sglang_v0.5.2/pytorch_2.8.0/third_party/fbgemm/bench/GEMMsBenchmark.cc

362 lines
11 KiB
C++

/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include <algorithm>
#include <chrono>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <vector>
#ifdef _OPENMP
#include <omp.h>
#endif
#ifdef USE_MKL
#include <mkl.h>
#endif
#include "./BenchUtils.h"
#include "fbgemm/Fbgemm.h"
#include "src/RefImplementations.h"
#include "test/QuantizationHelpers.h"
using namespace std;
using namespace fbgemm;
void performance_test(
const int M,
const int N,
const int K,
const bool timebreak) {
// clang-format off
const vector<vector<int>> shapes = {
// NOTE: clang-format wants to use a different formatting but the current
// formatting should be easier to read.
// m, n, k
{M?M:64, N?N:800, K?K:320},
{M?M:64, N?N:768, K?K:512},
{M?M:16, N?N:256, K?K:512},
{M?M:128, N?N:128, K?K:128},
{M?M:256, N?N:512, K?K:256},
{M?M:1024, N?N:1024, K?K:1024},
};
// clang-format on
bool flush = true;
std::vector<char> llc;
if (flush) {
llc.resize(128 * 1024 * 1024, 1.0);
}
constexpr int NWARMUP = 4;
constexpr int NITER = 10;
if (timebreak) {
cout
<< "WARNING: the timer may be inaccurate when used by multiple threads."
<< endl;
cout << setw(8) << "M, " << setw(8) << "N, " << setw(8) << "K, " << setw(18)
<< "Type, " << setw(18) << "Packing (us), " << setw(18)
<< "Kernel (us), " << setw(18) << "Postproc (us), " << setw(18)
<< "Computation (us)," << setw(18) << "Total (us), " << setw(5)
<< "GOPs" << endl;
} else {
cout << setw(8) << "M, " << setw(8) << "N, " << setw(8) << "K, " << setw(18)
<< "Type, " << setw(5) << "GOPS" << endl;
}
chrono::time_point<chrono::high_resolution_clock> start, end;
for (const auto& shape : shapes) {
int m = shape[0];
int n = shape[1];
int k = shape[2];
aligned_vector<uint8_t> Aint8(m * k);
aligned_vector<int8_t> Bint8(k * n);
aligned_vector<float> Cfp32_mkl(m * n);
aligned_vector<int32_t> Cint32_mkl(Cfp32_mkl.size());
aligned_vector<int32_t> Cint32_ref(Cfp32_mkl.size());
aligned_vector<int32_t> Cint32_fb_acc32(Cfp32_mkl.size());
aligned_vector<int32_t> Cint32_fb_acc16(Cfp32_mkl.size());
// A matrix
randFill<uint8_t>(Aint8, 0, 5);
aligned_vector<float> Afp32(Aint8.begin(), Aint8.end());
randFill<int8_t>(Bint8, -4, 4);
avoidOverflow(m, n, k, Aint8.data(), Bint8.data());
aligned_vector<float> Bfp32(Bint8.begin(), Bint8.end());
double nops = 2.0 * m * n * k;
double ttot = 0.0;
string runType;
#ifdef USE_MKL
const float alpha = 1.f;
const float beta = 0.f;
runType = "MKL_fp32";
ttot = measureWithWarmup(
[&]() {
cblas_sgemm(
CblasRowMajor,
CblasNoTrans,
CblasNoTrans,
m,
n,
k,
alpha,
Afp32.data(),
k,
Bfp32.data(),
n,
beta,
Cfp32_mkl.data(),
n);
},
NWARMUP,
NITER,
[&]() {
if (flush) {
llc_flush(llc);
}
});
ttot *= 1e9; // convert to ns
std::ignore = ((volatile char*)(llc.data()));
cout << setw(6) << m << ", " << setw(6) << n << ", " << setw(6) << k << ", "
<< setw(16) << runType << ", ";
if (timebreak) {
cout << setw(16) << 0 << ", " << setw(16) << 0 << ", " << setw(16) << 0
<< ", " << setw(16) << ttot / 1e3 << ", ";
}
cout << setw(5) << fixed << setw(5) << setprecision(1) << nops / ttot
<< endl;
for (size_t i = 0; i < Cfp32_mkl.size(); ++i) {
Cint32_mkl[i] = (int32_t)Cfp32_mkl[i];
}
#endif
vector<int32_t> row_offsets(m);
matmul_u8i8acc32_ref(
m, n, k, k, n, n, Aint8.data(), Bint8.data(), Cint32_ref.data());
// printMatrix(matrix_op_t::NoTranspose, Bint8.data(), k, n, n, "B
// unpacked");
// printMatrix(matrix_op_t::NoTranspose, Aint8.data(), m, k, k,
// "A unpacked");
// printMatrix(matrix_op_t::NoTranspose, Cint32_ref.data(),
// m, n, n, "C int32");
PackBMatrix<int8_t> packedB_int32(
matrix_op_t::NoTranspose, k, n, Bint8.data(), n, nullptr, 1);
ttot = 0.0;
runType = "FBGEMM_i8_acc32";
double packing_time = 0.0, total_packing_time = 0.0;
double computing_time = 0.0, total_computing_time = 0.0;
double kernel_time = 0.0, total_kernel_time = 0.0;
double postprocessing_time = 0.0;
double total_postprocessing_time = 0.0;
double run_time = 0.0, total_run_time = 0.0;
cout << setw(6) << m << ", " << setw(6) << n << ", " << setw(6) << k << ", "
<< setw(16) << runType;
for (auto i = 0; i < NWARMUP + NITER; ++i) {
if (timebreak) {
packing_time = 0.0;
computing_time = 0.0;
kernel_time = 0.0;
postprocessing_time = 0.0;
run_time = 0.0;
}
llc_flush(llc);
start = chrono::high_resolution_clock::now();
#ifdef _OPENMP
#pragma omp parallel
#endif
{
PackAMatrix<uint8_t> packA_int32(
matrix_op_t::NoTranspose, m, k, Aint8.data(), k, nullptr, 1);
DoNothing<int32_t, int32_t> doNothing32BitObj;
memCopy<> memcopyObj(doNothing32BitObj);
int num_threads = fbgemm_get_num_threads();
int tid = fbgemm_get_thread_num();
// printf ( "tid: %d, num_threads: %d\n", tid, num_threads );
fbgemmPacked(
packA_int32,
packedB_int32,
Cint32_fb_acc32.data(),
Cint32_fb_acc32.data(),
n,
memcopyObj,
tid,
num_threads);
}
end = chrono::high_resolution_clock::now();
if (i >= NWARMUP) {
auto dur = chrono::duration_cast<chrono::nanoseconds>(end - start);
ttot += dur.count();
run_time = dur.count();
if (timebreak) {
total_packing_time += packing_time;
total_computing_time += computing_time;
total_kernel_time += kernel_time;
total_postprocessing_time += postprocessing_time;
total_run_time += run_time;
}
}
}
if (flush) {
((volatile char*)(llc.data()))[0] = llc.data()[0] + 1;
}
// printMatrix(matrix_op_t::NoTranspose, Bint8.data(), k, n, n, "B
// unpacked");
// printMatrix(matrix_op_t::NoTranspose, Aint8.data(), m, k, k,
// "A unpacked");
// printMatrix(matrix_op_t::NoTranspose,
// Cint8_fb.data(), m, n, n, "C fb");
if (timebreak) {
cout << ", " << setw(16) << total_packing_time / (double)NITER / 1e3
<< ", " << setw(16) << total_kernel_time / (double)NITER / 1e3
<< ", " << setw(16)
<< total_postprocessing_time / (double)NITER / 1e3 << ", "
<< setw(16) << total_computing_time / (double)NITER / 1e3 << ", "
<< setw(16) << total_run_time / (double)NITER / 1e3;
}
cout << ", " << setw(5) << fixed << setw(5) << setprecision(1)
<< NITER * nops / ttot << endl;
compare_buffers(Cint32_ref.data(), Cint32_fb_acc32.data(), m, n, n, 5);
PackBMatrix<int8_t, int16_t> packedB_int16(
matrix_op_t::NoTranspose, k, n, Bint8.data(), n, nullptr, 1);
ttot = 0.0;
runType = "FBGEMM_i8_acc16";
if (timebreak) {
total_packing_time = 0.0;
total_computing_time = 0.0;
total_kernel_time = 0.0;
total_postprocessing_time = 0.0;
total_run_time = 0.0;
}
cout << setw(6) << m << ", " << setw(6) << n << ", " << setw(6) << k << ", "
<< setw(16) << runType;
for (auto i = 0; i < NWARMUP + NITER; ++i) {
if (timebreak) {
packing_time = 0.0;
computing_time = 0.0;
kernel_time = 0.0;
postprocessing_time = 0.0;
run_time = 0.0;
}
llc_flush(llc);
start = chrono::high_resolution_clock::now();
#ifdef _OPENMP
#pragma omp parallel
#endif
{
PackAMatrix<uint8_t, int16_t> packA_int16(
matrix_op_t::NoTranspose, m, k, Aint8.data(), k, nullptr, 1);
DoNothing<int32_t, int32_t> doNothing32BitObj;
memCopy<> memcopyObj(doNothing32BitObj);
int num_threads = fbgemm_get_num_threads();
int tid = fbgemm_get_thread_num();
// printf ( "tid: %d, num_threads: %d\n", tid, num_threads );
fbgemmPacked(
packA_int16,
packedB_int16,
Cint32_fb_acc16.data(),
Cint32_fb_acc16.data(),
n,
memcopyObj,
tid,
num_threads);
}
end = chrono::high_resolution_clock::now();
if (i >= NWARMUP) {
auto dur = chrono::duration_cast<chrono::nanoseconds>(end - start);
ttot += dur.count();
run_time = dur.count();
if (timebreak) {
total_packing_time += packing_time;
total_computing_time += computing_time;
total_kernel_time += kernel_time;
total_postprocessing_time += postprocessing_time;
total_run_time += run_time;
}
}
}
if (flush) {
((volatile char*)(llc.data()))[0] = llc.data()[0] + 1;
}
// printMatrix(matrix_op_t::NoTranspose, Bint8.data(), k, n, n, "B
// unpacked");
// printMatrix(matrix_op_t::NoTranspose, Aint8.data(), m, k, k,
// "A unpacked");
// printMatrix(matrix_op_t::NoTranspose,
// Cint8_fb.data(), m, n, n, "C fb");
// compare_buffers(row_offsets.data(), row_offset_buf.data(),
// row_offsets.size(), 5);
if (timebreak) {
cout << ", " << setw(16) << total_packing_time / (double)NITER / 1e3
<< ", " << setw(16) << total_kernel_time / (double)NITER / 1e3
<< ", " << setw(16)
<< total_postprocessing_time / (double)NITER / 1e3 << ", "
<< setw(16) << total_computing_time / (double)NITER / 1e3 << ", "
<< setw(16) << total_run_time / (double)NITER / 1e3;
}
cout << ", " << setw(5) << fixed << setw(5) << setprecision(1)
<< NITER * nops / ttot << endl;
cout << endl;
compare_buffers(Cint32_ref.data(), Cint32_fb_acc16.data(), m, n, n, 5);
}
}
int main(int argc, const char** argv) {
#ifdef _OPENMP
// Use 1 thread unless OMP_NUM_THREADS is explicit set.
const char* val = getenv("OMP_NUM_THREADS");
if (val == nullptr || !*val) {
omp_set_num_threads(1);
}
#endif
const int M = parseArgumentInt(argc, argv, "--M=", 0, 0);
const int N = parseArgumentInt(argc, argv, "--N=", 0, 0);
const int K = parseArgumentInt(argc, argv, "--K=", 0, 0);
const bool timebreak = parseArgumentBool(argc, argv, "--timebreak", false);
performance_test(M, N, K, timebreak);
return 0;
}