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 4
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
1625 #define GEMM_DEFAULT_OFFSET_A 64
1626 #define GEMM_DEFAULT_OFFSET_B 0
1627 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1632 #define SGEMM_DEFAULT_UNROLL_M 4
1633 #define DGEMM_DEFAULT_UNROLL_M 2
1634 #define QGEMM_DEFAULT_UNROLL_M 2
1635 #define CGEMM_DEFAULT_UNROLL_M 2
1636 #define ZGEMM_DEFAULT_UNROLL_M 1
1637 #define XGEMM_DEFAULT_UNROLL_M 1
1639 #define SGEMM_DEFAULT_UNROLL_M 8
1640 #define DGEMM_DEFAULT_UNROLL_M 4
1641 #define QGEMM_DEFAULT_UNROLL_M 2
1642 #define CGEMM_DEFAULT_UNROLL_M 4
1643 #define ZGEMM_DEFAULT_UNROLL_M 2
1644 #define XGEMM_DEFAULT_UNROLL_M 1
1647 #define SGEMM_DEFAULT_UNROLL_N 4
1648 #define DGEMM_DEFAULT_UNROLL_N 2
1649 #define QGEMM_DEFAULT_UNROLL_N 2
1650 #define CGEMM_DEFAULT_UNROLL_N 2
1651 #define ZGEMM_DEFAULT_UNROLL_N 1
1652 #define XGEMM_DEFAULT_UNROLL_N 1
1654 #define SGEMM_DEFAULT_P sgemm_p
1655 #define SGEMM_DEFAULT_R sgemm_r
1657 #define DGEMM_DEFAULT_P dgemm_p
1658 #define DGEMM_DEFAULT_R dgemm_r
1660 #define QGEMM_DEFAULT_P qgemm_p
1661 #define QGEMM_DEFAULT_R qgemm_r
1663 #define CGEMM_DEFAULT_P cgemm_p
1664 #define CGEMM_DEFAULT_R cgemm_r
1666 #define ZGEMM_DEFAULT_P zgemm_p
1667 #define ZGEMM_DEFAULT_R zgemm_r
1669 #define XGEMM_DEFAULT_P xgemm_p
1670 #define XGEMM_DEFAULT_R xgemm_r
1672 #define SGEMM_DEFAULT_Q 256
1673 #define DGEMM_DEFAULT_Q 256
1674 #define QGEMM_DEFAULT_Q 256
1675 #define CGEMM_DEFAULT_Q 256
1676 #define ZGEMM_DEFAULT_Q 256
1677 #define XGEMM_DEFAULT_Q 256
1687 #define GEMM_DEFAULT_OFFSET_A 0
1688 #define GEMM_DEFAULT_OFFSET_B 128
1689 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1691 #define SGEMM_DEFAULT_UNROLL_M 8
1692 #define SGEMM_DEFAULT_UNROLL_N 8
1693 #define DGEMM_DEFAULT_UNROLL_M 8
1694 #define DGEMM_DEFAULT_UNROLL_N 8
1695 #define QGEMM_DEFAULT_UNROLL_M 8
1696 #define QGEMM_DEFAULT_UNROLL_N 8
1697 #define CGEMM_DEFAULT_UNROLL_M 4
1698 #define CGEMM_DEFAULT_UNROLL_N 4
1699 #define ZGEMM_DEFAULT_UNROLL_M 4
1700 #define ZGEMM_DEFAULT_UNROLL_N 4
1701 #define XGEMM_DEFAULT_UNROLL_M 4
1702 #define XGEMM_DEFAULT_UNROLL_N 4
1704 #define SGEMM_DEFAULT_P sgemm_p
1705 #define DGEMM_DEFAULT_P dgemm_p
1706 #define QGEMM_DEFAULT_P qgemm_p
1707 #define CGEMM_DEFAULT_P cgemm_p
1708 #define ZGEMM_DEFAULT_P zgemm_p
1709 #define XGEMM_DEFAULT_P xgemm_p
1711 #define SGEMM_DEFAULT_Q 1024
1712 #define DGEMM_DEFAULT_Q 1024
1713 #define QGEMM_DEFAULT_Q 1024
1714 #define CGEMM_DEFAULT_Q 1024
1715 #define ZGEMM_DEFAULT_Q 1024
1716 #define XGEMM_DEFAULT_Q 1024
1718 #define SGEMM_DEFAULT_R sgemm_r
1719 #define DGEMM_DEFAULT_R dgemm_r
1720 #define QGEMM_DEFAULT_R qgemm_r
1721 #define CGEMM_DEFAULT_R cgemm_r
1722 #define ZGEMM_DEFAULT_R zgemm_r
1723 #define XGEMM_DEFAULT_R xgemm_r
1727 #define GETRF_FACTOR 0.65
1731 #if defined(EV4) || defined(EV5) || defined(EV6)
1741 #define GEMM_DEFAULT_OFFSET_A 512
1742 #define GEMM_DEFAULT_OFFSET_B 512
1743 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1745 #define SGEMM_DEFAULT_UNROLL_M 4
1746 #define SGEMM_DEFAULT_UNROLL_N 4
1747 #define DGEMM_DEFAULT_UNROLL_M 4
1748 #define DGEMM_DEFAULT_UNROLL_N 4
1749 #define CGEMM_DEFAULT_UNROLL_M 2
1750 #define CGEMM_DEFAULT_UNROLL_N 2
1751 #define ZGEMM_DEFAULT_UNROLL_M 2
1752 #define ZGEMM_DEFAULT_UNROLL_N 2
1757 #define SGEMM_DEFAULT_P 32
1758 #define SGEMM_DEFAULT_Q 112
1759 #define SGEMM_DEFAULT_R 256
1761 #define DGEMM_DEFAULT_P 32
1762 #define DGEMM_DEFAULT_Q 56
1763 #define DGEMM_DEFAULT_R 256
1765 #define CGEMM_DEFAULT_P 32
1766 #define CGEMM_DEFAULT_Q 64
1767 #define CGEMM_DEFAULT_R 240
1769 #define ZGEMM_DEFAULT_P 32
1770 #define ZGEMM_DEFAULT_Q 32
1771 #define ZGEMM_DEFAULT_R 240
1775 #define SGEMM_DEFAULT_P 64
1776 #define SGEMM_DEFAULT_Q 256
1778 #define DGEMM_DEFAULT_P 64
1779 #define DGEMM_DEFAULT_Q 128
1781 #define CGEMM_DEFAULT_P 64
1782 #define CGEMM_DEFAULT_Q 128
1784 #define ZGEMM_DEFAULT_P 64
1785 #define ZGEMM_DEFAULT_Q 64
1789 #define SGEMM_DEFAULT_P 256
1790 #define SGEMM_DEFAULT_Q 512
1792 #define DGEMM_DEFAULT_P 256
1793 #define DGEMM_DEFAULT_Q 256
1795 #define CGEMM_DEFAULT_P 256
1796 #define CGEMM_DEFAULT_Q 256
1798 #define ZGEMM_DEFAULT_P 128
1799 #define ZGEMM_DEFAULT_Q 256
1809 #define GEMM_DEFAULT_OFFSET_A 0
1810 #define GEMM_DEFAULT_OFFSET_B 8192
1811 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1813 #define SGEMM_DEFAULT_UNROLL_M 16
1814 #define SGEMM_DEFAULT_UNROLL_N 4
1815 #define DGEMM_DEFAULT_UNROLL_M 4
1816 #define DGEMM_DEFAULT_UNROLL_N 4
1817 #define CGEMM_DEFAULT_UNROLL_M 8
1818 #define CGEMM_DEFAULT_UNROLL_N 2
1819 #define ZGEMM_DEFAULT_UNROLL_M 2
1820 #define ZGEMM_DEFAULT_UNROLL_N 2
1822 #define SGEMM_DEFAULT_P 128
1823 #define DGEMM_DEFAULT_P 128
1824 #define CGEMM_DEFAULT_P 128
1825 #define ZGEMM_DEFAULT_P 128
1827 #define SGEMM_DEFAULT_Q 512
1828 #define DGEMM_DEFAULT_Q 256
1829 #define CGEMM_DEFAULT_Q 256
1830 #define ZGEMM_DEFAULT_Q 128
1836 #define GEMM_DEFAULT_OFFSET_A 0
1837 #define GEMM_DEFAULT_OFFSET_B 1024
1838 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1840 #define SGEMM_DEFAULT_UNROLL_M 16
1841 #define SGEMM_DEFAULT_UNROLL_N 4
1842 #define DGEMM_DEFAULT_UNROLL_M 4
1843 #define DGEMM_DEFAULT_UNROLL_N 4
1844 #define CGEMM_DEFAULT_UNROLL_M 8
1845 #define CGEMM_DEFAULT_UNROLL_N 2
1846 #define ZGEMM_DEFAULT_UNROLL_M 2
1847 #define ZGEMM_DEFAULT_UNROLL_N 2
1849 #define SGEMM_DEFAULT_P 256
1850 #define DGEMM_DEFAULT_P 128
1851 #define CGEMM_DEFAULT_P 128
1852 #define ZGEMM_DEFAULT_P 64
1854 #define SGEMM_DEFAULT_Q 256
1855 #define DGEMM_DEFAULT_Q 256
1856 #define CGEMM_DEFAULT_Q 256
1857 #define ZGEMM_DEFAULT_Q 256
1867 #define GEMM_DEFAULT_OFFSET_A 2688
1868 #define GEMM_DEFAULT_OFFSET_B 3072
1869 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1871 #define SGEMM_DEFAULT_UNROLL_M 16
1872 #define SGEMM_DEFAULT_UNROLL_N 4
1873 #define DGEMM_DEFAULT_UNROLL_M 4
1874 #define DGEMM_DEFAULT_UNROLL_N 4
1875 #define CGEMM_DEFAULT_UNROLL_M 8
1876 #define CGEMM_DEFAULT_UNROLL_N 2
1877 #define ZGEMM_DEFAULT_UNROLL_M 2
1878 #define ZGEMM_DEFAULT_UNROLL_N 2
1881 #if L2_SIZE == 1024976
1882 #define SGEMM_DEFAULT_P 320
1883 #define DGEMM_DEFAULT_P 256
1884 #define CGEMM_DEFAULT_P 256
1885 #define ZGEMM_DEFAULT_P 256
1887 #define SGEMM_DEFAULT_P 176
1888 #define DGEMM_DEFAULT_P 176
1889 #define CGEMM_DEFAULT_P 176
1890 #define ZGEMM_DEFAULT_P 176
1894 #define SGEMM_DEFAULT_Q 512
1895 #define DGEMM_DEFAULT_Q 256
1896 #define CGEMM_DEFAULT_Q 256
1897 #define ZGEMM_DEFAULT_Q 128
1908 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
1909 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
1910 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1912 #define SGEMM_DEFAULT_UNROLL_M 4
1913 #define SGEMM_DEFAULT_UNROLL_N 4
1914 #define DGEMM_DEFAULT_UNROLL_M 4
1915 #define DGEMM_DEFAULT_UNROLL_N 4
1916 #define CGEMM_DEFAULT_UNROLL_M 2
1917 #define CGEMM_DEFAULT_UNROLL_N 2
1918 #define ZGEMM_DEFAULT_UNROLL_M 2
1919 #define ZGEMM_DEFAULT_UNROLL_N 2
1921 #define SGEMM_DEFAULT_P 512
1922 #define DGEMM_DEFAULT_P 512
1923 #define CGEMM_DEFAULT_P 512
1924 #define ZGEMM_DEFAULT_P 512
1926 #define SGEMM_DEFAULT_Q 1024
1927 #define DGEMM_DEFAULT_Q 512
1928 #define CGEMM_DEFAULT_Q 512
1929 #define ZGEMM_DEFAULT_Q 256
1931 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
1932 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
1933 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
1934 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
1944 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
1945 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
1946 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1948 #define SGEMM_DEFAULT_UNROLL_M 8
1949 #define SGEMM_DEFAULT_UNROLL_N 4
1950 #define DGEMM_DEFAULT_UNROLL_M 8
1951 #define DGEMM_DEFAULT_UNROLL_N 4
1952 #define CGEMM_DEFAULT_UNROLL_M 4
1953 #define CGEMM_DEFAULT_UNROLL_N 2
1954 #define ZGEMM_DEFAULT_UNROLL_M 4
1955 #define ZGEMM_DEFAULT_UNROLL_N 2
1957 #define SGEMM_DEFAULT_P 128
1958 #define DGEMM_DEFAULT_P 128
1959 #define CGEMM_DEFAULT_P 128
1960 #define ZGEMM_DEFAULT_P 128
1962 #define SGEMM_DEFAULT_Q 4096
1963 #define DGEMM_DEFAULT_Q 3072
1964 #define CGEMM_DEFAULT_Q 2048
1965 #define ZGEMM_DEFAULT_Q 1024
1967 #define SGEMM_DEFAULT_Q 512
1968 #define DGEMM_DEFAULT_Q 256
1969 #define CGEMM_DEFAULT_Q 256
1970 #define ZGEMM_DEFAULT_Q 128
1978 #if defined(POWER3) || defined(POWER4) || defined(POWER5)
1979 #define GEMM_DEFAULT_OFFSET_A 0
1980 #define GEMM_DEFAULT_OFFSET_B 2048
1981 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
1983 #define SGEMM_DEFAULT_UNROLL_M 4
1984 #define SGEMM_DEFAULT_UNROLL_N 4
1985 #define DGEMM_DEFAULT_UNROLL_M 4
1986 #define DGEMM_DEFAULT_UNROLL_N 4
1987 #define CGEMM_DEFAULT_UNROLL_M 2
1988 #define CGEMM_DEFAULT_UNROLL_N 2
1989 #define ZGEMM_DEFAULT_UNROLL_M 2
1990 #define ZGEMM_DEFAULT_UNROLL_N 2
1997 #define SGEMM_DEFAULT_P 256
1998 #define SGEMM_DEFAULT_Q 432
1999 #define SGEMM_DEFAULT_R 1012
2001 #define DGEMM_DEFAULT_P 256
2002 #define DGEMM_DEFAULT_Q 216
2003 #define DGEMM_DEFAULT_R 1012
2005 #define ZGEMM_DEFAULT_P 256
2006 #define ZGEMM_DEFAULT_Q 104
2007 #define ZGEMM_DEFAULT_R 1012
2011 #ifdef ALLOC_HUGETLB
2012 #define SGEMM_DEFAULT_P 184
2013 #define DGEMM_DEFAULT_P 184
2014 #define CGEMM_DEFAULT_P 184
2015 #define ZGEMM_DEFAULT_P 184
2017 #define SGEMM_DEFAULT_P 144
2018 #define DGEMM_DEFAULT_P 144
2019 #define CGEMM_DEFAULT_P 144
2020 #define ZGEMM_DEFAULT_P 144
2025 #ifdef ALLOC_HUGETLB
2026 #define SGEMM_DEFAULT_P 512
2027 #define DGEMM_DEFAULT_P 256
2028 #define CGEMM_DEFAULT_P 256
2029 #define ZGEMM_DEFAULT_P 128
2031 #define SGEMM_DEFAULT_P 320
2032 #define DGEMM_DEFAULT_P 160
2033 #define CGEMM_DEFAULT_P 160
2034 #define ZGEMM_DEFAULT_P 80
2037 #define SGEMM_DEFAULT_Q 256
2038 #define CGEMM_DEFAULT_Q 256
2039 #define DGEMM_DEFAULT_Q 256
2040 #define ZGEMM_DEFAULT_Q 256
2052 #define GEMM_DEFAULT_OFFSET_A 384
2053 #define GEMM_DEFAULT_OFFSET_B 1024
2054 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2056 #define SGEMM_DEFAULT_UNROLL_M 4
2057 #define SGEMM_DEFAULT_UNROLL_N 4
2058 #define DGEMM_DEFAULT_UNROLL_M 4
2059 #define DGEMM_DEFAULT_UNROLL_N 4
2060 #define CGEMM_DEFAULT_UNROLL_M 2
2061 #define CGEMM_DEFAULT_UNROLL_N 4
2062 #define ZGEMM_DEFAULT_UNROLL_M 2
2063 #define ZGEMM_DEFAULT_UNROLL_N 4
2065 #define SGEMM_DEFAULT_P 992
2066 #define DGEMM_DEFAULT_P 480
2067 #define CGEMM_DEFAULT_P 488
2068 #define ZGEMM_DEFAULT_P 248
2070 #define SGEMM_DEFAULT_Q 504
2071 #define DGEMM_DEFAULT_Q 504
2072 #define CGEMM_DEFAULT_Q 400
2073 #define ZGEMM_DEFAULT_Q 400
2084 #define GEMM_DEFAULT_OFFSET_A 0
2085 #define GEMM_DEFAULT_OFFSET_B 65536
2086 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2088 #define SGEMM_DEFAULT_UNROLL_M 16
2089 #define SGEMM_DEFAULT_UNROLL_N 8
2090 #define DGEMM_DEFAULT_UNROLL_M 16
2091 #define DGEMM_DEFAULT_UNROLL_N 4
2092 #define CGEMM_DEFAULT_UNROLL_M 8
2093 #define CGEMM_DEFAULT_UNROLL_N 4
2094 #define ZGEMM_DEFAULT_UNROLL_M 8
2095 #define ZGEMM_DEFAULT_UNROLL_N 2
2097 #define SGEMM_DEFAULT_P 1280
2098 #define DGEMM_DEFAULT_P 640
2099 #define CGEMM_DEFAULT_P 640
2100 #define ZGEMM_DEFAULT_P 320
2102 #define SGEMM_DEFAULT_Q 640
2103 #define DGEMM_DEFAULT_Q 720
2104 #define CGEMM_DEFAULT_Q 640
2105 #define ZGEMM_DEFAULT_Q 640
2112 #if defined(SPARC) && defined(V7)
2117 #define GEMM_DEFAULT_OFFSET_A 0
2118 #define GEMM_DEFAULT_OFFSET_B 2048
2119 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2121 #define SGEMM_DEFAULT_UNROLL_M 2
2122 #define SGEMM_DEFAULT_UNROLL_N 8
2123 #define DGEMM_DEFAULT_UNROLL_M 2
2124 #define DGEMM_DEFAULT_UNROLL_N 8
2125 #define CGEMM_DEFAULT_UNROLL_M 1
2126 #define CGEMM_DEFAULT_UNROLL_N 4
2127 #define ZGEMM_DEFAULT_UNROLL_M 1
2128 #define ZGEMM_DEFAULT_UNROLL_N 4
2130 #define SGEMM_DEFAULT_P 256
2131 #define DGEMM_DEFAULT_P 256
2132 #define CGEMM_DEFAULT_P 256
2133 #define ZGEMM_DEFAULT_P 256
2135 #define SGEMM_DEFAULT_Q 512
2136 #define DGEMM_DEFAULT_Q 256
2137 #define CGEMM_DEFAULT_Q 256
2138 #define ZGEMM_DEFAULT_Q 128
2141 #define GEMM_THREAD gemm_thread_mn
2144 #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
2149 #define GEMM_DEFAULT_OFFSET_A 0
2150 #define GEMM_DEFAULT_OFFSET_B 2048
2151 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2153 #define SGEMM_DEFAULT_UNROLL_M 4
2154 #define SGEMM_DEFAULT_UNROLL_N 4
2155 #define DGEMM_DEFAULT_UNROLL_M 4
2156 #define DGEMM_DEFAULT_UNROLL_N 4
2157 #define CGEMM_DEFAULT_UNROLL_M 2
2158 #define CGEMM_DEFAULT_UNROLL_N 2
2159 #define ZGEMM_DEFAULT_UNROLL_M 2
2160 #define ZGEMM_DEFAULT_UNROLL_N 2
2162 #define SGEMM_DEFAULT_P 512
2163 #define DGEMM_DEFAULT_P 512
2164 #define CGEMM_DEFAULT_P 512
2165 #define ZGEMM_DEFAULT_P 512
2167 #define SGEMM_DEFAULT_Q 1024
2168 #define DGEMM_DEFAULT_Q 512
2169 #define CGEMM_DEFAULT_Q 512
2170 #define ZGEMM_DEFAULT_Q 256
2180 #define GEMM_DEFAULT_OFFSET_A 0
2181 #define GEMM_DEFAULT_OFFSET_B 0
2182 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2184 #define SGEMM_DEFAULT_UNROLL_M 2
2185 #define SGEMM_DEFAULT_UNROLL_N 8
2186 #define DGEMM_DEFAULT_UNROLL_M 2
2187 #define DGEMM_DEFAULT_UNROLL_N 8
2188 #define CGEMM_DEFAULT_UNROLL_M 1
2189 #define CGEMM_DEFAULT_UNROLL_N 4
2190 #define ZGEMM_DEFAULT_UNROLL_M 1
2191 #define ZGEMM_DEFAULT_UNROLL_N 4
2193 #define SGEMM_DEFAULT_P 108
2194 #define DGEMM_DEFAULT_P 112
2195 #define CGEMM_DEFAULT_P 108
2196 #define ZGEMM_DEFAULT_P 112
2198 #define SGEMM_DEFAULT_Q 288
2199 #define DGEMM_DEFAULT_Q 144
2200 #define CGEMM_DEFAULT_Q 144
2201 #define ZGEMM_DEFAULT_Q 72
2203 #define SGEMM_DEFAULT_R 2000
2204 #define DGEMM_DEFAULT_R 2000
2205 #define CGEMM_DEFAULT_R 2000
2206 #define ZGEMM_DEFAULT_R 2000
2212 ////Copy from SICORTEX
2216 #define GEMM_DEFAULT_OFFSET_A 0
2217 #define GEMM_DEFAULT_OFFSET_B 0
2218 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2220 #define SGEMM_DEFAULT_UNROLL_M 8
2221 #define SGEMM_DEFAULT_UNROLL_N 4
2223 #define DGEMM_DEFAULT_UNROLL_M 4
2224 #define DGEMM_DEFAULT_UNROLL_N 4
2226 #define CGEMM_DEFAULT_UNROLL_M 4
2227 #define CGEMM_DEFAULT_UNROLL_N 2
2229 #define ZGEMM_DEFAULT_UNROLL_M 2
2230 #define ZGEMM_DEFAULT_UNROLL_N 2
2232 #define SGEMM_DEFAULT_P 64
2233 #define DGEMM_DEFAULT_P 44
2234 #define CGEMM_DEFAULT_P 64
2235 #define ZGEMM_DEFAULT_P 32
2237 #define SGEMM_DEFAULT_Q 192
2238 #define DGEMM_DEFAULT_Q 92
2239 #define CGEMM_DEFAULT_Q 128
2240 #define ZGEMM_DEFAULT_Q 80
2242 #define SGEMM_DEFAULT_R 640
2243 #define DGEMM_DEFAULT_R dgemm_r
2244 #define CGEMM_DEFAULT_R 640
2245 #define ZGEMM_DEFAULT_R 640
2247 #define GEMM_OFFSET_A1 0x10000
2248 #define GEMM_OFFSET_B1 0x100000
2257 #define GEMM_DEFAULT_OFFSET_A 0
2258 #define GEMM_DEFAULT_OFFSET_B 0
2259 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2261 #define SGEMM_DEFAULT_UNROLL_M 2
2262 #define SGEMM_DEFAULT_UNROLL_N 2
2264 #define DGEMM_DEFAULT_UNROLL_M 2
2265 #define DGEMM_DEFAULT_UNROLL_N 2
2267 #define CGEMM_DEFAULT_UNROLL_M 2
2268 #define CGEMM_DEFAULT_UNROLL_N 2
2270 #define ZGEMM_DEFAULT_UNROLL_M 2
2271 #define ZGEMM_DEFAULT_UNROLL_N 2
2273 #define SGEMM_DEFAULT_P 64
2274 #define DGEMM_DEFAULT_P 24
2275 #define CGEMM_DEFAULT_P 24
2276 #define ZGEMM_DEFAULT_P 20
2278 #define SGEMM_DEFAULT_Q 192
2279 #define DGEMM_DEFAULT_Q 128
2280 #define CGEMM_DEFAULT_Q 128
2281 #define ZGEMM_DEFAULT_Q 64
2283 #define SGEMM_DEFAULT_R 512
2284 #define DGEMM_DEFAULT_R 512
2285 #define CGEMM_DEFAULT_R 512
2286 #define ZGEMM_DEFAULT_R 512
2288 #define GEMM_OFFSET_A1 0x10000
2289 #define GEMM_OFFSET_B1 0x100000
2294 #if defined(P5600) || defined(I6400) || defined(P6600)
2298 #define GEMM_DEFAULT_OFFSET_A 0
2299 #define GEMM_DEFAULT_OFFSET_B 0
2300 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2303 #define SGEMM_DEFAULT_UNROLL_M 8
2304 #define SGEMM_DEFAULT_UNROLL_N 8
2306 #define DGEMM_DEFAULT_UNROLL_M 8
2307 #define DGEMM_DEFAULT_UNROLL_N 4
2309 #define CGEMM_DEFAULT_UNROLL_M 8
2310 #define CGEMM_DEFAULT_UNROLL_N 4
2312 #define ZGEMM_DEFAULT_UNROLL_M 4
2313 #define ZGEMM_DEFAULT_UNROLL_N 4
2315 #define SGEMM_DEFAULT_UNROLL_M 2
2316 #define SGEMM_DEFAULT_UNROLL_N 2
2318 #define DGEMM_DEFAULT_UNROLL_M 2
2319 #define DGEMM_DEFAULT_UNROLL_N 2
2321 #define CGEMM_DEFAULT_UNROLL_M 2
2322 #define CGEMM_DEFAULT_UNROLL_N 2
2324 #define ZGEMM_DEFAULT_UNROLL_M 2
2325 #define ZGEMM_DEFAULT_UNROLL_N 2
2328 #define SGEMM_DEFAULT_P 128
2329 #define DGEMM_DEFAULT_P 128
2330 #define CGEMM_DEFAULT_P 96
2331 #define ZGEMM_DEFAULT_P 64
2333 #define SGEMM_DEFAULT_Q 240
2334 #define DGEMM_DEFAULT_Q 120
2335 #define CGEMM_DEFAULT_Q 120
2336 #define ZGEMM_DEFAULT_Q 120
2338 #define SGEMM_DEFAULT_R 12288
2339 #define DGEMM_DEFAULT_R 8192
2340 #define CGEMM_DEFAULT_R 4096
2341 #define ZGEMM_DEFAULT_R 4096
2350 #define GEMM_DEFAULT_OFFSET_A 0
2351 #define GEMM_DEFAULT_OFFSET_B 0
2352 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2354 #define SGEMM_DEFAULT_UNROLL_M 4
2355 #define SGEMM_DEFAULT_UNROLL_N 4
2357 #define DGEMM_DEFAULT_UNROLL_M 4
2358 #define DGEMM_DEFAULT_UNROLL_N 4
2360 #define CGEMM_DEFAULT_UNROLL_M 2
2361 #define CGEMM_DEFAULT_UNROLL_N 2
2363 #define ZGEMM_DEFAULT_UNROLL_M 2
2364 #define ZGEMM_DEFAULT_UNROLL_N 2
2366 #define SGEMM_DEFAULT_P 128
2367 #define DGEMM_DEFAULT_P 128
2368 #define CGEMM_DEFAULT_P 96
2369 #define ZGEMM_DEFAULT_P 64
2371 #define SGEMM_DEFAULT_Q 240
2372 #define DGEMM_DEFAULT_Q 120
2373 #define CGEMM_DEFAULT_Q 120
2374 #define ZGEMM_DEFAULT_Q 120
2376 #define SGEMM_DEFAULT_R 12288
2377 #define DGEMM_DEFAULT_R 8192
2378 #define CGEMM_DEFAULT_R 4096
2379 #define ZGEMM_DEFAULT_R 4096
2391 #define GEMM_DEFAULT_OFFSET_A 0
2392 #define GEMM_DEFAULT_OFFSET_B 0
2393 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2395 #define SGEMM_DEFAULT_UNROLL_M 4
2396 #define SGEMM_DEFAULT_UNROLL_N 2
2398 #define DGEMM_DEFAULT_UNROLL_M 4
2399 #define DGEMM_DEFAULT_UNROLL_N 2
2401 #define CGEMM_DEFAULT_UNROLL_M 2
2402 #define CGEMM_DEFAULT_UNROLL_N 2
2404 #define ZGEMM_DEFAULT_UNROLL_M 2
2405 #define ZGEMM_DEFAULT_UNROLL_N 2
2407 #define SGEMM_DEFAULT_P 128
2408 #define DGEMM_DEFAULT_P 128
2409 #define CGEMM_DEFAULT_P 96
2410 #define ZGEMM_DEFAULT_P 64
2412 #define SGEMM_DEFAULT_Q 240
2413 #define DGEMM_DEFAULT_Q 120
2414 #define CGEMM_DEFAULT_Q 120
2415 #define ZGEMM_DEFAULT_Q 120
2417 #define SGEMM_DEFAULT_R 12288
2418 #define DGEMM_DEFAULT_R 8192
2419 #define CGEMM_DEFAULT_R 4096
2420 #define ZGEMM_DEFAULT_R 4096
2427 #if defined(CORTEXA57)
2431 #define GEMM_DEFAULT_OFFSET_A 0
2432 #define GEMM_DEFAULT_OFFSET_B 0
2433 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2435 #define SGEMM_DEFAULT_UNROLL_M 16
2436 #define SGEMM_DEFAULT_UNROLL_N 4
2438 #define DGEMM_DEFAULT_UNROLL_M 8
2439 #define DGEMM_DEFAULT_UNROLL_N 4
2441 #define CGEMM_DEFAULT_UNROLL_M 8
2442 #define CGEMM_DEFAULT_UNROLL_N 4
2444 #define ZGEMM_DEFAULT_UNROLL_M 4
2445 #define ZGEMM_DEFAULT_UNROLL_N 4
2447 #define SGEMM_DEFAULT_P 512
2448 #define DGEMM_DEFAULT_P 256
2449 #define CGEMM_DEFAULT_P 256
2450 #define ZGEMM_DEFAULT_P 128
2452 #define SGEMM_DEFAULT_Q 1024
2453 #define DGEMM_DEFAULT_Q 512
2454 #define CGEMM_DEFAULT_Q 512
2455 #define ZGEMM_DEFAULT_Q 512
2457 #define SGEMM_DEFAULT_R 4096
2458 #define DGEMM_DEFAULT_R 4096
2459 #define CGEMM_DEFAULT_R 4096
2460 #define ZGEMM_DEFAULT_R 2048
2470 #define GEMM_DEFAULT_OFFSET_A 0
2471 #define GEMM_DEFAULT_OFFSET_B 0
2472 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2474 #define SGEMM_DEFAULT_UNROLL_M 4
2475 #define SGEMM_DEFAULT_UNROLL_N 4
2477 #define DGEMM_DEFAULT_UNROLL_M 2
2478 #define DGEMM_DEFAULT_UNROLL_N 2
2480 #define CGEMM_DEFAULT_UNROLL_M 2
2481 #define CGEMM_DEFAULT_UNROLL_N 2
2483 #define ZGEMM_DEFAULT_UNROLL_M 2
2484 #define ZGEMM_DEFAULT_UNROLL_N 2
2486 #define SGEMM_DEFAULT_P 128
2487 #define DGEMM_DEFAULT_P 128
2488 #define CGEMM_DEFAULT_P 96
2489 #define ZGEMM_DEFAULT_P 64
2491 #define SGEMM_DEFAULT_Q 240
2492 #define DGEMM_DEFAULT_Q 120
2493 #define CGEMM_DEFAULT_Q 120
2494 #define ZGEMM_DEFAULT_Q 120
2496 #define SGEMM_DEFAULT_R 12288
2497 #define DGEMM_DEFAULT_R 8192
2498 #define CGEMM_DEFAULT_R 4096
2499 #define ZGEMM_DEFAULT_R 4096
2505 #if defined(THUNDERX)
2509 #define GEMM_DEFAULT_OFFSET_A 0
2510 #define GEMM_DEFAULT_OFFSET_B 0
2511 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2513 #define SGEMM_DEFAULT_UNROLL_M 4
2514 #define SGEMM_DEFAULT_UNROLL_N 4
2516 #define DGEMM_DEFAULT_UNROLL_M 2
2517 #define DGEMM_DEFAULT_UNROLL_N 2
2519 #define CGEMM_DEFAULT_UNROLL_M 2
2520 #define CGEMM_DEFAULT_UNROLL_N 2
2522 #define ZGEMM_DEFAULT_UNROLL_M 2
2523 #define ZGEMM_DEFAULT_UNROLL_N 2
2525 #define SGEMM_DEFAULT_P 128
2526 #define DGEMM_DEFAULT_P 128
2527 #define CGEMM_DEFAULT_P 96
2528 #define ZGEMM_DEFAULT_P 64
2530 #define SGEMM_DEFAULT_Q 240
2531 #define DGEMM_DEFAULT_Q 120
2532 #define CGEMM_DEFAULT_Q 120
2533 #define ZGEMM_DEFAULT_Q 120
2535 #define SGEMM_DEFAULT_R 12288
2536 #define DGEMM_DEFAULT_R 8192
2537 #define CGEMM_DEFAULT_R 4096
2538 #define ZGEMM_DEFAULT_R 4096
2544 #if defined(THUNDERX2T99) || defined(VULCAN)
2548 #define GEMM_DEFAULT_OFFSET_A 0
2549 #define GEMM_DEFAULT_OFFSET_B 0
2550 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2552 #define SGEMM_DEFAULT_UNROLL_M 16
2553 #define SGEMM_DEFAULT_UNROLL_N 4
2555 #define DGEMM_DEFAULT_UNROLL_M 8
2556 #define DGEMM_DEFAULT_UNROLL_N 4
2558 #define CGEMM_DEFAULT_UNROLL_M 8
2559 #define CGEMM_DEFAULT_UNROLL_N 4
2561 #define ZGEMM_DEFAULT_UNROLL_M 4
2562 #define ZGEMM_DEFAULT_UNROLL_N 4
2564 #define SGEMM_DEFAULT_P sgemm_p
2565 #define DGEMM_DEFAULT_P dgemm_p
2566 #define CGEMM_DEFAULT_P cgemm_p
2567 #define ZGEMM_DEFAULT_P zgemm_p
2569 #define SGEMM_DEFAULT_Q sgemm_q
2570 #define DGEMM_DEFAULT_Q dgemm_q
2571 #define CGEMM_DEFAULT_Q cgemm_q
2572 #define ZGEMM_DEFAULT_Q zgemm_q
2574 #define SGEMM_DEFAULT_R sgemm_r
2575 #define DGEMM_DEFAULT_R dgemm_r
2576 #define CGEMM_DEFAULT_R cgemm_r
2577 #define ZGEMM_DEFAULT_R zgemm_r
2586 #define GEMM_DEFAULT_OFFSET_A 0
2587 #define GEMM_DEFAULT_OFFSET_B 0
2588 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2590 #define SGEMM_DEFAULT_UNROLL_M 2
2591 #define SGEMM_DEFAULT_UNROLL_N 2
2593 #define DGEMM_DEFAULT_UNROLL_M 2
2594 #define DGEMM_DEFAULT_UNROLL_N 2
2596 #define CGEMM_DEFAULT_UNROLL_M 2
2597 #define CGEMM_DEFAULT_UNROLL_N 2
2599 #define ZGEMM_DEFAULT_UNROLL_M 2
2600 #define ZGEMM_DEFAULT_UNROLL_N 2
2602 #define SGEMM_DEFAULT_P 128
2603 #define DGEMM_DEFAULT_P 128
2604 #define CGEMM_DEFAULT_P 96
2605 #define ZGEMM_DEFAULT_P 64
2607 #define SGEMM_DEFAULT_Q 240
2608 #define DGEMM_DEFAULT_Q 120
2609 #define CGEMM_DEFAULT_Q 120
2610 #define ZGEMM_DEFAULT_Q 120
2612 #define SGEMM_DEFAULT_R 12288
2613 #define DGEMM_DEFAULT_R 8192
2614 #define CGEMM_DEFAULT_R 4096
2615 #define ZGEMM_DEFAULT_R 4096
2627 #define GEMM_DEFAULT_OFFSET_A 0
2628 #define GEMM_DEFAULT_OFFSET_B 0
2629 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2631 #define SGEMM_DEFAULT_UNROLL_M 4
2632 #define SGEMM_DEFAULT_UNROLL_N 4
2634 #define DGEMM_DEFAULT_UNROLL_M 4
2635 #define DGEMM_DEFAULT_UNROLL_N 4
2637 #define CGEMM_DEFAULT_UNROLL_M 2
2638 #define CGEMM_DEFAULT_UNROLL_N 2
2640 #define ZGEMM_DEFAULT_UNROLL_M 2
2641 #define ZGEMM_DEFAULT_UNROLL_N 2
2643 #define SGEMM_DEFAULT_P 128
2644 #define DGEMM_DEFAULT_P 128
2645 #define CGEMM_DEFAULT_P 96
2646 #define ZGEMM_DEFAULT_P 64
2648 #define SGEMM_DEFAULT_Q 240
2649 #define DGEMM_DEFAULT_Q 120
2650 #define CGEMM_DEFAULT_Q 120
2651 #define ZGEMM_DEFAULT_Q 120
2653 #define SGEMM_DEFAULT_R 12288
2654 #define DGEMM_DEFAULT_R 8192
2655 #define CGEMM_DEFAULT_R 4096
2656 #define ZGEMM_DEFAULT_R 4096
2668 #define GEMM_DEFAULT_OFFSET_A 0
2669 #define GEMM_DEFAULT_OFFSET_B 0
2670 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2672 #define SGEMM_DEFAULT_UNROLL_M 4
2673 #define SGEMM_DEFAULT_UNROLL_N 4
2675 #define DGEMM_DEFAULT_UNROLL_M 4
2676 #define DGEMM_DEFAULT_UNROLL_N 4
2678 #define CGEMM_DEFAULT_UNROLL_M 2
2679 #define CGEMM_DEFAULT_UNROLL_N 2
2681 #define ZGEMM_DEFAULT_UNROLL_M 2
2682 #define ZGEMM_DEFAULT_UNROLL_N 2
2684 #define SGEMM_DEFAULT_P 128
2685 #define DGEMM_DEFAULT_P 128
2686 #define CGEMM_DEFAULT_P 96
2687 #define ZGEMM_DEFAULT_P 64
2689 #define SGEMM_DEFAULT_Q 240
2690 #define DGEMM_DEFAULT_Q 120
2691 #define CGEMM_DEFAULT_Q 120
2692 #define ZGEMM_DEFAULT_Q 120
2694 #define SGEMM_DEFAULT_R 12288
2695 #define DGEMM_DEFAULT_R 8192
2696 #define CGEMM_DEFAULT_R 4096
2697 #define ZGEMM_DEFAULT_R 4096
2705 #if defined(ZARCH_GENERIC)
2709 #define GEMM_DEFAULT_OFFSET_A 0
2710 #define GEMM_DEFAULT_OFFSET_B 0
2711 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2713 #define SGEMM_DEFAULT_UNROLL_M 2
2714 #define SGEMM_DEFAULT_UNROLL_N 2
2716 #define DGEMM_DEFAULT_UNROLL_M 2
2717 #define DGEMM_DEFAULT_UNROLL_N 2
2719 #define CGEMM_DEFAULT_UNROLL_M 2
2720 #define CGEMM_DEFAULT_UNROLL_N 2
2722 #define ZGEMM_DEFAULT_UNROLL_M 2
2723 #define ZGEMM_DEFAULT_UNROLL_N 2
2725 #define SGEMM_DEFAULT_P 128
2726 #define DGEMM_DEFAULT_P 128
2727 #define CGEMM_DEFAULT_P 96
2728 #define ZGEMM_DEFAULT_P 64
2730 #define SGEMM_DEFAULT_Q 240
2731 #define DGEMM_DEFAULT_Q 120
2732 #define CGEMM_DEFAULT_Q 120
2733 #define ZGEMM_DEFAULT_Q 120
2735 #define SGEMM_DEFAULT_R 12288
2736 #define DGEMM_DEFAULT_R 8192
2737 #define CGEMM_DEFAULT_R 4096
2738 #define ZGEMM_DEFAULT_R 4096
2748 #define GEMM_DEFAULT_OFFSET_A 0
2749 #define GEMM_DEFAULT_OFFSET_B 0
2750 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2752 #define SGEMM_DEFAULT_UNROLL_M 8
2753 #define SGEMM_DEFAULT_UNROLL_N 4
2755 #define DGEMM_DEFAULT_UNROLL_M 8
2756 #define DGEMM_DEFAULT_UNROLL_N 4
2758 #define CGEMM_DEFAULT_UNROLL_M 4
2759 #define CGEMM_DEFAULT_UNROLL_N 4
2761 #define ZGEMM_DEFAULT_UNROLL_M 4
2762 #define ZGEMM_DEFAULT_UNROLL_N 4
2764 #define SGEMM_DEFAULT_P 456
2765 #define DGEMM_DEFAULT_P 320
2766 #define CGEMM_DEFAULT_P 480
2767 #define ZGEMM_DEFAULT_P 224
2769 #define SGEMM_DEFAULT_Q 488
2770 #define DGEMM_DEFAULT_Q 384
2771 #define CGEMM_DEFAULT_Q 128
2772 #define ZGEMM_DEFAULT_Q 352
2774 #define SGEMM_DEFAULT_R 8192
2775 #define DGEMM_DEFAULT_R 4096
2776 #define CGEMM_DEFAULT_R 4096
2777 #define ZGEMM_DEFAULT_R 2048
2790 #define GEMM_DEFAULT_OFFSET_A 0
2791 #define GEMM_DEFAULT_OFFSET_B 0
2792 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2794 #define SGEMM_DEFAULT_UNROLL_N 2
2795 #define DGEMM_DEFAULT_UNROLL_N 2
2796 #define QGEMM_DEFAULT_UNROLL_N 2
2797 #define CGEMM_DEFAULT_UNROLL_N 2
2798 #define ZGEMM_DEFAULT_UNROLL_N 2
2799 #define XGEMM_DEFAULT_UNROLL_N 1
2802 #define SGEMM_DEFAULT_UNROLL_M 2
2803 #define DGEMM_DEFAULT_UNROLL_M 2
2804 #define QGEMM_DEFAULT_UNROLL_M 2
2805 #define CGEMM_DEFAULT_UNROLL_M 2
2806 #define ZGEMM_DEFAULT_UNROLL_M 2
2807 #define XGEMM_DEFAULT_UNROLL_M 1
2809 #define SGEMM_DEFAULT_UNROLL_M 2
2810 #define DGEMM_DEFAULT_UNROLL_M 2
2811 #define QGEMM_DEFAULT_UNROLL_M 2
2812 #define CGEMM_DEFAULT_UNROLL_M 2
2813 #define ZGEMM_DEFAULT_UNROLL_M 2
2814 #define XGEMM_DEFAULT_UNROLL_M 1
2817 #define SGEMM_DEFAULT_P sgemm_p
2818 #define DGEMM_DEFAULT_P dgemm_p
2819 #define QGEMM_DEFAULT_P qgemm_p
2820 #define CGEMM_DEFAULT_P cgemm_p
2821 #define ZGEMM_DEFAULT_P zgemm_p
2822 #define XGEMM_DEFAULT_P xgemm_p
2824 #define SGEMM_DEFAULT_R sgemm_r
2825 #define DGEMM_DEFAULT_R dgemm_r
2826 #define QGEMM_DEFAULT_R qgemm_r
2827 #define CGEMM_DEFAULT_R cgemm_r
2828 #define ZGEMM_DEFAULT_R zgemm_r
2829 #define XGEMM_DEFAULT_R xgemm_r
2831 #define SGEMM_DEFAULT_Q 128
2832 #define DGEMM_DEFAULT_Q 128
2833 #define QGEMM_DEFAULT_Q 128
2834 #define CGEMM_DEFAULT_Q 128
2835 #define ZGEMM_DEFAULT_Q 128
2836 #define XGEMM_DEFAULT_Q 128
2842 #ifndef QGEMM_DEFAULT_UNROLL_M
2843 #define QGEMM_DEFAULT_UNROLL_M 2
2846 #ifndef QGEMM_DEFAULT_UNROLL_N
2847 #define QGEMM_DEFAULT_UNROLL_N 2
2850 #ifndef XGEMM_DEFAULT_UNROLL_M
2851 #define XGEMM_DEFAULT_UNROLL_M 2
2854 #ifndef XGEMM_DEFAULT_UNROLL_N
2855 #define XGEMM_DEFAULT_UNROLL_N 2
2859 #define SHUFPD_0 shufps $0x44,
2860 #define SHUFPD_1 shufps $0x4e,
2861 #define SHUFPD_2 shufps $0xe4,
2862 #define SHUFPD_3 shufps $0xee,
2866 #define SHUFPD_0 shufpd $0,
2870 #define SHUFPD_1 shufpd $1,
2874 #define SHUFPD_2 shufpd $2,
2878 #define SHUFPD_3 shufpd $3,
2882 #define SHUFPS_39 shufps $0x39,