1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
|
/*
* Copyright (c) Facebook, Inc. and its 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 <chrono>
#include <cmath>
#include <random>
#ifdef USE_MKL
#include <mkl.h>
#endif
#ifdef _OPENMP
#include <omp.h>
#endif
#include "bench/BenchUtils.h"
#include "fbgemm/FbgemmFP16.h"
#include "AlignedVec.h"
using namespace std;
using namespace fbgemm2;
void performance_test() {
// cache flush
bool flush = true;
std::vector<char> llc;
if (flush) {
llc.resize(64L * 1024L * 1024L, 1.0);
}
float alpha = 1.f, beta = 1.f;
matrix_op_t btran = matrix_op_t::Transpose;
using btype = float16;
#define dataset 1
#if dataset == 1
const int NITER = (flush) ? 10 : 100;
std::vector<std::vector<int>> shapes;
for (auto m = 1; m < 120; m++) {
// shapes.push_back({m, 128, 512});
shapes.push_back({m, 512, 512});
}
#elif dataset == 2
const int NITER = (flush) ? 10 : 100;
#include "shapes_dataset.h"
#else
flush = false;
constexpr int NITER = 1;
std::vector<std::vector<int>> shapes;
std::random_device r;
std::default_random_engine generator(r());
std::uniform_int_distribution<int> dm(1, 100);
std::uniform_int_distribution<int> dnk(1, 1024);
for (int i = 0; i < 1000; i++) {
int m = dm(generator);
int n = dnk(generator);
int k = dnk(generator);
shapes.push_back({m, n, k});
}
#endif
std::string type;
double gflops, gbs, ttot;
for (auto s : shapes) {
int m = s[0];
int n = s[1];
int k = s[2];
aligned_vector<float> A(m * k, 0.f);
aligned_vector<float> B(k * n, 0.f);
aligned_vector<float> Cg(m * n, 1.f);
aligned_vector<float> Cp(m * n, NAN);
// initialize with small numbers
randFill(A, 0, 4);
randFill(B, 0, 4);
PackedGemmMatrixFP16 Bp(btran, k, n, alpha, B.data());
if (beta != 0.0f) {
randFill(Cg, 0, 4);
Cp = Cg;
}
double nflops = 2.0 * (double)m * (double)n * (double)k * (double)NITER;
double nbytes = (4.0 * (double)m * (double)k + 2.0 * (double)k * (double)n +
4.0 * (double)m * (double)n) *
NITER;
// warm up MKL and fbgemm
// check correctness at the same time
for (auto w = 0; w < 3; w++) {
#ifdef USE_MKL
cblas_sgemm(
CblasRowMajor,
CblasNoTrans,
btran == matrix_op_t::Transpose ? CblasTrans : CblasNoTrans,
m,
n,
k,
alpha,
A.data(),
k,
B.data(),
(btran == matrix_op_t::NoTranspose) ? n : k,
beta,
Cg.data(),
n);
#endif
cblas_gemm_compute(
matrix_op_t::NoTranspose, m, A.data(), Bp, beta, Cp.data());
#ifdef USE_MKL
// Compare results
for (auto i = 0; i < Cg.size(); i++) {
// printf("%f %f\n", Cg[i], Cp[i]);
assert(std::abs(Cg[i] - Cp[i]) < 1e-3);
}
#endif
}
chrono::time_point<chrono::system_clock> t_begin, t_end;
#ifdef USE_MKL
// Gold via MKL sgemm
type = "MKL_FP32";
ttot = 0;
for (auto it = -3; it < NITER; it++) {
if (flush) {
for (auto i = 0; i < llc.size(); i++) {
llc[i]++;
}
}
t_begin = chrono::system_clock::now();
cblas_sgemm(
CblasRowMajor,
CblasNoTrans,
btran == matrix_op_t::Transpose ? CblasTrans : CblasNoTrans,
m,
n,
k,
alpha,
A.data(),
k,
B.data(),
(btran == matrix_op_t::NoTranspose) ? n : k,
beta,
Cg.data(),
n);
t_end = chrono::system_clock::now();
if (it >= 0) {
double dt = chrono::duration<double>(t_end - t_begin).count();
ttot += dt;
}
}
gflops = nflops / ttot / 1e9;
gbs = nbytes / ttot / 1e9;
printf(
"\n%15s m = %5d n = %5d k = %5d Gflops = %8.4lf GBytes = %8.4lf\n",
type.c_str(),
m,
n,
k,
gflops,
gbs);
((volatile char*)(llc.data()));
#endif
type = "FBP_" + std::string(typeid(btype).name());
ttot = 0;
for (auto it = -3; it < NITER; it++) {
if (flush) {
for (auto i = 0; i < llc.size(); i++) {
llc[i]++;
}
}
t_begin = chrono::system_clock::now();
cblas_gemm_compute(
matrix_op_t::NoTranspose, m, A.data(), Bp, beta, Cp.data());
t_end = chrono::system_clock::now();
if (it >= 0) {
double dt = chrono::duration<double>(t_end - t_begin).count();
ttot += dt;
}
}
gflops = nflops / ttot / 1e9;
gbs = nbytes / ttot / 1e9;
printf(
"%15s m = %5d n = %5d k = %5d Gflops = %8.4lf GBytes = %8.4lf\n",
type.c_str(),
m,
n,
k,
gflops,
gbs);
((volatile char*)(llc.data()));
}
}
int main(int /*argc*/, char** /*argv*/) {
#ifdef _OPENMP
omp_set_num_threads(1);
#endif
performance_test();
}
|
