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
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
|
/******************************************************************************
* 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 *
******************************************************************************/
/* Alexander Heinecke (Intel Corp.)
******************************************************************************/
#include <libxsmm.h>
#include "common_edge_proxy.h"
LIBXSMM_INLINE
void qfma_fill_in( REALTYPE* rm_dense_data, unsigned int m, unsigned int n, unsigned int **colptr, unsigned int **rowidx, REALTYPE **values) {
REALTYPE* cm_dense = NULL;
REALTYPE* cm_dense_data = NULL;
unsigned int i = 0;
unsigned int j = 0;
unsigned int l_max_reg_block = 28;
unsigned int l_max_cols = 0;
unsigned int l_n_chunks = 0;
unsigned int l_n_chunksize = 0;
unsigned int l_n_limit = 0;
unsigned int l_n_processed = 0;
unsigned int l_nnz = 0;
unsigned int* l_colptr = NULL;
unsigned int* l_rowidx = NULL;
REALTYPE* l_values = NULL;
unsigned int l_count = 0;
unsigned int l_found_qmadd = 0;
cm_dense = (REALTYPE*)malloc( m*n*sizeof(REALTYPE) );
cm_dense_data = (REALTYPE*)malloc( m*n*sizeof(REALTYPE) );
/* set all values in copy to 1 or 0 */
for ( j = 0; j < n; ++j ) {
for ( i = 0; i < m; ++i ) {
cm_dense[(j*m)+i] = (REALTYPE)(LIBXSMM_FEQ(rm_dense_data[(i*n)+j], 0) ? 0 : 1);
cm_dense_data[(j*m)+i] = rm_dense_data[(i*n)+j];
}
}
#if 1
/* finding max. active columns */
l_max_cols = 0;
for ( j = 0; j < n; ++j ) {
for ( i = 0; i < m; ++i ) {
if (cm_dense[(j*m) + i] > 0.0) {
l_max_cols = j+1;
}
}
}
/* calculate n blocking as in the generator */
l_n_chunks = ( (l_max_cols % l_max_reg_block) == 0 ) ? (l_max_cols / l_max_reg_block) : (l_max_cols / l_max_reg_block) + 1;
assert(0 != l_n_chunks); /* mute static analysis (division-by-zero); such invalid input must be caught upfront */
l_n_chunksize = ( (l_max_cols % l_n_chunks) == 0 ) ? (l_max_cols / l_n_chunks) : (l_max_cols / l_n_chunks) + 1;
/* qmadd padding */
l_n_processed = 0;
l_n_limit = l_n_chunksize;
while ( l_n_processed < l_max_cols ) {
/* first pass look for qmadds and potential qmadds in the same rows */
for ( i = 0; i < m; ++i ) {
if ( i >= m-3 ) continue;
l_found_qmadd = 0;
for ( j = l_n_processed; j < l_n_limit - l_n_processed; ++j ) {
if ( LIBXSMM_FEQ(cm_dense[(j*m)+(i+0)], 1) &&
LIBXSMM_FEQ(cm_dense[(j*m)+(i+1)], 1) &&
LIBXSMM_FEQ(cm_dense[(j*m)+(i+2)], 1) &&
LIBXSMM_FEQ(cm_dense[(j*m)+(i+3)], 1) ) {
cm_dense[(j*m)+(i+0)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+1)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+2)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+3)] = (REALTYPE)10.0;
l_found_qmadd = 1;
}
}
/* if we found qmadd in at least one column, let's check the other columns in the current block for 3 nnz */
/* -> let's pad them to 4 nnz */
if (l_found_qmadd == 1) {
for ( j = l_n_processed; j < l_n_limit - l_n_processed; ++j ) {
if ( LIBXSMM_FEQ( cm_dense[(j*m)+(i+0)] +
cm_dense[(j*m)+(i+1)] +
cm_dense[(j*m)+(i+2)] +
cm_dense[(j*m)+(i+3)], 3) ) {
cm_dense[(j*m)+(i+0)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+1)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+2)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+3)] = (REALTYPE)10.0;
}
}
i += 3;
}
}
/* second pass look out for consecutive 4 rows which have 3 nnz in a specifc column */
for ( i = 0; i < m; ++i ) {
if ( i >= m-3 ) continue;
l_found_qmadd = 0;
/* first check if already a qmadd in that row */
for ( j = l_n_processed; j < l_n_limit - l_n_processed; ++j ) {
if ( LIBXSMM_FEQ(cm_dense[(j*m)+(i+0)], 10) ) {
l_found_qmadd = 1;
}
}
/* we are in a potential candidate row for padding 0 for qmadd */
if ( l_found_qmadd == 0 ) {
for ( j = l_n_processed; j < l_n_limit - l_n_processed; ++j ) {
if ( LIBXSMM_FEQ( cm_dense[(j*m)+(i+0)] +
cm_dense[(j*m)+(i+1)] +
cm_dense[(j*m)+(i+2)] +
cm_dense[(j*m)+(i+3)], 3) ) {
cm_dense[(j*m)+(i+0)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+1)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+2)] = (REALTYPE)10.0;
cm_dense[(j*m)+(i+3)] = (REALTYPE)10.0;
l_found_qmadd = 1;
}
}
}
if ( l_found_qmadd > 0 ) {
i += 3;
}
}
/* adjust n progression */
l_n_processed += l_n_chunksize;
l_n_limit = LIBXSMM_MIN(l_n_processed + l_n_chunksize, l_max_cols);
}
#endif
/* creating a new CSC matrix */
/* determining new number of NNZ */
l_nnz = 0;
for ( j = 0; j < n; ++j ) {
for ( i = 0; i < m; ++i ) {
if (cm_dense[(j*m) + i] > 0.0) {
l_nnz++;
}
}
}
(*colptr) = (unsigned int*) malloc( (n+1)*sizeof(unsigned int) );
(*rowidx) = (unsigned int*) malloc( l_nnz*sizeof(unsigned int) );
(*values) = (REALTYPE* ) malloc( l_nnz*sizeof(REALTYPE ) );
l_colptr = (*colptr);
l_rowidx = (*rowidx);
l_values = (*values);
/* generating CSC from dense padded structure */
l_count = 0;
for ( j = 0; j < n; ++j ) {
l_colptr[j] = l_count;
for ( i = 0; i < m; ++i ) {
if (cm_dense[(j*m) + i] > (REALTYPE)0.0) {
l_rowidx[l_count] = i;
l_values[l_count] = (REALTYPE)cm_dense_data[(j*m) + i];
l_count++;
}
}
}
l_colptr[n] = l_nnz;
free ( cm_dense );
free ( cm_dense_data );
}
int main(int argc, char* argv[]) {
int M = ( argc == 7 ) ? atoi(argv[1]) : 9;
int N = ( argc == 7 ) ? atoi(argv[2]) : 10;
int K = ( argc == 7 ) ? atoi(argv[3]) : 20;
unsigned int N_CRUNS = ( argc == 7 ) ? atoi(argv[4]) : 8;
unsigned int REPS = ( argc == 7 ) ? atoi(argv[5]) : 1;
char* l_csc_file = ( argc == 7 ) ? argv[6] : "file.csc";
const libxsmm_gemm_prefetch_type prefetch = LIBXSMM_GEMM_PREFETCH_NONE;
const int flags = LIBXSMM_GEMM_FLAGS('N', 'N');
const REALTYPE alpha = 1, beta = 1;
edge_mat_desc mat_desc = libxsmm_sparse_csc_reader_desc( l_csc_file );
unsigned int l_rowcount = mat_desc.row_count;
unsigned int l_colcount = mat_desc.col_count;
unsigned int l_elements = mat_desc.num_elements;
REALTYPE* l_a = (REALTYPE*)libxsmm_aligned_malloc(K * M * N_CRUNS * sizeof(REALTYPE), 64);
REALTYPE* l_b_de = (REALTYPE*)libxsmm_aligned_malloc(K * N * sizeof(REALTYPE), 64);
REALTYPE* l_b_sp = NULL;
unsigned int* l_colptr = NULL;
unsigned int* l_rowidx = NULL;
REALTYPE* l_b_sp_padded = NULL;
unsigned int* l_colptr_padded = NULL;
unsigned int* l_rowidx_padded = NULL;
REALTYPE* l_c_gold = (REALTYPE*)libxsmm_aligned_malloc(M * N * N_CRUNS * sizeof(REALTYPE), 64);
REALTYPE* l_c_asm = (REALTYPE*)libxsmm_aligned_malloc(M * N * N_CRUNS * sizeof(REALTYPE), 64);
REALTYPE l_max_error = 0.0;
unsigned int l_k, l_n;
int l_i, l_j, l_jj;
LIBXSMM_VLA_DECL(3, REALTYPE, l_p_a, l_a, K, N_CRUNS);
LIBXSMM_VLA_DECL(3, REALTYPE, l_p_c_asm, l_c_asm, N, N_CRUNS);
LIBXSMM_VLA_DECL(3, REALTYPE, l_p_c_gold, l_c_gold, N, N_CRUNS);
libxsmm_descriptor_blob l_xgemm_blob;
const libxsmm_gemm_descriptor* l_xgemm_desc = 0;
LIBXSMM_MMFUNCTION_TYPE(REALTYPE) mykernel = NULL;
unsigned long long l_start, l_end;
double l_total;
unsigned long long l_libxsmmflops;
libxsmm_kernel_info l_kinfo;
if (argc != 7) {
fprintf( stderr, "arguments: M CRUNS #iters csc-file!\n" );
exit(-1);
}
if ((unsigned int)K != l_rowcount) {
fprintf( stderr, "arguments K needs to match number of rows of the sparse matrix!\n" );
exit(-1);
}
if ((unsigned int)N != l_colcount) {
fprintf( stderr, "arguments N needs to match number of columns of the sparse matrix!\n" );
exit(-1);
}
if (M != 9) {
fprintf( stderr, "arguments M needs to match 9!\n" );
exit(-1);
}
/* touch A */
for ( l_i = 0; l_i < M; l_i++) {
for ( l_j = 0; l_j < K; l_j++) {
for ( l_k = 0; l_k < N_CRUNS; l_k++ ) {
LIBXSMM_VLA_ACCESS(3, l_p_a, l_i, l_j, l_k, K, N_CRUNS) = (REALTYPE)libxsmm_rng_f64();
}
}
}
/* touch C */
for ( l_i = 0; l_i < M; l_i++) {
for ( l_j = 0; l_j < N; l_j++) {
for ( l_k = 0; l_k < N_CRUNS; l_k++ ) {
LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_i, l_j, l_k, N, N_CRUNS) = (REALTYPE)0.0;
LIBXSMM_VLA_ACCESS(3, l_p_c_asm, l_i, l_j, l_k, N, N_CRUNS) = (REALTYPE)0.0;
}
}
}
/* read B, csc */
libxsmm_sparse_csc_reader( l_csc_file,
&l_colptr,
&l_rowidx,
&l_b_sp,
&l_rowcount, &l_colcount, &l_elements );
/* copy b to dense */
printf("csc matrix data structure we just read:\n");
printf("rows: %u, columns: %u, elements: %u\n", l_rowcount, l_colcount, l_elements);
for ( l_n = 0; l_n < (((unsigned int)K) * N); l_n++) {
l_b_de[l_n] = 0.0;
}
for ( l_n = 0; l_n < (unsigned int)N; l_n++) {
const unsigned int l_colelems = l_colptr[l_n+1] - l_colptr[l_n];
assert(l_colptr[l_n+1] >= l_colptr[l_n]);
for ( l_k = 0; l_k < l_colelems; l_k++) {
l_b_de[(l_rowidx[l_colptr[l_n] + l_k] * N) + l_n] = l_b_sp[l_colptr[l_n] + l_k];
}
}
/* pad B to a better qmadd matrix */
if ( libxsmm_get_target_archid() == LIBXSMM_X86_AVX512_KNM ) {
qfma_fill_in( l_b_de, K, N, &l_colptr_padded, &l_rowidx_padded, &l_b_sp_padded );
printf("qfma padded CSC matrix data structure we just read:\n");
printf("rows: %u, columns: %u, elements: %u\n", l_rowcount, l_colcount, l_colptr_padded[N]);
}
/* dense routine */
l_start = libxsmm_timer_tick();
#if 1
for ( l_n = 0; l_n < REPS; l_n++) {
for ( l_i = 0; l_i < M; l_i++) {
for ( l_j = 0; l_j < N; l_j++) {
for ( l_jj = 0; l_jj < K; l_jj++) {
LIBXSMM_PRAGMA_SIMD
for (l_k = 0; l_k < N_CRUNS; l_k++) {
LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_i, l_j, l_k, N, N_CRUNS)
+= LIBXSMM_VLA_ACCESS(3, l_p_a, l_i, l_jj, l_k, K, N_CRUNS)
* l_b_de[(l_jj*N)+l_j];
}
}
}
}
}
#endif
l_end = libxsmm_timer_tick();
l_total = libxsmm_timer_duration(l_start, l_end);
printf("%fs for dense\n", l_total);
printf("%f GFLOPS for dense\n", ((double)((double)REPS * (double)M * (double)N * (double)K * (double)N_CRUNS) * 2.0) / (l_total * 1.0e9));
l_xgemm_desc = libxsmm_gemm_descriptor_dinit(&l_xgemm_blob, LIBXSMM_GEMM_PRECISION(REALTYPE),
M, N, K, K, 0, N, alpha, beta, flags, prefetch);
/* sparse routine */
#if defined(__EDGE_EXECUTE_F32__)
if ( libxsmm_get_target_archid() == LIBXSMM_X86_AVX512_KNM ) {
mykernel = libxsmm_create_xcsc_soa(l_xgemm_desc, l_colptr_padded, l_rowidx_padded, (const void*)l_b_sp, N_CRUNS).smm;
} else {
mykernel = libxsmm_create_xcsc_soa(l_xgemm_desc, l_colptr, l_rowidx, (const void*)l_b_sp, N_CRUNS).smm;
}
#else
mykernel = libxsmm_create_xcsc_soa(l_xgemm_desc, l_colptr, l_rowidx, (const void*)l_b_sp, N_CRUNS).dmm;
#endif
if ( libxsmm_get_target_archid() == LIBXSMM_X86_AVX512_KNM ) {
l_start = libxsmm_timer_tick();
for ( l_n = 0; l_n < REPS; l_n++) {
#if defined(__EDGE_EXECUTE_F32__)
mykernel( l_a, l_b_sp_padded, l_c_asm );
#else
mykernel( l_a, l_b_sp, l_c_asm );
#endif
}
l_end = libxsmm_timer_tick();
} else {
l_start = libxsmm_timer_tick();
for ( l_n = 0; l_n < REPS; l_n++) {
mykernel( l_a, l_b_sp, l_c_asm );
}
l_end = libxsmm_timer_tick();
}
l_total = libxsmm_timer_duration(l_start, l_end);
libxsmm_get_kernel_info( LIBXSMM_CONST_VOID_PTR(mykernel), &l_kinfo);
l_libxsmmflops = l_kinfo.nflops;
printf("%fs for sparse (asm)\n", l_total);
printf("%f GFLOPS for sparse (asm), calculated\n", ((double)((double)REPS * (double)M * (double)l_elements * (double)N_CRUNS) * 2.0) / (l_total * 1.0e9));
printf("%f GFLOPS for sparse (asm), libxsmm \n", ((double)((double)REPS * (double)l_libxsmmflops)) / (l_total * 1.0e9));
/* check for errors */
l_max_error = (REALTYPE)0.0;
for ( l_i = 0; l_i < M; l_i++) {
for ( l_j = 0; l_j < N; l_j++) {
for ( l_k = 0; l_k < N_CRUNS; l_k++ ) {
if (fabs( LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_i, l_j, l_k, N, N_CRUNS)
- LIBXSMM_VLA_ACCESS(3, l_p_c_asm, l_i, l_j, l_k, N, N_CRUNS) ) > l_max_error ) {
l_max_error = (REALTYPE)fabs( LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_i, l_j, l_k, N, N_CRUNS)
-LIBXSMM_VLA_ACCESS(3, l_p_c_asm, l_i, l_j, l_k, N, N_CRUNS) );
}
}
}
}
printf("max error: %f\n", l_max_error);
printf("PERFDUMP,%s,%u,%i,%i,%i,%u,%u,%f,%f,%f\n", l_csc_file, REPS, M, N, K, l_elements, M * l_elements * N_CRUNS * 2, l_max_error, l_total, ((double)((double)REPS * (double)M * (double)l_elements * (double)N_CRUNS) * 2.0) / (l_total * 1.0e9) );
/* free */
libxsmm_free( l_b_de );
libxsmm_free( l_a );
libxsmm_free( l_c_gold );
libxsmm_free( l_c_asm );
free( l_b_sp );
free( l_colptr );
free( l_rowidx );
if ( l_b_sp_padded != NULL ) free( l_b_sp_padded );
if ( l_colptr_padded != NULL ) free( l_colptr_padded );
if ( l_rowidx_padded != NULL ) free( l_rowidx_padded );
return 0;
}
|
