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
|
/******************************************************************************
* 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 <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <libxsmm.h>
#include <libxsmm_intrinsics_x86.h>
int main(int argc, char* argv[]) {
int N = ( argc > 1 ) ? atoi(argv[1]) : 64;
int C = ( argc > 2 ) ? atoi(argv[2]) : 512;
int K = ( argc > 3 ) ? atoi(argv[3]) : 32;
double sparse_frac = ( argc > 4 ) ? atof(argv[4]) : 0.90;
unsigned int REPS = ( argc > 5 ) ? atoi(argv[5]) : 1;
unsigned int* l_rowptr = NULL;
unsigned int* l_colidx = NULL;
float* l_a_de = (float*)libxsmm_aligned_malloc(sizeof(float) * C * K, 64);
float* l_a_sp_csr = NULL;
float* l_b = (float*)libxsmm_aligned_malloc(sizeof(float) * N * C, 64);
float* l_c_gold = (float*)libxsmm_aligned_malloc(sizeof(float) * N * K, 64);
float* l_c_asm_csr = (float*)libxsmm_aligned_malloc(sizeof(float) * N * K, 64);
float l_max_error = 0.0;
int l_i, l_j, l_k, l_jj;
unsigned int l_n;
LIBXSMM_VLA_DECL(2, float, l_p_a_de, l_a_de, K);
LIBXSMM_VLA_DECL(3, float, l_p_b, l_b, N/16, 16);
LIBXSMM_VLA_DECL(3, float, l_p_c_asm_csr, l_c_asm_csr, N/16, 16);
LIBXSMM_VLA_DECL(3, float, l_p_c_gold, l_c_gold, N/16, 16);
unsigned long long l_start, l_end;
double l_total;
int NB, nb;
int nnz = 0;
if (argc != 6 && argc != 1) {
fprintf( stderr, "arguments failure\n" );
return -1;
}
if ( N % 64 != 0 ) {
fprintf( stderr, "N needs to be disable by 64\n" );
return -1;
}
NB = N / 16;
nb = 16;
/* touch B */
for ( l_i = 0; l_i < C; l_i++) {
for ( l_j = 0; l_j < NB; l_j++) {
for ( l_k = 0; l_k < nb; l_k++ ) {
LIBXSMM_VLA_ACCESS(3, l_p_b, l_i, l_j, l_k, NB, nb) = (float)libxsmm_rng_f64();
}
}
}
/* touch dense A */
for ( l_i = 0; l_i < K; l_i++ ) {
for ( l_j = 0; l_j < C; l_j++ ) {
float tmp = (float)libxsmm_rng_f64();
if ( tmp < sparse_frac ) {
tmp = 0;
} else {
nnz++;
}
LIBXSMM_VLA_ACCESS(2, l_p_a_de, l_i, l_j, C) = tmp;
}
}
printf("we just generated a %i x %i matrix with %i NZ entries\n", K, C, nnz);
/* touch C */
for ( l_i = 0; l_i < K; l_i++) {
for ( l_j = 0; l_j < NB; l_j++) {
for ( l_k = 0; l_k < nb; l_k++ ) {
LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_i, l_j, l_k, NB, nb) = 0.f;
LIBXSMM_VLA_ACCESS(3, l_p_c_asm_csr, l_i, l_j, l_k, NB, nb) = 0.f;
}
}
}
/* create A, csr */
l_rowptr = (unsigned int*) libxsmm_aligned_malloc( (K+1)*sizeof(unsigned int), 64 );
l_colidx = (unsigned int*) libxsmm_aligned_malloc( nnz*sizeof(unsigned int), 64 );
l_a_sp_csr = (float* ) libxsmm_aligned_malloc( nnz*sizeof(float), 64 );
l_k = 0;
l_rowptr[K] = nnz;
for ( l_i = 0; l_i < K; l_i++ ) {
l_rowptr[l_i] = l_k;
for ( l_j = 0; l_j < C; l_j++ ) {
if ( LIBXSMM_VLA_ACCESS(2, l_p_a_de, l_i, l_j, C) != 0.0 ) {
l_colidx[l_k] = l_j;
l_a_sp_csr[l_k] = LIBXSMM_VLA_ACCESS(2, l_p_a_de, l_i, l_j, C);
l_k++;
}
}
}
/* dense routine */
l_start = libxsmm_timer_tick();
#if 1
for ( l_n = 0; l_n < REPS; l_n++) {
# pragma omp parallel for private(l_j, l_jj, l_i, l_k)
for ( l_j = 0; l_j < K; l_j++) {
for ( l_jj = 0; l_jj < C; l_jj++) {
for ( l_i = 0; l_i < NB; l_i++) {
LIBXSMM_PRAGMA_SIMD
for (l_k = 0; l_k < nb; l_k++) {
LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_j, l_i, l_k, NB, nb)
+= LIBXSMM_VLA_ACCESS(3, l_p_b, l_jj, l_i, l_k, NB, nb)
* l_a_de[(l_j*C)+l_jj];
}
}
}
}
}
#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)N * (double)C * (double)K) * 2.0) / (l_total * 1.0e9));
/* sparse routine */
l_start = libxsmm_timer_tick();
for ( l_n = 0; l_n < REPS; l_n++) {
for ( l_i = 0; l_i < N; l_i+= 64 ) {
#if defined(_OPENMP)
# pragma omp parallel for private(l_j,l_k)
#endif
for ( l_k = 0; l_k < K; l_k++) {
#if defined(__AVX512F__)
__m512 c0 = _mm512_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i ] );
__m512 c1 = _mm512_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+16] );
__m512 c2 = _mm512_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+32] );
__m512 c3 = _mm512_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+48] );
#elif defined(__AVX2__)
__m256 c0 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i ] );
__m256 c1 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+ 8] );
__m256 c2 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+16] );
__m256 c3 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+24] );
__m256 c4 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+32] );
__m256 c5 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+40] );
__m256 c6 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+48] );
__m256 c7 = _mm256_loadu_ps( &l_c_asm_csr[(l_k*N)+l_i+56] );
#endif
for ( l_j = 0; l_j < (int)(l_rowptr[l_k+1] - l_rowptr[l_k]); l_j++) {
#if defined(__AVX512F__)
c0 = _mm512_fmadd_ps( _mm512_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm512_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i ] ), c0 );
c1 = _mm512_fmadd_ps( _mm512_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm512_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+16] ), c1 );
c2 = _mm512_fmadd_ps( _mm512_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm512_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+32] ), c2 );
c3 = _mm512_fmadd_ps( _mm512_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm512_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+48] ), c3 );
#elif defined(__AVX2__)
c0 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i ] ), c0 );
c1 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+ 8] ), c1 );
c2 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+16] ), c2 );
c3 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+24] ), c3 );
c4 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+32] ), c4 );
c5 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+40] ), c5 );
c6 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+48] ), c6 );
c7 = _mm256_fmadd_ps( _mm256_set1_ps( l_a_sp_csr[l_rowptr[l_k] + l_j] ), _mm256_loadu_ps( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+56] ), c7 );
#else
unsigned int l_ii;
LIBXSMM_PRAGMA_SIMD
for ( l_ii = 0; l_ii < 64; l_ii++ ) {
l_c_asm_csr[(l_k*N)+l_i+l_ii] += l_a_sp_csr[l_rowptr[l_k]+l_j] * l_b[(l_colidx[l_rowptr[l_k]+l_j]*N)+l_i+l_ii];
}
#endif
#if 0
_mm_prefetch( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+ 64], _MM_HINT_T1 );
_mm_prefetch( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+ 80], _MM_HINT_T1 );
_mm_prefetch( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+ 96], _MM_HINT_T1 );
_mm_prefetch( &l_b[(l_colidx[l_rowptr[l_k] + l_j]*N) + l_i+112], _MM_HINT_T1 );
#endif
}
#if defined(__AVX512F__)
_mm512_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i] , c0 );
_mm512_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+16], c1 );
_mm512_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+32], c2 );
_mm512_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+48], c3 );
#elif defined(__AVX2__)
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i] , c0 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+ 8], c1 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+16], c2 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+24], c3 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+32], c4 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+40], c5 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+48], c6 );
_mm256_storeu_ps( &l_c_asm_csr[(l_k*N)+l_i+56], c7 );
#endif
}
}
}
l_end = libxsmm_timer_tick();
l_total = libxsmm_timer_duration(l_start, l_end);
printf("%fs for sparse (asm, csr)\n", l_total);
printf("%f GFLOPS for sparse (asm, csr)\n", ((double)((double)REPS * (double)N * (double)C * (double)K) * 2.0) / (l_total * 1.0e9));
/* check for errors */
l_max_error = 0.f;
for ( l_i = 0; l_i < NB; l_i++) {
for ( l_j = 0; l_j < K; l_j++) {
for ( l_k = 0; l_k < nb; l_k++ ) {
if (fabs( LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_j, l_i, l_k, NB, nb)
- LIBXSMM_VLA_ACCESS(3, l_p_c_asm_csr, l_j, l_i, l_k, NB, nb) ) > l_max_error ) {
l_max_error = (float)fabs( LIBXSMM_VLA_ACCESS(3, l_p_c_gold, l_j, l_i, l_k, NB, nb)
-LIBXSMM_VLA_ACCESS(3, l_p_c_asm_csr, l_j, l_i, l_k, NB, nb) );
}
}
}
}
printf("max error (csr): %f\n", l_max_error);
/* free */
libxsmm_free( l_a_de );
libxsmm_free( l_b );
libxsmm_free( l_c_gold );
libxsmm_free( l_c_asm_csr );
libxsmm_free( l_a_sp_csr );
libxsmm_free( l_rowptr );
libxsmm_free( l_colidx );
return 0;
}
|
