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
|
/******************************************************************************
* Copyright (c) Intel Corporation - All rights reserved. *
* This file is part of the LIBXSMM library. *
* *
* For information on the license, see the LICENSE file. *
* Further information: https://github.com/hfp/libxsmm/ *
* SPDX-License-Identifier: BSD-3-Clause *
******************************************************************************/
/* Xing Liu (Intel Corp.)
******************************************************************************/
#include <libxsmm.h>
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
void BlockSpMatStep1(int K, int C, int KB, int CB, unsigned int *colptr,
unsigned int *rowidx, unsigned int *b_colptr[],
int *nnzb) {
int num_blocks = K / KB * C / CB;
int blk_idx, i, k;
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
nnzb[blk_idx] = 0;
for (i = 0; i <= KB; ++i) {
b_colptr[blk_idx][i] = 0;
}
}
for (k = 0; k < K; ++k) {
int k_blk_idx = k / KB;
int k_blk_offset = k % KB;
unsigned colstart = colptr[k];
unsigned colend = colptr[k + 1];
for (i = colstart; i < (int)colend; ++i) {
int c = rowidx[i];
int c_blk_idx = c / CB;
blk_idx = k_blk_idx * C / CB + c_blk_idx;
nnzb[blk_idx]++;
b_colptr[blk_idx][k_blk_offset + 1]++;
}
}
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
for (i = 0; i < KB; ++i) {
b_colptr[blk_idx][i + 1] += b_colptr[blk_idx][i];
}
}
}
void BlockSpMatStep2(int K, int C, int KB, int CB, unsigned int *colptr,
unsigned int *rowidx, float *values,
unsigned int *b_colptr[], unsigned int *b_rowidx[],
float *b_values[]) {
int num_blocks = K / KB * C / CB;
int blk_idx, k, i;
for (k = 0; k < K; ++k) {
int k_blk_idx = k / KB;
int k_blk_offset = k % KB;
unsigned colstart = colptr[k];
unsigned colend = colptr[k + 1];
for (i = colstart; i < (int)colend; ++i) {
int c = rowidx[i];
int c_blk_idx = c / CB;
int c_blk_offset = c % CB;
blk_idx = k_blk_idx * C / CB + c_blk_idx;
b_rowidx[blk_idx][b_colptr[blk_idx][k_blk_offset]] = c_blk_offset;
b_values[blk_idx][b_colptr[blk_idx][k_blk_offset]] = values[i];
b_colptr[blk_idx][k_blk_offset]++;
}
}
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
for (i = KB; i > 0; --i) {
b_colptr[blk_idx][i] = b_colptr[blk_idx][i - 1];
}
b_colptr[blk_idx][0] = 0;
}
}
int main(int argc, char **argv) {
int N = (argc > 1) ? atoi(argv[1]) : 2048;
int C = (argc > 2) ? atoi(argv[2]) : 512;
int K = (argc > 3) ? atoi(argv[3]) : 512;
int NB = (argc > 4) ? atoi(argv[4]) : 32;
int CB = (argc > 5) ? atoi(argv[5]) : 32;
int KB = (argc > 6) ? atoi(argv[6]) : 32;
int nb = (argc > 7) ? atoi(argv[7]) : 16;
double sparse_frac = (argc > 8) ? atof(argv[8]) : 0.90;
unsigned int REPS = (argc > 9) ? atoi(argv[9]) : 10;
if (N < NB ||
K < KB ||
C < CB ||
NB < nb ||
C % CB != 0 ||
N % NB != 0 ||
nb % 16 != 0 ||
NB % nb != 0 ||
sparse_frac <= 0.0 ||
sparse_frac >= 1.0 ||
REPS <= 0) {
return -1;
}
int l_n, l_c, l_nn, l_cc, l_nnn, l_k, l_kk, blk_idx;
int i, k, n, c;
libxsmm_gemm_prefetch_type prefetch = LIBXSMM_GEMM_PREFETCH_NONE;
int flags = LIBXSMM_GEMM_FLAGS('N', 'N');
float *l_A = (float *)libxsmm_aligned_malloc(sizeof(float) * N * C, 64);
float *l_B = (float *)libxsmm_aligned_malloc(sizeof(float) * C * K, 64);
float *l_C = (float *)libxsmm_aligned_malloc(sizeof(float) * N * K, 64);
float *l_C_gold =
(float *)libxsmm_aligned_malloc(sizeof(float) * N * K, 64);
LIBXSMM_VLA_DECL(5, float, l_p_A, l_A, C / CB, NB / nb, CB, nb);
LIBXSMM_VLA_DECL(5, float, l_p_C, l_C, K / KB, NB / nb, KB, nb);
LIBXSMM_VLA_DECL(5, float, l_p_C_gold, l_C_gold, K / KB, NB / nb, KB, 16);
/* touch A */
for (l_n = 0; l_n < N / NB; ++l_n) {
for (l_c = 0; l_c < C / CB; ++l_c) {
for (l_nn = 0; l_nn < NB / nb; ++l_nn) {
for (l_cc = 0; l_cc < CB; ++l_cc) {
for (l_nnn = 0; l_nnn < nb; ++l_nnn) {
LIBXSMM_VLA_ACCESS(5, l_p_A, l_n, l_c, l_nn, l_cc,
l_nnn, C / CB, NB / nb, CB, nb) =
(float)libxsmm_rng_f64();
}
}
}
}
}
/* touch dense B and init sparse B*/
int nnz = 0;
unsigned int *colptr = (unsigned int *)libxsmm_aligned_malloc(
(K + 1) * sizeof(unsigned int), 64);
colptr[0] = 0;
for (l_k = 0; l_k < K; l_k++) {
colptr[l_k + 1] = 0;
for (l_c = 0; l_c < C; l_c++) {
double tmp = libxsmm_rng_f64();
if (tmp < sparse_frac) {
tmp = 0.0;
} else {
nnz++;
colptr[l_k + 1]++;
}
l_B[l_k * C + l_c] = (float)tmp;
}
}
for (l_k = 0; l_k < K; l_k++) {
colptr[l_k + 1] += colptr[l_k];
}
unsigned int *rowidx =
(unsigned int *)libxsmm_aligned_malloc(nnz * sizeof(unsigned int), 64);
float *values = (float *)libxsmm_aligned_malloc(nnz * sizeof(float), 64);
for (l_k = 0; l_k < K; l_k++) {
int offset = colptr[l_k];
for (l_c = 0; l_c < C; l_c++) {
if (l_B[l_k * C + l_c] != 0) {
rowidx[offset] = l_c;
values[offset] = l_B[l_k * C + l_c];
offset++;
}
}
}
unsigned num_k_blocks = K / KB;
unsigned num_c_blocks = C / CB;
int num_blocks = num_k_blocks * num_c_blocks;
unsigned int **b_colptr = (unsigned int **)libxsmm_aligned_malloc(
num_blocks * sizeof(unsigned int *), 64);
unsigned int **b_rowidx = (unsigned int **)libxsmm_aligned_malloc(
num_blocks * sizeof(unsigned int *), 64);
float **b_values =
(float **)libxsmm_aligned_malloc(num_blocks * sizeof(float *), 64);
int *nnzb = (int *)libxsmm_aligned_malloc(num_blocks * sizeof(int), 64);
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
b_colptr[blk_idx] = (unsigned int *)libxsmm_aligned_malloc(
(KB + 1) * sizeof(unsigned int), 64);
}
BlockSpMatStep1(K, C, KB, CB, colptr, rowidx, b_colptr, nnzb);
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
b_rowidx[blk_idx] = (unsigned int *)libxsmm_aligned_malloc(
nnzb[blk_idx] * sizeof(unsigned int), 64);
b_values[blk_idx] =
(float *)libxsmm_aligned_malloc(nnzb[blk_idx] * sizeof(float), 64);
}
BlockSpMatStep2(K, C, KB, CB, colptr, rowidx, values, b_colptr, b_rowidx,
b_values);
/* touch C */
for (l_n = 0; l_n < N / NB; ++l_n) {
for (l_k = 0; l_k < K / KB; ++l_k) {
for (l_nn = 0; l_nn < NB / nb; ++l_nn) {
for (l_kk = 0; l_kk < KB; ++l_kk) {
for (l_nnn = 0; l_nnn < nb; ++l_nnn) {
LIBXSMM_VLA_ACCESS(5, l_p_C_gold, l_n, l_k, l_nn, l_kk,
l_nnn, K / KB, NB / nb, KB, nb) =
0.0f;
LIBXSMM_VLA_ACCESS(5, l_p_C, l_n, l_k, l_nn, l_kk,
l_nnn, K / KB, NB / nb, KB, nb) =
0.0f;
}
}
}
}
}
/* dense routine */
for (l_n = 0; l_n < N / NB; ++l_n) {
for (l_k = 0; l_k < K / KB; ++l_k) {
for (l_c = 0; l_c < C / CB; ++l_c) {
for (l_nn = 0; l_nn < NB / nb; ++l_nn) {
for (l_kk = 0; l_kk < KB; ++l_kk) {
k = l_k * KB + l_kk;
for (l_cc = 0; l_cc < CB; ++l_cc) {
c = l_c * CB + l_cc;
for (l_nnn = 0; l_nnn < nb; ++l_nnn) {
LIBXSMM_VLA_ACCESS(5, l_p_C_gold, l_n, l_k,
l_nn, l_kk, l_nnn, K / KB,
NB / nb, KB, nb) +=
LIBXSMM_VLA_ACCESS(5, l_p_A, l_n, l_c, l_nn,
l_cc, l_nnn, C / CB,
NB / nb, CB, nb) *
l_B[k * C + c];
}
}
}
}
}
}
}
/* FWD */
float alpha = 1.0;
float beta = 1.0;
libxsmm_descriptor_blob l_xgemm_blob;
libxsmm_gemm_descriptor **l_xgemm_desc =
(libxsmm_gemm_descriptor **)libxsmm_aligned_malloc(
num_blocks * sizeof(libxsmm_gemm_descriptor *), 64);
libxsmm_smmfunction *mykernel =
(libxsmm_smmfunction *)libxsmm_aligned_malloc(
num_blocks * sizeof(libxsmm_smmfunction), 64);
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
l_xgemm_desc[blk_idx] = libxsmm_gemm_descriptor_dinit(
&l_xgemm_blob, LIBXSMM_GEMM_PRECISION(float), NB / nb, KB, CB, CB,
0, KB, alpha, beta, flags, prefetch);
mykernel[blk_idx] =
libxsmm_create_xcsc_soa(l_xgemm_desc[blk_idx], b_colptr[blk_idx],
b_rowidx[blk_idx],
(const void *)b_values[blk_idx], nb).smm;
}
#ifdef _OPENMP
# pragma omp parallel for LIBXSMM_OPENMP_COLLAPSE(2) private(k,n,c)
#endif
for (k = 0; k < K / KB; ++k) {
for (n = 0; n < N / NB; ++n) {
for (c = 0; c < C / CB; ++c) {
mykernel[k * C / CB + c](&(l_A[(n * C / CB + c) * CB * NB]),
b_values[k * C / CB + c],
&(l_C[(n * K / KB + k) * NB * KB]));
}
}
}
/* check error */
float l_max_error = 0.0f;
for (i = 0; i < N * K; ++i) {
if (fabs(l_C[i] - l_C_gold[i]) > l_max_error) {
l_max_error = (float)fabs(l_C[i] - l_C_gold[i]);
}
}
printf("max error = %f\n", l_max_error);
/* check performace */
unsigned long long l_start = libxsmm_timer_tick();
for (i = 0; i < (int)REPS; ++i) {
#ifdef _OPENMP
# pragma omp parallel for LIBXSMM_OPENMP_COLLAPSE(2) private(k,n,c)
#endif
for (k = 0; k < K / KB; ++k) {
for (n = 0; n < N / NB; ++n) {
for (c = 0; c < C / CB; ++c) {
mykernel[k * C / CB + c](
&(l_A[(n * C / CB + c) * CB * NB]),
b_values[k * C / CB + c],
&(l_C[(n * K / KB + k) * NB * KB]));
}
}
}
}
unsigned long long l_end = libxsmm_timer_tick();
double l_total = libxsmm_timer_duration(l_start, l_end);
printf("%fs for sparse (asm)\n", l_total);
printf("%f GFLOPS for sparse (asm)\n",
((double)((double)REPS * (double)N * (double)C * (double)K) * 2.0) /
(l_total * 1.0e9));
/* clean up */
libxsmm_free(l_A);
libxsmm_free(l_B);
libxsmm_free(l_C);
libxsmm_free(l_C_gold);
for (blk_idx = 0; blk_idx < num_blocks; ++blk_idx) {
libxsmm_free(b_values[blk_idx]);
libxsmm_free(b_colptr[blk_idx]);
libxsmm_free(b_rowidx[blk_idx]);
}
libxsmm_free(b_values);
libxsmm_free(b_colptr);
libxsmm_free(b_rowidx);
return 0;
}
|
