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
|
/******************************************************************************
* 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 *
******************************************************************************/
/* Hans Pabst (Intel Corp.)
******************************************************************************/
#if !defined(INCLUDE_LIBXSMM_LAST)
# include <libxsmm.h>
#endif
#include <inttypes.h>
#if defined(_OPENMP)
# include <omp.h>
#endif
#if defined(INCLUDE_LIBXSMM_LAST)
# include <libxsmm.h>
#endif
#if !defined(MAX_NKERNELS)
# define MAX_NKERNELS 800
#endif
#if !defined(CHECK_PARALLEL_INIT)
# define CHECK_PARALLEL_INIT
#endif
#if !defined(CHECK_PARALLEL_JIT)
# define CHECK_PARALLEL_JIT
#endif
#if !defined(CHECK_SEPARATE)
# define CHECK_SEPARATE
#endif
#if !defined(USE_VERBOSE)
# define USE_VERBOSE
#endif
#if !defined(ITYPE)
# define ITYPE float
#endif
#if !defined(OTYPE)
# define OTYPE ITYPE
#endif
#if defined(CHECK_SEPARATE)
int test(libxsmm_blasint /*m*/, libxsmm_blasint /*n*/, libxsmm_blasint /*k*/);
int test(libxsmm_blasint m, libxsmm_blasint n, libxsmm_blasint k)
{
const OTYPE alpha = 1, beta = 0;
LIBXSMM_MMFUNCTION_TYPE2(ITYPE,OTYPE) kernel;
int result = EXIT_FAILURE;
#if defined(_OPENMP) && !defined(CHECK_PARALLEL_JIT)
# pragma omp single
#endif
kernel = LIBXSMM_MMDISPATCH_SYMBOL2(ITYPE,OTYPE)(m, n, k,
NULL/*lda*/, NULL/*ldb*/, NULL/*ldc*/, &alpha, &beta,
NULL/*flags*/, NULL/*prefetch*/);
if (NULL != kernel) {
libxsmm_mmkernel_info info;
libxsmm_xmmfunction xmm;
xmm.LIBXSMM_TPREFIX2(ITYPE,OTYPE,mm) = kernel;
result = libxsmm_get_mmkernel_info(xmm, &info);
if (EXIT_SUCCESS == result) {
const unsigned int um = (unsigned int)m, un = (unsigned int)n, uk = (unsigned int)k;
if ( um != info.m || un != info.n || uk != info.k
|| um != info.lda || uk != info.ldb || um != info.ldc
|| LIBXSMM_GEMM_PRECISION(ITYPE) != info.iprecision
|| LIBXSMM_GEMM_PRECISION(OTYPE) != info.oprecision)
{
#if defined(_DEBUG) || defined(USE_VERBOSE)
fprintf(stderr, "Error: the %" PRIuPTR "x%" PRIuPTR "x%" PRIuPTR "-kernel does not match!\n",
(uintptr_t)m, (uintptr_t)n, (uintptr_t)k);
#endif
result = EXIT_FAILURE;
}
}
#if defined(_DEBUG) || defined(USE_VERBOSE)
else {
fprintf(stderr, "Error: the %" PRIuPTR "x%" PRIuPTR "x%" PRIuPTR "-kernel is corrupted!\n",
(uintptr_t)m, (uintptr_t)n, (uintptr_t)k);
}
#endif
}
#if !defined(LIBXSMM_JIT) || (0 == LIBXSMM_JIT)
else result = EXIT_SUCCESS;
#endif
return result;
}
#endif /*defined(CHECK_SEPARATE)*/
int main(void)
{
union { libxsmm_xmmfunction x; void* p; } f[MAX_NKERNELS];
const OTYPE alpha = LIBXSMM_ALPHA, beta = LIBXSMM_BETA;
const int prefetch = LIBXSMM_PREFETCH_AUTO;
libxsmm_registry_info registry_info;
const int max_shape = LIBXSMM_MAX_M, flags = LIBXSMM_FLAGS;
int result = EXIT_SUCCESS, nkernels = MAX_NKERNELS, ndup = 0, i;
#if defined(CHECK_SEPARATE)
int mnk[3*MAX_NKERNELS] = { 8,8,8, 16,16,8 };
const int shift = 1, nr = 2; /* nr: predefined triplets */
#endif
int r[3*MAX_NKERNELS];
#if defined(_OPENMP)
const int nthreads = omp_get_max_threads();
#else
const int nthreads = 1;
#endif
/* generate set of random number for parallel region */
for (i = 0; i < (3 * nkernels); i += 3) {
r[i+0] = rand(); r[i+1] = rand(); r[i+2] = rand();
}
#if defined(CHECK_SEPARATE)
/* fill-up set of (m,n,k) for distinct test set */
for (i = 3 * nr; i < (3 * nkernels); ++i) {
mnk[i] = (r[i] + shift) % max_shape + 1;
}
#endif
#if defined(CHECK_PARALLEL_INIT)
# if defined(_OPENMP)
# pragma omp parallel for num_threads(nthreads) private(i)
# endif
for (i = 0; i < MAX_NKERNELS; ++i) {
if (0 == (i % 2)) {
libxsmm_init();
}
else {
libxsmm_finalize();
}
}
#endif
libxsmm_init();
result = libxsmm_get_registry_info(®istry_info);
if (EXIT_SUCCESS == result) {
nkernels = (int)LIBXSMM_MIN((size_t)nkernels, registry_info.capacity);
#if defined(CHECK_SEPARATE)
for (i = 0; i < nkernels; i += nthreads) {
#if defined(_OPENMP) && defined(CHECK_PARALLEL_JIT)
# pragma omp parallel num_threads(nthreads)
#endif
{
#if defined(_OPENMP) && defined(CHECK_PARALLEL_JIT)
const int tid = omp_get_thread_num();
#else
const int tid = 0;
#endif
const int j = LIBXSMM_MIN(3 * (i + tid), nkernels - 3);
const int ri = test(mnk[j+0], mnk[j+1], mnk[j+2]);
if (EXIT_SUCCESS != ri) {
#if defined(_OPENMP) && defined(CHECK_PARALLEL_JIT)
# if (201107 <= _OPENMP)
# pragma omp atomic write
# else
# pragma omp critical
# endif
#endif
result = ri;
}
}
}
#endif
}
if (EXIT_SUCCESS == result) {
#if defined(_OPENMP) && defined(CHECK_PARALLEL_JIT)
# pragma omp parallel for num_threads(nthreads) private(i)
#endif
for (i = 0; i < nkernels; ++i) {
const libxsmm_blasint m = r[3*i+0] % max_shape + 1;
const libxsmm_blasint n = r[3*i+1] % max_shape + 1;
const libxsmm_blasint k = r[3*i+2] % max_shape + 1;
f[i].x.LIBXSMM_TPREFIX2(ITYPE,OTYPE,mm) = LIBXSMM_MMDISPATCH_SYMBOL2(ITYPE,OTYPE)(
m, n, k, &m/*lda*/, &k/*ldb*/, &m/*ldc*/, &alpha, &beta, &flags, &prefetch);
}
}
#if defined(_OPENMP) && !defined(CHECK_PARALLEL_JIT)
# pragma omp parallel for num_threads(nthreads) private(i)
#endif
for (i = 0; i < nkernels; ++i) {
if (EXIT_SUCCESS == result) {
const libxsmm_blasint m = r[3*i+0] % max_shape + 1;
const libxsmm_blasint n = r[3*i+1] % max_shape + 1;
const libxsmm_blasint k = r[3*i+2] % max_shape + 1;
union { libxsmm_xmmfunction x; void* p; } fi;
libxsmm_descriptor_blob blob;
const libxsmm_gemm_descriptor *const desc = libxsmm_gemm_descriptor_init(&blob, LIBXSMM_GEMM_PRECISION2(ITYPE,OTYPE),
m, n, k, m/*lda*/, k/*ldb*/, m/*ldc*/, &alpha, &beta, flags, prefetch);
fi.x = libxsmm_xmmdispatch(desc);
if (NULL != fi.p && NULL != f[i].p) {
if (fi.p != f[i].p) {
libxsmm_kernel_info a_info, b_info;
const int ra = libxsmm_get_kernel_info(f[i].p, &a_info);
const int rb = libxsmm_get_kernel_info(fi.p, &b_info);
/* perform deeper check based on another code generation (used as reference) */
if (EXIT_SUCCESS == ra && EXIT_SUCCESS == rb && (a_info.code_size != b_info.code_size ||
0 != memcmp(f[i].p, fi.p, a_info.code_size)))
{
#if defined(_DEBUG) || defined(USE_VERBOSE)
fprintf(stderr, "Error: the %" PRIuPTR "x%" PRIuPTR "x%" PRIuPTR "-kernel does not match!\n",
(uintptr_t)m, (uintptr_t)n, (uintptr_t)k);
#endif
#if defined(_OPENMP) && !defined(CHECK_PARALLEL_JIT)
# if (201107 <= _OPENMP)
# pragma omp atomic write
# else
# pragma omp critical
# endif
#endif
result = EXIT_FAILURE;
}
#if defined(_OPENMP) && !defined(CHECK_PARALLEL_JIT)
# if (201107 <= _OPENMP)
# pragma omp atomic write
# else
# pragma omp critical
# endif
#endif
++ndup;
}
}
#if (0 != LIBXSMM_JIT)
else {
# if defined(_DEBUG) || defined(USE_VERBOSE)
fprintf(stderr, "Error: no code generated for %" PRIuPTR "x%" PRIuPTR "x%" PRIuPTR "-kernel!\n",
(uintptr_t)m, (uintptr_t)n, (uintptr_t)k);
# endif
# if defined(_OPENMP) && !defined(CHECK_PARALLEL_JIT)
# if (201107 <= _OPENMP)
# pragma omp atomic write
# else
# pragma omp critical
# endif
# endif
result = EXIT_FAILURE;
}
#endif
}
}
#if defined(_DEBUG) || defined(USE_VERBOSE)
if (0 != ndup) fprintf(stderr, "Info: %i kernel%s duplicated.\n", ndup, 1 != ndup ? "s" : "");
#endif
/* test unregistering and freeing kernels */
if (EXIT_SUCCESS == result) {
for (i = 0; i < nkernels; ++i) {
int j = i + 1;
/* avoid to double-release kernels */
for (; j < nkernels; ++j) {
if (f[i].p == f[j].p) f[j].p = NULL;
}
libxsmm_release_kernel(f[i].p);
}
}
libxsmm_finalize();
return result;
}
|
