1 /*****************************************************************************
2 Copyright (c) 2011-2014, The OpenBLAS Project
5 Redistribution and use in source and binary forms, with or without
6 modification, are permitted provided that the following conditions are
9 1. Redistributions of source code must retain the above copyright
10 notice, this list of conditions and the following disclaimer.
12 2. Redistributions in binary form must reproduce the above copyright
13 notice, this list of conditions and the following disclaimer in
14 the documentation and/or other materials provided with the
16 3. Neither the name of the OpenBLAS project nor the names of
17 its contributors may be used to endorse or promote products
18 derived from this software without specific prior written
21 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
22 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
25 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
27 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
28 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
30 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 **********************************************************************************/
34 /*********************************************************************/
35 /* Copyright 2009, 2010 The University of Texas at Austin. */
36 /* All rights reserved. */
38 /* Redistribution and use in source and binary forms, with or */
39 /* without modification, are permitted provided that the following */
40 /* conditions are met: */
42 /* 1. Redistributions of source code must retain the above */
43 /* copyright notice, this list of conditions and the following */
46 /* 2. Redistributions in binary form must reproduce the above */
47 /* copyright notice, this list of conditions and the following */
48 /* disclaimer in the documentation and/or other materials */
49 /* provided with the distribution. */
51 /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
52 /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
53 /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
54 /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
55 /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
56 /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
57 /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
58 /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
59 /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
60 /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
61 /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
62 /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
63 /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
64 /* POSSIBILITY OF SUCH DAMAGE. */
66 /* The views and conclusions contained in the software and */
67 /* documentation are those of the authors and should not be */
68 /* interpreted as representing official policies, either expressed */
69 /* or implied, of The University of Texas at Austin. */
70 /*********************************************************************/
80 #define GEMM_DEFAULT_OFFSET_A 64
81 #define GEMM_DEFAULT_OFFSET_B 256
82 #define GEMM_DEFAULT_ALIGN 0x01ffffUL
84 #define SGEMM_DEFAULT_UNROLL_N 4
85 #define DGEMM_DEFAULT_UNROLL_N 4
86 #define QGEMM_DEFAULT_UNROLL_N 2
87 #define CGEMM_DEFAULT_UNROLL_N 2
88 #define ZGEMM_DEFAULT_UNROLL_N 2
89 #define XGEMM_DEFAULT_UNROLL_N 1
92 #define SGEMM_DEFAULT_UNROLL_M 4
93 #define DGEMM_DEFAULT_UNROLL_M 2
94 #define QGEMM_DEFAULT_UNROLL_M 2
95 #define CGEMM_DEFAULT_UNROLL_M 2
96 #define ZGEMM_DEFAULT_UNROLL_M 1
97 #define XGEMM_DEFAULT_UNROLL_M 1
99 #define SGEMM_DEFAULT_UNROLL_M 8
100 #define DGEMM_DEFAULT_UNROLL_M 4
101 #define QGEMM_DEFAULT_UNROLL_M 2
102 #define CGEMM_DEFAULT_UNROLL_M 4
103 #define ZGEMM_DEFAULT_UNROLL_M 2
104 #define XGEMM_DEFAULT_UNROLL_M 1
107 #define SGEMM_DEFAULT_P sgemm_p
108 #define DGEMM_DEFAULT_P dgemm_p
109 #define QGEMM_DEFAULT_P qgemm_p
110 #define CGEMM_DEFAULT_P cgemm_p
111 #define ZGEMM_DEFAULT_P zgemm_p
112 #define XGEMM_DEFAULT_P xgemm_p
114 #define SGEMM_DEFAULT_R sgemm_r
115 #define DGEMM_DEFAULT_R dgemm_r
116 #define QGEMM_DEFAULT_R qgemm_r
117 #define CGEMM_DEFAULT_R cgemm_r
118 #define ZGEMM_DEFAULT_R zgemm_r
119 #define XGEMM_DEFAULT_R xgemm_r
123 #define SGEMM_DEFAULT_Q 248
124 #define DGEMM_DEFAULT_Q 248
125 #define QGEMM_DEFAULT_Q 248
126 #define CGEMM_DEFAULT_Q 248
127 #define ZGEMM_DEFAULT_Q 248
128 #define XGEMM_DEFAULT_Q 248
132 #define SGEMM_DEFAULT_Q 240
133 #define DGEMM_DEFAULT_Q 240
134 #define QGEMM_DEFAULT_Q 240
135 #define CGEMM_DEFAULT_Q 240
136 #define ZGEMM_DEFAULT_Q 240
137 #define XGEMM_DEFAULT_Q 240
143 #define HAVE_EXCLUSIVE_CACHE
147 #if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
152 #define GEMM_DEFAULT_OFFSET_A 64
153 #define GEMM_DEFAULT_OFFSET_B 832
154 #define GEMM_DEFAULT_ALIGN 0x0fffUL
156 #define SGEMM_DEFAULT_UNROLL_N 4
157 #define DGEMM_DEFAULT_UNROLL_N 4
158 #define QGEMM_DEFAULT_UNROLL_N 2
159 #define CGEMM_DEFAULT_UNROLL_N 2
160 #define ZGEMM_DEFAULT_UNROLL_N 2
161 #define XGEMM_DEFAULT_UNROLL_N 1
164 #define SGEMM_DEFAULT_UNROLL_M 4
165 #define DGEMM_DEFAULT_UNROLL_M 2
166 #define QGEMM_DEFAULT_UNROLL_M 2
167 #define CGEMM_DEFAULT_UNROLL_M 2
168 #define ZGEMM_DEFAULT_UNROLL_M 1
169 #define XGEMM_DEFAULT_UNROLL_M 1
171 #define SGEMM_DEFAULT_UNROLL_M 8
172 #define DGEMM_DEFAULT_UNROLL_M 4
173 #define QGEMM_DEFAULT_UNROLL_M 2
174 #define CGEMM_DEFAULT_UNROLL_M 4
175 #define ZGEMM_DEFAULT_UNROLL_M 2
176 #define XGEMM_DEFAULT_UNROLL_M 1
180 #define SGEMM_DEFAULT_P 496
181 #define DGEMM_DEFAULT_P 248
182 #define QGEMM_DEFAULT_P 124
183 #define CGEMM_DEFAULT_P 248
184 #define ZGEMM_DEFAULT_P 124
185 #define XGEMM_DEFAULT_P 62
187 #define SGEMM_DEFAULT_Q 248
188 #define DGEMM_DEFAULT_Q 248
189 #define QGEMM_DEFAULT_Q 248
190 #define CGEMM_DEFAULT_Q 248
191 #define ZGEMM_DEFAULT_Q 248
192 #define XGEMM_DEFAULT_Q 248
196 #define SGEMM_DEFAULT_P 448
197 #define DGEMM_DEFAULT_P 224
198 #define QGEMM_DEFAULT_P 112
199 #define CGEMM_DEFAULT_P 224
200 #define ZGEMM_DEFAULT_P 112
201 #define XGEMM_DEFAULT_P 56
203 #define SGEMM_DEFAULT_Q 224
204 #define DGEMM_DEFAULT_Q 224
205 #define QGEMM_DEFAULT_Q 224
206 #define CGEMM_DEFAULT_Q 224
207 #define ZGEMM_DEFAULT_Q 224
208 #define XGEMM_DEFAULT_Q 224
212 #define SGEMM_DEFAULT_R sgemm_r
213 #define QGEMM_DEFAULT_R qgemm_r
214 #define DGEMM_DEFAULT_R dgemm_r
215 #define CGEMM_DEFAULT_R cgemm_r
216 #define ZGEMM_DEFAULT_R zgemm_r
217 #define XGEMM_DEFAULT_R xgemm_r
220 #define HAVE_EXCLUSIVE_CACHE
222 #define GEMM_THREAD gemm_thread_mn
232 #define GEMM_DEFAULT_OFFSET_A 64
233 #define GEMM_DEFAULT_OFFSET_B 832
234 #define GEMM_DEFAULT_ALIGN 0x0fffUL
238 #define QGEMM_DEFAULT_UNROLL_N 2
239 #define CGEMM_DEFAULT_UNROLL_N 2
240 #define ZGEMM_DEFAULT_UNROLL_N 2
241 #define XGEMM_DEFAULT_UNROLL_N 1
244 #define SGEMM_DEFAULT_UNROLL_N 4
245 #define DGEMM_DEFAULT_UNROLL_N 4
246 #define SGEMM_DEFAULT_UNROLL_M 4
247 #define DGEMM_DEFAULT_UNROLL_M 2
248 #define QGEMM_DEFAULT_UNROLL_M 2
249 #define CGEMM_DEFAULT_UNROLL_M 2
250 #define ZGEMM_DEFAULT_UNROLL_M 1
251 #define XGEMM_DEFAULT_UNROLL_M 1
253 #define SGEMM_DEFAULT_UNROLL_N 2
254 #define DGEMM_DEFAULT_UNROLL_N 2
255 #define SGEMM_DEFAULT_UNROLL_M 16
256 #define DGEMM_DEFAULT_UNROLL_M 8
257 #define QGEMM_DEFAULT_UNROLL_M 2
258 #define CGEMM_DEFAULT_UNROLL_M 4
259 #define ZGEMM_DEFAULT_UNROLL_M 2
260 #define XGEMM_DEFAULT_UNROLL_M 1
261 #define CGEMM3M_DEFAULT_UNROLL_N 4
262 #define CGEMM3M_DEFAULT_UNROLL_M 8
263 #define ZGEMM3M_DEFAULT_UNROLL_N 4
264 #define ZGEMM3M_DEFAULT_UNROLL_M 4
266 #define DGEMM_DEFAULT_UNROLL_MN 16
267 #define GEMV_UNROLL 8
271 #if defined(ARCH_X86_64)
272 #define SGEMM_DEFAULT_P 768
273 #define DGEMM_DEFAULT_P 384
275 #define SGEMM_DEFAULT_P 448
276 #define DGEMM_DEFAULT_P 224
279 #define QGEMM_DEFAULT_P 112
280 #define CGEMM_DEFAULT_P 224
281 #define ZGEMM_DEFAULT_P 112
282 #define XGEMM_DEFAULT_P 56
284 #if defined(ARCH_X86_64)
285 #define SGEMM_DEFAULT_Q 168
286 #define DGEMM_DEFAULT_Q 168
288 #define SGEMM_DEFAULT_Q 224
289 #define DGEMM_DEFAULT_Q 224
292 #define QGEMM_DEFAULT_Q 224
293 #define CGEMM_DEFAULT_Q 224
294 #define ZGEMM_DEFAULT_Q 224
295 #define XGEMM_DEFAULT_Q 224
297 #define CGEMM3M_DEFAULT_P 448
298 #define ZGEMM3M_DEFAULT_P 224
299 #define XGEMM3M_DEFAULT_P 112
300 #define CGEMM3M_DEFAULT_Q 224
301 #define ZGEMM3M_DEFAULT_Q 224
302 #define XGEMM3M_DEFAULT_Q 224
303 #define CGEMM3M_DEFAULT_R 12288
304 #define ZGEMM3M_DEFAULT_R 12288
305 #define XGEMM3M_DEFAULT_R 12288
307 #define SGEMM_DEFAULT_R sgemm_r
308 #define QGEMM_DEFAULT_R qgemm_r
309 #define DGEMM_DEFAULT_R dgemm_r
310 #define CGEMM_DEFAULT_R cgemm_r
311 #define ZGEMM_DEFAULT_R zgemm_r
312 #define XGEMM_DEFAULT_R xgemm_r
315 #define HAVE_EXCLUSIVE_CACHE
317 #define GEMM_THREAD gemm_thread_mn
325 #define GEMM_DEFAULT_OFFSET_A 64
326 #define GEMM_DEFAULT_OFFSET_B 832
327 #define GEMM_DEFAULT_ALIGN 0x0fffUL
331 #define QGEMM_DEFAULT_UNROLL_N 2
332 #define CGEMM_DEFAULT_UNROLL_N 2
333 #define ZGEMM_DEFAULT_UNROLL_N 2
334 #define XGEMM_DEFAULT_UNROLL_N 1
337 #define SGEMM_DEFAULT_UNROLL_N 4
338 #define DGEMM_DEFAULT_UNROLL_N 4
339 #define SGEMM_DEFAULT_UNROLL_M 4
340 #define DGEMM_DEFAULT_UNROLL_M 2
341 #define QGEMM_DEFAULT_UNROLL_M 2
342 #define CGEMM_DEFAULT_UNROLL_M 2
343 #define ZGEMM_DEFAULT_UNROLL_M 1
344 #define XGEMM_DEFAULT_UNROLL_M 1
346 #define SGEMM_DEFAULT_UNROLL_N 2
347 #define DGEMM_DEFAULT_UNROLL_N 2
348 #define SGEMM_DEFAULT_UNROLL_M 16
349 #define DGEMM_DEFAULT_UNROLL_M 8
350 #define QGEMM_DEFAULT_UNROLL_M 2
351 #define CGEMM_DEFAULT_UNROLL_M 4
352 #define ZGEMM_DEFAULT_UNROLL_M 2
353 #define XGEMM_DEFAULT_UNROLL_M 1
354 #define CGEMM3M_DEFAULT_UNROLL_N 4
355 #define CGEMM3M_DEFAULT_UNROLL_M 8
356 #define ZGEMM3M_DEFAULT_UNROLL_N 4
357 #define ZGEMM3M_DEFAULT_UNROLL_M 4
358 #define GEMV_UNROLL 8
361 #if defined(ARCH_X86_64)
362 #define SGEMM_DEFAULT_P 768
363 #define DGEMM_DEFAULT_P 768
364 #define ZGEMM_DEFAULT_P 384
365 #define CGEMM_DEFAULT_P 768
367 #define SGEMM_DEFAULT_P 448
368 #define DGEMM_DEFAULT_P 480
369 #define ZGEMM_DEFAULT_P 112
370 #define CGEMM_DEFAULT_P 224
372 #define QGEMM_DEFAULT_P 112
373 #define XGEMM_DEFAULT_P 56
375 #if defined(ARCH_X86_64)
376 #define SGEMM_DEFAULT_Q 192
377 #define DGEMM_DEFAULT_Q 168
378 #define ZGEMM_DEFAULT_Q 168
379 #define CGEMM_DEFAULT_Q 168
381 #define SGEMM_DEFAULT_Q 224
382 #define DGEMM_DEFAULT_Q 224
383 #define ZGEMM_DEFAULT_Q 224
384 #define CGEMM_DEFAULT_Q 224
386 #define QGEMM_DEFAULT_Q 224
387 #define XGEMM_DEFAULT_Q 224
389 #define CGEMM3M_DEFAULT_P 448
390 #define ZGEMM3M_DEFAULT_P 224
391 #define XGEMM3M_DEFAULT_P 112
392 #define CGEMM3M_DEFAULT_Q 224
393 #define ZGEMM3M_DEFAULT_Q 224
394 #define XGEMM3M_DEFAULT_Q 224
395 #define CGEMM3M_DEFAULT_R 12288
396 #define ZGEMM3M_DEFAULT_R 12288
397 #define XGEMM3M_DEFAULT_R 12288
399 #define SGEMM_DEFAULT_R 12288
400 #define QGEMM_DEFAULT_R qgemm_r
401 #define DGEMM_DEFAULT_R 12288
402 #define CGEMM_DEFAULT_R cgemm_r
403 #define ZGEMM_DEFAULT_R zgemm_r
404 #define XGEMM_DEFAULT_R xgemm_r
407 #define HAVE_EXCLUSIVE_CACHE
409 #define GEMM_THREAD gemm_thread_mn
417 #define GEMM_DEFAULT_OFFSET_A 64
418 #define GEMM_DEFAULT_OFFSET_B 832
419 #define GEMM_DEFAULT_ALIGN 0x0fffUL
423 #define QGEMM_DEFAULT_UNROLL_N 2
424 #define CGEMM_DEFAULT_UNROLL_N 2
425 #define ZGEMM_DEFAULT_UNROLL_N 2
426 #define XGEMM_DEFAULT_UNROLL_N 1
429 #define SGEMM_DEFAULT_UNROLL_N 4
430 #define DGEMM_DEFAULT_UNROLL_N 4
431 #define SGEMM_DEFAULT_UNROLL_M 4
432 #define DGEMM_DEFAULT_UNROLL_M 2
433 #define QGEMM_DEFAULT_UNROLL_M 2
434 #define CGEMM_DEFAULT_UNROLL_M 2
435 #define ZGEMM_DEFAULT_UNROLL_M 1
436 #define XGEMM_DEFAULT_UNROLL_M 1
438 #define SGEMM_DEFAULT_UNROLL_N 2
439 #define DGEMM_DEFAULT_UNROLL_N 2
440 #define SGEMM_DEFAULT_UNROLL_M 16
441 #define DGEMM_DEFAULT_UNROLL_M 8
442 #define QGEMM_DEFAULT_UNROLL_M 2
443 #define CGEMM_DEFAULT_UNROLL_M 4
444 #define ZGEMM_DEFAULT_UNROLL_M 2
445 #define XGEMM_DEFAULT_UNROLL_M 1
446 #define CGEMM3M_DEFAULT_UNROLL_N 4
447 #define CGEMM3M_DEFAULT_UNROLL_M 8
448 #define ZGEMM3M_DEFAULT_UNROLL_N 4
449 #define ZGEMM3M_DEFAULT_UNROLL_M 4
450 #define GEMV_UNROLL 8
453 #if defined(ARCH_X86_64)
454 #define SGEMM_DEFAULT_P 768
455 #define DGEMM_DEFAULT_P 576
456 #define ZGEMM_DEFAULT_P 288
457 #define CGEMM_DEFAULT_P 576
459 #define SGEMM_DEFAULT_P 448
460 #define DGEMM_DEFAULT_P 480
461 #define ZGEMM_DEFAULT_P 112
462 #define CGEMM_DEFAULT_P 224
464 #define QGEMM_DEFAULT_P 112
465 #define XGEMM_DEFAULT_P 56
467 #if defined(ARCH_X86_64)
468 #define SGEMM_DEFAULT_Q 192
469 #define DGEMM_DEFAULT_Q 160
470 #define ZGEMM_DEFAULT_Q 160
471 #define CGEMM_DEFAULT_Q 160
473 #define SGEMM_DEFAULT_Q 224
474 #define DGEMM_DEFAULT_Q 224
475 #define ZGEMM_DEFAULT_Q 224
476 #define CGEMM_DEFAULT_Q 224
478 #define QGEMM_DEFAULT_Q 224
479 #define XGEMM_DEFAULT_Q 224
481 #define CGEMM3M_DEFAULT_P 448
482 #define ZGEMM3M_DEFAULT_P 224
483 #define XGEMM3M_DEFAULT_P 112
484 #define CGEMM3M_DEFAULT_Q 224
485 #define ZGEMM3M_DEFAULT_Q 224
486 #define XGEMM3M_DEFAULT_Q 224
487 #define CGEMM3M_DEFAULT_R 12288
488 #define ZGEMM3M_DEFAULT_R 12288
489 #define XGEMM3M_DEFAULT_R 12288
491 #define SGEMM_DEFAULT_R 12288
492 #define QGEMM_DEFAULT_R qgemm_r
493 #define DGEMM_DEFAULT_R 12288
494 #define CGEMM_DEFAULT_R cgemm_r
495 #define ZGEMM_DEFAULT_R zgemm_r
496 #define XGEMM_DEFAULT_R xgemm_r
499 #define HAVE_EXCLUSIVE_CACHE
501 #define GEMM_THREAD gemm_thread_mn
510 #define GEMM_DEFAULT_OFFSET_A 64
511 #define GEMM_DEFAULT_OFFSET_B 832
512 #define GEMM_DEFAULT_ALIGN 0x0fffUL
516 #define QGEMM_DEFAULT_UNROLL_N 2
517 #define CGEMM_DEFAULT_UNROLL_N 2
518 #define ZGEMM_DEFAULT_UNROLL_N 2
519 #define XGEMM_DEFAULT_UNROLL_N 1
522 #define SGEMM_DEFAULT_UNROLL_N 4
523 #define DGEMM_DEFAULT_UNROLL_N 4
524 #define SGEMM_DEFAULT_UNROLL_M 4
525 #define DGEMM_DEFAULT_UNROLL_M 2
526 #define QGEMM_DEFAULT_UNROLL_M 2
527 #define CGEMM_DEFAULT_UNROLL_M 2
528 #define ZGEMM_DEFAULT_UNROLL_M 1
529 #define XGEMM_DEFAULT_UNROLL_M 1
531 #define SGEMM_DEFAULT_UNROLL_N 2
532 #define DGEMM_DEFAULT_UNROLL_N 2
533 #define SGEMM_DEFAULT_UNROLL_M 16
534 #define DGEMM_DEFAULT_UNROLL_M 8
535 #define QGEMM_DEFAULT_UNROLL_M 2
536 #define CGEMM_DEFAULT_UNROLL_M 4
537 #define ZGEMM_DEFAULT_UNROLL_M 2
538 #define XGEMM_DEFAULT_UNROLL_M 1
539 #define CGEMM3M_DEFAULT_UNROLL_N 4
540 #define CGEMM3M_DEFAULT_UNROLL_M 8
541 #define ZGEMM3M_DEFAULT_UNROLL_N 4
542 #define ZGEMM3M_DEFAULT_UNROLL_M 4
543 #define GEMV_UNROLL 8
546 #if defined(ARCH_X86_64)
547 #define SGEMM_DEFAULT_P 768
548 #define DGEMM_DEFAULT_P 576
549 #define ZGEMM_DEFAULT_P 288
550 #define CGEMM_DEFAULT_P 576
552 #define SGEMM_DEFAULT_P 448
553 #define DGEMM_DEFAULT_P 480
554 #define ZGEMM_DEFAULT_P 112
555 #define CGEMM_DEFAULT_P 224
557 #define QGEMM_DEFAULT_P 112
558 #define XGEMM_DEFAULT_P 56
560 #if defined(ARCH_X86_64)
561 #define SGEMM_DEFAULT_Q 192
562 #define DGEMM_DEFAULT_Q 160
563 #define ZGEMM_DEFAULT_Q 160
564 #define CGEMM_DEFAULT_Q 160
566 #define SGEMM_DEFAULT_Q 224
567 #define DGEMM_DEFAULT_Q 224
568 #define ZGEMM_DEFAULT_Q 224
569 #define CGEMM_DEFAULT_Q 224
571 #define QGEMM_DEFAULT_Q 224
572 #define XGEMM_DEFAULT_Q 224
574 #define CGEMM3M_DEFAULT_P 448
575 #define ZGEMM3M_DEFAULT_P 224
576 #define XGEMM3M_DEFAULT_P 112
577 #define CGEMM3M_DEFAULT_Q 224
578 #define ZGEMM3M_DEFAULT_Q 224
579 #define XGEMM3M_DEFAULT_Q 224
580 #define CGEMM3M_DEFAULT_R 12288
581 #define ZGEMM3M_DEFAULT_R 12288
582 #define XGEMM3M_DEFAULT_R 12288
584 #define SGEMM_DEFAULT_R 12288
585 #define QGEMM_DEFAULT_R qgemm_r
586 #define DGEMM_DEFAULT_R 12288
587 #define CGEMM_DEFAULT_R cgemm_r
588 #define ZGEMM_DEFAULT_R zgemm_r
589 #define XGEMM_DEFAULT_R xgemm_r
592 #define HAVE_EXCLUSIVE_CACHE
594 #define GEMM_THREAD gemm_thread_mn
602 #define GEMM_DEFAULT_OFFSET_A 0
603 #define GEMM_DEFAULT_OFFSET_B 0
604 #define GEMM_DEFAULT_ALIGN 0x03fffUL
608 #define SWITCH_RATIO 4
612 #define SGEMM_DEFAULT_UNROLL_M 4
613 #define DGEMM_DEFAULT_UNROLL_M 2
614 #define QGEMM_DEFAULT_UNROLL_M 2
615 #define CGEMM_DEFAULT_UNROLL_M 2
616 #define ZGEMM_DEFAULT_UNROLL_M 1
617 #define XGEMM_DEFAULT_UNROLL_M 1
619 #define SGEMM_DEFAULT_UNROLL_N 4
620 #define DGEMM_DEFAULT_UNROLL_N 4
621 #define QGEMM_DEFAULT_UNROLL_N 2
622 #define CGEMM_DEFAULT_UNROLL_N 2
623 #define ZGEMM_DEFAULT_UNROLL_N 2
624 #define XGEMM_DEFAULT_UNROLL_N 1
628 #define SGEMM_DEFAULT_UNROLL_M 16
629 #define DGEMM_DEFAULT_UNROLL_M 4
630 #define QGEMM_DEFAULT_UNROLL_M 2
631 #define CGEMM_DEFAULT_UNROLL_M 8
632 #define ZGEMM_DEFAULT_UNROLL_M 4
633 #define XGEMM_DEFAULT_UNROLL_M 1
635 #define SGEMM_DEFAULT_UNROLL_N 4
636 #define DGEMM_DEFAULT_UNROLL_N 8
637 #define QGEMM_DEFAULT_UNROLL_N 2
638 #define CGEMM_DEFAULT_UNROLL_N 2
639 #define ZGEMM_DEFAULT_UNROLL_N 2
640 #define XGEMM_DEFAULT_UNROLL_N 1
642 #define SGEMM_DEFAULT_UNROLL_MN 32
643 #define DGEMM_DEFAULT_UNROLL_MN 32
648 #define SGEMM_DEFAULT_P 512
649 #define SGEMM_DEFAULT_R sgemm_r
650 #define DGEMM_DEFAULT_P 512
651 #define DGEMM_DEFAULT_R dgemm_r
652 #define QGEMM_DEFAULT_P 504
653 #define QGEMM_DEFAULT_R qgemm_r
654 #define CGEMM_DEFAULT_P 128
655 #define CGEMM_DEFAULT_R 1024
656 #define ZGEMM_DEFAULT_P 512
657 #define ZGEMM_DEFAULT_R zgemm_r
658 #define XGEMM_DEFAULT_P 252
659 #define XGEMM_DEFAULT_R xgemm_r
660 #define SGEMM_DEFAULT_Q 256
661 #define DGEMM_DEFAULT_Q 256
662 #define QGEMM_DEFAULT_Q 128
663 #define CGEMM_DEFAULT_Q 256
664 #define ZGEMM_DEFAULT_Q 192
665 #define XGEMM_DEFAULT_Q 128
669 #define SGEMM_DEFAULT_P 768
670 #define DGEMM_DEFAULT_P 512
671 #define CGEMM_DEFAULT_P 384
672 #define ZGEMM_DEFAULT_P 256
675 #define SGEMM_DEFAULT_Q 320
676 #define DGEMM_DEFAULT_Q 128
678 #define SGEMM_DEFAULT_Q 384
679 #define DGEMM_DEFAULT_Q 256
681 #define CGEMM_DEFAULT_Q 192
682 #define ZGEMM_DEFAULT_Q 128
684 #define SGEMM_DEFAULT_R sgemm_r
685 #define DGEMM_DEFAULT_R 13824
686 #define CGEMM_DEFAULT_R cgemm_r
687 #define ZGEMM_DEFAULT_R zgemm_r
689 #define QGEMM_DEFAULT_Q 128
690 #define QGEMM_DEFAULT_P 504
691 #define QGEMM_DEFAULT_R qgemm_r
692 #define XGEMM_DEFAULT_P 252
693 #define XGEMM_DEFAULT_R xgemm_r
694 #define XGEMM_DEFAULT_Q 128
696 #define CGEMM3M_DEFAULT_UNROLL_N 8
697 #define CGEMM3M_DEFAULT_UNROLL_M 4
698 #define ZGEMM3M_DEFAULT_UNROLL_N 8
699 #define ZGEMM3M_DEFAULT_UNROLL_M 2
701 #define CGEMM3M_DEFAULT_P 448
702 #define ZGEMM3M_DEFAULT_P 224
703 #define XGEMM3M_DEFAULT_P 112
704 #define CGEMM3M_DEFAULT_Q 224
705 #define ZGEMM3M_DEFAULT_Q 224
706 #define XGEMM3M_DEFAULT_Q 224
707 #define CGEMM3M_DEFAULT_R 12288
708 #define ZGEMM3M_DEFAULT_R 12288
709 #define XGEMM3M_DEFAULT_R 12288
720 #define GEMM_DEFAULT_OFFSET_A 0
721 #define GEMM_DEFAULT_OFFSET_B 384
722 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
724 #define SGEMM_DEFAULT_UNROLL_N 4
725 #define DGEMM_DEFAULT_UNROLL_N 4
726 #define QGEMM_DEFAULT_UNROLL_N 2
727 #define CGEMM_DEFAULT_UNROLL_N 2
728 #define ZGEMM_DEFAULT_UNROLL_N 2
729 #define XGEMM_DEFAULT_UNROLL_N 1
731 #define SGEMM_DEFAULT_UNROLL_M 2
732 #define DGEMM_DEFAULT_UNROLL_M 1
733 #define QGEMM_DEFAULT_UNROLL_M 2
734 #define CGEMM_DEFAULT_UNROLL_M 1
735 #define ZGEMM_DEFAULT_UNROLL_M 1
736 #define XGEMM_DEFAULT_UNROLL_M 1
738 #define SGEMM_DEFAULT_R sgemm_r
739 #define DGEMM_DEFAULT_R dgemm_r
740 #define QGEMM_DEFAULT_R qgemm_r
741 #define CGEMM_DEFAULT_R cgemm_r
742 #define ZGEMM_DEFAULT_R zgemm_r
743 #define XGEMM_DEFAULT_R xgemm_r
745 #define SGEMM_DEFAULT_P 208
746 #define DGEMM_DEFAULT_P 104
747 #define QGEMM_DEFAULT_P 56
748 #define CGEMM_DEFAULT_P 104
749 #define ZGEMM_DEFAULT_P 56
750 #define XGEMM_DEFAULT_P 28
752 #define SGEMM_DEFAULT_Q 208
753 #define DGEMM_DEFAULT_Q 208
754 #define QGEMM_DEFAULT_Q 208
755 #define CGEMM_DEFAULT_Q 208
756 #define ZGEMM_DEFAULT_Q 208
757 #define XGEMM_DEFAULT_Q 208
760 #define HAVE_EXCLUSIVE_CACHE
768 #define GEMM_DEFAULT_OFFSET_A 0
769 #define GEMM_DEFAULT_OFFSET_B 256
770 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
772 #define SGEMM_DEFAULT_UNROLL_N 4
773 #define DGEMM_DEFAULT_UNROLL_N 4
774 #define QGEMM_DEFAULT_UNROLL_N 2
775 #define CGEMM_DEFAULT_UNROLL_N 2
776 #define ZGEMM_DEFAULT_UNROLL_N 2
777 #define XGEMM_DEFAULT_UNROLL_N 1
779 #define SGEMM_DEFAULT_UNROLL_M 2
780 #define DGEMM_DEFAULT_UNROLL_M 1
781 #define QGEMM_DEFAULT_UNROLL_M 2
782 #define CGEMM_DEFAULT_UNROLL_M 1
783 #define ZGEMM_DEFAULT_UNROLL_M 1
784 #define XGEMM_DEFAULT_UNROLL_M 1
786 #define SGEMM_DEFAULT_R sgemm_r
787 #define DGEMM_DEFAULT_R dgemm_r
788 #define QGEMM_DEFAULT_R qgemm_r
789 #define CGEMM_DEFAULT_R cgemm_r
790 #define ZGEMM_DEFAULT_R zgemm_r
791 #define XGEMM_DEFAULT_R xgemm_r
793 #define SGEMM_DEFAULT_P 128
794 #define DGEMM_DEFAULT_P 128
795 #define QGEMM_DEFAULT_P 128
796 #define CGEMM_DEFAULT_P 128
797 #define ZGEMM_DEFAULT_P 128
798 #define XGEMM_DEFAULT_P 128
800 #define SGEMM_DEFAULT_Q 512
801 #define DGEMM_DEFAULT_Q 256
802 #define QGEMM_DEFAULT_Q 256
803 #define CGEMM_DEFAULT_Q 256
804 #define ZGEMM_DEFAULT_Q 128
805 #define XGEMM_DEFAULT_Q 128
815 #define GEMM_DEFAULT_OFFSET_A 64
816 #define GEMM_DEFAULT_OFFSET_B 256
817 #define GEMM_DEFAULT_ALIGN 0x01ffffUL
820 #define SGEMM_DEFAULT_UNROLL_N 4
821 #define DGEMM_DEFAULT_UNROLL_N 4
822 #define QGEMM_DEFAULT_UNROLL_N 2
823 #define CGEMM_DEFAULT_UNROLL_N 2
824 #define ZGEMM_DEFAULT_UNROLL_N 2
825 #define XGEMM_DEFAULT_UNROLL_N 1
827 #define SGEMM_DEFAULT_UNROLL_M 4
828 #define DGEMM_DEFAULT_UNROLL_M 2
829 #define QGEMM_DEFAULT_UNROLL_M 2
830 #define CGEMM_DEFAULT_UNROLL_M 2
831 #define ZGEMM_DEFAULT_UNROLL_M 1
832 #define XGEMM_DEFAULT_UNROLL_M 1
834 #define SGEMM_DEFAULT_UNROLL_N 8
835 #define DGEMM_DEFAULT_UNROLL_N 4
836 #define QGEMM_DEFAULT_UNROLL_N 2
837 #define CGEMM_DEFAULT_UNROLL_N 4
838 #define ZGEMM_DEFAULT_UNROLL_N 2
839 #define XGEMM_DEFAULT_UNROLL_N 1
841 #define SGEMM_DEFAULT_UNROLL_M 4
842 #define DGEMM_DEFAULT_UNROLL_M 4
843 #define QGEMM_DEFAULT_UNROLL_M 2
844 #define CGEMM_DEFAULT_UNROLL_M 2
845 #define ZGEMM_DEFAULT_UNROLL_M 2
846 #define XGEMM_DEFAULT_UNROLL_M 1
849 #define SGEMM_DEFAULT_P 288
850 #define DGEMM_DEFAULT_P 288
851 #define QGEMM_DEFAULT_P 288
852 #define CGEMM_DEFAULT_P 288
853 #define ZGEMM_DEFAULT_P 288
854 #define XGEMM_DEFAULT_P 288
856 #define SGEMM_DEFAULT_R sgemm_r
857 #define DGEMM_DEFAULT_R dgemm_r
858 #define QGEMM_DEFAULT_R qgemm_r
859 #define CGEMM_DEFAULT_R cgemm_r
860 #define ZGEMM_DEFAULT_R zgemm_r
861 #define XGEMM_DEFAULT_R xgemm_r
863 #define SGEMM_DEFAULT_Q 256
864 #define DGEMM_DEFAULT_Q 128
865 #define QGEMM_DEFAULT_Q 64
866 #define CGEMM_DEFAULT_Q 128
867 #define ZGEMM_DEFAULT_Q 64
868 #define XGEMM_DEFAULT_Q 32
871 #define HAVE_EXCLUSIVE_CACHE
875 #if defined(PENTIUM) || defined(PENTIUM2) || defined(PENTIUM3)
884 #define GEMM_DEFAULT_OFFSET_A 0
885 #define GEMM_DEFAULT_OFFSET_B 0
886 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
889 #define SGEMM_DEFAULT_UNROLL_M 8
890 #define CGEMM_DEFAULT_UNROLL_M 4
892 #define SGEMM_DEFAULT_UNROLL_M 4
893 #define CGEMM_DEFAULT_UNROLL_M 2
895 #define DGEMM_DEFAULT_UNROLL_M 2
896 #define SGEMM_DEFAULT_UNROLL_N 2
897 #define DGEMM_DEFAULT_UNROLL_N 2
898 #define QGEMM_DEFAULT_UNROLL_M 2
899 #define QGEMM_DEFAULT_UNROLL_N 2
900 #define CGEMM_DEFAULT_UNROLL_N 1
901 #define ZGEMM_DEFAULT_UNROLL_M 1
902 #define ZGEMM_DEFAULT_UNROLL_N 1
903 #define XGEMM_DEFAULT_UNROLL_M 1
904 #define XGEMM_DEFAULT_UNROLL_N 1
906 #define SGEMM_DEFAULT_P sgemm_p
907 #define SGEMM_DEFAULT_Q 256
908 #define SGEMM_DEFAULT_R sgemm_r
910 #define DGEMM_DEFAULT_P dgemm_p
911 #define DGEMM_DEFAULT_Q 256
912 #define DGEMM_DEFAULT_R dgemm_r
914 #define QGEMM_DEFAULT_P qgemm_p
915 #define QGEMM_DEFAULT_Q 256
916 #define QGEMM_DEFAULT_R qgemm_r
918 #define CGEMM_DEFAULT_P cgemm_p
919 #define CGEMM_DEFAULT_Q 256
920 #define CGEMM_DEFAULT_R cgemm_r
922 #define ZGEMM_DEFAULT_P zgemm_p
923 #define ZGEMM_DEFAULT_Q 256
924 #define ZGEMM_DEFAULT_R zgemm_r
926 #define XGEMM_DEFAULT_P xgemm_p
927 #define XGEMM_DEFAULT_Q 256
928 #define XGEMM_DEFAULT_R xgemm_r
939 #define GEMM_DEFAULT_OFFSET_A 0
940 #define GEMM_DEFAULT_OFFSET_B 0
941 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
944 #define SGEMM_DEFAULT_UNROLL_M 4
945 #define SGEMM_DEFAULT_UNROLL_N 4
946 #define DGEMM_DEFAULT_UNROLL_M 2
947 #define DGEMM_DEFAULT_UNROLL_N 4
948 #define QGEMM_DEFAULT_UNROLL_M 2
949 #define QGEMM_DEFAULT_UNROLL_N 2
950 #define CGEMM_DEFAULT_UNROLL_M 2
951 #define CGEMM_DEFAULT_UNROLL_N 2
952 #define ZGEMM_DEFAULT_UNROLL_M 1
953 #define ZGEMM_DEFAULT_UNROLL_N 2
954 #define XGEMM_DEFAULT_UNROLL_M 1
955 #define XGEMM_DEFAULT_UNROLL_N 1
957 #define SGEMM_DEFAULT_UNROLL_M 8
958 #define SGEMM_DEFAULT_UNROLL_N 2
959 #define DGEMM_DEFAULT_UNROLL_M 2
960 #define DGEMM_DEFAULT_UNROLL_N 2
961 #define QGEMM_DEFAULT_UNROLL_M 2
962 #define QGEMM_DEFAULT_UNROLL_N 2
963 #define CGEMM_DEFAULT_UNROLL_M 4
964 #define CGEMM_DEFAULT_UNROLL_N 1
965 #define ZGEMM_DEFAULT_UNROLL_M 1
966 #define ZGEMM_DEFAULT_UNROLL_N 1
967 #define XGEMM_DEFAULT_UNROLL_M 1
968 #define XGEMM_DEFAULT_UNROLL_N 1
972 #define SGEMM_DEFAULT_P sgemm_p
973 #define SGEMM_DEFAULT_Q 256
974 #define SGEMM_DEFAULT_R sgemm_r
976 #define DGEMM_DEFAULT_P dgemm_p
977 #define DGEMM_DEFAULT_Q 256
978 #define DGEMM_DEFAULT_R dgemm_r
980 #define QGEMM_DEFAULT_P qgemm_p
981 #define QGEMM_DEFAULT_Q 256
982 #define QGEMM_DEFAULT_R qgemm_r
984 #define CGEMM_DEFAULT_P cgemm_p
985 #define CGEMM_DEFAULT_Q 256
986 #define CGEMM_DEFAULT_R cgemm_r
988 #define ZGEMM_DEFAULT_P zgemm_p
989 #define ZGEMM_DEFAULT_Q 256
990 #define ZGEMM_DEFAULT_R zgemm_r
992 #define XGEMM_DEFAULT_P xgemm_p
993 #define XGEMM_DEFAULT_Q 256
994 #define XGEMM_DEFAULT_R xgemm_r
999 #ifdef CORE_NORTHWOOD
1004 #define GEMM_DEFAULT_OFFSET_A 0
1005 #define GEMM_DEFAULT_OFFSET_B 32
1007 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1011 #define SGEMM_DEFAULT_UNROLL_M 8
1012 #define DGEMM_DEFAULT_UNROLL_M 4
1013 #define QGEMM_DEFAULT_UNROLL_M 2
1014 #define CGEMM_DEFAULT_UNROLL_M 4
1015 #define ZGEMM_DEFAULT_UNROLL_M 2
1016 #define XGEMM_DEFAULT_UNROLL_M 1
1018 #define SGEMM_DEFAULT_UNROLL_N 2
1019 #define DGEMM_DEFAULT_UNROLL_N 2
1020 #define QGEMM_DEFAULT_UNROLL_N 2
1021 #define CGEMM_DEFAULT_UNROLL_N 1
1022 #define ZGEMM_DEFAULT_UNROLL_N 1
1023 #define XGEMM_DEFAULT_UNROLL_N 1
1025 #define SGEMM_DEFAULT_P sgemm_p
1026 #define SGEMM_DEFAULT_R sgemm_r
1028 #define DGEMM_DEFAULT_P dgemm_p
1029 #define DGEMM_DEFAULT_R dgemm_r
1031 #define QGEMM_DEFAULT_P qgemm_p
1032 #define QGEMM_DEFAULT_R qgemm_r
1034 #define CGEMM_DEFAULT_P cgemm_p
1035 #define CGEMM_DEFAULT_R cgemm_r
1037 #define ZGEMM_DEFAULT_P zgemm_p
1038 #define ZGEMM_DEFAULT_R zgemm_r
1040 #define XGEMM_DEFAULT_P xgemm_p
1041 #define XGEMM_DEFAULT_R xgemm_r
1043 #define SGEMM_DEFAULT_Q 128
1044 #define DGEMM_DEFAULT_Q 128
1045 #define QGEMM_DEFAULT_Q 128
1046 #define CGEMM_DEFAULT_Q 128
1047 #define ZGEMM_DEFAULT_Q 128
1048 #define XGEMM_DEFAULT_Q 128
1051 #ifdef CORE_PRESCOTT
1057 #define GEMM_DEFAULT_OFFSET_A 128
1058 #define GEMM_DEFAULT_OFFSET_B 192
1060 #define GEMM_DEFAULT_OFFSET_A 0
1061 #define GEMM_DEFAULT_OFFSET_B 256
1064 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1069 #define SGEMM_DEFAULT_UNROLL_M 4
1070 #define DGEMM_DEFAULT_UNROLL_M 2
1071 #define QGEMM_DEFAULT_UNROLL_M 2
1072 #define CGEMM_DEFAULT_UNROLL_M 2
1073 #define ZGEMM_DEFAULT_UNROLL_M 1
1074 #define XGEMM_DEFAULT_UNROLL_M 1
1076 #define SGEMM_DEFAULT_UNROLL_M 8
1077 #define DGEMM_DEFAULT_UNROLL_M 4
1078 #define QGEMM_DEFAULT_UNROLL_M 2
1079 #define CGEMM_DEFAULT_UNROLL_M 4
1080 #define ZGEMM_DEFAULT_UNROLL_M 2
1081 #define XGEMM_DEFAULT_UNROLL_M 1
1084 #define SGEMM_DEFAULT_UNROLL_N 4
1085 #define DGEMM_DEFAULT_UNROLL_N 4
1086 #define QGEMM_DEFAULT_UNROLL_N 2
1087 #define CGEMM_DEFAULT_UNROLL_N 2
1088 #define ZGEMM_DEFAULT_UNROLL_N 2
1089 #define XGEMM_DEFAULT_UNROLL_N 1
1091 #define SGEMM_DEFAULT_P sgemm_p
1092 #define SGEMM_DEFAULT_R sgemm_r
1094 #define DGEMM_DEFAULT_P dgemm_p
1095 #define DGEMM_DEFAULT_R dgemm_r
1097 #define QGEMM_DEFAULT_P qgemm_p
1098 #define QGEMM_DEFAULT_R qgemm_r
1100 #define CGEMM_DEFAULT_P cgemm_p
1101 #define CGEMM_DEFAULT_R cgemm_r
1103 #define ZGEMM_DEFAULT_P zgemm_p
1104 #define ZGEMM_DEFAULT_R zgemm_r
1106 #define XGEMM_DEFAULT_P xgemm_p
1107 #define XGEMM_DEFAULT_R xgemm_r
1109 #define SGEMM_DEFAULT_Q 128
1110 #define DGEMM_DEFAULT_Q 128
1111 #define QGEMM_DEFAULT_Q 128
1112 #define CGEMM_DEFAULT_Q 128
1113 #define ZGEMM_DEFAULT_Q 128
1114 #define XGEMM_DEFAULT_Q 128
1122 #define GEMM_DEFAULT_OFFSET_A 448
1123 #define GEMM_DEFAULT_OFFSET_B 128
1124 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1128 #define SWITCH_RATIO 4
1131 #define SGEMM_DEFAULT_UNROLL_M 8
1132 #define DGEMM_DEFAULT_UNROLL_M 4
1133 #define QGEMM_DEFAULT_UNROLL_M 2
1134 #define CGEMM_DEFAULT_UNROLL_M 4
1135 #define ZGEMM_DEFAULT_UNROLL_M 2
1136 #define XGEMM_DEFAULT_UNROLL_M 1
1138 #define SGEMM_DEFAULT_UNROLL_N 2
1139 #define DGEMM_DEFAULT_UNROLL_N 2
1140 #define QGEMM_DEFAULT_UNROLL_N 2
1141 #define CGEMM_DEFAULT_UNROLL_N 1
1142 #define ZGEMM_DEFAULT_UNROLL_N 1
1143 #define XGEMM_DEFAULT_UNROLL_N 1
1145 #define MASK(a, b) ((((a) + (b) - 1) / (b)) * (b))
1148 #define SGEMM_DEFAULT_UNROLL_M 8
1149 #define DGEMM_DEFAULT_UNROLL_M 4
1150 #define QGEMM_DEFAULT_UNROLL_M 2
1151 #define CGEMM_DEFAULT_UNROLL_M 4
1152 #define ZGEMM_DEFAULT_UNROLL_M 2
1153 #define XGEMM_DEFAULT_UNROLL_M 1
1155 #define SGEMM_DEFAULT_UNROLL_N 4
1156 #define DGEMM_DEFAULT_UNROLL_N 4
1157 #define QGEMM_DEFAULT_UNROLL_N 2
1158 #define CGEMM_DEFAULT_UNROLL_N 2
1159 #define ZGEMM_DEFAULT_UNROLL_N 2
1160 #define XGEMM_DEFAULT_UNROLL_N 1
1163 #define SGEMM_DEFAULT_P sgemm_p
1164 #define SGEMM_DEFAULT_R sgemm_r
1166 #define DGEMM_DEFAULT_P dgemm_p
1167 #define DGEMM_DEFAULT_R dgemm_r
1169 #define QGEMM_DEFAULT_P qgemm_p
1170 #define QGEMM_DEFAULT_R qgemm_r
1172 #define CGEMM_DEFAULT_P cgemm_p
1173 #define CGEMM_DEFAULT_R cgemm_r
1175 #define ZGEMM_DEFAULT_P zgemm_p
1176 #define ZGEMM_DEFAULT_R zgemm_r
1178 #define XGEMM_DEFAULT_P xgemm_p
1179 #define XGEMM_DEFAULT_R xgemm_r
1181 #define SGEMM_DEFAULT_Q 256
1182 #define DGEMM_DEFAULT_Q 256
1183 #define QGEMM_DEFAULT_Q 256
1184 #define CGEMM_DEFAULT_Q 256
1185 #define ZGEMM_DEFAULT_Q 256
1186 #define XGEMM_DEFAULT_Q 256
1195 #define GEMM_DEFAULT_OFFSET_A 128
1196 #define GEMM_DEFAULT_OFFSET_B 0
1197 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1201 #define SWITCH_RATIO 4
1204 #define SGEMM_DEFAULT_UNROLL_M 4
1205 #define DGEMM_DEFAULT_UNROLL_M 2
1206 #define QGEMM_DEFAULT_UNROLL_M 2
1207 #define CGEMM_DEFAULT_UNROLL_M 2
1208 #define ZGEMM_DEFAULT_UNROLL_M 1
1209 #define XGEMM_DEFAULT_UNROLL_M 1
1211 #define SGEMM_DEFAULT_UNROLL_N 4
1212 #define DGEMM_DEFAULT_UNROLL_N 4
1213 #define QGEMM_DEFAULT_UNROLL_N 2
1214 #define CGEMM_DEFAULT_UNROLL_N 2
1215 #define ZGEMM_DEFAULT_UNROLL_N 2
1216 #define XGEMM_DEFAULT_UNROLL_N 1
1218 #define SGEMM_DEFAULT_UNROLL_M 8
1219 #define DGEMM_DEFAULT_UNROLL_M 4
1220 #define QGEMM_DEFAULT_UNROLL_M 2
1221 #define CGEMM_DEFAULT_UNROLL_M 4
1222 #define ZGEMM_DEFAULT_UNROLL_M 2
1223 #define XGEMM_DEFAULT_UNROLL_M 1
1225 #define SGEMM_DEFAULT_UNROLL_N 4
1226 #define DGEMM_DEFAULT_UNROLL_N 4
1227 #define QGEMM_DEFAULT_UNROLL_N 2
1228 #define CGEMM_DEFAULT_UNROLL_N 2
1229 #define ZGEMM_DEFAULT_UNROLL_N 2
1230 #define XGEMM_DEFAULT_UNROLL_N 1
1233 #define SGEMM_DEFAULT_P sgemm_p
1234 #define SGEMM_DEFAULT_R sgemm_r
1236 #define DGEMM_DEFAULT_P dgemm_p
1237 #define DGEMM_DEFAULT_R dgemm_r
1239 #define QGEMM_DEFAULT_P qgemm_p
1240 #define QGEMM_DEFAULT_R qgemm_r
1242 #define CGEMM_DEFAULT_P cgemm_p
1243 #define CGEMM_DEFAULT_R cgemm_r
1245 #define ZGEMM_DEFAULT_P zgemm_p
1246 #define ZGEMM_DEFAULT_R zgemm_r
1248 #define XGEMM_DEFAULT_P xgemm_p
1249 #define XGEMM_DEFAULT_R xgemm_r
1251 #define SGEMM_DEFAULT_Q 512
1252 #define DGEMM_DEFAULT_Q 256
1253 #define QGEMM_DEFAULT_Q 128
1254 #define CGEMM_DEFAULT_Q 512
1255 #define ZGEMM_DEFAULT_Q 256
1256 #define XGEMM_DEFAULT_Q 128
1258 #define GETRF_FACTOR 0.75
1266 #define GEMM_DEFAULT_OFFSET_A 128
1267 #define GEMM_DEFAULT_OFFSET_B 0
1268 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1272 #define SWITCH_RATIO 4
1275 #define SGEMM_DEFAULT_UNROLL_M 4
1276 #define DGEMM_DEFAULT_UNROLL_M 2
1277 #define QGEMM_DEFAULT_UNROLL_M 2
1278 #define CGEMM_DEFAULT_UNROLL_M 2
1279 #define ZGEMM_DEFAULT_UNROLL_M 1
1280 #define XGEMM_DEFAULT_UNROLL_M 1
1282 #define SGEMM_DEFAULT_UNROLL_N 4
1283 #define DGEMM_DEFAULT_UNROLL_N 4
1284 #define QGEMM_DEFAULT_UNROLL_N 2
1285 #define CGEMM_DEFAULT_UNROLL_N 2
1286 #define ZGEMM_DEFAULT_UNROLL_N 2
1287 #define XGEMM_DEFAULT_UNROLL_N 1
1289 #define SGEMM_DEFAULT_UNROLL_M 8
1290 #define DGEMM_DEFAULT_UNROLL_M 4
1291 #define QGEMM_DEFAULT_UNROLL_M 2
1292 #define CGEMM_DEFAULT_UNROLL_M 4
1293 #define ZGEMM_DEFAULT_UNROLL_M 2
1294 #define XGEMM_DEFAULT_UNROLL_M 1
1296 #define SGEMM_DEFAULT_UNROLL_N 4
1297 #define DGEMM_DEFAULT_UNROLL_N 4
1298 #define QGEMM_DEFAULT_UNROLL_N 2
1299 #define CGEMM_DEFAULT_UNROLL_N 2
1300 #define ZGEMM_DEFAULT_UNROLL_N 2
1301 #define XGEMM_DEFAULT_UNROLL_N 1
1304 #define SGEMM_DEFAULT_P sgemm_p
1305 #define SGEMM_DEFAULT_R sgemm_r
1307 #define DGEMM_DEFAULT_P dgemm_p
1308 #define DGEMM_DEFAULT_R dgemm_r
1310 #define QGEMM_DEFAULT_P qgemm_p
1311 #define QGEMM_DEFAULT_R qgemm_r
1313 #define CGEMM_DEFAULT_P cgemm_p
1314 #define CGEMM_DEFAULT_R cgemm_r
1316 #define ZGEMM_DEFAULT_P zgemm_p
1317 #define ZGEMM_DEFAULT_R zgemm_r
1319 #define XGEMM_DEFAULT_P xgemm_p
1320 #define XGEMM_DEFAULT_R xgemm_r
1322 #define SGEMM_DEFAULT_Q 768
1323 #define DGEMM_DEFAULT_Q 384
1324 #define QGEMM_DEFAULT_Q 192
1325 #define CGEMM_DEFAULT_Q 768
1326 #define ZGEMM_DEFAULT_Q 384
1327 #define XGEMM_DEFAULT_Q 192
1329 #define GETRF_FACTOR 0.75
1330 #define GEMM_THREAD gemm_thread_mn
1338 #define GEMM_DEFAULT_OFFSET_A 32
1339 #define GEMM_DEFAULT_OFFSET_B 0
1340 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1344 #define SWITCH_RATIO 4
1347 #define SGEMM_DEFAULT_UNROLL_M 4
1348 #define DGEMM_DEFAULT_UNROLL_M 2
1349 #define QGEMM_DEFAULT_UNROLL_M 2
1350 #define CGEMM_DEFAULT_UNROLL_M 2
1351 #define ZGEMM_DEFAULT_UNROLL_M 1
1352 #define XGEMM_DEFAULT_UNROLL_M 1
1354 #define SGEMM_DEFAULT_UNROLL_N 4
1355 #define DGEMM_DEFAULT_UNROLL_N 4
1356 #define QGEMM_DEFAULT_UNROLL_N 2
1357 #define CGEMM_DEFAULT_UNROLL_N 2
1358 #define ZGEMM_DEFAULT_UNROLL_N 2
1359 #define XGEMM_DEFAULT_UNROLL_N 1
1361 #define SGEMM_DEFAULT_UNROLL_M 4
1362 #define DGEMM_DEFAULT_UNROLL_M 2
1363 #define QGEMM_DEFAULT_UNROLL_M 2
1364 #define CGEMM_DEFAULT_UNROLL_M 2
1365 #define ZGEMM_DEFAULT_UNROLL_M 1
1366 #define XGEMM_DEFAULT_UNROLL_M 1
1368 #define SGEMM_DEFAULT_UNROLL_N 8
1369 #define DGEMM_DEFAULT_UNROLL_N 8
1370 #define QGEMM_DEFAULT_UNROLL_N 2
1371 #define CGEMM_DEFAULT_UNROLL_N 4
1372 #define ZGEMM_DEFAULT_UNROLL_N 4
1373 #define XGEMM_DEFAULT_UNROLL_N 1
1376 #define SGEMM_DEFAULT_P 504
1377 #define SGEMM_DEFAULT_R sgemm_r
1379 #define DGEMM_DEFAULT_P 504
1380 #define DGEMM_DEFAULT_R dgemm_r
1382 #define QGEMM_DEFAULT_P 504
1383 #define QGEMM_DEFAULT_R qgemm_r
1385 #define CGEMM_DEFAULT_P 252
1386 #define CGEMM_DEFAULT_R cgemm_r
1388 #define ZGEMM_DEFAULT_P 252
1389 #define ZGEMM_DEFAULT_R zgemm_r
1391 #define XGEMM_DEFAULT_P 252
1392 #define XGEMM_DEFAULT_R xgemm_r
1394 #define SGEMM_DEFAULT_Q 512
1395 #define DGEMM_DEFAULT_Q 256
1396 #define QGEMM_DEFAULT_Q 128
1397 #define CGEMM_DEFAULT_Q 512
1398 #define ZGEMM_DEFAULT_Q 256
1399 #define XGEMM_DEFAULT_Q 128
1401 #define GETRF_FACTOR 0.72
1411 #define GEMM_DEFAULT_OFFSET_A 0
1412 #define GEMM_DEFAULT_OFFSET_B 0
1413 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1417 #define SWITCH_RATIO 4
1420 #define SGEMM_DEFAULT_UNROLL_M 4
1421 #define DGEMM_DEFAULT_UNROLL_M 2
1422 #define QGEMM_DEFAULT_UNROLL_M 2
1423 #define CGEMM_DEFAULT_UNROLL_M 2
1424 #define ZGEMM_DEFAULT_UNROLL_M 1
1425 #define XGEMM_DEFAULT_UNROLL_M 1
1427 #define SGEMM_DEFAULT_UNROLL_N 4
1428 #define DGEMM_DEFAULT_UNROLL_N 4
1429 #define QGEMM_DEFAULT_UNROLL_N 2
1430 #define CGEMM_DEFAULT_UNROLL_N 2
1431 #define ZGEMM_DEFAULT_UNROLL_N 2
1432 #define XGEMM_DEFAULT_UNROLL_N 1
1434 #define SGEMM_DEFAULT_UNROLL_M 16
1435 #define DGEMM_DEFAULT_UNROLL_M 8
1436 #define QGEMM_DEFAULT_UNROLL_M 2
1437 #define CGEMM_DEFAULT_UNROLL_M 8
1438 #define ZGEMM_DEFAULT_UNROLL_M 1
1439 #define XGEMM_DEFAULT_UNROLL_M 1
1441 #define SGEMM_DEFAULT_UNROLL_N 4
1442 #define DGEMM_DEFAULT_UNROLL_N 4
1443 #define QGEMM_DEFAULT_UNROLL_N 2
1444 #define CGEMM_DEFAULT_UNROLL_N 2
1445 #define ZGEMM_DEFAULT_UNROLL_N 4
1446 #define XGEMM_DEFAULT_UNROLL_N 1
1449 #define SGEMM_DEFAULT_P 768
1450 #define SGEMM_DEFAULT_R sgemm_r
1451 //#define SGEMM_DEFAULT_R 1024
1453 #define DGEMM_DEFAULT_P 512
1454 #define DGEMM_DEFAULT_R dgemm_r
1455 //#define DGEMM_DEFAULT_R 1024
1457 #define QGEMM_DEFAULT_P 504
1458 #define QGEMM_DEFAULT_R qgemm_r
1460 #define CGEMM_DEFAULT_P 768
1461 #define CGEMM_DEFAULT_R cgemm_r
1462 //#define CGEMM_DEFAULT_R 1024
1464 #define ZGEMM_DEFAULT_P 512
1465 #define ZGEMM_DEFAULT_R zgemm_r
1466 //#define ZGEMM_DEFAULT_R 1024
1468 #define XGEMM_DEFAULT_P 252
1469 #define XGEMM_DEFAULT_R xgemm_r
1471 #define SGEMM_DEFAULT_Q 384
1472 #define DGEMM_DEFAULT_Q 256
1473 #define QGEMM_DEFAULT_Q 128
1474 #define CGEMM_DEFAULT_Q 512
1475 #define ZGEMM_DEFAULT_Q 192
1476 #define XGEMM_DEFAULT_Q 128
1478 #define CGEMM3M_DEFAULT_UNROLL_N 8
1479 #define CGEMM3M_DEFAULT_UNROLL_M 4
1480 #define ZGEMM3M_DEFAULT_UNROLL_N 8
1481 #define ZGEMM3M_DEFAULT_UNROLL_M 2
1483 #define CGEMM3M_DEFAULT_P 448
1484 #define ZGEMM3M_DEFAULT_P 224
1485 #define XGEMM3M_DEFAULT_P 112
1486 #define CGEMM3M_DEFAULT_Q 224
1487 #define ZGEMM3M_DEFAULT_Q 224
1488 #define XGEMM3M_DEFAULT_Q 224
1489 #define CGEMM3M_DEFAULT_R 12288
1490 #define ZGEMM3M_DEFAULT_R 12288
1491 #define XGEMM3M_DEFAULT_R 12288
1495 #define GETRF_FACTOR 0.72
1504 #define GEMM_DEFAULT_OFFSET_A 0
1505 #define GEMM_DEFAULT_OFFSET_B 0
1506 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1510 #define SWITCH_RATIO 32
1514 #define SGEMM_DEFAULT_UNROLL_M 4
1515 #define DGEMM_DEFAULT_UNROLL_M 2
1516 #define QGEMM_DEFAULT_UNROLL_M 2
1517 #define CGEMM_DEFAULT_UNROLL_M 2
1518 #define ZGEMM_DEFAULT_UNROLL_M 1
1519 #define XGEMM_DEFAULT_UNROLL_M 1
1521 #define SGEMM_DEFAULT_UNROLL_N 4
1522 #define DGEMM_DEFAULT_UNROLL_N 4
1523 #define QGEMM_DEFAULT_UNROLL_N 2
1524 #define CGEMM_DEFAULT_UNROLL_N 2
1525 #define ZGEMM_DEFAULT_UNROLL_N 2
1526 #define XGEMM_DEFAULT_UNROLL_N 1
1530 #define SGEMM_DEFAULT_UNROLL_M 16
1531 #define DGEMM_DEFAULT_UNROLL_M 4
1532 #define QGEMM_DEFAULT_UNROLL_M 2
1533 #define CGEMM_DEFAULT_UNROLL_M 8
1534 #define ZGEMM_DEFAULT_UNROLL_M 4
1535 #define XGEMM_DEFAULT_UNROLL_M 1
1537 #define SGEMM_DEFAULT_UNROLL_N 4
1538 #define DGEMM_DEFAULT_UNROLL_N 8
1539 #define QGEMM_DEFAULT_UNROLL_N 2
1540 #define CGEMM_DEFAULT_UNROLL_N 2
1541 #define ZGEMM_DEFAULT_UNROLL_N 2
1542 #define XGEMM_DEFAULT_UNROLL_N 1
1544 #define SGEMM_DEFAULT_UNROLL_MN 32
1545 #define DGEMM_DEFAULT_UNROLL_MN 32
1550 #define SGEMM_DEFAULT_P 512
1551 #define SGEMM_DEFAULT_R sgemm_r
1552 #define DGEMM_DEFAULT_P 512
1553 #define DGEMM_DEFAULT_R dgemm_r
1554 #define QGEMM_DEFAULT_P 504
1555 #define QGEMM_DEFAULT_R qgemm_r
1556 #define CGEMM_DEFAULT_P 128
1557 #define CGEMM_DEFAULT_R 1024
1558 #define ZGEMM_DEFAULT_P 512
1559 #define ZGEMM_DEFAULT_R zgemm_r
1560 #define XGEMM_DEFAULT_P 252
1561 #define XGEMM_DEFAULT_R xgemm_r
1562 #define SGEMM_DEFAULT_Q 256
1563 #define DGEMM_DEFAULT_Q 256
1564 #define QGEMM_DEFAULT_Q 128
1565 #define CGEMM_DEFAULT_Q 256
1566 #define ZGEMM_DEFAULT_Q 192
1567 #define XGEMM_DEFAULT_Q 128
1571 #define SGEMM_DEFAULT_P 768
1572 #define DGEMM_DEFAULT_P 512
1573 #define CGEMM_DEFAULT_P 384
1574 #define ZGEMM_DEFAULT_P 256
1577 #define SGEMM_DEFAULT_Q 320
1578 #define DGEMM_DEFAULT_Q 128
1580 #define SGEMM_DEFAULT_Q 384
1581 #define DGEMM_DEFAULT_Q 256
1583 #define CGEMM_DEFAULT_Q 192
1584 #define ZGEMM_DEFAULT_Q 128
1586 #define SGEMM_DEFAULT_R sgemm_r
1587 #define DGEMM_DEFAULT_R 13824
1588 #define CGEMM_DEFAULT_R cgemm_r
1589 #define ZGEMM_DEFAULT_R zgemm_r
1591 #define QGEMM_DEFAULT_Q 128
1592 #define QGEMM_DEFAULT_P 504
1593 #define QGEMM_DEFAULT_R qgemm_r
1594 #define XGEMM_DEFAULT_P 252
1595 #define XGEMM_DEFAULT_R xgemm_r
1596 #define XGEMM_DEFAULT_Q 128
1598 #define CGEMM3M_DEFAULT_UNROLL_N 8
1599 #define CGEMM3M_DEFAULT_UNROLL_M 4
1600 #define ZGEMM3M_DEFAULT_UNROLL_N 8
1601 #define ZGEMM3M_DEFAULT_UNROLL_M 2
1603 #define CGEMM3M_DEFAULT_P 448
1604 #define ZGEMM3M_DEFAULT_P 224
1605 #define XGEMM3M_DEFAULT_P 112
1606 #define CGEMM3M_DEFAULT_Q 224
1607 #define ZGEMM3M_DEFAULT_Q 224
1608 #define XGEMM3M_DEFAULT_Q 224
1609 #define CGEMM3M_DEFAULT_R 12288
1610 #define ZGEMM3M_DEFAULT_R 12288
1611 #define XGEMM3M_DEFAULT_R 12288
1623 #define GEMM_DEFAULT_OFFSET_A 0
1624 #define GEMM_DEFAULT_OFFSET_B 0
1625 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1629 #define SWITCH_RATIO 32
1633 #define SGEMM_DEFAULT_UNROLL_M 4
1634 #define DGEMM_DEFAULT_UNROLL_M 2
1635 #define QGEMM_DEFAULT_UNROLL_M 2
1636 #define CGEMM_DEFAULT_UNROLL_M 2
1637 #define ZGEMM_DEFAULT_UNROLL_M 1
1638 #define XGEMM_DEFAULT_UNROLL_M 1
1640 #define SGEMM_DEFAULT_UNROLL_N 4
1641 #define DGEMM_DEFAULT_UNROLL_N 4
1642 #define QGEMM_DEFAULT_UNROLL_N 2
1643 #define CGEMM_DEFAULT_UNROLL_N 2
1644 #define ZGEMM_DEFAULT_UNROLL_N 2
1645 #define XGEMM_DEFAULT_UNROLL_N 1
1649 #define SGEMM_DEFAULT_UNROLL_M 16
1650 #define DGEMM_DEFAULT_UNROLL_M 4
1651 #define QGEMM_DEFAULT_UNROLL_M 2
1652 #define CGEMM_DEFAULT_UNROLL_M 8
1653 #define ZGEMM_DEFAULT_UNROLL_M 4
1654 #define XGEMM_DEFAULT_UNROLL_M 1
1656 #define SGEMM_DEFAULT_UNROLL_N 4
1657 #define DGEMM_DEFAULT_UNROLL_N 8
1658 #define QGEMM_DEFAULT_UNROLL_N 2
1659 #define CGEMM_DEFAULT_UNROLL_N 2
1660 #define ZGEMM_DEFAULT_UNROLL_N 2
1661 #define XGEMM_DEFAULT_UNROLL_N 1
1663 #define SGEMM_DEFAULT_UNROLL_MN 32
1664 #define DGEMM_DEFAULT_UNROLL_MN 32
1669 #define SGEMM_DEFAULT_P 512
1670 #define SGEMM_DEFAULT_R sgemm_r
1671 #define DGEMM_DEFAULT_P 512
1672 #define DGEMM_DEFAULT_R dgemm_r
1673 #define QGEMM_DEFAULT_P 504
1674 #define QGEMM_DEFAULT_R qgemm_r
1675 #define CGEMM_DEFAULT_P 128
1676 #define CGEMM_DEFAULT_R 1024
1677 #define ZGEMM_DEFAULT_P 512
1678 #define ZGEMM_DEFAULT_R zgemm_r
1679 #define XGEMM_DEFAULT_P 252
1680 #define XGEMM_DEFAULT_R xgemm_r
1681 #define SGEMM_DEFAULT_Q 256
1682 #define DGEMM_DEFAULT_Q 256
1683 #define QGEMM_DEFAULT_Q 128
1684 #define CGEMM_DEFAULT_Q 256
1685 #define ZGEMM_DEFAULT_Q 192
1686 #define XGEMM_DEFAULT_Q 128
1690 #define SGEMM_DEFAULT_P 768
1691 #define DGEMM_DEFAULT_P 512
1692 #define CGEMM_DEFAULT_P 384
1693 #define ZGEMM_DEFAULT_P 256
1696 #define SGEMM_DEFAULT_Q 320
1697 #define DGEMM_DEFAULT_Q 128
1699 #define SGEMM_DEFAULT_Q 384
1700 #define DGEMM_DEFAULT_Q 256
1702 #define CGEMM_DEFAULT_Q 192
1703 #define ZGEMM_DEFAULT_Q 128
1705 #define SGEMM_DEFAULT_R sgemm_r
1706 #define DGEMM_DEFAULT_R 13824
1707 #define CGEMM_DEFAULT_R cgemm_r
1708 #define ZGEMM_DEFAULT_R zgemm_r
1710 #define QGEMM_DEFAULT_Q 128
1711 #define QGEMM_DEFAULT_P 504
1712 #define QGEMM_DEFAULT_R qgemm_r
1713 #define XGEMM_DEFAULT_P 252
1714 #define XGEMM_DEFAULT_R xgemm_r
1715 #define XGEMM_DEFAULT_Q 128
1717 #define CGEMM3M_DEFAULT_UNROLL_N 8
1718 #define CGEMM3M_DEFAULT_UNROLL_M 4
1719 #define ZGEMM3M_DEFAULT_UNROLL_N 8
1720 #define ZGEMM3M_DEFAULT_UNROLL_M 2
1722 #define CGEMM3M_DEFAULT_P 448
1723 #define ZGEMM3M_DEFAULT_P 224
1724 #define XGEMM3M_DEFAULT_P 112
1725 #define CGEMM3M_DEFAULT_Q 224
1726 #define ZGEMM3M_DEFAULT_Q 224
1727 #define XGEMM3M_DEFAULT_Q 224
1728 #define CGEMM3M_DEFAULT_R 12288
1729 #define ZGEMM3M_DEFAULT_R 12288
1730 #define XGEMM3M_DEFAULT_R 12288
1744 #define GEMM_DEFAULT_OFFSET_A 64
1745 #define GEMM_DEFAULT_OFFSET_B 0
1746 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1751 #define SGEMM_DEFAULT_UNROLL_M 4
1752 #define DGEMM_DEFAULT_UNROLL_M 2
1753 #define QGEMM_DEFAULT_UNROLL_M 2
1754 #define CGEMM_DEFAULT_UNROLL_M 2
1755 #define ZGEMM_DEFAULT_UNROLL_M 1
1756 #define XGEMM_DEFAULT_UNROLL_M 1
1758 #define SGEMM_DEFAULT_UNROLL_M 8
1759 #define DGEMM_DEFAULT_UNROLL_M 4
1760 #define QGEMM_DEFAULT_UNROLL_M 2
1761 #define CGEMM_DEFAULT_UNROLL_M 4
1762 #define ZGEMM_DEFAULT_UNROLL_M 2
1763 #define XGEMM_DEFAULT_UNROLL_M 1
1766 #define SGEMM_DEFAULT_UNROLL_N 4
1767 #define DGEMM_DEFAULT_UNROLL_N 2
1768 #define QGEMM_DEFAULT_UNROLL_N 2
1769 #define CGEMM_DEFAULT_UNROLL_N 2
1770 #define ZGEMM_DEFAULT_UNROLL_N 1
1771 #define XGEMM_DEFAULT_UNROLL_N 1
1773 #define SGEMM_DEFAULT_P sgemm_p
1774 #define SGEMM_DEFAULT_R sgemm_r
1776 #define DGEMM_DEFAULT_P dgemm_p
1777 #define DGEMM_DEFAULT_R dgemm_r
1779 #define QGEMM_DEFAULT_P qgemm_p
1780 #define QGEMM_DEFAULT_R qgemm_r
1782 #define CGEMM_DEFAULT_P cgemm_p
1783 #define CGEMM_DEFAULT_R cgemm_r
1785 #define ZGEMM_DEFAULT_P zgemm_p
1786 #define ZGEMM_DEFAULT_R zgemm_r
1788 #define XGEMM_DEFAULT_P xgemm_p
1789 #define XGEMM_DEFAULT_R xgemm_r
1791 #define SGEMM_DEFAULT_Q 256
1792 #define DGEMM_DEFAULT_Q 256
1793 #define QGEMM_DEFAULT_Q 256
1794 #define CGEMM_DEFAULT_Q 256
1795 #define ZGEMM_DEFAULT_Q 256
1796 #define XGEMM_DEFAULT_Q 256
1806 #define GEMM_DEFAULT_OFFSET_A 0
1807 #define GEMM_DEFAULT_OFFSET_B 128
1808 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1810 #define SGEMM_DEFAULT_UNROLL_M 8
1811 #define SGEMM_DEFAULT_UNROLL_N 8
1812 #define DGEMM_DEFAULT_UNROLL_M 8
1813 #define DGEMM_DEFAULT_UNROLL_N 8
1814 #define QGEMM_DEFAULT_UNROLL_M 8
1815 #define QGEMM_DEFAULT_UNROLL_N 8
1816 #define CGEMM_DEFAULT_UNROLL_M 4
1817 #define CGEMM_DEFAULT_UNROLL_N 4
1818 #define ZGEMM_DEFAULT_UNROLL_M 4
1819 #define ZGEMM_DEFAULT_UNROLL_N 4
1820 #define XGEMM_DEFAULT_UNROLL_M 4
1821 #define XGEMM_DEFAULT_UNROLL_N 4
1823 #define SGEMM_DEFAULT_P sgemm_p
1824 #define DGEMM_DEFAULT_P dgemm_p
1825 #define QGEMM_DEFAULT_P qgemm_p
1826 #define CGEMM_DEFAULT_P cgemm_p
1827 #define ZGEMM_DEFAULT_P zgemm_p
1828 #define XGEMM_DEFAULT_P xgemm_p
1830 #define SGEMM_DEFAULT_Q 1024
1831 #define DGEMM_DEFAULT_Q 1024
1832 #define QGEMM_DEFAULT_Q 1024
1833 #define CGEMM_DEFAULT_Q 1024
1834 #define ZGEMM_DEFAULT_Q 1024
1835 #define XGEMM_DEFAULT_Q 1024
1837 #define SGEMM_DEFAULT_R sgemm_r
1838 #define DGEMM_DEFAULT_R dgemm_r
1839 #define QGEMM_DEFAULT_R qgemm_r
1840 #define CGEMM_DEFAULT_R cgemm_r
1841 #define ZGEMM_DEFAULT_R zgemm_r
1842 #define XGEMM_DEFAULT_R xgemm_r
1846 #define GETRF_FACTOR 0.65
1850 #if defined(EV4) || defined(EV5) || defined(EV6)
1860 #define GEMM_DEFAULT_OFFSET_A 512
1861 #define GEMM_DEFAULT_OFFSET_B 512
1862 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1864 #define SGEMM_DEFAULT_UNROLL_M 4
1865 #define SGEMM_DEFAULT_UNROLL_N 4
1866 #define DGEMM_DEFAULT_UNROLL_M 4
1867 #define DGEMM_DEFAULT_UNROLL_N 4
1868 #define CGEMM_DEFAULT_UNROLL_M 2
1869 #define CGEMM_DEFAULT_UNROLL_N 2
1870 #define ZGEMM_DEFAULT_UNROLL_M 2
1871 #define ZGEMM_DEFAULT_UNROLL_N 2
1876 #define SGEMM_DEFAULT_P 32
1877 #define SGEMM_DEFAULT_Q 112
1878 #define SGEMM_DEFAULT_R 256
1880 #define DGEMM_DEFAULT_P 32
1881 #define DGEMM_DEFAULT_Q 56
1882 #define DGEMM_DEFAULT_R 256
1884 #define CGEMM_DEFAULT_P 32
1885 #define CGEMM_DEFAULT_Q 64
1886 #define CGEMM_DEFAULT_R 240
1888 #define ZGEMM_DEFAULT_P 32
1889 #define ZGEMM_DEFAULT_Q 32
1890 #define ZGEMM_DEFAULT_R 240
1894 #define SGEMM_DEFAULT_P 64
1895 #define SGEMM_DEFAULT_Q 256
1897 #define DGEMM_DEFAULT_P 64
1898 #define DGEMM_DEFAULT_Q 128
1900 #define CGEMM_DEFAULT_P 64
1901 #define CGEMM_DEFAULT_Q 128
1903 #define ZGEMM_DEFAULT_P 64
1904 #define ZGEMM_DEFAULT_Q 64
1908 #define SGEMM_DEFAULT_P 256
1909 #define SGEMM_DEFAULT_Q 512
1911 #define DGEMM_DEFAULT_P 256
1912 #define DGEMM_DEFAULT_Q 256
1914 #define CGEMM_DEFAULT_P 256
1915 #define CGEMM_DEFAULT_Q 256
1917 #define ZGEMM_DEFAULT_P 128
1918 #define ZGEMM_DEFAULT_Q 256
1928 #define GEMM_DEFAULT_OFFSET_A 0
1929 #define GEMM_DEFAULT_OFFSET_B 8192
1930 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1932 #define SGEMM_DEFAULT_UNROLL_M 16
1933 #define SGEMM_DEFAULT_UNROLL_N 4
1934 #define DGEMM_DEFAULT_UNROLL_M 4
1935 #define DGEMM_DEFAULT_UNROLL_N 4
1936 #define CGEMM_DEFAULT_UNROLL_M 8
1937 #define CGEMM_DEFAULT_UNROLL_N 2
1938 #define ZGEMM_DEFAULT_UNROLL_M 2
1939 #define ZGEMM_DEFAULT_UNROLL_N 2
1941 #define SGEMM_DEFAULT_P 128
1942 #define DGEMM_DEFAULT_P 128
1943 #define CGEMM_DEFAULT_P 128
1944 #define ZGEMM_DEFAULT_P 128
1946 #define SGEMM_DEFAULT_Q 512
1947 #define DGEMM_DEFAULT_Q 256
1948 #define CGEMM_DEFAULT_Q 256
1949 #define ZGEMM_DEFAULT_Q 128
1955 #define GEMM_DEFAULT_OFFSET_A 0
1956 #define GEMM_DEFAULT_OFFSET_B 1024
1957 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1959 #define SGEMM_DEFAULT_UNROLL_M 16
1960 #define SGEMM_DEFAULT_UNROLL_N 4
1961 #define DGEMM_DEFAULT_UNROLL_M 4
1962 #define DGEMM_DEFAULT_UNROLL_N 4
1963 #define CGEMM_DEFAULT_UNROLL_M 8
1964 #define CGEMM_DEFAULT_UNROLL_N 2
1965 #define ZGEMM_DEFAULT_UNROLL_M 2
1966 #define ZGEMM_DEFAULT_UNROLL_N 2
1968 #define SGEMM_DEFAULT_P 256
1969 #define DGEMM_DEFAULT_P 128
1970 #define CGEMM_DEFAULT_P 128
1971 #define ZGEMM_DEFAULT_P 64
1973 #define SGEMM_DEFAULT_Q 256
1974 #define DGEMM_DEFAULT_Q 256
1975 #define CGEMM_DEFAULT_Q 256
1976 #define ZGEMM_DEFAULT_Q 256
1986 #define GEMM_DEFAULT_OFFSET_A 2688
1987 #define GEMM_DEFAULT_OFFSET_B 3072
1988 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1990 #define SGEMM_DEFAULT_UNROLL_M 16
1991 #define SGEMM_DEFAULT_UNROLL_N 4
1992 #define DGEMM_DEFAULT_UNROLL_M 4
1993 #define DGEMM_DEFAULT_UNROLL_N 4
1994 #define CGEMM_DEFAULT_UNROLL_M 8
1995 #define CGEMM_DEFAULT_UNROLL_N 2
1996 #define ZGEMM_DEFAULT_UNROLL_M 2
1997 #define ZGEMM_DEFAULT_UNROLL_N 2
2000 #if L2_SIZE == 1024976
2001 #define SGEMM_DEFAULT_P 320
2002 #define DGEMM_DEFAULT_P 256
2003 #define CGEMM_DEFAULT_P 256
2004 #define ZGEMM_DEFAULT_P 256
2006 #define SGEMM_DEFAULT_P 176
2007 #define DGEMM_DEFAULT_P 176
2008 #define CGEMM_DEFAULT_P 176
2009 #define ZGEMM_DEFAULT_P 176
2013 #define SGEMM_DEFAULT_Q 512
2014 #define DGEMM_DEFAULT_Q 256
2015 #define CGEMM_DEFAULT_Q 256
2016 #define ZGEMM_DEFAULT_Q 128
2027 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2028 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2029 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2031 #define SGEMM_DEFAULT_UNROLL_M 4
2032 #define SGEMM_DEFAULT_UNROLL_N 4
2033 #define DGEMM_DEFAULT_UNROLL_M 4
2034 #define DGEMM_DEFAULT_UNROLL_N 4
2035 #define CGEMM_DEFAULT_UNROLL_M 2
2036 #define CGEMM_DEFAULT_UNROLL_N 2
2037 #define ZGEMM_DEFAULT_UNROLL_M 2
2038 #define ZGEMM_DEFAULT_UNROLL_N 2
2040 #define SGEMM_DEFAULT_P 512
2041 #define DGEMM_DEFAULT_P 512
2042 #define CGEMM_DEFAULT_P 512
2043 #define ZGEMM_DEFAULT_P 512
2045 #define SGEMM_DEFAULT_Q 1024
2046 #define DGEMM_DEFAULT_Q 512
2047 #define CGEMM_DEFAULT_Q 512
2048 #define ZGEMM_DEFAULT_Q 256
2050 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2051 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2052 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2053 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2063 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2064 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2065 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2067 #define SGEMM_DEFAULT_UNROLL_M 8
2068 #define SGEMM_DEFAULT_UNROLL_N 4
2069 #define DGEMM_DEFAULT_UNROLL_M 8
2070 #define DGEMM_DEFAULT_UNROLL_N 4
2071 #define CGEMM_DEFAULT_UNROLL_M 4
2072 #define CGEMM_DEFAULT_UNROLL_N 2
2073 #define ZGEMM_DEFAULT_UNROLL_M 4
2074 #define ZGEMM_DEFAULT_UNROLL_N 2
2076 #define SGEMM_DEFAULT_P 128
2077 #define DGEMM_DEFAULT_P 128
2078 #define CGEMM_DEFAULT_P 128
2079 #define ZGEMM_DEFAULT_P 128
2081 #define SGEMM_DEFAULT_Q 4096
2082 #define DGEMM_DEFAULT_Q 3072
2083 #define CGEMM_DEFAULT_Q 2048
2084 #define ZGEMM_DEFAULT_Q 1024
2086 #define SGEMM_DEFAULT_Q 512
2087 #define DGEMM_DEFAULT_Q 256
2088 #define CGEMM_DEFAULT_Q 256
2089 #define ZGEMM_DEFAULT_Q 128
2097 #if defined(POWER3) || defined(POWER4) || defined(POWER5)
2098 #define GEMM_DEFAULT_OFFSET_A 0
2099 #define GEMM_DEFAULT_OFFSET_B 2048
2100 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2102 #define SGEMM_DEFAULT_UNROLL_M 4
2103 #define SGEMM_DEFAULT_UNROLL_N 4
2104 #define DGEMM_DEFAULT_UNROLL_M 4
2105 #define DGEMM_DEFAULT_UNROLL_N 4
2106 #define CGEMM_DEFAULT_UNROLL_M 2
2107 #define CGEMM_DEFAULT_UNROLL_N 2
2108 #define ZGEMM_DEFAULT_UNROLL_M 2
2109 #define ZGEMM_DEFAULT_UNROLL_N 2
2116 #define SGEMM_DEFAULT_P 256
2117 #define SGEMM_DEFAULT_Q 432
2118 #define SGEMM_DEFAULT_R 1012
2120 #define DGEMM_DEFAULT_P 256
2121 #define DGEMM_DEFAULT_Q 216
2122 #define DGEMM_DEFAULT_R 1012
2124 #define ZGEMM_DEFAULT_P 256
2125 #define ZGEMM_DEFAULT_Q 104
2126 #define ZGEMM_DEFAULT_R 1012
2130 #ifdef ALLOC_HUGETLB
2131 #define SGEMM_DEFAULT_P 184
2132 #define DGEMM_DEFAULT_P 184
2133 #define CGEMM_DEFAULT_P 184
2134 #define ZGEMM_DEFAULT_P 184
2136 #define SGEMM_DEFAULT_P 144
2137 #define DGEMM_DEFAULT_P 144
2138 #define CGEMM_DEFAULT_P 144
2139 #define ZGEMM_DEFAULT_P 144
2144 #ifdef ALLOC_HUGETLB
2145 #define SGEMM_DEFAULT_P 512
2146 #define DGEMM_DEFAULT_P 256
2147 #define CGEMM_DEFAULT_P 256
2148 #define ZGEMM_DEFAULT_P 128
2150 #define SGEMM_DEFAULT_P 320
2151 #define DGEMM_DEFAULT_P 160
2152 #define CGEMM_DEFAULT_P 160
2153 #define ZGEMM_DEFAULT_P 80
2156 #define SGEMM_DEFAULT_Q 256
2157 #define CGEMM_DEFAULT_Q 256
2158 #define DGEMM_DEFAULT_Q 256
2159 #define ZGEMM_DEFAULT_Q 256
2171 #define GEMM_DEFAULT_OFFSET_A 384
2172 #define GEMM_DEFAULT_OFFSET_B 1024
2173 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2175 #define SGEMM_DEFAULT_UNROLL_M 4
2176 #define SGEMM_DEFAULT_UNROLL_N 4
2177 #define DGEMM_DEFAULT_UNROLL_M 4
2178 #define DGEMM_DEFAULT_UNROLL_N 4
2179 #define CGEMM_DEFAULT_UNROLL_M 2
2180 #define CGEMM_DEFAULT_UNROLL_N 4
2181 #define ZGEMM_DEFAULT_UNROLL_M 2
2182 #define ZGEMM_DEFAULT_UNROLL_N 4
2184 #define SGEMM_DEFAULT_P 992
2185 #define DGEMM_DEFAULT_P 480
2186 #define CGEMM_DEFAULT_P 488
2187 #define ZGEMM_DEFAULT_P 248
2189 #define SGEMM_DEFAULT_Q 504
2190 #define DGEMM_DEFAULT_Q 504
2191 #define CGEMM_DEFAULT_Q 400
2192 #define ZGEMM_DEFAULT_Q 400
2203 #define GEMM_DEFAULT_OFFSET_A 0
2204 #define GEMM_DEFAULT_OFFSET_B 65536
2205 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2207 #define SGEMM_DEFAULT_UNROLL_M 16
2208 #define SGEMM_DEFAULT_UNROLL_N 8
2209 #define DGEMM_DEFAULT_UNROLL_M 16
2210 #define DGEMM_DEFAULT_UNROLL_N 4
2211 #define CGEMM_DEFAULT_UNROLL_M 8
2212 #define CGEMM_DEFAULT_UNROLL_N 4
2213 #define ZGEMM_DEFAULT_UNROLL_M 8
2214 #define ZGEMM_DEFAULT_UNROLL_N 2
2216 #define SGEMM_DEFAULT_P 1280
2217 #define DGEMM_DEFAULT_P 640
2218 #define CGEMM_DEFAULT_P 640
2219 #define ZGEMM_DEFAULT_P 320
2221 #define SGEMM_DEFAULT_Q 640
2222 #define DGEMM_DEFAULT_Q 720
2223 #define CGEMM_DEFAULT_Q 640
2224 #define ZGEMM_DEFAULT_Q 640
2231 #if defined(SPARC) && defined(V7)
2236 #define GEMM_DEFAULT_OFFSET_A 0
2237 #define GEMM_DEFAULT_OFFSET_B 2048
2238 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2240 #define SGEMM_DEFAULT_UNROLL_M 2
2241 #define SGEMM_DEFAULT_UNROLL_N 8
2242 #define DGEMM_DEFAULT_UNROLL_M 2
2243 #define DGEMM_DEFAULT_UNROLL_N 8
2244 #define CGEMM_DEFAULT_UNROLL_M 1
2245 #define CGEMM_DEFAULT_UNROLL_N 4
2246 #define ZGEMM_DEFAULT_UNROLL_M 1
2247 #define ZGEMM_DEFAULT_UNROLL_N 4
2249 #define SGEMM_DEFAULT_P 256
2250 #define DGEMM_DEFAULT_P 256
2251 #define CGEMM_DEFAULT_P 256
2252 #define ZGEMM_DEFAULT_P 256
2254 #define SGEMM_DEFAULT_Q 512
2255 #define DGEMM_DEFAULT_Q 256
2256 #define CGEMM_DEFAULT_Q 256
2257 #define ZGEMM_DEFAULT_Q 128
2260 #define GEMM_THREAD gemm_thread_mn
2263 #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
2268 #define GEMM_DEFAULT_OFFSET_A 0
2269 #define GEMM_DEFAULT_OFFSET_B 2048
2270 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2272 #define SGEMM_DEFAULT_UNROLL_M 4
2273 #define SGEMM_DEFAULT_UNROLL_N 4
2274 #define DGEMM_DEFAULT_UNROLL_M 4
2275 #define DGEMM_DEFAULT_UNROLL_N 4
2276 #define CGEMM_DEFAULT_UNROLL_M 2
2277 #define CGEMM_DEFAULT_UNROLL_N 2
2278 #define ZGEMM_DEFAULT_UNROLL_M 2
2279 #define ZGEMM_DEFAULT_UNROLL_N 2
2281 #define SGEMM_DEFAULT_P 512
2282 #define DGEMM_DEFAULT_P 512
2283 #define CGEMM_DEFAULT_P 512
2284 #define ZGEMM_DEFAULT_P 512
2286 #define SGEMM_DEFAULT_Q 1024
2287 #define DGEMM_DEFAULT_Q 512
2288 #define CGEMM_DEFAULT_Q 512
2289 #define ZGEMM_DEFAULT_Q 256
2299 #define GEMM_DEFAULT_OFFSET_A 0
2300 #define GEMM_DEFAULT_OFFSET_B 0
2301 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2303 #define SGEMM_DEFAULT_UNROLL_M 2
2304 #define SGEMM_DEFAULT_UNROLL_N 8
2305 #define DGEMM_DEFAULT_UNROLL_M 2
2306 #define DGEMM_DEFAULT_UNROLL_N 8
2307 #define CGEMM_DEFAULT_UNROLL_M 1
2308 #define CGEMM_DEFAULT_UNROLL_N 4
2309 #define ZGEMM_DEFAULT_UNROLL_M 1
2310 #define ZGEMM_DEFAULT_UNROLL_N 4
2312 #define SGEMM_DEFAULT_P 108
2313 #define DGEMM_DEFAULT_P 112
2314 #define CGEMM_DEFAULT_P 108
2315 #define ZGEMM_DEFAULT_P 112
2317 #define SGEMM_DEFAULT_Q 288
2318 #define DGEMM_DEFAULT_Q 144
2319 #define CGEMM_DEFAULT_Q 144
2320 #define ZGEMM_DEFAULT_Q 72
2322 #define SGEMM_DEFAULT_R 2000
2323 #define DGEMM_DEFAULT_R 2000
2324 #define CGEMM_DEFAULT_R 2000
2325 #define ZGEMM_DEFAULT_R 2000
2331 ////Copy from SICORTEX
2335 #define GEMM_DEFAULT_OFFSET_A 0
2336 #define GEMM_DEFAULT_OFFSET_B 0
2337 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2339 #define SGEMM_DEFAULT_UNROLL_M 8
2340 #define SGEMM_DEFAULT_UNROLL_N 4
2342 #define DGEMM_DEFAULT_UNROLL_M 4
2343 #define DGEMM_DEFAULT_UNROLL_N 4
2345 #define CGEMM_DEFAULT_UNROLL_M 4
2346 #define CGEMM_DEFAULT_UNROLL_N 2
2348 #define ZGEMM_DEFAULT_UNROLL_M 2
2349 #define ZGEMM_DEFAULT_UNROLL_N 2
2351 #define SGEMM_DEFAULT_P 64
2352 #define DGEMM_DEFAULT_P 44
2353 #define CGEMM_DEFAULT_P 64
2354 #define ZGEMM_DEFAULT_P 32
2356 #define SGEMM_DEFAULT_Q 192
2357 #define DGEMM_DEFAULT_Q 92
2358 #define CGEMM_DEFAULT_Q 128
2359 #define ZGEMM_DEFAULT_Q 80
2361 #define SGEMM_DEFAULT_R 640
2362 #define DGEMM_DEFAULT_R dgemm_r
2363 #define CGEMM_DEFAULT_R 640
2364 #define ZGEMM_DEFAULT_R 640
2366 #define GEMM_OFFSET_A1 0x10000
2367 #define GEMM_OFFSET_B1 0x100000
2376 #define GEMM_DEFAULT_OFFSET_A 0
2377 #define GEMM_DEFAULT_OFFSET_B 0
2378 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2380 #define SGEMM_DEFAULT_UNROLL_M 2
2381 #define SGEMM_DEFAULT_UNROLL_N 2
2383 #define DGEMM_DEFAULT_UNROLL_M 2
2384 #define DGEMM_DEFAULT_UNROLL_N 2
2386 #define CGEMM_DEFAULT_UNROLL_M 2
2387 #define CGEMM_DEFAULT_UNROLL_N 2
2389 #define ZGEMM_DEFAULT_UNROLL_M 2
2390 #define ZGEMM_DEFAULT_UNROLL_N 2
2392 #define SGEMM_DEFAULT_P 64
2393 #define DGEMM_DEFAULT_P 24
2394 #define CGEMM_DEFAULT_P 24
2395 #define ZGEMM_DEFAULT_P 20
2397 #define SGEMM_DEFAULT_Q 192
2398 #define DGEMM_DEFAULT_Q 128
2399 #define CGEMM_DEFAULT_Q 128
2400 #define ZGEMM_DEFAULT_Q 64
2402 #define SGEMM_DEFAULT_R 512
2403 #define DGEMM_DEFAULT_R 512
2404 #define CGEMM_DEFAULT_R 512
2405 #define ZGEMM_DEFAULT_R 512
2407 #define GEMM_OFFSET_A1 0x10000
2408 #define GEMM_OFFSET_B1 0x100000
2413 #if defined(P5600) || defined(MIPS1004K) || defined(I6400) || defined(P6600) || defined(I6500)
2417 #define GEMM_DEFAULT_OFFSET_A 0
2418 #define GEMM_DEFAULT_OFFSET_B 0
2419 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2422 #define SGEMM_DEFAULT_UNROLL_M 8
2423 #define SGEMM_DEFAULT_UNROLL_N 8
2425 #define DGEMM_DEFAULT_UNROLL_M 8
2426 #define DGEMM_DEFAULT_UNROLL_N 4
2428 #define CGEMM_DEFAULT_UNROLL_M 8
2429 #define CGEMM_DEFAULT_UNROLL_N 4
2431 #define ZGEMM_DEFAULT_UNROLL_M 4
2432 #define ZGEMM_DEFAULT_UNROLL_N 4
2434 #define SGEMM_DEFAULT_UNROLL_M 2
2435 #define SGEMM_DEFAULT_UNROLL_N 2
2437 #define DGEMM_DEFAULT_UNROLL_M 2
2438 #define DGEMM_DEFAULT_UNROLL_N 2
2440 #define CGEMM_DEFAULT_UNROLL_M 2
2441 #define CGEMM_DEFAULT_UNROLL_N 2
2443 #define ZGEMM_DEFAULT_UNROLL_M 2
2444 #define ZGEMM_DEFAULT_UNROLL_N 2
2447 #define SGEMM_DEFAULT_P 128
2448 #define DGEMM_DEFAULT_P 128
2449 #define CGEMM_DEFAULT_P 96
2450 #define ZGEMM_DEFAULT_P 64
2452 #define SGEMM_DEFAULT_Q 240
2453 #define DGEMM_DEFAULT_Q 120
2454 #define CGEMM_DEFAULT_Q 120
2455 #define ZGEMM_DEFAULT_Q 120
2457 #define SGEMM_DEFAULT_R 12288
2458 #define DGEMM_DEFAULT_R 8192
2459 #define CGEMM_DEFAULT_R 4096
2460 #define ZGEMM_DEFAULT_R 4096
2469 #define GEMM_DEFAULT_OFFSET_A 0
2470 #define GEMM_DEFAULT_OFFSET_B 0
2471 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2473 #define SGEMM_DEFAULT_UNROLL_M 4
2474 #define SGEMM_DEFAULT_UNROLL_N 4
2476 #define DGEMM_DEFAULT_UNROLL_M 4
2477 #define DGEMM_DEFAULT_UNROLL_N 4
2479 #define CGEMM_DEFAULT_UNROLL_M 2
2480 #define CGEMM_DEFAULT_UNROLL_N 2
2482 #define ZGEMM_DEFAULT_UNROLL_M 2
2483 #define ZGEMM_DEFAULT_UNROLL_N 2
2485 #define SGEMM_DEFAULT_P 128
2486 #define DGEMM_DEFAULT_P 128
2487 #define CGEMM_DEFAULT_P 96
2488 #define ZGEMM_DEFAULT_P 64
2490 #define SGEMM_DEFAULT_Q 240
2491 #define DGEMM_DEFAULT_Q 120
2492 #define CGEMM_DEFAULT_Q 120
2493 #define ZGEMM_DEFAULT_Q 120
2495 #define SGEMM_DEFAULT_R 12288
2496 #define DGEMM_DEFAULT_R 8192
2497 #define CGEMM_DEFAULT_R 4096
2498 #define ZGEMM_DEFAULT_R 4096
2510 #define GEMM_DEFAULT_OFFSET_A 0
2511 #define GEMM_DEFAULT_OFFSET_B 0
2512 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2514 #define SGEMM_DEFAULT_UNROLL_M 4
2515 #define SGEMM_DEFAULT_UNROLL_N 2
2517 #define DGEMM_DEFAULT_UNROLL_M 4
2518 #define DGEMM_DEFAULT_UNROLL_N 2
2520 #define CGEMM_DEFAULT_UNROLL_M 2
2521 #define CGEMM_DEFAULT_UNROLL_N 2
2523 #define ZGEMM_DEFAULT_UNROLL_M 2
2524 #define ZGEMM_DEFAULT_UNROLL_N 2
2526 #define SGEMM_DEFAULT_P 128
2527 #define DGEMM_DEFAULT_P 128
2528 #define CGEMM_DEFAULT_P 96
2529 #define ZGEMM_DEFAULT_P 64
2531 #define SGEMM_DEFAULT_Q 240
2532 #define DGEMM_DEFAULT_Q 120
2533 #define CGEMM_DEFAULT_Q 120
2534 #define ZGEMM_DEFAULT_Q 120
2536 #define SGEMM_DEFAULT_R 12288
2537 #define DGEMM_DEFAULT_R 8192
2538 #define CGEMM_DEFAULT_R 4096
2539 #define ZGEMM_DEFAULT_R 4096
2546 #if defined(CORTEXA57)
2550 #define GEMM_DEFAULT_OFFSET_A 0
2551 #define GEMM_DEFAULT_OFFSET_B 0
2552 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2554 #define SGEMM_DEFAULT_UNROLL_M 16
2555 #define SGEMM_DEFAULT_UNROLL_N 4
2557 #define DGEMM_DEFAULT_UNROLL_M 8
2558 #define DGEMM_DEFAULT_UNROLL_N 4
2560 #define CGEMM_DEFAULT_UNROLL_M 8
2561 #define CGEMM_DEFAULT_UNROLL_N 4
2563 #define ZGEMM_DEFAULT_UNROLL_M 4
2564 #define ZGEMM_DEFAULT_UNROLL_N 4
2566 #define SGEMM_DEFAULT_P 512
2567 #define DGEMM_DEFAULT_P 256
2568 #define CGEMM_DEFAULT_P 256
2569 #define ZGEMM_DEFAULT_P 128
2571 #define SGEMM_DEFAULT_Q 1024
2572 #define DGEMM_DEFAULT_Q 512
2573 #define CGEMM_DEFAULT_Q 512
2574 #define ZGEMM_DEFAULT_Q 512
2576 #define SGEMM_DEFAULT_R 4096
2577 #define DGEMM_DEFAULT_R 4096
2578 #define CGEMM_DEFAULT_R 4096
2579 #define ZGEMM_DEFAULT_R 2048
2589 #define GEMM_DEFAULT_OFFSET_A 0
2590 #define GEMM_DEFAULT_OFFSET_B 0
2591 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2593 #define SGEMM_DEFAULT_UNROLL_M 4
2594 #define SGEMM_DEFAULT_UNROLL_N 4
2596 #define DGEMM_DEFAULT_UNROLL_M 2
2597 #define DGEMM_DEFAULT_UNROLL_N 2
2599 #define CGEMM_DEFAULT_UNROLL_M 2
2600 #define CGEMM_DEFAULT_UNROLL_N 2
2602 #define ZGEMM_DEFAULT_UNROLL_M 2
2603 #define ZGEMM_DEFAULT_UNROLL_N 2
2605 #define SGEMM_DEFAULT_P 128
2606 #define DGEMM_DEFAULT_P 128
2607 #define CGEMM_DEFAULT_P 96
2608 #define ZGEMM_DEFAULT_P 64
2610 #define SGEMM_DEFAULT_Q 240
2611 #define DGEMM_DEFAULT_Q 120
2612 #define CGEMM_DEFAULT_Q 120
2613 #define ZGEMM_DEFAULT_Q 120
2615 #define SGEMM_DEFAULT_R 12288
2616 #define DGEMM_DEFAULT_R 8192
2617 #define CGEMM_DEFAULT_R 4096
2618 #define ZGEMM_DEFAULT_R 4096
2624 #if defined(THUNDERX)
2628 #define GEMM_DEFAULT_OFFSET_A 0
2629 #define GEMM_DEFAULT_OFFSET_B 0
2630 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2632 #define SGEMM_DEFAULT_UNROLL_M 4
2633 #define SGEMM_DEFAULT_UNROLL_N 4
2635 #define DGEMM_DEFAULT_UNROLL_M 2
2636 #define DGEMM_DEFAULT_UNROLL_N 2
2638 #define CGEMM_DEFAULT_UNROLL_M 2
2639 #define CGEMM_DEFAULT_UNROLL_N 2
2641 #define ZGEMM_DEFAULT_UNROLL_M 2
2642 #define ZGEMM_DEFAULT_UNROLL_N 2
2644 #define SGEMM_DEFAULT_P 128
2645 #define DGEMM_DEFAULT_P 128
2646 #define CGEMM_DEFAULT_P 96
2647 #define ZGEMM_DEFAULT_P 64
2649 #define SGEMM_DEFAULT_Q 240
2650 #define DGEMM_DEFAULT_Q 120
2651 #define CGEMM_DEFAULT_Q 120
2652 #define ZGEMM_DEFAULT_Q 120
2654 #define SGEMM_DEFAULT_R 12288
2655 #define DGEMM_DEFAULT_R 8192
2656 #define CGEMM_DEFAULT_R 4096
2657 #define ZGEMM_DEFAULT_R 4096
2663 #if defined(THUNDERX2T99) || defined(VULCAN)
2667 #define GEMM_DEFAULT_OFFSET_A 0
2668 #define GEMM_DEFAULT_OFFSET_B 0
2669 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2671 #define SGEMM_DEFAULT_UNROLL_M 16
2672 #define SGEMM_DEFAULT_UNROLL_N 4
2674 #define DGEMM_DEFAULT_UNROLL_M 8
2675 #define DGEMM_DEFAULT_UNROLL_N 4
2677 #define CGEMM_DEFAULT_UNROLL_M 8
2678 #define CGEMM_DEFAULT_UNROLL_N 4
2680 #define ZGEMM_DEFAULT_UNROLL_M 4
2681 #define ZGEMM_DEFAULT_UNROLL_N 4
2683 #define SGEMM_DEFAULT_P sgemm_p
2684 #define DGEMM_DEFAULT_P dgemm_p
2685 #define CGEMM_DEFAULT_P cgemm_p
2686 #define ZGEMM_DEFAULT_P zgemm_p
2688 #define SGEMM_DEFAULT_Q sgemm_q
2689 #define DGEMM_DEFAULT_Q dgemm_q
2690 #define CGEMM_DEFAULT_Q cgemm_q
2691 #define ZGEMM_DEFAULT_Q zgemm_q
2693 #define SGEMM_DEFAULT_R sgemm_r
2694 #define DGEMM_DEFAULT_R dgemm_r
2695 #define CGEMM_DEFAULT_R cgemm_r
2696 #define ZGEMM_DEFAULT_R zgemm_r
2705 #define GEMM_DEFAULT_OFFSET_A 0
2706 #define GEMM_DEFAULT_OFFSET_B 0
2707 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2709 #define SGEMM_DEFAULT_UNROLL_M 2
2710 #define SGEMM_DEFAULT_UNROLL_N 2
2712 #define DGEMM_DEFAULT_UNROLL_M 2
2713 #define DGEMM_DEFAULT_UNROLL_N 2
2715 #define CGEMM_DEFAULT_UNROLL_M 2
2716 #define CGEMM_DEFAULT_UNROLL_N 2
2718 #define ZGEMM_DEFAULT_UNROLL_M 2
2719 #define ZGEMM_DEFAULT_UNROLL_N 2
2721 #define SGEMM_DEFAULT_P 128
2722 #define DGEMM_DEFAULT_P 128
2723 #define CGEMM_DEFAULT_P 96
2724 #define ZGEMM_DEFAULT_P 64
2726 #define SGEMM_DEFAULT_Q 240
2727 #define DGEMM_DEFAULT_Q 120
2728 #define CGEMM_DEFAULT_Q 120
2729 #define ZGEMM_DEFAULT_Q 120
2731 #define SGEMM_DEFAULT_R 12288
2732 #define DGEMM_DEFAULT_R 8192
2733 #define CGEMM_DEFAULT_R 4096
2734 #define ZGEMM_DEFAULT_R 4096
2746 #define GEMM_DEFAULT_OFFSET_A 0
2747 #define GEMM_DEFAULT_OFFSET_B 0
2748 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2750 #define SGEMM_DEFAULT_UNROLL_M 4
2751 #define SGEMM_DEFAULT_UNROLL_N 4
2753 #define DGEMM_DEFAULT_UNROLL_M 4
2754 #define DGEMM_DEFAULT_UNROLL_N 4
2756 #define CGEMM_DEFAULT_UNROLL_M 2
2757 #define CGEMM_DEFAULT_UNROLL_N 2
2759 #define ZGEMM_DEFAULT_UNROLL_M 2
2760 #define ZGEMM_DEFAULT_UNROLL_N 2
2762 #define SGEMM_DEFAULT_P 128
2763 #define DGEMM_DEFAULT_P 128
2764 #define CGEMM_DEFAULT_P 96
2765 #define ZGEMM_DEFAULT_P 64
2767 #define SGEMM_DEFAULT_Q 240
2768 #define DGEMM_DEFAULT_Q 120
2769 #define CGEMM_DEFAULT_Q 120
2770 #define ZGEMM_DEFAULT_Q 120
2772 #define SGEMM_DEFAULT_R 12288
2773 #define DGEMM_DEFAULT_R 8192
2774 #define CGEMM_DEFAULT_R 4096
2775 #define ZGEMM_DEFAULT_R 4096
2787 #define GEMM_DEFAULT_OFFSET_A 0
2788 #define GEMM_DEFAULT_OFFSET_B 0
2789 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2791 #define SGEMM_DEFAULT_UNROLL_M 4
2792 #define SGEMM_DEFAULT_UNROLL_N 4
2794 #define DGEMM_DEFAULT_UNROLL_M 4
2795 #define DGEMM_DEFAULT_UNROLL_N 4
2797 #define CGEMM_DEFAULT_UNROLL_M 2
2798 #define CGEMM_DEFAULT_UNROLL_N 2
2800 #define ZGEMM_DEFAULT_UNROLL_M 2
2801 #define ZGEMM_DEFAULT_UNROLL_N 2
2803 #define SGEMM_DEFAULT_P 128
2804 #define DGEMM_DEFAULT_P 128
2805 #define CGEMM_DEFAULT_P 96
2806 #define ZGEMM_DEFAULT_P 64
2808 #define SGEMM_DEFAULT_Q 240
2809 #define DGEMM_DEFAULT_Q 120
2810 #define CGEMM_DEFAULT_Q 120
2811 #define ZGEMM_DEFAULT_Q 120
2813 #define SGEMM_DEFAULT_R 12288
2814 #define DGEMM_DEFAULT_R 8192
2815 #define CGEMM_DEFAULT_R 4096
2816 #define ZGEMM_DEFAULT_R 4096
2824 #if defined(ZARCH_GENERIC)
2828 #define GEMM_DEFAULT_OFFSET_A 0
2829 #define GEMM_DEFAULT_OFFSET_B 0
2830 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2832 #define SGEMM_DEFAULT_UNROLL_M 2
2833 #define SGEMM_DEFAULT_UNROLL_N 2
2835 #define DGEMM_DEFAULT_UNROLL_M 2
2836 #define DGEMM_DEFAULT_UNROLL_N 2
2838 #define CGEMM_DEFAULT_UNROLL_M 2
2839 #define CGEMM_DEFAULT_UNROLL_N 2
2841 #define ZGEMM_DEFAULT_UNROLL_M 2
2842 #define ZGEMM_DEFAULT_UNROLL_N 2
2844 #define SGEMM_DEFAULT_P 128
2845 #define DGEMM_DEFAULT_P 128
2846 #define CGEMM_DEFAULT_P 96
2847 #define ZGEMM_DEFAULT_P 64
2849 #define SGEMM_DEFAULT_Q 240
2850 #define DGEMM_DEFAULT_Q 120
2851 #define CGEMM_DEFAULT_Q 120
2852 #define ZGEMM_DEFAULT_Q 120
2854 #define SGEMM_DEFAULT_R 12288
2855 #define DGEMM_DEFAULT_R 8192
2856 #define CGEMM_DEFAULT_R 4096
2857 #define ZGEMM_DEFAULT_R 4096
2867 #define GEMM_DEFAULT_OFFSET_A 0
2868 #define GEMM_DEFAULT_OFFSET_B 0
2869 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2871 #define SGEMM_DEFAULT_UNROLL_M 8
2872 #define SGEMM_DEFAULT_UNROLL_N 4
2874 #define DGEMM_DEFAULT_UNROLL_M 8
2875 #define DGEMM_DEFAULT_UNROLL_N 4
2877 #define CGEMM_DEFAULT_UNROLL_M 4
2878 #define CGEMM_DEFAULT_UNROLL_N 4
2880 #define ZGEMM_DEFAULT_UNROLL_M 4
2881 #define ZGEMM_DEFAULT_UNROLL_N 4
2883 #define SGEMM_DEFAULT_P 456
2884 #define DGEMM_DEFAULT_P 320
2885 #define CGEMM_DEFAULT_P 480
2886 #define ZGEMM_DEFAULT_P 224
2888 #define SGEMM_DEFAULT_Q 488
2889 #define DGEMM_DEFAULT_Q 384
2890 #define CGEMM_DEFAULT_Q 128
2891 #define ZGEMM_DEFAULT_Q 352
2893 #define SGEMM_DEFAULT_R 8192
2894 #define DGEMM_DEFAULT_R 4096
2895 #define CGEMM_DEFAULT_R 4096
2896 #define ZGEMM_DEFAULT_R 2048
2909 #define GEMM_DEFAULT_OFFSET_A 0
2910 #define GEMM_DEFAULT_OFFSET_B 0
2911 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2913 #define SGEMM_DEFAULT_UNROLL_N 2
2914 #define DGEMM_DEFAULT_UNROLL_N 2
2915 #define QGEMM_DEFAULT_UNROLL_N 2
2916 #define CGEMM_DEFAULT_UNROLL_N 2
2917 #define ZGEMM_DEFAULT_UNROLL_N 2
2918 #define XGEMM_DEFAULT_UNROLL_N 1
2921 #define SGEMM_DEFAULT_UNROLL_M 2
2922 #define DGEMM_DEFAULT_UNROLL_M 2
2923 #define QGEMM_DEFAULT_UNROLL_M 2
2924 #define CGEMM_DEFAULT_UNROLL_M 2
2925 #define ZGEMM_DEFAULT_UNROLL_M 2
2926 #define XGEMM_DEFAULT_UNROLL_M 1
2928 #define SGEMM_DEFAULT_UNROLL_M 2
2929 #define DGEMM_DEFAULT_UNROLL_M 2
2930 #define QGEMM_DEFAULT_UNROLL_M 2
2931 #define CGEMM_DEFAULT_UNROLL_M 2
2932 #define ZGEMM_DEFAULT_UNROLL_M 2
2933 #define XGEMM_DEFAULT_UNROLL_M 1
2936 #define SGEMM_DEFAULT_P sgemm_p
2937 #define DGEMM_DEFAULT_P dgemm_p
2938 #define QGEMM_DEFAULT_P qgemm_p
2939 #define CGEMM_DEFAULT_P cgemm_p
2940 #define ZGEMM_DEFAULT_P zgemm_p
2941 #define XGEMM_DEFAULT_P xgemm_p
2943 #define SGEMM_DEFAULT_R sgemm_r
2944 #define DGEMM_DEFAULT_R dgemm_r
2945 #define QGEMM_DEFAULT_R qgemm_r
2946 #define CGEMM_DEFAULT_R cgemm_r
2947 #define ZGEMM_DEFAULT_R zgemm_r
2948 #define XGEMM_DEFAULT_R xgemm_r
2950 #define SGEMM_DEFAULT_Q 128
2951 #define DGEMM_DEFAULT_Q 128
2952 #define QGEMM_DEFAULT_Q 128
2953 #define CGEMM_DEFAULT_Q 128
2954 #define ZGEMM_DEFAULT_Q 128
2955 #define XGEMM_DEFAULT_Q 128
2961 #ifndef QGEMM_DEFAULT_UNROLL_M
2962 #define QGEMM_DEFAULT_UNROLL_M 2
2965 #ifndef QGEMM_DEFAULT_UNROLL_N
2966 #define QGEMM_DEFAULT_UNROLL_N 2
2969 #ifndef XGEMM_DEFAULT_UNROLL_M
2970 #define XGEMM_DEFAULT_UNROLL_M 2
2973 #ifndef XGEMM_DEFAULT_UNROLL_N
2974 #define XGEMM_DEFAULT_UNROLL_N 2
2978 #define SHUFPD_0 shufps $0x44,
2979 #define SHUFPD_1 shufps $0x4e,
2980 #define SHUFPD_2 shufps $0xe4,
2981 #define SHUFPD_3 shufps $0xee,
2985 #define SHUFPD_0 shufpd $0,
2989 #define SHUFPD_1 shufpd $1,
2993 #define SHUFPD_2 shufpd $2,
2997 #define SHUFPD_3 shufpd $3,
3001 #define SHUFPS_39 shufps $0x39,