-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.cpp
351 lines (333 loc) · 15.8 KB
/
main.cpp
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
//
// Created by FdyCN on 2024/2/22.
//
#include <cuda.h>
#include <cuda_runtime.h>
#include <iostream>
#include <fstream>
#include <string>
#include "gemm.hpp"
#include "test_utils.hpp"
#include "cublas_v2.h"
#define CHECK_CUBLAS(expr) \
if((expr) != CUBLAS_STATUS_SUCCESS) \
{ \
std::cout << "cublas function: [ " << #expr << " ] error!" << std::endl; \
return -1; \
}
#define CHECK_CUDA(expr) \
if((expr) != cudaSuccess) \
{ \
std::cout << "cuda function: [ " << #expr << " ] error!" << std::endl; \
return -1; \
}
int get_gpu_properties() {
int device_count;
CHECK_CUDA(cudaGetDeviceCount(&device_count));
for (int i = 0; i < device_count; i++) {
cudaDeviceProp prop;
printf("=================GPU [%d]=================\n", i);
CHECK_CUDA(cudaGetDeviceProperties(&prop, i));
printf("GPU Name = %s\n", prop.name);
printf("Compute Capability = %d.%d\n", prop.major, prop.minor);
printf("GPU SMs = %d\n", prop.multiProcessorCount);
printf("GPU CUDA cores = %d\n", query_cudacore_per_sm(prop.major, prop.minor) * prop.multiProcessorCount);
printf("GPU Mem clock rate = %.3f GHz\n", prop.memoryClockRate / 1e6);
printf("[In Boost]GPU SM clock rate = %.3f GHz\n", prop.clockRate / 1e6);
printf("[In Boost]FP32 Peak Performance = %.3f GFLOPS\n",
query_cudacore_per_sm(prop.major, prop.minor) * prop.multiProcessorCount * (prop.clockRate / 1e6) * 2);
int total_cudacore = query_cudacore_per_sm(prop.major, prop.minor) * prop.multiProcessorCount;
if (support_fp16(prop.major, prop.minor)) {
printf("[In Boost]FP16 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 2);
}
if (support_int8(prop.major, prop.minor)) {
printf("[In Boost]INT8 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 4);
}
if (support_tensorcore_v1(prop.major, prop.minor)) {
printf("[In Boost]Tensor Core V1 FP16 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 8);
}
if (support_tensorcore_v2(prop.major, prop.minor)) {
printf("[In Boost]Tensor Core V2 FP16 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 8);
printf("[In Boost]Tensor Core V2 INT8 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 16);
}
if (support_tensorcore_v3(prop.major, prop.minor)) {
printf("[In Boost]Tensor Core V3 FP16 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 16);
printf("[In Boost]Tensor Core V3 INT8 Peak Performance = %.3f GFLOPS\n",
total_cudacore * (prop.clockRate / 1e6) * 2 * 32);
}
printf("\n");
}
return 0;
}
int
test_gemm_kernels(const int M, const int N, const int K, GEMM_OP op, float *perf = nullptr, bool perf_only = false) {
float *a = (float *) malloc(M * K * sizeof(float));
generate_random_float(a, M * K);
float *b = (float *) malloc(N * K * sizeof(float));
generate_random_float(b, N * K);
float *c = (float *) malloc(M * N * sizeof(float));
memset(c, 0, M * N * sizeof(float));
float *d = (float *) malloc(M * N * sizeof(float));
memset(d, 0, M * N * sizeof(float));
float *dev_a;
cudaMalloc((void **) &dev_a, M * K * sizeof(float));
cudaMemcpy(dev_a, a, M * K * sizeof(float), cudaMemcpyKind::cudaMemcpyHostToDevice);
float *dev_b;
cudaMalloc((void **) &dev_b, N * K * sizeof(float));
cudaMemcpy(dev_b, b, N * K * sizeof(float), cudaMemcpyKind::cudaMemcpyHostToDevice);
float *dev_c;
cudaMalloc((void **) &dev_c, M * N * sizeof(float));
int iter = 10;
cublasHandle_t handle;
CHECK_CUBLAS(cublasCreate(&handle));
switch (op) {
case GEMM_OP::FLOAT_NAIVE_GEMM_N_T: {
CHECK_RETURN(gemm_float(dev_a, dev_b, dev_c, M, N, K, iter, op, perf), "FLOAT_NAIVE_GEMM_N_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, false, true);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_NAIVE_GEMM_N_T");
}
break;
}
case GEMM_OP::FLOAT_SRAM_GEMM_N_N: {
CHECK_RETURN(gemm_float(dev_a, dev_b, dev_c, M, N, K, iter, op, perf), "FLOAT_SRAM_GEMM_N_N");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, false, false);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_SRAM_GEMM_N_N");
}
break;
}
case GEMM_OP::FLOAT_CUBLAS_GEMM_N_N: {
CHECK_RETURN(cublas_gemm_float((void **) &handle, dev_a, dev_b, dev_c, M, N, K, iter, op, perf),
"FLOAT_CUBLAS_GEMM_N_N");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, false, false);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_CUBLAS_GEMM_N_N");
}
break;
}
case GEMM_OP::FLOAT_CUBLAS_GEMM_N_T: {
CHECK_RETURN(cublas_gemm_float((void **) &handle, dev_a, dev_b, dev_c, M, N, K, iter, op, perf),
"FLOAT_CUBLAS_GEMM_N_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, false, true);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_CUBLAS_GEMM_N_T");
}
break;
}
case GEMM_OP::FLOAT_CUBLAS_GEMM_T_N: {
CHECK_RETURN(cublas_gemm_float((void **) &handle, dev_a, dev_b, dev_c, M, N, K, iter, op, perf),
"FLOAT_CUBLAS_GEMM_T_N");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, true, false);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_CUBLAS_GEMM_T_N");
}
break;
}
case GEMM_OP::FLOAT_CUBLAS_GEMM_T_T: {
CHECK_RETURN(cublas_gemm_float((void **) &handle, dev_a, dev_b, dev_c, M, N, K, iter, op, perf),
"FLOAT_CUBLAS_GEMM_T_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(float), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_float(a, b, d, M, N, K, true, true);
CHECK_TEST(compare_results<float>(d, c, M * N), "FLOAT_CUBLAS_GEMM_T_T");
}
break;
}
default: {
std::cout << "unsupported test op type!" << std::endl;
break;
}
}
// Destroy the handle
CHECK_CUBLAS(cublasDestroy(handle));
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
free(a);
free(b);
free(c);
free(d);
return 0;
}
int
test_gemm_kernels_half(const int M, const int N, const int K, GEMM_OP op, float *perf = nullptr,
bool perf_only = false) {
half_float::half *a = (half_float::half *) malloc(M * K * sizeof(half_float::half));
generate_random_half(a, M * K);
half_float::half *b = (half_float::half *) malloc(N * K * sizeof(half_float::half));
generate_random_half(b, N * K);
half_float::half *c = (half_float::half *) malloc(M * N * sizeof(half_float::half));
memset(c, 0, M * N * sizeof(half_float::half));
half_float::half *d = (half_float::half *) malloc(M * N * sizeof(half_float::half));
memset(d, 0, M * N * sizeof(half_float::half));
void *dev_a;
cudaMalloc((void **) &dev_a, M * K * sizeof(half_float::half));
cudaMemcpy(dev_a, a, M * K * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyHostToDevice);
void *dev_b;
cudaMalloc((void **) &dev_b, N * K * sizeof(half_float::half));
cudaMemcpy(dev_b, b, N * K * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyHostToDevice);
void *dev_c;
cudaMalloc((void **) &dev_c, M * N * sizeof(half_float::half));
int iter = 10;
cublasHandle_t handle;
CHECK_CUBLAS(cublasCreate(&handle));
switch (op) {
case GEMM_OP::HALF_NAIVE_TENSORCORE_N_T: {
CHECK_RETURN(gemm_half((void *) dev_a, (void *) dev_b, (void *) dev_c, M, N, K, iter, op, perf),
"HALF_NAIVE_TENSORCORE_N_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_half(a, b, d, M, N, K, false, true);
CHECK_TEST(compare_results_half(d, c, M * N), "HALF_NAIVE_TENSORCORE_N_T");
}
break;
}
case GEMM_OP::HALF_CUBLAS_GEMM_N_N: {
CHECK_RETURN(
cublas_gemm_half((void **) &handle, (void *) dev_a, (void *) dev_b, (void *) dev_c, M, N, K, iter,
op,
perf),
"HALF_CUBLAS_GEMM_N_N");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_half(a, b, d, M, N, K, false, false);
CHECK_TEST(compare_results_half(d, c, M * N), "HALF_CUBLAS_GEMM_N_N");
}
break;
}
case GEMM_OP::HALF_CUBLAS_GEMM_N_T: {
CHECK_RETURN(
cublas_gemm_half((void **) &handle, (void *) dev_a, (void *) dev_b, (void *) dev_c, M, N, K, iter,
op,
perf),
"HALF_CUBLAS_GEMM_N_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_half(a, b, d, M, N, K, false, true);
CHECK_TEST(compare_results_half(d, c, M * N), "HALF_CUBLAS_GEMM_N_T");
}
break;
}
case GEMM_OP::HALF_CUBLAS_GEMM_T_N: {
CHECK_RETURN(
cublas_gemm_half((void **) &handle, (void *) dev_a, (void *) dev_b, (void *) dev_c, M, N, K, iter,
op,
perf),
"HALF_CUBLAS_GEMM_T_N");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_half(a, b, d, M, N, K, true, false);
CHECK_TEST(compare_results_half(d, c, M * N), "HALF_CUBLAS_GEMM_T_N");
}
break;
}
case GEMM_OP::HALF_CUBLAS_GEMM_T_T: {
CHECK_RETURN(
cublas_gemm_half((void **) &handle, (void *) dev_a, (void *) dev_b, (void *) dev_c, M, N, K, iter,
op,
perf),
"HALF_CUBLAS_GEMM_T_T");
if (!perf_only) {
cudaMemcpy(c, dev_c, M * N * sizeof(half_float::half), cudaMemcpyKind::cudaMemcpyDeviceToHost);
standard_gemm_host_half(a, b, d, M, N, K, true, true);
CHECK_TEST(compare_results_half(d, c, M * N), "HALF_CUBLAS_GEMM_T_T");
}
break;
}
default: {
std::cout << "unsupported test op type!" << std::endl;
break;
}
}
// Destroy the handle
CHECK_CUBLAS(cublasDestroy(handle));
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
free(a);
free(b);
free(c);
free(d);
return 0;
}
int main(int argc, char **argv) {
get_gpu_properties();
if (argc != 5 && argc != 1) {
printf("usage: ./main [M] [N] [K] [Performance only] or ./main (means batch tests performace) \n");
exit(0);
}
if (argc == 5) {
size_t M = atoi(argv[1]);
size_t N = atoi(argv[2]);
size_t K = atoi(argv[3]);
bool performance_only = atoi(argv[4]) != 0 ? true : false;
printf("=====FLOAT OP=====\n");
test_gemm_kernels(M, N, K, GEMM_OP::FLOAT_SRAM_GEMM_N_N, nullptr, performance_only);
test_gemm_kernels(M, N, K, GEMM_OP::FLOAT_CUBLAS_GEMM_N_N, nullptr, performance_only);
test_gemm_kernels(M, N, K, GEMM_OP::FLOAT_CUBLAS_GEMM_T_N, nullptr, performance_only);
test_gemm_kernels(M, N, K, GEMM_OP::FLOAT_CUBLAS_GEMM_N_T, nullptr, performance_only);
test_gemm_kernels(M, N, K, GEMM_OP::FLOAT_CUBLAS_GEMM_T_T, nullptr, performance_only);
printf("\n=====HALF OP=====\n");
test_gemm_kernels_half(M, N, K, GEMM_OP::HALF_NAIVE_TENSORCORE_N_T, nullptr, performance_only);
test_gemm_kernels_half(M, N, K, GEMM_OP::HALF_CUBLAS_GEMM_N_N, nullptr, performance_only);
test_gemm_kernels_half(M, N, K, GEMM_OP::HALF_CUBLAS_GEMM_N_T, nullptr, performance_only);
test_gemm_kernels_half(M, N, K, GEMM_OP::HALF_CUBLAS_GEMM_T_N, nullptr, performance_only);
test_gemm_kernels_half(M, N, K, GEMM_OP::HALF_CUBLAS_GEMM_T_T, nullptr, performance_only);
}
if (argc == 1) {
std::vector<int> MNK(16, 256);
int index = 1;
for (auto &v: MNK) {
v *= index;
index++;
}
std::string fname = "gemm_batch_test.csv";
std::ofstream out_file(fname, std::ios::out);
if (out_file.is_open()) {
// titles
out_file << "name" << ','
<< "M" << ','
<< "N" << ','
<< "K" << ','
<< "Perf(GFLOPS)" << std::endl;
for (const auto &v: MNK) {
float perf = 0.0f; // returned in GFLOPS
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_SRAM_GEMM_N_N, &perf, true);
out_file << "FLOAT_SRAM_GEMM_N_N" << ',' << v << ',' << v << ',' << v << ',' << perf << std::endl;
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_N_N, &perf, true);
out_file << "FLOAT_CUBLAS_GEMM_N_N" << ',' << v << ',' << v << ',' << v << ',' << perf << std::endl;
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_N_T, &perf, true);
out_file << "FLOAT_CUBLAS_GEMM_N_T" << ',' << v << ',' << v << ',' << v << ',' << perf << std::endl;
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_T_N, &perf, true);
out_file << "FLOAT_CUBLAS_GEMM_T_N" << ',' << v << ',' << v << ',' << v << ',' << perf << std::endl;
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_T_T, &perf, true);
out_file << "FLOAT_CUBLAS_GEMM_T_T" << ',' << v << ',' << v << ',' << v << ',' << perf << std::endl;
}
out_file.close();
} else {
std::cout << "Cannot save to file: " << fname << std::endl;
std::cout << "ONLY print in terminal...." << std::endl;
for (const auto &v: MNK) {
float perf = 0.0f; // returned in GFLOPS
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_SRAM_GEMM_N_N, &perf, true);
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_N_N, &perf, true);
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_T_N, &perf, true);
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_N_T, &perf, true);
test_gemm_kernels(v, v, v, GEMM_OP::FLOAT_CUBLAS_GEMM_T_T, &perf, true);
}
}
}
return 0;
}