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
|
/*
Copyright (c) 2018-2023, Intel Corporation
SPDX-License-Identifier: BSD-3-Clause
*/
// Simple harness for testing SGEMM implementations in ISPC
// Junkins, September 2018
// clang-format off
/**
Matrix layout, rows x col, row major storage:
N K K
--- --- --- --- --- ---
| | | | | |
M | | X N | | = M | |
| | | | | |
--- --- --- --- --- ---
A X B = C
**/
// clang-format on
#ifdef _MSC_VER
#define _CRT_SECURE_NO_WARNINGS
#endif
#include "../../common/timing.h"
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
// OS Independent millisec wall timers
#ifdef _WIN32
#define WINDOWS
#endif
#ifdef _WIN64
#define WINDOWS
#endif
#ifdef __linux__
#define LINUX
#endif
#ifdef WINDOWS
#define TIMER_DECLARE_AND_INIT() \
LARGE_INTEGER beginClock, endClock, cpuClockFreq; \
QueryPerformanceFrequency(&cpuClockFreq);
#define TIMER_RESET_AND_START() QueryPerformanceCounter(&beginClock);
#define TIMER_GET_ELAPSED_MSEC() \
(QueryPerformanceCounter(&endClock) & 0) + (QueryPerformanceFrequency(&cpuClockFreq) & 0) + \
(double(endClock.QuadPart - beginClock.QuadPart) * 1000.0f / cpuClockFreq.QuadPart)
#else
#define TIMER_DECLARE_AND_INIT()
#define TIMER_RESET_AND_START() reset_and_start_timer();
#define TIMER_GET_ELAPSED_MSEC() get_elapsed_msec();
#endif
// Include the header file that the ispc compiler generates
#include "SGEMM_kernels_ispc.h"
using namespace ispc;
void init_matrix(float M[], unsigned int rows, unsigned int cols, float value) {
for (unsigned int r = 0; r < rows; r++)
for (unsigned int c = 0; c < cols; c++)
M[r * cols + c] = value;
}
void init_matrix_rand(float M[], unsigned int rows, unsigned int cols, float rangeValue) {
for (unsigned int r = 0; r < rows; r++)
for (unsigned int c = 0; c < cols; c++) {
float rnd = ((float)rand() / (float)(RAND_MAX)) * rangeValue;
M[r * cols + c] = rnd;
}
}
void print_matrix(char *pcName, float M[], unsigned int rows, unsigned int cols) {
printf("%s:\n", pcName);
for (unsigned int r = 0; r < rows; r++) {
for (unsigned int c = 0; c < cols; c++) {
printf("%8.2f ", M[r * cols + c]);
}
printf("\n");
}
printf("\n");
}
void SGEMM_CPU_validation(float matrixA[], float matrixB[], float matrixC[], unsigned int M, unsigned int N,
unsigned int K) {
for (unsigned int m = 0; m < M; m++) {
unsigned int n, k;
float sum;
for (k = 0; k < K; k++) {
sum = 0.0f;
for (n = 0; n < N; n++) {
sum += matrixA[m * N + n] * matrixB[n * K + k];
}
matrixC[m * K + k] = sum;
}
}
}
#define EPSILON 0.01f
bool Validate_result(float matrixC[], float matrixValid[], unsigned int M, unsigned int K) {
for (unsigned int m = 0; m < M; m++) {
for (unsigned int k = 0; k < K; k++) {
float delta = (float)fabs(matrixC[m * K + k] - matrixValid[m * K + k]);
if (delta > EPSILON)
return false;
}
}
return true;
}
typedef void (*SGEMMFuncPtr)(void);
typedef void (*SGEMMFuncPtr_SingleThreaded)(float matrixA[], float matrixB[], float matrixC[], unsigned int M,
unsigned int N, unsigned int K);
typedef void (*SGEMMFuncPtr_MultiThreaded)(float matrixA[], float matrixB[], float matrixC[], unsigned int M,
unsigned int N, unsigned int K);
void Test_SGEMM(SGEMMFuncPtr SGEMMFunc, char *pcFuncName, float matrixA[], float matrixB[], float matrixC[],
unsigned int M, unsigned int N, unsigned int K, bool tasks, unsigned int numIterations,
float matrixValid[]) {
double totalWallTime;
float avgTime;
unsigned int i;
bool bValid;
TIMER_DECLARE_AND_INIT();
// Total = MNK mults + MN(K-1) adds.
float fFlopsPerGEMM = (float)(M * N * K) + (M * N * (K - 1));
if (tasks == false) {
// type cast
SGEMMFuncPtr_SingleThreaded SGEMMFunc_ST = (SGEMMFuncPtr_SingleThreaded)SGEMMFunc;
TIMER_RESET_AND_START();
for (i = 0; i < numIterations; i++)
SGEMMFunc_ST(matrixA, matrixB, matrixC, M, N, K);
totalWallTime = TIMER_GET_ELAPSED_MSEC();
} else {
// type cast
SGEMMFuncPtr_MultiThreaded SGEMMFunc_MT = (SGEMMFuncPtr_MultiThreaded)SGEMMFunc;
TIMER_RESET_AND_START();
for (i = 0; i < numIterations; i++)
SGEMMFunc_MT(matrixA, matrixB, matrixC, M, N, K);
totalWallTime = TIMER_GET_ELAPSED_MSEC();
}
avgTime = (float)totalWallTime / (float)numIterations;
bValid = Validate_result(matrixC, matrixValid, M, K);
printf("%40s %10.4f millisecs %10.4f GFLOPs Validation: %s.\n", pcFuncName, avgTime,
(fFlopsPerGEMM / (avgTime / 1000.0f)) / 1000000000.0f, (bValid ? "valid" : "ERROR"));
init_matrix(matrixC, M, K, 0.0f);
}
int main(int argc, char **argv) {
// Random number filled matrix test case:
// Default values for input parameters
int ITERATIONS = 500;
int M = 256;
int N = 64;
int K = 512;
printf("\nUsage: SGEMM (optional)[ispc iterations] (optional)[[Matrix A Rows] [Matrix A Columns/ matrix B Rows] "
"[Matrix B Columns]]\n");
if (argc < 2) {
printf("ispc iterations = %d[default],\t Matrix A Rows = %d[default],\t Matrix A Columns/ matrix B Rows = "
"%d[default], Matrix B Columns = %d[default]\n",
ITERATIONS, M, N, K);
} else if (argc == 2) {
ITERATIONS = atoi(argv[1]);
printf("%s\n", argv[0]);
printf("ispc iterations = %d,\t Matrix A Rows = %d[default],\t Matrix A Columns/ matrix B Rows = %d[default], "
"Matrix B Columns = %d[default]\n",
ITERATIONS, M, N, K);
} else if (argc == 4) {
M = atoi(argv[1]);
N = atoi(argv[2]);
K = atoi(argv[3]);
printf("%s\n", argv[0]);
printf("ispc iterations = %d[default],\t Matrix A Rows = %d,\t Matrix A Columns/ matrix B Rows = %d, Matrix B "
"Columns = %d\n",
ITERATIONS, M, N, K);
} else if (argc == 5) {
ITERATIONS = atoi(argv[1]);
M = atoi(argv[2]);
N = atoi(argv[3]);
K = atoi(argv[4]);
printf("%s\n", argv[0]);
printf("ispc iterations = %d,\t Matrix A Rows = %d,\t Matrix A Columns/ matrix B Rows = %d, Matrix B Columns = "
"%d\n",
ITERATIONS, M, N, K);
} else {
printf("%s\n", argv[0]);
printf("\nInvalid number of inputs\n");
exit(-1);
}
int programCount = SGEMM_get_program_count();
int tileSize = SGEMM_get_tile_size();
if (K % programCount != 0 || K % tileSize != 0) {
printf("\nNumber of columns in Matrix B (K) must be a multiple of %d (target width) and %d (tile size)!\n",
programCount, tileSize);
exit(-1);
}
if (M % programCount != 0) {
printf("\nNumber of rows in Matrix A (M) must be a multiple of %d (target width)!\n", programCount);
exit(-1);
}
if (N % programCount != 0) {
printf("\nNumber of columns in Matrix A (N), which is also number of rows in Matrix B, "
"must be a multiple of %d (target width)!\n",
programCount);
exit(-1);
}
float *matrixA;
matrixA = (float *)malloc(M * N * sizeof(float));
init_matrix_rand(matrixA, M, N, 10.0f);
float *matrixB;
matrixB = (float *)malloc(N * K * sizeof(float));
init_matrix_rand(matrixB, N, K, 10.0f);
float *matrixC;
matrixC = (float *)malloc(M * K * sizeof(float));
init_matrix(matrixC, M, K, 0.0f);
float *matrixValid;
matrixValid = (float *)malloc(M * K * sizeof(float));
bool tasks = false;
// Generate a validation matrix using CPU code:
SGEMM_CPU_validation(matrixA, matrixB, matrixValid, M, N, K);
// Single threaded test cases:
Test_SGEMM((SGEMMFuncPtr)SGEMM_naive, (char *)"SGEMM_naive", matrixA, matrixB, matrixC, M, N, K, tasks, ITERATIONS,
matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileShuffle, (char *)"SGEMM_tileShuffle", matrixA, matrixB, matrixC, M, N, K, tasks,
ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileReduceAdd, (char *)"SGEMM_tileReduceAdd", matrixA, matrixB, matrixC, M, N, K,
tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileAtomicAdd, (char *)"SGEMM_tileAtomicAdd", matrixA, matrixB, matrixC, M, N, K,
tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileNoSIMDIntrin, (char *)"SGEMM_tileNoSIMDIntrin", matrixA, matrixB, matrixC, M, N,
K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileBlockNoSIMDIntrin, (char *)"SGEMM_tileBlockNoSIMDIntrin", matrixA, matrixB,
matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileBlockNoSIMDIntrin_2, (char *)"SGEMM_tileBlockNoSIMDIntrin_2", matrixA, matrixB,
matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
printf("\n");
// Multi-threaded test cases:
tasks = true;
Test_SGEMM((SGEMMFuncPtr)SGEMM_naive_withTasks, (char *)"SGEMM_naive_withTasks", matrixA, matrixB, matrixC, M, N, K,
tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileShuffle_withTasks, (char *)"SGEMM_tileShuffle_withTasks", matrixA, matrixB,
matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileReduceAdd_withTasks, (char *)"SGEMM_tileReduceAdd_withTasks", matrixA, matrixB,
matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileAtomicAdd_withTasks, (char *)"SGEMM_tileAtomicAdd_withTasks", matrixA, matrixB,
matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileNoSIMDIntrin_withTasks, (char *)"SGEMM_tileNoSIMDIntrin_withTasks", matrixA,
matrixB, matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileBlockNoSIMDIntrin_withTasks, (char *)"SGEMM_tileBlockNoSIMDIntrin_withTasks",
matrixA, matrixB, matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
Test_SGEMM((SGEMMFuncPtr)SGEMM_tileBlockNoSIMDIntrin_2_withTasks, (char *)"SGEMM_tileBlockNoSIMDIntrin_2_withTasks",
matrixA, matrixB, matrixC, M, N, K, tasks, ITERATIONS, matrixValid);
free(matrixA);
free(matrixB);
free(matrixC);
free(matrixValid);
return 0;
}
|
