue
Block a user