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 /*********************************************************************/
76 #define SBGEMM_DEFAULT_UNROLL_N 4
77 #define SBGEMM_DEFAULT_UNROLL_M 8
78 #define SBGEMM_DEFAULT_UNROLL_MN 32
79 #define SBGEMM_DEFAULT_P 256
80 #define SBGEMM_DEFAULT_R 256
81 #define SBGEMM_DEFAULT_Q 256
87 #define GEMM_DEFAULT_OFFSET_A 64
88 #define GEMM_DEFAULT_OFFSET_B 256
89 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x01ffffUL
91 #define SGEMM_DEFAULT_UNROLL_N 4
92 #define DGEMM_DEFAULT_UNROLL_N 4
93 #define QGEMM_DEFAULT_UNROLL_N 2
94 #define CGEMM_DEFAULT_UNROLL_N 2
95 #define ZGEMM_DEFAULT_UNROLL_N 2
96 #define XGEMM_DEFAULT_UNROLL_N 1
99 #define SGEMM_DEFAULT_UNROLL_M 4
100 #define DGEMM_DEFAULT_UNROLL_M 2
101 #define QGEMM_DEFAULT_UNROLL_M 2
102 #define CGEMM_DEFAULT_UNROLL_M 2
103 #define ZGEMM_DEFAULT_UNROLL_M 1
104 #define XGEMM_DEFAULT_UNROLL_M 1
106 #define SGEMM_DEFAULT_UNROLL_M 8
107 #define DGEMM_DEFAULT_UNROLL_M 4
108 #define QGEMM_DEFAULT_UNROLL_M 2
109 #define CGEMM_DEFAULT_UNROLL_M 4
110 #define ZGEMM_DEFAULT_UNROLL_M 2
111 #define XGEMM_DEFAULT_UNROLL_M 1
114 #define SGEMM_DEFAULT_P sgemm_p
115 #define DGEMM_DEFAULT_P dgemm_p
116 #define QGEMM_DEFAULT_P qgemm_p
117 #define CGEMM_DEFAULT_P cgemm_p
118 #define ZGEMM_DEFAULT_P zgemm_p
119 #define XGEMM_DEFAULT_P xgemm_p
121 #define SGEMM_DEFAULT_R sgemm_r
122 #define DGEMM_DEFAULT_R dgemm_r
123 #define QGEMM_DEFAULT_R qgemm_r
124 #define CGEMM_DEFAULT_R cgemm_r
125 #define ZGEMM_DEFAULT_R zgemm_r
126 #define XGEMM_DEFAULT_R xgemm_r
130 #define SGEMM_DEFAULT_Q 248
131 #define DGEMM_DEFAULT_Q 248
132 #define QGEMM_DEFAULT_Q 248
133 #define CGEMM_DEFAULT_Q 248
134 #define ZGEMM_DEFAULT_Q 248
135 #define XGEMM_DEFAULT_Q 248
139 #define SGEMM_DEFAULT_Q 240
140 #define DGEMM_DEFAULT_Q 240
141 #define QGEMM_DEFAULT_Q 240
142 #define CGEMM_DEFAULT_Q 240
143 #define ZGEMM_DEFAULT_Q 240
144 #define XGEMM_DEFAULT_Q 240
150 #define HAVE_EXCLUSIVE_CACHE
154 #if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
159 #define GEMM_DEFAULT_OFFSET_A 64
160 #define GEMM_DEFAULT_OFFSET_B 832
161 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0fffUL
163 #define SGEMM_DEFAULT_UNROLL_N 4
164 #define DGEMM_DEFAULT_UNROLL_N 4
165 #define QGEMM_DEFAULT_UNROLL_N 2
166 #define CGEMM_DEFAULT_UNROLL_N 2
167 #define ZGEMM_DEFAULT_UNROLL_N 2
168 #define XGEMM_DEFAULT_UNROLL_N 1
171 #define SGEMM_DEFAULT_UNROLL_M 4
172 #define DGEMM_DEFAULT_UNROLL_M 2
173 #define QGEMM_DEFAULT_UNROLL_M 2
174 #define CGEMM_DEFAULT_UNROLL_M 2
175 #define ZGEMM_DEFAULT_UNROLL_M 1
176 #define XGEMM_DEFAULT_UNROLL_M 1
178 #define SGEMM_DEFAULT_UNROLL_M 8
179 #define DGEMM_DEFAULT_UNROLL_M 4
180 #define QGEMM_DEFAULT_UNROLL_M 2
181 #define CGEMM_DEFAULT_UNROLL_M 4
182 #define ZGEMM_DEFAULT_UNROLL_M 2
183 #define XGEMM_DEFAULT_UNROLL_M 1
187 #define SGEMM_DEFAULT_P 496
188 #define DGEMM_DEFAULT_P 248
189 #define QGEMM_DEFAULT_P 124
190 #define CGEMM_DEFAULT_P 248
191 #define ZGEMM_DEFAULT_P 124
192 #define XGEMM_DEFAULT_P 62
194 #define SGEMM_DEFAULT_Q 248
195 #define DGEMM_DEFAULT_Q 248
196 #define QGEMM_DEFAULT_Q 248
197 #define CGEMM_DEFAULT_Q 248
198 #define ZGEMM_DEFAULT_Q 248
199 #define XGEMM_DEFAULT_Q 248
203 #define SGEMM_DEFAULT_P 448
204 #define DGEMM_DEFAULT_P 224
205 #define QGEMM_DEFAULT_P 112
206 #define CGEMM_DEFAULT_P 224
207 #define ZGEMM_DEFAULT_P 112
208 #define XGEMM_DEFAULT_P 56
210 #define SGEMM_DEFAULT_Q 224
211 #define DGEMM_DEFAULT_Q 224
212 #define QGEMM_DEFAULT_Q 224
213 #define CGEMM_DEFAULT_Q 224
214 #define ZGEMM_DEFAULT_Q 224
215 #define XGEMM_DEFAULT_Q 224
219 #define SGEMM_DEFAULT_R sgemm_r
220 #define QGEMM_DEFAULT_R qgemm_r
221 #define DGEMM_DEFAULT_R dgemm_r
222 #define CGEMM_DEFAULT_R cgemm_r
223 #define ZGEMM_DEFAULT_R zgemm_r
224 #define XGEMM_DEFAULT_R xgemm_r
227 #define HAVE_EXCLUSIVE_CACHE
229 #define GEMM_THREAD gemm_thread_mn
239 #define GEMM_DEFAULT_OFFSET_A 64
240 #define GEMM_DEFAULT_OFFSET_B 832
241 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0fffUL
245 #define QGEMM_DEFAULT_UNROLL_N 2
246 #define CGEMM_DEFAULT_UNROLL_N 2
247 #define ZGEMM_DEFAULT_UNROLL_N 2
248 #define XGEMM_DEFAULT_UNROLL_N 1
251 #define SGEMM_DEFAULT_UNROLL_N 4
252 #define DGEMM_DEFAULT_UNROLL_N 4
253 #define SGEMM_DEFAULT_UNROLL_M 4
254 #define DGEMM_DEFAULT_UNROLL_M 2
255 #define QGEMM_DEFAULT_UNROLL_M 2
256 #define CGEMM_DEFAULT_UNROLL_M 2
257 #define ZGEMM_DEFAULT_UNROLL_M 1
258 #define XGEMM_DEFAULT_UNROLL_M 1
260 #define SGEMM_DEFAULT_UNROLL_N 2
261 #define DGEMM_DEFAULT_UNROLL_N 2
262 #define SGEMM_DEFAULT_UNROLL_M 16
263 #define DGEMM_DEFAULT_UNROLL_M 8
264 #define QGEMM_DEFAULT_UNROLL_M 2
265 #define CGEMM_DEFAULT_UNROLL_M 4
266 #define ZGEMM_DEFAULT_UNROLL_M 2
267 #define XGEMM_DEFAULT_UNROLL_M 1
268 #define CGEMM3M_DEFAULT_UNROLL_N 4
269 #define CGEMM3M_DEFAULT_UNROLL_M 8
270 #define ZGEMM3M_DEFAULT_UNROLL_N 4
271 #define ZGEMM3M_DEFAULT_UNROLL_M 4
273 #define DGEMM_DEFAULT_UNROLL_MN 16
274 #define GEMV_UNROLL 8
278 #if defined(ARCH_X86_64)
279 #define SGEMM_DEFAULT_P 768
280 #define DGEMM_DEFAULT_P 384
282 #define SGEMM_DEFAULT_P 448
283 #define DGEMM_DEFAULT_P 224
286 #define QGEMM_DEFAULT_P 112
287 #define CGEMM_DEFAULT_P 224
288 #define ZGEMM_DEFAULT_P 112
289 #define XGEMM_DEFAULT_P 56
291 #if defined(ARCH_X86_64)
292 #define SGEMM_DEFAULT_Q 168
293 #define DGEMM_DEFAULT_Q 168
295 #define SGEMM_DEFAULT_Q 224
296 #define DGEMM_DEFAULT_Q 224
299 #define QGEMM_DEFAULT_Q 224
300 #define CGEMM_DEFAULT_Q 224
301 #define ZGEMM_DEFAULT_Q 224
302 #define XGEMM_DEFAULT_Q 224
304 #define CGEMM3M_DEFAULT_P 448
305 #define ZGEMM3M_DEFAULT_P 224
306 #define XGEMM3M_DEFAULT_P 112
307 #define CGEMM3M_DEFAULT_Q 224
308 #define ZGEMM3M_DEFAULT_Q 224
309 #define XGEMM3M_DEFAULT_Q 224
310 #define CGEMM3M_DEFAULT_R 12288
311 #define ZGEMM3M_DEFAULT_R 12288
312 #define XGEMM3M_DEFAULT_R 12288
314 #define SGEMM_DEFAULT_R sgemm_r
315 #define QGEMM_DEFAULT_R qgemm_r
316 #define DGEMM_DEFAULT_R dgemm_r
317 #define CGEMM_DEFAULT_R cgemm_r
318 #define ZGEMM_DEFAULT_R zgemm_r
319 #define XGEMM_DEFAULT_R xgemm_r
322 #define HAVE_EXCLUSIVE_CACHE
324 #define GEMM_THREAD gemm_thread_mn
332 #define GEMM_DEFAULT_OFFSET_A 64
333 #define GEMM_DEFAULT_OFFSET_B 832
334 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0fffUL
338 #define QGEMM_DEFAULT_UNROLL_N 2
339 #define CGEMM_DEFAULT_UNROLL_N 2
340 #define ZGEMM_DEFAULT_UNROLL_N 2
341 #define XGEMM_DEFAULT_UNROLL_N 1
344 #define SGEMM_DEFAULT_UNROLL_N 4
345 #define DGEMM_DEFAULT_UNROLL_N 4
346 #define SGEMM_DEFAULT_UNROLL_M 4
347 #define DGEMM_DEFAULT_UNROLL_M 2
348 #define QGEMM_DEFAULT_UNROLL_M 2
349 #define CGEMM_DEFAULT_UNROLL_M 2
350 #define ZGEMM_DEFAULT_UNROLL_M 1
351 #define XGEMM_DEFAULT_UNROLL_M 1
353 #define SGEMM_DEFAULT_UNROLL_N 2
354 #define DGEMM_DEFAULT_UNROLL_N 2
355 #define SGEMM_DEFAULT_UNROLL_M 16
356 #define DGEMM_DEFAULT_UNROLL_M 8
357 #define QGEMM_DEFAULT_UNROLL_M 2
358 #define CGEMM_DEFAULT_UNROLL_M 4
359 #define ZGEMM_DEFAULT_UNROLL_M 2
360 #define XGEMM_DEFAULT_UNROLL_M 1
361 #define CGEMM3M_DEFAULT_UNROLL_N 4
362 #define CGEMM3M_DEFAULT_UNROLL_M 8
363 #define ZGEMM3M_DEFAULT_UNROLL_N 4
364 #define ZGEMM3M_DEFAULT_UNROLL_M 4
365 #define GEMV_UNROLL 8
368 #if defined(ARCH_X86_64)
369 #define SGEMM_DEFAULT_P 768
370 #define DGEMM_DEFAULT_P 768
371 #define ZGEMM_DEFAULT_P 384
372 #define CGEMM_DEFAULT_P 768
374 #define SGEMM_DEFAULT_P 448
375 #define DGEMM_DEFAULT_P 480
376 #define ZGEMM_DEFAULT_P 112
377 #define CGEMM_DEFAULT_P 224
379 #define QGEMM_DEFAULT_P 112
380 #define XGEMM_DEFAULT_P 56
382 #if defined(ARCH_X86_64)
383 #define SGEMM_DEFAULT_Q 192
384 #define DGEMM_DEFAULT_Q 168
385 #define ZGEMM_DEFAULT_Q 168
386 #define CGEMM_DEFAULT_Q 168
388 #define SGEMM_DEFAULT_Q 224
389 #define DGEMM_DEFAULT_Q 224
390 #define ZGEMM_DEFAULT_Q 224
391 #define CGEMM_DEFAULT_Q 224
393 #define QGEMM_DEFAULT_Q 224
394 #define XGEMM_DEFAULT_Q 224
396 #define CGEMM3M_DEFAULT_P 448
397 #define ZGEMM3M_DEFAULT_P 224
398 #define XGEMM3M_DEFAULT_P 112
399 #define CGEMM3M_DEFAULT_Q 224
400 #define ZGEMM3M_DEFAULT_Q 224
401 #define XGEMM3M_DEFAULT_Q 224
402 #define CGEMM3M_DEFAULT_R 12288
403 #define ZGEMM3M_DEFAULT_R 12288
404 #define XGEMM3M_DEFAULT_R 12288
406 #define SGEMM_DEFAULT_R 12288
407 #define QGEMM_DEFAULT_R qgemm_r
408 #define DGEMM_DEFAULT_R 12288
409 #define CGEMM_DEFAULT_R cgemm_r
410 #define ZGEMM_DEFAULT_R zgemm_r
411 #define XGEMM_DEFAULT_R xgemm_r
414 #define HAVE_EXCLUSIVE_CACHE
416 #define GEMM_THREAD gemm_thread_mn
424 #define GEMM_DEFAULT_OFFSET_A 64
425 #define GEMM_DEFAULT_OFFSET_B 832
426 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0fffUL
430 #define QGEMM_DEFAULT_UNROLL_N 2
431 #define CGEMM_DEFAULT_UNROLL_N 2
432 #define ZGEMM_DEFAULT_UNROLL_N 2
433 #define XGEMM_DEFAULT_UNROLL_N 1
436 #define SGEMM_DEFAULT_UNROLL_N 4
437 #define DGEMM_DEFAULT_UNROLL_N 4
438 #define SGEMM_DEFAULT_UNROLL_M 4
439 #define DGEMM_DEFAULT_UNROLL_M 2
440 #define QGEMM_DEFAULT_UNROLL_M 2
441 #define CGEMM_DEFAULT_UNROLL_M 2
442 #define ZGEMM_DEFAULT_UNROLL_M 1
443 #define XGEMM_DEFAULT_UNROLL_M 1
445 #define SGEMM_DEFAULT_UNROLL_N 2
446 #define DGEMM_DEFAULT_UNROLL_N 2
447 #define SGEMM_DEFAULT_UNROLL_M 16
448 #define DGEMM_DEFAULT_UNROLL_M 8
449 #define QGEMM_DEFAULT_UNROLL_M 2
450 #define CGEMM_DEFAULT_UNROLL_M 4
451 #define ZGEMM_DEFAULT_UNROLL_M 2
452 #define XGEMM_DEFAULT_UNROLL_M 1
453 #define CGEMM3M_DEFAULT_UNROLL_N 4
454 #define CGEMM3M_DEFAULT_UNROLL_M 8
455 #define ZGEMM3M_DEFAULT_UNROLL_N 4
456 #define ZGEMM3M_DEFAULT_UNROLL_M 4
457 #define GEMV_UNROLL 8
460 #if defined(ARCH_X86_64)
461 #define SGEMM_DEFAULT_P 768
462 #define DGEMM_DEFAULT_P 576
463 #define ZGEMM_DEFAULT_P 288
464 #define CGEMM_DEFAULT_P 576
466 #define SGEMM_DEFAULT_P 448
467 #define DGEMM_DEFAULT_P 480
468 #define ZGEMM_DEFAULT_P 112
469 #define CGEMM_DEFAULT_P 224
471 #define QGEMM_DEFAULT_P 112
472 #define XGEMM_DEFAULT_P 56
474 #if defined(ARCH_X86_64)
475 #define SGEMM_DEFAULT_Q 192
476 #define DGEMM_DEFAULT_Q 160
477 #define ZGEMM_DEFAULT_Q 160
478 #define CGEMM_DEFAULT_Q 160
480 #define SGEMM_DEFAULT_Q 224
481 #define DGEMM_DEFAULT_Q 224
482 #define ZGEMM_DEFAULT_Q 224
483 #define CGEMM_DEFAULT_Q 224
485 #define QGEMM_DEFAULT_Q 224
486 #define XGEMM_DEFAULT_Q 224
488 #define CGEMM3M_DEFAULT_P 448
489 #define ZGEMM3M_DEFAULT_P 224
490 #define XGEMM3M_DEFAULT_P 112
491 #define CGEMM3M_DEFAULT_Q 224
492 #define ZGEMM3M_DEFAULT_Q 224
493 #define XGEMM3M_DEFAULT_Q 224
494 #define CGEMM3M_DEFAULT_R 12288
495 #define ZGEMM3M_DEFAULT_R 12288
496 #define XGEMM3M_DEFAULT_R 12288
498 #define SGEMM_DEFAULT_R 12288
499 #define QGEMM_DEFAULT_R qgemm_r
500 #define DGEMM_DEFAULT_R 12288
501 #define CGEMM_DEFAULT_R cgemm_r
502 #define ZGEMM_DEFAULT_R zgemm_r
503 #define XGEMM_DEFAULT_R xgemm_r
506 #define HAVE_EXCLUSIVE_CACHE
508 #define GEMM_THREAD gemm_thread_mn
517 #define GEMM_DEFAULT_OFFSET_A 64
518 #define GEMM_DEFAULT_OFFSET_B 832
519 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0fffUL
523 #define QGEMM_DEFAULT_UNROLL_N 2
524 #define CGEMM_DEFAULT_UNROLL_N 2
525 #define ZGEMM_DEFAULT_UNROLL_N 2
526 #define XGEMM_DEFAULT_UNROLL_N 1
529 #define SGEMM_DEFAULT_UNROLL_N 4
530 #define DGEMM_DEFAULT_UNROLL_N 4
531 #define SGEMM_DEFAULT_UNROLL_M 4
532 #define DGEMM_DEFAULT_UNROLL_M 2
533 #define QGEMM_DEFAULT_UNROLL_M 2
534 #define CGEMM_DEFAULT_UNROLL_M 2
535 #define ZGEMM_DEFAULT_UNROLL_M 1
536 #define XGEMM_DEFAULT_UNROLL_M 1
538 #define SGEMM_DEFAULT_UNROLL_N 2
539 #define DGEMM_DEFAULT_UNROLL_N 2
540 #define SGEMM_DEFAULT_UNROLL_M 16
541 #define DGEMM_DEFAULT_UNROLL_M 8
542 #define QGEMM_DEFAULT_UNROLL_M 2
543 #define CGEMM_DEFAULT_UNROLL_M 4
544 #define ZGEMM_DEFAULT_UNROLL_M 2
545 #define XGEMM_DEFAULT_UNROLL_M 1
546 #define CGEMM3M_DEFAULT_UNROLL_N 4
547 #define CGEMM3M_DEFAULT_UNROLL_M 8
548 #define ZGEMM3M_DEFAULT_UNROLL_N 4
549 #define ZGEMM3M_DEFAULT_UNROLL_M 4
550 #define GEMV_UNROLL 8
553 #if defined(ARCH_X86_64)
554 #define SGEMM_DEFAULT_P 768
555 #define DGEMM_DEFAULT_P 576
556 #define ZGEMM_DEFAULT_P 288
557 #define CGEMM_DEFAULT_P 576
559 #define SGEMM_DEFAULT_P 448
560 #define DGEMM_DEFAULT_P 480
561 #define ZGEMM_DEFAULT_P 112
562 #define CGEMM_DEFAULT_P 224
564 #define QGEMM_DEFAULT_P 112
565 #define XGEMM_DEFAULT_P 56
567 #if defined(ARCH_X86_64)
568 #define SGEMM_DEFAULT_Q 192
569 #define DGEMM_DEFAULT_Q 160
570 #define ZGEMM_DEFAULT_Q 160
571 #define CGEMM_DEFAULT_Q 160
573 #define SGEMM_DEFAULT_Q 224
574 #define DGEMM_DEFAULT_Q 224
575 #define ZGEMM_DEFAULT_Q 224
576 #define CGEMM_DEFAULT_Q 224
578 #define QGEMM_DEFAULT_Q 224
579 #define XGEMM_DEFAULT_Q 224
581 #define CGEMM3M_DEFAULT_P 448
582 #define ZGEMM3M_DEFAULT_P 224
583 #define XGEMM3M_DEFAULT_P 112
584 #define CGEMM3M_DEFAULT_Q 224
585 #define ZGEMM3M_DEFAULT_Q 224
586 #define XGEMM3M_DEFAULT_Q 224
587 #define CGEMM3M_DEFAULT_R 12288
588 #define ZGEMM3M_DEFAULT_R 12288
589 #define XGEMM3M_DEFAULT_R 12288
591 #define SGEMM_DEFAULT_R 12288
592 #define QGEMM_DEFAULT_R qgemm_r
593 #define DGEMM_DEFAULT_R 12288
594 #define CGEMM_DEFAULT_R cgemm_r
595 #define ZGEMM_DEFAULT_R zgemm_r
596 #define XGEMM_DEFAULT_R xgemm_r
599 #define HAVE_EXCLUSIVE_CACHE
601 #define GEMM_THREAD gemm_thread_mn
609 #define GEMM_DEFAULT_OFFSET_A 0
610 #define GEMM_DEFAULT_OFFSET_B 0
611 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
615 #define SWITCH_RATIO 16
619 #define SGEMM_DEFAULT_UNROLL_M 4
620 #define DGEMM_DEFAULT_UNROLL_M 2
621 #define QGEMM_DEFAULT_UNROLL_M 2
622 #define CGEMM_DEFAULT_UNROLL_M 2
623 #define ZGEMM_DEFAULT_UNROLL_M 1
624 #define XGEMM_DEFAULT_UNROLL_M 1
626 #define SGEMM_DEFAULT_UNROLL_N 4
627 #define DGEMM_DEFAULT_UNROLL_N 4
628 #define QGEMM_DEFAULT_UNROLL_N 2
629 #define CGEMM_DEFAULT_UNROLL_N 2
630 #define ZGEMM_DEFAULT_UNROLL_N 2
631 #define XGEMM_DEFAULT_UNROLL_N 1
635 #define SGEMM_DEFAULT_UNROLL_M 8
636 #define DGEMM_DEFAULT_UNROLL_M 4
637 #define QGEMM_DEFAULT_UNROLL_M 2
638 #define CGEMM_DEFAULT_UNROLL_M 8
639 #define ZGEMM_DEFAULT_UNROLL_M 4
640 #define XGEMM_DEFAULT_UNROLL_M 1
642 #define SGEMM_DEFAULT_UNROLL_N 4
643 #define DGEMM_DEFAULT_UNROLL_N 8
644 #define QGEMM_DEFAULT_UNROLL_N 2
645 #define CGEMM_DEFAULT_UNROLL_N 2
646 #define ZGEMM_DEFAULT_UNROLL_N 2
647 #define XGEMM_DEFAULT_UNROLL_N 1
649 #define SGEMM_DEFAULT_UNROLL_MN 32
650 #define DGEMM_DEFAULT_UNROLL_MN 32
656 #define SGEMM_DEFAULT_P 512
657 #define SGEMM_DEFAULT_R sgemm_r
658 #define DGEMM_DEFAULT_P 512
659 #define DGEMM_DEFAULT_R dgemm_r
660 #define QGEMM_DEFAULT_P 504
661 #define QGEMM_DEFAULT_R qgemm_r
662 #define CGEMM_DEFAULT_P 128
663 #define CGEMM_DEFAULT_R 1024
664 #define ZGEMM_DEFAULT_P 512
665 #define ZGEMM_DEFAULT_R zgemm_r
666 #define XGEMM_DEFAULT_P 252
667 #define XGEMM_DEFAULT_R xgemm_r
668 #define SGEMM_DEFAULT_Q 256
669 #define DGEMM_DEFAULT_Q 256
670 #define QGEMM_DEFAULT_Q 128
671 #define CGEMM_DEFAULT_Q 256
672 #define ZGEMM_DEFAULT_Q 192
673 #define XGEMM_DEFAULT_Q 128
677 #define SGEMM_DEFAULT_P 320
678 #define DGEMM_DEFAULT_P 512
679 #define CGEMM_DEFAULT_P 256
680 #define ZGEMM_DEFAULT_P 192
683 #define SGEMM_DEFAULT_Q 320
684 #define DGEMM_DEFAULT_Q 128
686 #define SGEMM_DEFAULT_Q 320
687 #define DGEMM_DEFAULT_Q 256
689 #define CGEMM_DEFAULT_Q 256
690 #define ZGEMM_DEFAULT_Q 192
692 #define SGEMM_DEFAULT_R sgemm_r
693 #define DGEMM_DEFAULT_R 13824
694 #define CGEMM_DEFAULT_R cgemm_r
695 #define ZGEMM_DEFAULT_R zgemm_r
697 #define QGEMM_DEFAULT_Q 128
698 #define QGEMM_DEFAULT_P 504
699 #define QGEMM_DEFAULT_R qgemm_r
700 #define XGEMM_DEFAULT_P 252
701 #define XGEMM_DEFAULT_R xgemm_r
702 #define XGEMM_DEFAULT_Q 128
704 #define CGEMM3M_DEFAULT_UNROLL_N 4
705 #define CGEMM3M_DEFAULT_UNROLL_M 8
706 #define ZGEMM3M_DEFAULT_UNROLL_N 4
707 #define ZGEMM3M_DEFAULT_UNROLL_M 4
709 #define CGEMM3M_DEFAULT_P 320
710 #define ZGEMM3M_DEFAULT_P 256
711 #define XGEMM3M_DEFAULT_P 112
712 #define CGEMM3M_DEFAULT_Q 320
713 #define ZGEMM3M_DEFAULT_Q 256
714 #define XGEMM3M_DEFAULT_Q 224
715 #define CGEMM3M_DEFAULT_R 12288
716 #define ZGEMM3M_DEFAULT_R 12288
717 #define XGEMM3M_DEFAULT_R 12288
728 #define GEMM_DEFAULT_OFFSET_A 0
729 #define GEMM_DEFAULT_OFFSET_B 384
730 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
732 #define SGEMM_DEFAULT_UNROLL_N 4
733 #define DGEMM_DEFAULT_UNROLL_N 4
734 #define QGEMM_DEFAULT_UNROLL_N 2
735 #define CGEMM_DEFAULT_UNROLL_N 2
736 #define ZGEMM_DEFAULT_UNROLL_N 2
737 #define XGEMM_DEFAULT_UNROLL_N 1
739 #define SGEMM_DEFAULT_UNROLL_M 2
740 #define DGEMM_DEFAULT_UNROLL_M 1
741 #define QGEMM_DEFAULT_UNROLL_M 2
742 #define CGEMM_DEFAULT_UNROLL_M 1
743 #define ZGEMM_DEFAULT_UNROLL_M 1
744 #define XGEMM_DEFAULT_UNROLL_M 1
746 #define SGEMM_DEFAULT_R sgemm_r
747 #define DGEMM_DEFAULT_R dgemm_r
748 #define QGEMM_DEFAULT_R qgemm_r
749 #define CGEMM_DEFAULT_R cgemm_r
750 #define ZGEMM_DEFAULT_R zgemm_r
751 #define XGEMM_DEFAULT_R xgemm_r
753 #define SGEMM_DEFAULT_P 208
754 #define DGEMM_DEFAULT_P 104
755 #define QGEMM_DEFAULT_P 56
756 #define CGEMM_DEFAULT_P 104
757 #define ZGEMM_DEFAULT_P 56
758 #define XGEMM_DEFAULT_P 28
760 #define SGEMM_DEFAULT_Q 208
761 #define DGEMM_DEFAULT_Q 208
762 #define QGEMM_DEFAULT_Q 208
763 #define CGEMM_DEFAULT_Q 208
764 #define ZGEMM_DEFAULT_Q 208
765 #define XGEMM_DEFAULT_Q 208
768 #define HAVE_EXCLUSIVE_CACHE
776 #define GEMM_DEFAULT_OFFSET_A 0
777 #define GEMM_DEFAULT_OFFSET_B 256
778 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
780 #define SGEMM_DEFAULT_UNROLL_N 4
781 #define DGEMM_DEFAULT_UNROLL_N 4
782 #define QGEMM_DEFAULT_UNROLL_N 2
783 #define CGEMM_DEFAULT_UNROLL_N 2
784 #define ZGEMM_DEFAULT_UNROLL_N 2
785 #define XGEMM_DEFAULT_UNROLL_N 1
787 #define SGEMM_DEFAULT_UNROLL_M 2
788 #define DGEMM_DEFAULT_UNROLL_M 1
789 #define QGEMM_DEFAULT_UNROLL_M 2
790 #define CGEMM_DEFAULT_UNROLL_M 1
791 #define ZGEMM_DEFAULT_UNROLL_M 1
792 #define XGEMM_DEFAULT_UNROLL_M 1
794 #define SGEMM_DEFAULT_R sgemm_r
795 #define DGEMM_DEFAULT_R dgemm_r
796 #define QGEMM_DEFAULT_R qgemm_r
797 #define CGEMM_DEFAULT_R cgemm_r
798 #define ZGEMM_DEFAULT_R zgemm_r
799 #define XGEMM_DEFAULT_R xgemm_r
801 #define SGEMM_DEFAULT_P 128
802 #define DGEMM_DEFAULT_P 128
803 #define QGEMM_DEFAULT_P 128
804 #define CGEMM_DEFAULT_P 128
805 #define ZGEMM_DEFAULT_P 128
806 #define XGEMM_DEFAULT_P 128
808 #define SGEMM_DEFAULT_Q 512
809 #define DGEMM_DEFAULT_Q 256
810 #define QGEMM_DEFAULT_Q 256
811 #define CGEMM_DEFAULT_Q 256
812 #define ZGEMM_DEFAULT_Q 128
813 #define XGEMM_DEFAULT_Q 128
823 #define GEMM_DEFAULT_OFFSET_A 64
824 #define GEMM_DEFAULT_OFFSET_B 256
825 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x01ffffUL
828 #define SGEMM_DEFAULT_UNROLL_N 4
829 #define DGEMM_DEFAULT_UNROLL_N 4
830 #define QGEMM_DEFAULT_UNROLL_N 2
831 #define CGEMM_DEFAULT_UNROLL_N 2
832 #define ZGEMM_DEFAULT_UNROLL_N 2
833 #define XGEMM_DEFAULT_UNROLL_N 1
835 #define SGEMM_DEFAULT_UNROLL_M 4
836 #define DGEMM_DEFAULT_UNROLL_M 2
837 #define QGEMM_DEFAULT_UNROLL_M 2
838 #define CGEMM_DEFAULT_UNROLL_M 2
839 #define ZGEMM_DEFAULT_UNROLL_M 1
840 #define XGEMM_DEFAULT_UNROLL_M 1
842 #define SGEMM_DEFAULT_UNROLL_N 8
843 #define DGEMM_DEFAULT_UNROLL_N 4
844 #define QGEMM_DEFAULT_UNROLL_N 2
845 #define CGEMM_DEFAULT_UNROLL_N 4
846 #define ZGEMM_DEFAULT_UNROLL_N 2
847 #define XGEMM_DEFAULT_UNROLL_N 1
849 #define SGEMM_DEFAULT_UNROLL_M 4
850 #define DGEMM_DEFAULT_UNROLL_M 4
851 #define QGEMM_DEFAULT_UNROLL_M 2
852 #define CGEMM_DEFAULT_UNROLL_M 2
853 #define ZGEMM_DEFAULT_UNROLL_M 2
854 #define XGEMM_DEFAULT_UNROLL_M 1
857 #define SGEMM_DEFAULT_P 288
858 #define DGEMM_DEFAULT_P 288
859 #define QGEMM_DEFAULT_P 288
860 #define CGEMM_DEFAULT_P 288
861 #define ZGEMM_DEFAULT_P 288
862 #define XGEMM_DEFAULT_P 288
864 #define SGEMM_DEFAULT_R sgemm_r
865 #define DGEMM_DEFAULT_R dgemm_r
866 #define QGEMM_DEFAULT_R qgemm_r
867 #define CGEMM_DEFAULT_R cgemm_r
868 #define ZGEMM_DEFAULT_R zgemm_r
869 #define XGEMM_DEFAULT_R xgemm_r
871 #define SGEMM_DEFAULT_Q 256
872 #define DGEMM_DEFAULT_Q 128
873 #define QGEMM_DEFAULT_Q 64
874 #define CGEMM_DEFAULT_Q 128
875 #define ZGEMM_DEFAULT_Q 64
876 #define XGEMM_DEFAULT_Q 32
879 #define HAVE_EXCLUSIVE_CACHE
883 #if defined(PENTIUM) || defined(PENTIUM2) || defined(PENTIUM3)
892 #define GEMM_DEFAULT_OFFSET_A 0
893 #define GEMM_DEFAULT_OFFSET_B 0
894 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
897 #define SGEMM_DEFAULT_UNROLL_M 8
898 #define CGEMM_DEFAULT_UNROLL_M 4
900 #define SGEMM_DEFAULT_UNROLL_M 4
901 #define CGEMM_DEFAULT_UNROLL_M 2
903 #define DGEMM_DEFAULT_UNROLL_M 2
904 #define SGEMM_DEFAULT_UNROLL_N 2
905 #define DGEMM_DEFAULT_UNROLL_N 2
906 #define QGEMM_DEFAULT_UNROLL_M 2
907 #define QGEMM_DEFAULT_UNROLL_N 2
908 #define CGEMM_DEFAULT_UNROLL_N 1
909 #define ZGEMM_DEFAULT_UNROLL_M 1
910 #define ZGEMM_DEFAULT_UNROLL_N 1
911 #define XGEMM_DEFAULT_UNROLL_M 1
912 #define XGEMM_DEFAULT_UNROLL_N 1
914 #define SGEMM_DEFAULT_P sgemm_p
915 #define SGEMM_DEFAULT_Q 256
916 #define SGEMM_DEFAULT_R sgemm_r
918 #define DGEMM_DEFAULT_P dgemm_p
919 #define DGEMM_DEFAULT_Q 256
920 #define DGEMM_DEFAULT_R dgemm_r
922 #define QGEMM_DEFAULT_P qgemm_p
923 #define QGEMM_DEFAULT_Q 256
924 #define QGEMM_DEFAULT_R qgemm_r
926 #define CGEMM_DEFAULT_P cgemm_p
927 #define CGEMM_DEFAULT_Q 256
928 #define CGEMM_DEFAULT_R cgemm_r
930 #define ZGEMM_DEFAULT_P zgemm_p
931 #define ZGEMM_DEFAULT_Q 256
932 #define ZGEMM_DEFAULT_R zgemm_r
934 #define XGEMM_DEFAULT_P xgemm_p
935 #define XGEMM_DEFAULT_Q 256
936 #define XGEMM_DEFAULT_R xgemm_r
947 #define GEMM_DEFAULT_OFFSET_A 0
948 #define GEMM_DEFAULT_OFFSET_B 0
949 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
952 #define SGEMM_DEFAULT_UNROLL_M 4
953 #define SGEMM_DEFAULT_UNROLL_N 4
954 #define DGEMM_DEFAULT_UNROLL_M 2
955 #define DGEMM_DEFAULT_UNROLL_N 4
956 #define QGEMM_DEFAULT_UNROLL_M 2
957 #define QGEMM_DEFAULT_UNROLL_N 2
958 #define CGEMM_DEFAULT_UNROLL_M 2
959 #define CGEMM_DEFAULT_UNROLL_N 2
960 #define ZGEMM_DEFAULT_UNROLL_M 1
961 #define ZGEMM_DEFAULT_UNROLL_N 2
962 #define XGEMM_DEFAULT_UNROLL_M 1
963 #define XGEMM_DEFAULT_UNROLL_N 1
965 #define SGEMM_DEFAULT_UNROLL_M 8
966 #define SGEMM_DEFAULT_UNROLL_N 2
967 #define DGEMM_DEFAULT_UNROLL_M 2
968 #define DGEMM_DEFAULT_UNROLL_N 2
969 #define QGEMM_DEFAULT_UNROLL_M 2
970 #define QGEMM_DEFAULT_UNROLL_N 2
971 #define CGEMM_DEFAULT_UNROLL_M 4
972 #define CGEMM_DEFAULT_UNROLL_N 1
973 #define ZGEMM_DEFAULT_UNROLL_M 1
974 #define ZGEMM_DEFAULT_UNROLL_N 1
975 #define XGEMM_DEFAULT_UNROLL_M 1
976 #define XGEMM_DEFAULT_UNROLL_N 1
980 #define SGEMM_DEFAULT_P sgemm_p
981 #define SGEMM_DEFAULT_Q 256
982 #define SGEMM_DEFAULT_R sgemm_r
984 #define DGEMM_DEFAULT_P dgemm_p
985 #define DGEMM_DEFAULT_Q 256
986 #define DGEMM_DEFAULT_R dgemm_r
988 #define QGEMM_DEFAULT_P qgemm_p
989 #define QGEMM_DEFAULT_Q 256
990 #define QGEMM_DEFAULT_R qgemm_r
992 #define CGEMM_DEFAULT_P cgemm_p
993 #define CGEMM_DEFAULT_Q 256
994 #define CGEMM_DEFAULT_R cgemm_r
996 #define ZGEMM_DEFAULT_P zgemm_p
997 #define ZGEMM_DEFAULT_Q 256
998 #define ZGEMM_DEFAULT_R zgemm_r
1000 #define XGEMM_DEFAULT_P xgemm_p
1001 #define XGEMM_DEFAULT_Q 256
1002 #define XGEMM_DEFAULT_R xgemm_r
1007 #ifdef CORE_NORTHWOOD
1012 #define GEMM_DEFAULT_OFFSET_A 0
1013 #define GEMM_DEFAULT_OFFSET_B 32
1015 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
1019 #define SGEMM_DEFAULT_UNROLL_M 8
1020 #define DGEMM_DEFAULT_UNROLL_M 4
1021 #define QGEMM_DEFAULT_UNROLL_M 2
1022 #define CGEMM_DEFAULT_UNROLL_M 4
1023 #define ZGEMM_DEFAULT_UNROLL_M 2
1024 #define XGEMM_DEFAULT_UNROLL_M 1
1026 #define SGEMM_DEFAULT_UNROLL_N 2
1027 #define DGEMM_DEFAULT_UNROLL_N 2
1028 #define QGEMM_DEFAULT_UNROLL_N 2
1029 #define CGEMM_DEFAULT_UNROLL_N 1
1030 #define ZGEMM_DEFAULT_UNROLL_N 1
1031 #define XGEMM_DEFAULT_UNROLL_N 1
1033 #define SGEMM_DEFAULT_P sgemm_p
1034 #define SGEMM_DEFAULT_R sgemm_r
1036 #define DGEMM_DEFAULT_P dgemm_p
1037 #define DGEMM_DEFAULT_R dgemm_r
1039 #define QGEMM_DEFAULT_P qgemm_p
1040 #define QGEMM_DEFAULT_R qgemm_r
1042 #define CGEMM_DEFAULT_P cgemm_p
1043 #define CGEMM_DEFAULT_R cgemm_r
1045 #define ZGEMM_DEFAULT_P zgemm_p
1046 #define ZGEMM_DEFAULT_R zgemm_r
1048 #define XGEMM_DEFAULT_P xgemm_p
1049 #define XGEMM_DEFAULT_R xgemm_r
1051 #define SGEMM_DEFAULT_Q 128
1052 #define DGEMM_DEFAULT_Q 128
1053 #define QGEMM_DEFAULT_Q 128
1054 #define CGEMM_DEFAULT_Q 128
1055 #define ZGEMM_DEFAULT_Q 128
1056 #define XGEMM_DEFAULT_Q 128
1059 #ifdef CORE_PRESCOTT
1065 #define GEMM_DEFAULT_OFFSET_A 128
1066 #define GEMM_DEFAULT_OFFSET_B 192
1068 #define GEMM_DEFAULT_OFFSET_A 0
1069 #define GEMM_DEFAULT_OFFSET_B 256
1072 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
1077 #define SGEMM_DEFAULT_UNROLL_M 4
1078 #define DGEMM_DEFAULT_UNROLL_M 2
1079 #define QGEMM_DEFAULT_UNROLL_M 2
1080 #define CGEMM_DEFAULT_UNROLL_M 2
1081 #define ZGEMM_DEFAULT_UNROLL_M 1
1082 #define XGEMM_DEFAULT_UNROLL_M 1
1084 #define SGEMM_DEFAULT_UNROLL_M 8
1085 #define DGEMM_DEFAULT_UNROLL_M 4
1086 #define QGEMM_DEFAULT_UNROLL_M 2
1087 #define CGEMM_DEFAULT_UNROLL_M 4
1088 #define ZGEMM_DEFAULT_UNROLL_M 2
1089 #define XGEMM_DEFAULT_UNROLL_M 1
1092 #define SGEMM_DEFAULT_UNROLL_N 4
1093 #define DGEMM_DEFAULT_UNROLL_N 4
1094 #define QGEMM_DEFAULT_UNROLL_N 2
1095 #define CGEMM_DEFAULT_UNROLL_N 2
1096 #define ZGEMM_DEFAULT_UNROLL_N 2
1097 #define XGEMM_DEFAULT_UNROLL_N 1
1099 #define SGEMM_DEFAULT_P sgemm_p
1100 #define SGEMM_DEFAULT_R sgemm_r
1102 #define DGEMM_DEFAULT_P dgemm_p
1103 #define DGEMM_DEFAULT_R dgemm_r
1105 #define QGEMM_DEFAULT_P qgemm_p
1106 #define QGEMM_DEFAULT_R qgemm_r
1108 #define CGEMM_DEFAULT_P cgemm_p
1109 #define CGEMM_DEFAULT_R cgemm_r
1111 #define ZGEMM_DEFAULT_P zgemm_p
1112 #define ZGEMM_DEFAULT_R zgemm_r
1114 #define XGEMM_DEFAULT_P xgemm_p
1115 #define XGEMM_DEFAULT_R xgemm_r
1117 #define SGEMM_DEFAULT_Q 128
1118 #define DGEMM_DEFAULT_Q 128
1119 #define QGEMM_DEFAULT_Q 128
1120 #define CGEMM_DEFAULT_Q 128
1121 #define ZGEMM_DEFAULT_Q 128
1122 #define XGEMM_DEFAULT_Q 128
1130 #define GEMM_DEFAULT_OFFSET_A 448
1131 #define GEMM_DEFAULT_OFFSET_B 128
1132 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1136 #define SWITCH_RATIO 4
1139 #define SGEMM_DEFAULT_UNROLL_M 8
1140 #define DGEMM_DEFAULT_UNROLL_M 4
1141 #define QGEMM_DEFAULT_UNROLL_M 2
1142 #define CGEMM_DEFAULT_UNROLL_M 4
1143 #define ZGEMM_DEFAULT_UNROLL_M 2
1144 #define XGEMM_DEFAULT_UNROLL_M 1
1146 #define SGEMM_DEFAULT_UNROLL_N 2
1147 #define DGEMM_DEFAULT_UNROLL_N 2
1148 #define QGEMM_DEFAULT_UNROLL_N 2
1149 #define CGEMM_DEFAULT_UNROLL_N 1
1150 #define ZGEMM_DEFAULT_UNROLL_N 1
1151 #define XGEMM_DEFAULT_UNROLL_N 1
1153 #define MASK(a, b) ((((a) + (b) - 1) / (b)) * (b))
1156 #define SGEMM_DEFAULT_UNROLL_M 8
1157 #define DGEMM_DEFAULT_UNROLL_M 4
1158 #define QGEMM_DEFAULT_UNROLL_M 2
1159 #define CGEMM_DEFAULT_UNROLL_M 4
1160 #define ZGEMM_DEFAULT_UNROLL_M 2
1161 #define XGEMM_DEFAULT_UNROLL_M 1
1163 #define SGEMM_DEFAULT_UNROLL_N 4
1164 #define DGEMM_DEFAULT_UNROLL_N 4
1165 #define QGEMM_DEFAULT_UNROLL_N 2
1166 #define CGEMM_DEFAULT_UNROLL_N 2
1167 #define ZGEMM_DEFAULT_UNROLL_N 2
1168 #define XGEMM_DEFAULT_UNROLL_N 1
1171 #define SGEMM_DEFAULT_P sgemm_p
1172 #define SGEMM_DEFAULT_R sgemm_r
1174 #define DGEMM_DEFAULT_P dgemm_p
1175 #define DGEMM_DEFAULT_R dgemm_r
1177 #define QGEMM_DEFAULT_P qgemm_p
1178 #define QGEMM_DEFAULT_R qgemm_r
1180 #define CGEMM_DEFAULT_P cgemm_p
1181 #define CGEMM_DEFAULT_R cgemm_r
1183 #define ZGEMM_DEFAULT_P zgemm_p
1184 #define ZGEMM_DEFAULT_R zgemm_r
1186 #define XGEMM_DEFAULT_P xgemm_p
1187 #define XGEMM_DEFAULT_R xgemm_r
1189 #define SGEMM_DEFAULT_Q 256
1190 #define DGEMM_DEFAULT_Q 256
1191 #define QGEMM_DEFAULT_Q 256
1192 #define CGEMM_DEFAULT_Q 256
1193 #define ZGEMM_DEFAULT_Q 256
1194 #define XGEMM_DEFAULT_Q 256
1203 #define GEMM_DEFAULT_OFFSET_A 128
1204 #define GEMM_DEFAULT_OFFSET_B 0
1205 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1209 #define SWITCH_RATIO 4
1212 #define SGEMM_DEFAULT_UNROLL_M 4
1213 #define DGEMM_DEFAULT_UNROLL_M 2
1214 #define QGEMM_DEFAULT_UNROLL_M 2
1215 #define CGEMM_DEFAULT_UNROLL_M 2
1216 #define ZGEMM_DEFAULT_UNROLL_M 1
1217 #define XGEMM_DEFAULT_UNROLL_M 1
1219 #define SGEMM_DEFAULT_UNROLL_N 4
1220 #define DGEMM_DEFAULT_UNROLL_N 4
1221 #define QGEMM_DEFAULT_UNROLL_N 2
1222 #define CGEMM_DEFAULT_UNROLL_N 2
1223 #define ZGEMM_DEFAULT_UNROLL_N 2
1224 #define XGEMM_DEFAULT_UNROLL_N 1
1226 #define SGEMM_DEFAULT_UNROLL_M 8
1227 #define DGEMM_DEFAULT_UNROLL_M 4
1228 #define QGEMM_DEFAULT_UNROLL_M 2
1229 #define CGEMM_DEFAULT_UNROLL_M 4
1230 #define ZGEMM_DEFAULT_UNROLL_M 2
1231 #define XGEMM_DEFAULT_UNROLL_M 1
1233 #define SGEMM_DEFAULT_UNROLL_N 4
1234 #define DGEMM_DEFAULT_UNROLL_N 4
1235 #define QGEMM_DEFAULT_UNROLL_N 2
1236 #define CGEMM_DEFAULT_UNROLL_N 2
1237 #define ZGEMM_DEFAULT_UNROLL_N 2
1238 #define XGEMM_DEFAULT_UNROLL_N 1
1241 #define SGEMM_DEFAULT_P sgemm_p
1242 #define SGEMM_DEFAULT_R sgemm_r
1244 #define DGEMM_DEFAULT_P dgemm_p
1245 #define DGEMM_DEFAULT_R dgemm_r
1247 #define QGEMM_DEFAULT_P qgemm_p
1248 #define QGEMM_DEFAULT_R qgemm_r
1250 #define CGEMM_DEFAULT_P cgemm_p
1251 #define CGEMM_DEFAULT_R cgemm_r
1253 #define ZGEMM_DEFAULT_P zgemm_p
1254 #define ZGEMM_DEFAULT_R zgemm_r
1256 #define XGEMM_DEFAULT_P xgemm_p
1257 #define XGEMM_DEFAULT_R xgemm_r
1259 #define SGEMM_DEFAULT_Q 512
1260 #define DGEMM_DEFAULT_Q 256
1261 #define QGEMM_DEFAULT_Q 128
1262 #define CGEMM_DEFAULT_Q 512
1263 #define ZGEMM_DEFAULT_Q 256
1264 #define XGEMM_DEFAULT_Q 128
1266 #define GETRF_FACTOR 0.75
1274 #define GEMM_DEFAULT_OFFSET_A 128
1275 #define GEMM_DEFAULT_OFFSET_B 0
1276 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1280 #define SWITCH_RATIO 4
1283 #define SGEMM_DEFAULT_UNROLL_M 4
1284 #define DGEMM_DEFAULT_UNROLL_M 2
1285 #define QGEMM_DEFAULT_UNROLL_M 2
1286 #define CGEMM_DEFAULT_UNROLL_M 2
1287 #define ZGEMM_DEFAULT_UNROLL_M 1
1288 #define XGEMM_DEFAULT_UNROLL_M 1
1290 #define SGEMM_DEFAULT_UNROLL_N 4
1291 #define DGEMM_DEFAULT_UNROLL_N 4
1292 #define QGEMM_DEFAULT_UNROLL_N 2
1293 #define CGEMM_DEFAULT_UNROLL_N 2
1294 #define ZGEMM_DEFAULT_UNROLL_N 2
1295 #define XGEMM_DEFAULT_UNROLL_N 1
1297 #define SGEMM_DEFAULT_UNROLL_M 8
1298 #define DGEMM_DEFAULT_UNROLL_M 4
1299 #define QGEMM_DEFAULT_UNROLL_M 2
1300 #define CGEMM_DEFAULT_UNROLL_M 4
1301 #define ZGEMM_DEFAULT_UNROLL_M 2
1302 #define XGEMM_DEFAULT_UNROLL_M 1
1304 #define SGEMM_DEFAULT_UNROLL_N 4
1305 #define DGEMM_DEFAULT_UNROLL_N 4
1306 #define QGEMM_DEFAULT_UNROLL_N 2
1307 #define CGEMM_DEFAULT_UNROLL_N 2
1308 #define ZGEMM_DEFAULT_UNROLL_N 2
1309 #define XGEMM_DEFAULT_UNROLL_N 1
1312 #define SGEMM_DEFAULT_P sgemm_p
1313 #define SGEMM_DEFAULT_R sgemm_r
1315 #define DGEMM_DEFAULT_P dgemm_p
1316 #define DGEMM_DEFAULT_R dgemm_r
1318 #define QGEMM_DEFAULT_P qgemm_p
1319 #define QGEMM_DEFAULT_R qgemm_r
1321 #define CGEMM_DEFAULT_P cgemm_p
1322 #define CGEMM_DEFAULT_R cgemm_r
1324 #define ZGEMM_DEFAULT_P zgemm_p
1325 #define ZGEMM_DEFAULT_R zgemm_r
1327 #define XGEMM_DEFAULT_P xgemm_p
1328 #define XGEMM_DEFAULT_R xgemm_r
1330 #define SGEMM_DEFAULT_Q 768
1331 #define DGEMM_DEFAULT_Q 384
1332 #define QGEMM_DEFAULT_Q 192
1333 #define CGEMM_DEFAULT_Q 768
1334 #define ZGEMM_DEFAULT_Q 384
1335 #define XGEMM_DEFAULT_Q 192
1337 #define GETRF_FACTOR 0.75
1338 #define GEMM_THREAD gemm_thread_mn
1346 #define GEMM_DEFAULT_OFFSET_A 32
1347 #define GEMM_DEFAULT_OFFSET_B 0
1348 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1352 #define SWITCH_RATIO 4
1355 #define SGEMM_DEFAULT_UNROLL_M 4
1356 #define DGEMM_DEFAULT_UNROLL_M 2
1357 #define QGEMM_DEFAULT_UNROLL_M 2
1358 #define CGEMM_DEFAULT_UNROLL_M 2
1359 #define ZGEMM_DEFAULT_UNROLL_M 1
1360 #define XGEMM_DEFAULT_UNROLL_M 1
1362 #define SGEMM_DEFAULT_UNROLL_N 4
1363 #define DGEMM_DEFAULT_UNROLL_N 4
1364 #define QGEMM_DEFAULT_UNROLL_N 2
1365 #define CGEMM_DEFAULT_UNROLL_N 2
1366 #define ZGEMM_DEFAULT_UNROLL_N 2
1367 #define XGEMM_DEFAULT_UNROLL_N 1
1369 #define SGEMM_DEFAULT_UNROLL_M 4
1370 #define DGEMM_DEFAULT_UNROLL_M 2
1371 #define QGEMM_DEFAULT_UNROLL_M 2
1372 #define CGEMM_DEFAULT_UNROLL_M 2
1373 #define ZGEMM_DEFAULT_UNROLL_M 1
1374 #define XGEMM_DEFAULT_UNROLL_M 1
1376 #define SGEMM_DEFAULT_UNROLL_N 8
1377 #define DGEMM_DEFAULT_UNROLL_N 8
1378 #define QGEMM_DEFAULT_UNROLL_N 2
1379 #define CGEMM_DEFAULT_UNROLL_N 4
1380 #define ZGEMM_DEFAULT_UNROLL_N 4
1381 #define XGEMM_DEFAULT_UNROLL_N 1
1384 #define SGEMM_DEFAULT_P 504
1385 #define SGEMM_DEFAULT_R sgemm_r
1387 #define DGEMM_DEFAULT_P 504
1388 #define DGEMM_DEFAULT_R dgemm_r
1390 #define QGEMM_DEFAULT_P 504
1391 #define QGEMM_DEFAULT_R qgemm_r
1393 #define CGEMM_DEFAULT_P 252
1394 #define CGEMM_DEFAULT_R cgemm_r
1396 #define ZGEMM_DEFAULT_P 252
1397 #define ZGEMM_DEFAULT_R zgemm_r
1399 #define XGEMM_DEFAULT_P 252
1400 #define XGEMM_DEFAULT_R xgemm_r
1402 #define SGEMM_DEFAULT_Q 512
1403 #define DGEMM_DEFAULT_Q 256
1404 #define QGEMM_DEFAULT_Q 128
1405 #define CGEMM_DEFAULT_Q 512
1406 #define ZGEMM_DEFAULT_Q 256
1407 #define XGEMM_DEFAULT_Q 128
1409 #define GETRF_FACTOR 0.72
1419 #define GEMM_DEFAULT_OFFSET_A 0
1420 #define GEMM_DEFAULT_OFFSET_B 0
1421 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1425 #define SWITCH_RATIO 4
1428 #define SGEMM_DEFAULT_UNROLL_M 4
1429 #define DGEMM_DEFAULT_UNROLL_M 2
1430 #define QGEMM_DEFAULT_UNROLL_M 2
1431 #define CGEMM_DEFAULT_UNROLL_M 2
1432 #define ZGEMM_DEFAULT_UNROLL_M 1
1433 #define XGEMM_DEFAULT_UNROLL_M 1
1435 #define SGEMM_DEFAULT_UNROLL_N 4
1436 #define DGEMM_DEFAULT_UNROLL_N 4
1437 #define QGEMM_DEFAULT_UNROLL_N 2
1438 #define CGEMM_DEFAULT_UNROLL_N 2
1439 #define ZGEMM_DEFAULT_UNROLL_N 2
1440 #define XGEMM_DEFAULT_UNROLL_N 1
1442 #define SGEMM_DEFAULT_UNROLL_M 16
1443 #define DGEMM_DEFAULT_UNROLL_M 8
1444 #define QGEMM_DEFAULT_UNROLL_M 2
1445 #define CGEMM_DEFAULT_UNROLL_M 8
1446 #define ZGEMM_DEFAULT_UNROLL_M 1
1447 #define XGEMM_DEFAULT_UNROLL_M 1
1449 #define SGEMM_DEFAULT_UNROLL_N 4
1450 #define DGEMM_DEFAULT_UNROLL_N 4
1451 #define QGEMM_DEFAULT_UNROLL_N 2
1452 #define CGEMM_DEFAULT_UNROLL_N 2
1453 #define ZGEMM_DEFAULT_UNROLL_N 4
1454 #define XGEMM_DEFAULT_UNROLL_N 1
1457 #define SGEMM_DEFAULT_P 768
1458 #define SGEMM_DEFAULT_R sgemm_r
1459 /*#define SGEMM_DEFAULT_R 1024*/
1461 #define DGEMM_DEFAULT_P 512
1462 #define DGEMM_DEFAULT_R dgemm_r
1463 /*#define DGEMM_DEFAULT_R 1024*/
1465 #define QGEMM_DEFAULT_P 504
1466 #define QGEMM_DEFAULT_R qgemm_r
1468 #define CGEMM_DEFAULT_P 768
1469 #define CGEMM_DEFAULT_R cgemm_r
1470 /*#define CGEMM_DEFAULT_R 1024*/
1472 #define ZGEMM_DEFAULT_P 512
1473 #define ZGEMM_DEFAULT_R zgemm_r
1474 /*#define ZGEMM_DEFAULT_R 1024*/
1476 #define XGEMM_DEFAULT_P 252
1477 #define XGEMM_DEFAULT_R xgemm_r
1479 #define SGEMM_DEFAULT_Q 384
1480 #define DGEMM_DEFAULT_Q 256
1481 #define QGEMM_DEFAULT_Q 128
1482 #define CGEMM_DEFAULT_Q 512
1483 #define ZGEMM_DEFAULT_Q 192
1484 #define XGEMM_DEFAULT_Q 128
1486 #define CGEMM3M_DEFAULT_UNROLL_N 8
1487 #define CGEMM3M_DEFAULT_UNROLL_M 4
1488 #define ZGEMM3M_DEFAULT_UNROLL_N 8
1489 #define ZGEMM3M_DEFAULT_UNROLL_M 2
1491 #define CGEMM3M_DEFAULT_P 448
1492 #define ZGEMM3M_DEFAULT_P 224
1493 #define XGEMM3M_DEFAULT_P 112
1494 #define CGEMM3M_DEFAULT_Q 224
1495 #define ZGEMM3M_DEFAULT_Q 224
1496 #define XGEMM3M_DEFAULT_Q 224
1497 #define CGEMM3M_DEFAULT_R 12288
1498 #define ZGEMM3M_DEFAULT_R 12288
1499 #define XGEMM3M_DEFAULT_R 12288
1503 #define GETRF_FACTOR 0.72
1512 #define GEMM_DEFAULT_OFFSET_A 0
1513 #define GEMM_DEFAULT_OFFSET_B 0
1514 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1518 #if defined(XDOUBLE) || defined(DOUBLE)
1519 #define SWITCH_RATIO 4
1520 #define GEMM_PREFERED_SIZE 4
1522 #define SWITCH_RATIO 8
1523 #define GEMM_PREFERED_SIZE 8
1528 #define SGEMM_DEFAULT_UNROLL_M 4
1529 #define DGEMM_DEFAULT_UNROLL_M 2
1530 #define QGEMM_DEFAULT_UNROLL_M 2
1531 #define CGEMM_DEFAULT_UNROLL_M 2
1532 #define ZGEMM_DEFAULT_UNROLL_M 1
1533 #define XGEMM_DEFAULT_UNROLL_M 1
1535 #define SGEMM_DEFAULT_UNROLL_N 4
1536 #define DGEMM_DEFAULT_UNROLL_N 4
1537 #define QGEMM_DEFAULT_UNROLL_N 2
1538 #define CGEMM_DEFAULT_UNROLL_N 2
1539 #define ZGEMM_DEFAULT_UNROLL_N 2
1540 #define XGEMM_DEFAULT_UNROLL_N 1
1544 #define SGEMM_DEFAULT_UNROLL_M 8
1545 #define DGEMM_DEFAULT_UNROLL_M 4
1546 #define QGEMM_DEFAULT_UNROLL_M 2
1547 #define CGEMM_DEFAULT_UNROLL_M 8
1548 #define ZGEMM_DEFAULT_UNROLL_M 4
1549 #define XGEMM_DEFAULT_UNROLL_M 1
1551 #define SGEMM_DEFAULT_UNROLL_N 4
1552 #define DGEMM_DEFAULT_UNROLL_N 8
1553 #define QGEMM_DEFAULT_UNROLL_N 2
1554 #define CGEMM_DEFAULT_UNROLL_N 2
1555 #define ZGEMM_DEFAULT_UNROLL_N 2
1556 #define XGEMM_DEFAULT_UNROLL_N 1
1558 #define SGEMM_DEFAULT_UNROLL_MN 32
1559 #define DGEMM_DEFAULT_UNROLL_MN 32
1565 #define SGEMM_DEFAULT_P 512
1566 #define SGEMM_DEFAULT_R sgemm_r
1567 #define DGEMM_DEFAULT_P 512
1568 #define DGEMM_DEFAULT_R dgemm_r
1569 #define QGEMM_DEFAULT_P 504
1570 #define QGEMM_DEFAULT_R qgemm_r
1571 #define CGEMM_DEFAULT_P 128
1572 #define CGEMM_DEFAULT_R 1024
1573 #define ZGEMM_DEFAULT_P 512
1574 #define ZGEMM_DEFAULT_R zgemm_r
1575 #define XGEMM_DEFAULT_P 252
1576 #define XGEMM_DEFAULT_R xgemm_r
1577 #define SGEMM_DEFAULT_Q 256
1578 #define DGEMM_DEFAULT_Q 256
1579 #define QGEMM_DEFAULT_Q 128
1580 #define CGEMM_DEFAULT_Q 256
1581 #define ZGEMM_DEFAULT_Q 192
1582 #define XGEMM_DEFAULT_Q 128
1586 #define SGEMM_DEFAULT_P 320
1587 #define DGEMM_DEFAULT_P 512
1588 #define CGEMM_DEFAULT_P 256
1589 #define ZGEMM_DEFAULT_P 192
1592 #define SGEMM_DEFAULT_Q 320
1593 #define DGEMM_DEFAULT_Q 128
1595 #define SGEMM_DEFAULT_Q 320
1596 #define DGEMM_DEFAULT_Q 256
1598 #define CGEMM_DEFAULT_Q 256
1599 #define ZGEMM_DEFAULT_Q 192
1601 #define SGEMM_DEFAULT_R sgemm_r
1602 #define DGEMM_DEFAULT_R 13824
1603 #define CGEMM_DEFAULT_R cgemm_r
1604 #define ZGEMM_DEFAULT_R zgemm_r
1606 #define QGEMM_DEFAULT_Q 128
1607 #define QGEMM_DEFAULT_P 504
1608 #define QGEMM_DEFAULT_R qgemm_r
1609 #define XGEMM_DEFAULT_P 252
1610 #define XGEMM_DEFAULT_R xgemm_r
1611 #define XGEMM_DEFAULT_Q 128
1613 #define CGEMM3M_DEFAULT_UNROLL_N 4
1614 #define CGEMM3M_DEFAULT_UNROLL_M 8
1615 #define ZGEMM3M_DEFAULT_UNROLL_N 4
1616 #define ZGEMM3M_DEFAULT_UNROLL_M 4
1618 #define CGEMM3M_DEFAULT_P 320
1619 #define ZGEMM3M_DEFAULT_P 256
1620 #define XGEMM3M_DEFAULT_P 112
1621 #define CGEMM3M_DEFAULT_Q 320
1622 #define ZGEMM3M_DEFAULT_Q 256
1623 #define XGEMM3M_DEFAULT_Q 224
1624 #define CGEMM3M_DEFAULT_R 12288
1625 #define ZGEMM3M_DEFAULT_R 12288
1626 #define XGEMM3M_DEFAULT_R 12288
1638 #define GEMM_DEFAULT_OFFSET_A 0
1639 #define GEMM_DEFAULT_OFFSET_B 0
1640 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1644 #if defined(XDOUBLE) || defined(DOUBLE)
1645 #define SWITCH_RATIO 8
1646 #define GEMM_PREFERED_SIZE 8
1648 #define SWITCH_RATIO 16
1649 #define GEMM_PREFERED_SIZE 16
1651 #define USE_SGEMM_KERNEL_DIRECT 1
1655 #define SGEMM_DEFAULT_UNROLL_M 4
1656 #define DGEMM_DEFAULT_UNROLL_M 2
1657 #define QGEMM_DEFAULT_UNROLL_M 2
1658 #define CGEMM_DEFAULT_UNROLL_M 2
1659 #define ZGEMM_DEFAULT_UNROLL_M 1
1660 #define XGEMM_DEFAULT_UNROLL_M 1
1662 #define SGEMM_DEFAULT_UNROLL_N 4
1663 #define DGEMM_DEFAULT_UNROLL_N 4
1664 #define QGEMM_DEFAULT_UNROLL_N 2
1665 #define CGEMM_DEFAULT_UNROLL_N 2
1666 #define ZGEMM_DEFAULT_UNROLL_N 2
1667 #define XGEMM_DEFAULT_UNROLL_N 1
1671 #define SGEMM_DEFAULT_UNROLL_M 16
1672 #define DGEMM_DEFAULT_UNROLL_M 16
1673 #define QGEMM_DEFAULT_UNROLL_M 2
1674 #define CGEMM_DEFAULT_UNROLL_M 8
1675 #define ZGEMM_DEFAULT_UNROLL_M 4
1676 #define XGEMM_DEFAULT_UNROLL_M 1
1678 #define SGEMM_DEFAULT_UNROLL_N 4
1679 #define DGEMM_DEFAULT_UNROLL_N 2
1680 #define QGEMM_DEFAULT_UNROLL_N 2
1681 #define CGEMM_DEFAULT_UNROLL_N 2
1682 #define ZGEMM_DEFAULT_UNROLL_N 2
1683 #define XGEMM_DEFAULT_UNROLL_N 1
1685 #define SGEMM_DEFAULT_UNROLL_MN 32
1686 #define DGEMM_DEFAULT_UNROLL_MN 32
1691 #define SGEMM_DEFAULT_P 512
1692 #define SGEMM_DEFAULT_R sgemm_r
1693 #define DGEMM_DEFAULT_P 512
1694 #define DGEMM_DEFAULT_R dgemm_r
1695 #define QGEMM_DEFAULT_P 504
1696 #define QGEMM_DEFAULT_R qgemm_r
1697 #define CGEMM_DEFAULT_P 128
1698 #define CGEMM_DEFAULT_R 1024
1699 #define ZGEMM_DEFAULT_P 512
1700 #define ZGEMM_DEFAULT_R zgemm_r
1701 #define XGEMM_DEFAULT_P 252
1702 #define XGEMM_DEFAULT_R xgemm_r
1703 #define SGEMM_DEFAULT_Q 256
1704 #define DGEMM_DEFAULT_Q 256
1705 #define QGEMM_DEFAULT_Q 128
1706 #define CGEMM_DEFAULT_Q 256
1707 #define ZGEMM_DEFAULT_Q 192
1708 #define XGEMM_DEFAULT_Q 128
1712 #define SGEMM_DEFAULT_P 448
1713 #define DGEMM_DEFAULT_P 192
1714 #define CGEMM_DEFAULT_P 384
1715 #define ZGEMM_DEFAULT_P 256
1717 #define SGEMM_DEFAULT_Q 448
1718 #define DGEMM_DEFAULT_Q 384
1719 #define CGEMM_DEFAULT_Q 192
1720 #define ZGEMM_DEFAULT_Q 128
1722 #define SGEMM_DEFAULT_R sgemm_r
1723 #define DGEMM_DEFAULT_R 8640
1724 #define CGEMM_DEFAULT_R cgemm_r
1725 #define ZGEMM_DEFAULT_R zgemm_r
1727 #define QGEMM_DEFAULT_Q 128
1728 #define QGEMM_DEFAULT_P 504
1729 #define QGEMM_DEFAULT_R qgemm_r
1730 #define XGEMM_DEFAULT_P 252
1731 #define XGEMM_DEFAULT_R xgemm_r
1732 #define XGEMM_DEFAULT_Q 128
1734 #define CGEMM3M_DEFAULT_UNROLL_N 4
1735 #define CGEMM3M_DEFAULT_UNROLL_M 8
1736 #define ZGEMM3M_DEFAULT_UNROLL_N 4
1737 #define ZGEMM3M_DEFAULT_UNROLL_M 4
1739 #define CGEMM3M_DEFAULT_P 320
1740 #define ZGEMM3M_DEFAULT_P 256
1741 #define XGEMM3M_DEFAULT_P 112
1742 #define CGEMM3M_DEFAULT_Q 320
1743 #define ZGEMM3M_DEFAULT_Q 256
1744 #define XGEMM3M_DEFAULT_Q 224
1745 #define CGEMM3M_DEFAULT_R 12288
1746 #define ZGEMM3M_DEFAULT_R 12288
1747 #define XGEMM3M_DEFAULT_R 12288
1754 #ifdef SAPPHIRERAPIDS
1759 #define GEMM_DEFAULT_OFFSET_A 0
1760 #define GEMM_DEFAULT_OFFSET_B 0
1761 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1765 #if defined(XDOUBLE) || defined(DOUBLE)
1766 #define SWITCH_RATIO 8
1767 #define GEMM_PREFERED_SIZE 8
1769 #define SWITCH_RATIO 16
1770 #define GEMM_PREFERED_SIZE 16
1772 #define USE_SGEMM_KERNEL_DIRECT 1
1776 #define SGEMM_DEFAULT_UNROLL_M 4
1777 #define DGEMM_DEFAULT_UNROLL_M 2
1778 #define QGEMM_DEFAULT_UNROLL_M 2
1779 #define CGEMM_DEFAULT_UNROLL_M 2
1780 #define ZGEMM_DEFAULT_UNROLL_M 1
1781 #define XGEMM_DEFAULT_UNROLL_M 1
1783 #define SGEMM_DEFAULT_UNROLL_N 4
1784 #define DGEMM_DEFAULT_UNROLL_N 4
1785 #define QGEMM_DEFAULT_UNROLL_N 2
1786 #define CGEMM_DEFAULT_UNROLL_N 2
1787 #define ZGEMM_DEFAULT_UNROLL_N 2
1788 #define XGEMM_DEFAULT_UNROLL_N 1
1792 #define SGEMM_DEFAULT_UNROLL_M 16
1793 #define DGEMM_DEFAULT_UNROLL_M 16
1794 #define QGEMM_DEFAULT_UNROLL_M 2
1795 #define CGEMM_DEFAULT_UNROLL_M 8
1796 #define ZGEMM_DEFAULT_UNROLL_M 4
1797 #define XGEMM_DEFAULT_UNROLL_M 1
1799 #define SGEMM_DEFAULT_UNROLL_N 4
1800 #define DGEMM_DEFAULT_UNROLL_N 2
1801 #define QGEMM_DEFAULT_UNROLL_N 2
1802 #define CGEMM_DEFAULT_UNROLL_N 2
1803 #define ZGEMM_DEFAULT_UNROLL_N 2
1804 #define XGEMM_DEFAULT_UNROLL_N 1
1806 #define SGEMM_DEFAULT_UNROLL_MN 32
1807 #define DGEMM_DEFAULT_UNROLL_MN 32
1812 #define SGEMM_DEFAULT_P 512
1813 #define SGEMM_DEFAULT_R sgemm_r
1814 #define DGEMM_DEFAULT_P 512
1815 #define DGEMM_DEFAULT_R dgemm_r
1816 #define QGEMM_DEFAULT_P 504
1817 #define QGEMM_DEFAULT_R qgemm_r
1818 #define CGEMM_DEFAULT_P 128
1819 #define CGEMM_DEFAULT_R 1024
1820 #define ZGEMM_DEFAULT_P 512
1821 #define ZGEMM_DEFAULT_R zgemm_r
1822 #define XGEMM_DEFAULT_P 252
1823 #define XGEMM_DEFAULT_R xgemm_r
1824 #define SGEMM_DEFAULT_Q 256
1825 #define DGEMM_DEFAULT_Q 256
1826 #define QGEMM_DEFAULT_Q 128
1827 #define CGEMM_DEFAULT_Q 256
1828 #define ZGEMM_DEFAULT_Q 192
1829 #define XGEMM_DEFAULT_Q 128
1833 #define SGEMM_DEFAULT_P 640
1834 #define DGEMM_DEFAULT_P 192
1835 #define CGEMM_DEFAULT_P 384
1836 #define ZGEMM_DEFAULT_P 256
1838 #define SGEMM_DEFAULT_Q 320
1839 #define DGEMM_DEFAULT_Q 384
1840 #define CGEMM_DEFAULT_Q 192
1841 #define ZGEMM_DEFAULT_Q 128
1843 #define SGEMM_DEFAULT_R sgemm_r
1844 #define DGEMM_DEFAULT_R 8640
1845 #define CGEMM_DEFAULT_R cgemm_r
1846 #define ZGEMM_DEFAULT_R zgemm_r
1848 #define QGEMM_DEFAULT_Q 128
1849 #define QGEMM_DEFAULT_P 504
1850 #define QGEMM_DEFAULT_R qgemm_r
1851 #define XGEMM_DEFAULT_P 252
1852 #define XGEMM_DEFAULT_R xgemm_r
1853 #define XGEMM_DEFAULT_Q 128
1855 #define CGEMM3M_DEFAULT_UNROLL_N 4
1856 #define CGEMM3M_DEFAULT_UNROLL_M 8
1857 #define ZGEMM3M_DEFAULT_UNROLL_N 4
1858 #define ZGEMM3M_DEFAULT_UNROLL_M 4
1860 #define CGEMM3M_DEFAULT_P 320
1861 #define ZGEMM3M_DEFAULT_P 256
1862 #define XGEMM3M_DEFAULT_P 112
1863 #define CGEMM3M_DEFAULT_Q 320
1864 #define ZGEMM3M_DEFAULT_Q 256
1865 #define XGEMM3M_DEFAULT_Q 224
1866 #define CGEMM3M_DEFAULT_R 12288
1867 #define ZGEMM3M_DEFAULT_R 12288
1868 #define XGEMM3M_DEFAULT_R 12288
1878 #define GEMM_DEFAULT_OFFSET_A 0
1879 #define GEMM_DEFAULT_OFFSET_B 0
1880 #define GEMM_DEFAULT_ALIGN 0x03fffUL
1884 #if defined(XDOUBLE) || defined(DOUBLE)
1885 #define SWITCH_RATIO 8
1886 #define GEMM_PREFERED_SIZE 8
1888 #define SWITCH_RATIO 16
1889 #define GEMM_PREFERED_SIZE 16
1891 #define USE_SGEMM_KERNEL_DIRECT 1
1893 #undef SBGEMM_DEFAULT_UNROLL_N
1894 #undef SBGEMM_DEFAULT_UNROLL_M
1895 #undef SBGEMM_DEFAULT_P
1896 #undef SBGEMM_DEFAULT_R
1897 #undef SBGEMM_DEFAULT_Q
1898 #define SBGEMM_DEFAULT_UNROLL_N 4
1899 #define SBGEMM_DEFAULT_UNROLL_M 16
1900 #define SBGEMM_DEFAULT_P 384
1901 #define SBGEMM_DEFAULT_Q 768
1902 #define SBGEMM_DEFAULT_R sbgemm_r
1906 #define SGEMM_DEFAULT_UNROLL_M 4
1907 #define DGEMM_DEFAULT_UNROLL_M 2
1908 #define QGEMM_DEFAULT_UNROLL_M 2
1909 #define CGEMM_DEFAULT_UNROLL_M 2
1910 #define ZGEMM_DEFAULT_UNROLL_M 1
1911 #define XGEMM_DEFAULT_UNROLL_M 1
1913 #define SGEMM_DEFAULT_UNROLL_N 4
1914 #define DGEMM_DEFAULT_UNROLL_N 4
1915 #define QGEMM_DEFAULT_UNROLL_N 2
1916 #define CGEMM_DEFAULT_UNROLL_N 2
1917 #define ZGEMM_DEFAULT_UNROLL_N 2
1918 #define XGEMM_DEFAULT_UNROLL_N 1
1922 #define SGEMM_DEFAULT_UNROLL_M 16
1923 #define DGEMM_DEFAULT_UNROLL_M 16
1924 #define QGEMM_DEFAULT_UNROLL_M 2
1925 #define CGEMM_DEFAULT_UNROLL_M 8
1926 #define ZGEMM_DEFAULT_UNROLL_M 4
1927 #define XGEMM_DEFAULT_UNROLL_M 1
1929 #define SGEMM_DEFAULT_UNROLL_N 4
1930 #define DGEMM_DEFAULT_UNROLL_N 2
1931 #define QGEMM_DEFAULT_UNROLL_N 2
1932 #define CGEMM_DEFAULT_UNROLL_N 2
1933 #define ZGEMM_DEFAULT_UNROLL_N 2
1934 #define XGEMM_DEFAULT_UNROLL_N 1
1936 #define SGEMM_DEFAULT_UNROLL_MN 32
1937 #define DGEMM_DEFAULT_UNROLL_MN 32
1942 #define SGEMM_DEFAULT_P 512
1943 #define SGEMM_DEFAULT_R sgemm_r
1944 #define DGEMM_DEFAULT_P 512
1945 #define DGEMM_DEFAULT_R dgemm_r
1946 #define QGEMM_DEFAULT_P 504
1947 #define QGEMM_DEFAULT_R qgemm_r
1948 #define CGEMM_DEFAULT_P 128
1949 #define CGEMM_DEFAULT_R 1024
1950 #define ZGEMM_DEFAULT_P 512
1951 #define ZGEMM_DEFAULT_R zgemm_r
1952 #define XGEMM_DEFAULT_P 252
1953 #define XGEMM_DEFAULT_R xgemm_r
1954 #define SGEMM_DEFAULT_Q 256
1955 #define DGEMM_DEFAULT_Q 256
1956 #define QGEMM_DEFAULT_Q 128
1957 #define CGEMM_DEFAULT_Q 256
1958 #define ZGEMM_DEFAULT_Q 192
1959 #define XGEMM_DEFAULT_Q 128
1963 #define SGEMM_DEFAULT_P 640
1964 #define DGEMM_DEFAULT_P 192
1965 #define CGEMM_DEFAULT_P 384
1966 #define ZGEMM_DEFAULT_P 256
1968 #define SGEMM_DEFAULT_Q 320
1969 #define DGEMM_DEFAULT_Q 384
1970 #define CGEMM_DEFAULT_Q 192
1971 #define ZGEMM_DEFAULT_Q 128
1973 #define SGEMM_DEFAULT_R sgemm_r
1974 #define DGEMM_DEFAULT_R 8640
1975 #define CGEMM_DEFAULT_R cgemm_r
1976 #define ZGEMM_DEFAULT_R zgemm_r
1978 #define QGEMM_DEFAULT_Q 128
1979 #define QGEMM_DEFAULT_P 504
1980 #define QGEMM_DEFAULT_R qgemm_r
1981 #define XGEMM_DEFAULT_P 252
1982 #define XGEMM_DEFAULT_R xgemm_r
1983 #define XGEMM_DEFAULT_Q 128
1985 #define CGEMM3M_DEFAULT_UNROLL_N 4
1986 #define CGEMM3M_DEFAULT_UNROLL_M 8
1987 #define ZGEMM3M_DEFAULT_UNROLL_N 4
1988 #define ZGEMM3M_DEFAULT_UNROLL_M 4
1990 #define CGEMM3M_DEFAULT_P 320
1991 #define ZGEMM3M_DEFAULT_P 256
1992 #define XGEMM3M_DEFAULT_P 112
1993 #define CGEMM3M_DEFAULT_Q 320
1994 #define ZGEMM3M_DEFAULT_Q 256
1995 #define XGEMM3M_DEFAULT_Q 224
1996 #define CGEMM3M_DEFAULT_R 12288
1997 #define ZGEMM3M_DEFAULT_R 12288
1998 #define XGEMM3M_DEFAULT_R 12288
2009 #define GEMM_DEFAULT_OFFSET_A 64
2010 #define GEMM_DEFAULT_OFFSET_B 0
2011 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2016 #define SGEMM_DEFAULT_UNROLL_M 4
2017 #define DGEMM_DEFAULT_UNROLL_M 2
2018 #define QGEMM_DEFAULT_UNROLL_M 2
2019 #define CGEMM_DEFAULT_UNROLL_M 2
2020 #define ZGEMM_DEFAULT_UNROLL_M 1
2021 #define XGEMM_DEFAULT_UNROLL_M 1
2023 #define SGEMM_DEFAULT_UNROLL_M 8
2024 #define DGEMM_DEFAULT_UNROLL_M 4
2025 #define QGEMM_DEFAULT_UNROLL_M 2
2026 #define CGEMM_DEFAULT_UNROLL_M 4
2027 #define ZGEMM_DEFAULT_UNROLL_M 2
2028 #define XGEMM_DEFAULT_UNROLL_M 1
2031 #define SGEMM_DEFAULT_UNROLL_N 4
2032 #define DGEMM_DEFAULT_UNROLL_N 2
2033 #define QGEMM_DEFAULT_UNROLL_N 2
2034 #define CGEMM_DEFAULT_UNROLL_N 2
2035 #define ZGEMM_DEFAULT_UNROLL_N 1
2036 #define XGEMM_DEFAULT_UNROLL_N 1
2038 #define SGEMM_DEFAULT_P sgemm_p
2039 #define SGEMM_DEFAULT_R sgemm_r
2041 #define DGEMM_DEFAULT_P dgemm_p
2042 #define DGEMM_DEFAULT_R dgemm_r
2044 #define QGEMM_DEFAULT_P qgemm_p
2045 #define QGEMM_DEFAULT_R qgemm_r
2047 #define CGEMM_DEFAULT_P cgemm_p
2048 #define CGEMM_DEFAULT_R cgemm_r
2050 #define ZGEMM_DEFAULT_P zgemm_p
2051 #define ZGEMM_DEFAULT_R zgemm_r
2053 #define XGEMM_DEFAULT_P xgemm_p
2054 #define XGEMM_DEFAULT_R xgemm_r
2056 #define SGEMM_DEFAULT_Q 256
2057 #define DGEMM_DEFAULT_Q 256
2058 #define QGEMM_DEFAULT_Q 256
2059 #define CGEMM_DEFAULT_Q 256
2060 #define ZGEMM_DEFAULT_Q 256
2061 #define XGEMM_DEFAULT_Q 256
2071 #define GEMM_DEFAULT_OFFSET_A 0
2072 #define GEMM_DEFAULT_OFFSET_B 128
2073 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2075 #define SGEMM_DEFAULT_UNROLL_M 8
2076 #define SGEMM_DEFAULT_UNROLL_N 8
2077 #define DGEMM_DEFAULT_UNROLL_M 8
2078 #define DGEMM_DEFAULT_UNROLL_N 8
2079 #define QGEMM_DEFAULT_UNROLL_M 8
2080 #define QGEMM_DEFAULT_UNROLL_N 8
2081 #define CGEMM_DEFAULT_UNROLL_M 4
2082 #define CGEMM_DEFAULT_UNROLL_N 4
2083 #define ZGEMM_DEFAULT_UNROLL_M 4
2084 #define ZGEMM_DEFAULT_UNROLL_N 4
2085 #define XGEMM_DEFAULT_UNROLL_M 4
2086 #define XGEMM_DEFAULT_UNROLL_N 4
2088 #define SGEMM_DEFAULT_P sgemm_p
2089 #define DGEMM_DEFAULT_P dgemm_p
2090 #define QGEMM_DEFAULT_P qgemm_p
2091 #define CGEMM_DEFAULT_P cgemm_p
2092 #define ZGEMM_DEFAULT_P zgemm_p
2093 #define XGEMM_DEFAULT_P xgemm_p
2095 #define SGEMM_DEFAULT_Q 1024
2096 #define DGEMM_DEFAULT_Q 1024
2097 #define QGEMM_DEFAULT_Q 1024
2098 #define CGEMM_DEFAULT_Q 1024
2099 #define ZGEMM_DEFAULT_Q 1024
2100 #define XGEMM_DEFAULT_Q 1024
2102 #define SGEMM_DEFAULT_R sgemm_r
2103 #define DGEMM_DEFAULT_R dgemm_r
2104 #define QGEMM_DEFAULT_R qgemm_r
2105 #define CGEMM_DEFAULT_R cgemm_r
2106 #define ZGEMM_DEFAULT_R zgemm_r
2107 #define XGEMM_DEFAULT_R xgemm_r
2111 #define GETRF_FACTOR 0.65
2115 #if defined(EV4) || defined(EV5) || defined(EV6)
2125 #define GEMM_DEFAULT_OFFSET_A 512
2126 #define GEMM_DEFAULT_OFFSET_B 512
2127 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2129 #define SGEMM_DEFAULT_UNROLL_M 4
2130 #define SGEMM_DEFAULT_UNROLL_N 4
2131 #define DGEMM_DEFAULT_UNROLL_M 4
2132 #define DGEMM_DEFAULT_UNROLL_N 4
2133 #define CGEMM_DEFAULT_UNROLL_M 2
2134 #define CGEMM_DEFAULT_UNROLL_N 2
2135 #define ZGEMM_DEFAULT_UNROLL_M 2
2136 #define ZGEMM_DEFAULT_UNROLL_N 2
2141 #define SGEMM_DEFAULT_P 32
2142 #define SGEMM_DEFAULT_Q 112
2143 #define SGEMM_DEFAULT_R 256
2145 #define DGEMM_DEFAULT_P 32
2146 #define DGEMM_DEFAULT_Q 56
2147 #define DGEMM_DEFAULT_R 256
2149 #define CGEMM_DEFAULT_P 32
2150 #define CGEMM_DEFAULT_Q 64
2151 #define CGEMM_DEFAULT_R 240
2153 #define ZGEMM_DEFAULT_P 32
2154 #define ZGEMM_DEFAULT_Q 32
2155 #define ZGEMM_DEFAULT_R 240
2159 #define SGEMM_DEFAULT_P 64
2160 #define SGEMM_DEFAULT_Q 256
2162 #define DGEMM_DEFAULT_P 64
2163 #define DGEMM_DEFAULT_Q 128
2165 #define CGEMM_DEFAULT_P 64
2166 #define CGEMM_DEFAULT_Q 128
2168 #define ZGEMM_DEFAULT_P 64
2169 #define ZGEMM_DEFAULT_Q 64
2173 #define SGEMM_DEFAULT_P 256
2174 #define SGEMM_DEFAULT_Q 512
2176 #define DGEMM_DEFAULT_P 256
2177 #define DGEMM_DEFAULT_Q 256
2179 #define CGEMM_DEFAULT_P 256
2180 #define CGEMM_DEFAULT_Q 256
2182 #define ZGEMM_DEFAULT_P 128
2183 #define ZGEMM_DEFAULT_Q 256
2193 #define GEMM_DEFAULT_OFFSET_A 0
2194 #define GEMM_DEFAULT_OFFSET_B 8192
2195 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2197 #define SGEMM_DEFAULT_UNROLL_M 16
2198 #define SGEMM_DEFAULT_UNROLL_N 4
2199 #define DGEMM_DEFAULT_UNROLL_M 4
2200 #define DGEMM_DEFAULT_UNROLL_N 4
2201 #define CGEMM_DEFAULT_UNROLL_M 8
2202 #define CGEMM_DEFAULT_UNROLL_N 2
2203 #define ZGEMM_DEFAULT_UNROLL_M 2
2204 #define ZGEMM_DEFAULT_UNROLL_N 2
2206 #define SGEMM_DEFAULT_P 128
2207 #define DGEMM_DEFAULT_P 128
2208 #define CGEMM_DEFAULT_P 128
2209 #define ZGEMM_DEFAULT_P 128
2211 #define SGEMM_DEFAULT_Q 512
2212 #define DGEMM_DEFAULT_Q 256
2213 #define CGEMM_DEFAULT_Q 256
2214 #define ZGEMM_DEFAULT_Q 128
2220 #define GEMM_DEFAULT_OFFSET_A 0
2221 #define GEMM_DEFAULT_OFFSET_B 1024
2222 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2224 #define SGEMM_DEFAULT_UNROLL_M 16
2225 #define SGEMM_DEFAULT_UNROLL_N 4
2226 #define DGEMM_DEFAULT_UNROLL_M 4
2227 #define DGEMM_DEFAULT_UNROLL_N 4
2228 #define CGEMM_DEFAULT_UNROLL_M 2
2229 #define CGEMM_DEFAULT_UNROLL_N 2
2230 #define ZGEMM_DEFAULT_UNROLL_M 2
2231 #define ZGEMM_DEFAULT_UNROLL_N 2
2233 #define SGEMM_DEFAULT_P 256
2234 #define DGEMM_DEFAULT_P 128
2235 #define CGEMM_DEFAULT_P 128
2236 #define ZGEMM_DEFAULT_P 64
2238 #define SGEMM_DEFAULT_Q 256
2239 #define DGEMM_DEFAULT_Q 256
2240 #define CGEMM_DEFAULT_Q 256
2241 #define ZGEMM_DEFAULT_Q 256
2251 #define GEMM_DEFAULT_OFFSET_A 2688
2252 #define GEMM_DEFAULT_OFFSET_B 3072
2253 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2255 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2256 #define SGEMM_DEFAULT_UNROLL_M 4
2258 #define SGEMM_DEFAULT_UNROLL_M 16
2260 #define SGEMM_DEFAULT_UNROLL_N 4
2261 #define DGEMM_DEFAULT_UNROLL_M 4
2262 #define DGEMM_DEFAULT_UNROLL_N 4
2263 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2264 #define CGEMM_DEFAULT_UNROLL_M 2
2266 #define CGEMM_DEFAULT_UNROLL_M 8
2268 #define CGEMM_DEFAULT_UNROLL_N 2
2269 #define ZGEMM_DEFAULT_UNROLL_M 2
2270 #define ZGEMM_DEFAULT_UNROLL_N 2
2272 #if defined(OS_LINUX) || defined(OS_DARWIN) || defined(OS_FREEBSD)
2273 #if L2_SIZE == 1024976
2274 #define SGEMM_DEFAULT_P 320
2275 #define DGEMM_DEFAULT_P 256
2276 #define CGEMM_DEFAULT_P 256
2277 #define ZGEMM_DEFAULT_P 256
2279 #define SGEMM_DEFAULT_P 176
2280 #define DGEMM_DEFAULT_P 176
2281 #define CGEMM_DEFAULT_P 176
2282 #define ZGEMM_DEFAULT_P 176
2286 #define SGEMM_DEFAULT_Q 512
2287 #define DGEMM_DEFAULT_Q 256
2288 #define CGEMM_DEFAULT_Q 256
2289 #define ZGEMM_DEFAULT_Q 128
2300 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2301 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2302 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2304 #define SGEMM_DEFAULT_UNROLL_M 4
2305 #define SGEMM_DEFAULT_UNROLL_N 4
2306 #define DGEMM_DEFAULT_UNROLL_M 4
2307 #define DGEMM_DEFAULT_UNROLL_N 4
2308 #define CGEMM_DEFAULT_UNROLL_M 2
2309 #define CGEMM_DEFAULT_UNROLL_N 2
2310 #define ZGEMM_DEFAULT_UNROLL_M 2
2311 #define ZGEMM_DEFAULT_UNROLL_N 2
2313 #define SGEMM_DEFAULT_P 512
2314 #define DGEMM_DEFAULT_P 512
2315 #define CGEMM_DEFAULT_P 512
2316 #define ZGEMM_DEFAULT_P 512
2318 #define SGEMM_DEFAULT_Q 1024
2319 #define DGEMM_DEFAULT_Q 512
2320 #define CGEMM_DEFAULT_Q 512
2321 #define ZGEMM_DEFAULT_Q 256
2323 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2324 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2325 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2326 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2336 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2337 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2338 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2340 #define SGEMM_DEFAULT_UNROLL_M 8
2341 #define SGEMM_DEFAULT_UNROLL_N 4
2342 #define DGEMM_DEFAULT_UNROLL_M 8
2343 #define DGEMM_DEFAULT_UNROLL_N 4
2344 #define CGEMM_DEFAULT_UNROLL_M 4
2345 #define CGEMM_DEFAULT_UNROLL_N 2
2346 #define ZGEMM_DEFAULT_UNROLL_M 4
2347 #define ZGEMM_DEFAULT_UNROLL_N 2
2349 #define SGEMM_DEFAULT_P 128
2350 #define DGEMM_DEFAULT_P 128
2351 #define CGEMM_DEFAULT_P 128
2352 #define ZGEMM_DEFAULT_P 128
2354 #define SGEMM_DEFAULT_Q 4096
2355 #define DGEMM_DEFAULT_Q 3072
2356 #define CGEMM_DEFAULT_Q 2048
2357 #define ZGEMM_DEFAULT_Q 1024
2359 #define SGEMM_DEFAULT_Q 512
2360 #define DGEMM_DEFAULT_Q 256
2361 #define CGEMM_DEFAULT_Q 256
2362 #define ZGEMM_DEFAULT_Q 128
2370 #if defined(POWER3) || defined(POWER4) || defined(POWER5)
2371 #define GEMM_DEFAULT_OFFSET_A 0
2372 #define GEMM_DEFAULT_OFFSET_B 2048
2373 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2375 #define SGEMM_DEFAULT_UNROLL_M 4
2376 #define SGEMM_DEFAULT_UNROLL_N 4
2377 #define DGEMM_DEFAULT_UNROLL_M 4
2378 #define DGEMM_DEFAULT_UNROLL_N 4
2379 #define CGEMM_DEFAULT_UNROLL_M 2
2380 #define CGEMM_DEFAULT_UNROLL_N 2
2381 #define ZGEMM_DEFAULT_UNROLL_M 2
2382 #define ZGEMM_DEFAULT_UNROLL_N 2
2389 #define SGEMM_DEFAULT_P 256
2390 #define SGEMM_DEFAULT_Q 432
2391 #define SGEMM_DEFAULT_R 1012
2393 #define DGEMM_DEFAULT_P 256
2394 #define DGEMM_DEFAULT_Q 216
2395 #define DGEMM_DEFAULT_R 1012
2397 #define CGEMM_DEFAULT_P 256
2398 #define CGEMM_DEFAULT_Q 104
2399 #define CGEMM_DEFAULT_R 1012
2401 #define ZGEMM_DEFAULT_P 256
2402 #define ZGEMM_DEFAULT_Q 104
2403 #define ZGEMM_DEFAULT_R 1012
2407 #ifdef ALLOC_HUGETLB
2408 #define SGEMM_DEFAULT_P 184
2409 #define DGEMM_DEFAULT_P 184
2410 #define CGEMM_DEFAULT_P 184
2411 #define ZGEMM_DEFAULT_P 184
2413 #define SGEMM_DEFAULT_P 144
2414 #define DGEMM_DEFAULT_P 144
2415 #define CGEMM_DEFAULT_P 144
2416 #define ZGEMM_DEFAULT_P 144
2419 #define SGEMM_DEFAULT_Q 256
2420 #define CGEMM_DEFAULT_Q 256
2421 #define DGEMM_DEFAULT_Q 256
2422 #define ZGEMM_DEFAULT_Q 256
2426 #ifdef ALLOC_HUGETLB
2427 #define SGEMM_DEFAULT_P 512
2428 #define DGEMM_DEFAULT_P 256
2429 #define CGEMM_DEFAULT_P 256
2430 #define ZGEMM_DEFAULT_P 128
2432 #define SGEMM_DEFAULT_P 320
2433 #define DGEMM_DEFAULT_P 160
2434 #define CGEMM_DEFAULT_P 160
2435 #define ZGEMM_DEFAULT_P 80
2438 #define SGEMM_DEFAULT_Q 256
2439 #define CGEMM_DEFAULT_Q 256
2440 #define DGEMM_DEFAULT_Q 256
2441 #define ZGEMM_DEFAULT_Q 256
2453 #define GEMM_DEFAULT_OFFSET_A 384
2454 #define GEMM_DEFAULT_OFFSET_B 1024
2455 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2457 #define SGEMM_DEFAULT_UNROLL_M 4
2458 #define SGEMM_DEFAULT_UNROLL_N 4
2459 #define DGEMM_DEFAULT_UNROLL_M 4
2460 #define DGEMM_DEFAULT_UNROLL_N 4
2461 #define CGEMM_DEFAULT_UNROLL_M 2
2462 #define CGEMM_DEFAULT_UNROLL_N 4
2463 #define ZGEMM_DEFAULT_UNROLL_M 2
2464 #define ZGEMM_DEFAULT_UNROLL_N 4
2466 #define SGEMM_DEFAULT_P 992
2467 #define DGEMM_DEFAULT_P 480
2468 #define CGEMM_DEFAULT_P 488
2469 #define ZGEMM_DEFAULT_P 248
2471 #define SGEMM_DEFAULT_Q 504
2472 #define DGEMM_DEFAULT_Q 504
2473 #define CGEMM_DEFAULT_Q 400
2474 #define ZGEMM_DEFAULT_Q 400
2485 #define GEMM_DEFAULT_OFFSET_A 0
2486 #define GEMM_DEFAULT_OFFSET_B 65536
2488 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2489 #if defined(__32BIT__)
2490 #warning using BINARY32==POWER6
2491 #define SGEMM_DEFAULT_UNROLL_M 4
2492 #define SGEMM_DEFAULT_UNROLL_N 4
2493 #define DGEMM_DEFAULT_UNROLL_M 4
2494 #define DGEMM_DEFAULT_UNROLL_N 4
2495 #define CGEMM_DEFAULT_UNROLL_M 2
2496 #define CGEMM_DEFAULT_UNROLL_N 4
2497 #define ZGEMM_DEFAULT_UNROLL_M 2
2498 #define ZGEMM_DEFAULT_UNROLL_N 4
2500 #define SGEMM_DEFAULT_UNROLL_M 16
2501 #define SGEMM_DEFAULT_UNROLL_N 8
2502 #define DGEMM_DEFAULT_UNROLL_M 16
2503 #define DGEMM_DEFAULT_UNROLL_N 4
2504 #define CGEMM_DEFAULT_UNROLL_M 8
2505 #define CGEMM_DEFAULT_UNROLL_N 4
2506 #define ZGEMM_DEFAULT_UNROLL_M 8
2507 #define ZGEMM_DEFAULT_UNROLL_N 2
2509 #define SGEMM_DEFAULT_P 1280UL
2510 #define DGEMM_DEFAULT_P 640UL
2511 #define CGEMM_DEFAULT_P 640UL
2512 #define ZGEMM_DEFAULT_P 320UL
2514 #define SGEMM_DEFAULT_Q 640UL
2515 #define DGEMM_DEFAULT_Q 720UL
2516 #define CGEMM_DEFAULT_Q 640UL
2517 #define ZGEMM_DEFAULT_Q 640UL
2520 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2521 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2522 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2523 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2525 #define SGEMM_DEFAULT_R 4096
2526 #define DGEMM_DEFAULT_R 4096
2527 #define CGEMM_DEFAULT_R 4096
2528 #define ZGEMM_DEFAULT_R 4096
2539 #define GEMM_DEFAULT_OFFSET_A 0
2540 #define GEMM_DEFAULT_OFFSET_B 65536
2541 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2543 #define SWITCH_RATIO 16
2544 #define GEMM_PREFERED_SIZE 16
2546 #define SGEMM_DEFAULT_UNROLL_M 16
2547 #define SGEMM_DEFAULT_UNROLL_N 8
2548 #define DGEMM_DEFAULT_UNROLL_M 16
2549 #define DGEMM_DEFAULT_UNROLL_N 4
2550 #define CGEMM_DEFAULT_UNROLL_M 8
2551 #define CGEMM_DEFAULT_UNROLL_N 4
2552 #define ZGEMM_DEFAULT_UNROLL_M 8
2553 #define ZGEMM_DEFAULT_UNROLL_N 2
2555 #define SGEMM_DEFAULT_P 832
2556 #define DGEMM_DEFAULT_P 128
2557 #define CGEMM_DEFAULT_P 512
2558 #define ZGEMM_DEFAULT_P 256
2560 #define SGEMM_DEFAULT_Q 1026
2561 #define DGEMM_DEFAULT_Q 384
2562 #define CGEMM_DEFAULT_Q 1026
2563 #define ZGEMM_DEFAULT_Q 1026
2565 #define SGEMM_DEFAULT_R 4096
2566 #define DGEMM_DEFAULT_R 4096
2567 #define CGEMM_DEFAULT_R 4096
2568 #define ZGEMM_DEFAULT_R 4096
2574 #if defined(POWER10)
2578 #define GEMM_DEFAULT_OFFSET_A 0
2579 #define GEMM_DEFAULT_OFFSET_B 65536
2580 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2582 #define SWITCH_RATIO 16
2583 #define GEMM_PREFERED_SIZE 16
2585 #define SGEMM_DEFAULT_UNROLL_M 16
2586 #define SGEMM_DEFAULT_UNROLL_N 8
2587 #define DGEMM_DEFAULT_UNROLL_M 8
2588 #define DGEMM_DEFAULT_UNROLL_N 8
2589 #define CGEMM_DEFAULT_UNROLL_M 8
2590 #define CGEMM_DEFAULT_UNROLL_N 4
2591 #define ZGEMM_DEFAULT_UNROLL_M 8
2592 #define ZGEMM_DEFAULT_UNROLL_N 2
2594 #define SGEMM_DEFAULT_P 512
2595 #define DGEMM_DEFAULT_P 384
2596 #define CGEMM_DEFAULT_P 512
2597 #define ZGEMM_DEFAULT_P 256
2599 #define SGEMM_DEFAULT_Q 512
2600 #define DGEMM_DEFAULT_Q 512
2601 #define CGEMM_DEFAULT_Q 384
2602 #define ZGEMM_DEFAULT_Q 384
2604 #define SGEMM_DEFAULT_R 4096
2605 #define DGEMM_DEFAULT_R 4096
2606 #define CGEMM_DEFAULT_R 4096
2607 #define ZGEMM_DEFAULT_R 4096
2611 #undef SBGEMM_DEFAULT_UNROLL_N
2612 #undef SBGEMM_DEFAULT_UNROLL_M
2613 #undef SBGEMM_DEFAULT_P
2614 #undef SBGEMM_DEFAULT_R
2615 #undef SBGEMM_DEFAULT_Q
2616 #define SBGEMM_DEFAULT_UNROLL_M 16
2617 #define SBGEMM_DEFAULT_UNROLL_N 8
2618 #define SBGEMM_DEFAULT_P 832
2619 #define SBGEMM_DEFAULT_Q 1026
2620 #define SBGEMM_DEFAULT_R 4096
2623 #if defined(SPARC) && defined(V7)
2628 #define GEMM_DEFAULT_OFFSET_A 0
2629 #define GEMM_DEFAULT_OFFSET_B 2048
2630 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2632 #define SGEMM_DEFAULT_UNROLL_M 2
2633 #define SGEMM_DEFAULT_UNROLL_N 8
2634 #define DGEMM_DEFAULT_UNROLL_M 2
2635 #define DGEMM_DEFAULT_UNROLL_N 8
2636 #define CGEMM_DEFAULT_UNROLL_M 1
2637 #define CGEMM_DEFAULT_UNROLL_N 4
2638 #define ZGEMM_DEFAULT_UNROLL_M 1
2639 #define ZGEMM_DEFAULT_UNROLL_N 4
2641 #define SGEMM_DEFAULT_P 256
2642 #define DGEMM_DEFAULT_P 256
2643 #define CGEMM_DEFAULT_P 256
2644 #define ZGEMM_DEFAULT_P 256
2646 #define SGEMM_DEFAULT_Q 512
2647 #define DGEMM_DEFAULT_Q 256
2648 #define CGEMM_DEFAULT_Q 256
2649 #define ZGEMM_DEFAULT_Q 128
2652 #define GEMM_THREAD gemm_thread_mn
2655 #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
2660 #define GEMM_DEFAULT_OFFSET_A 0
2661 #define GEMM_DEFAULT_OFFSET_B 2048
2662 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2664 #define SGEMM_DEFAULT_UNROLL_M 4
2665 #define SGEMM_DEFAULT_UNROLL_N 4
2666 #define DGEMM_DEFAULT_UNROLL_M 4
2667 #define DGEMM_DEFAULT_UNROLL_N 4
2668 #define CGEMM_DEFAULT_UNROLL_M 2
2669 #define CGEMM_DEFAULT_UNROLL_N 2
2670 #define ZGEMM_DEFAULT_UNROLL_M 2
2671 #define ZGEMM_DEFAULT_UNROLL_N 2
2673 #define SGEMM_DEFAULT_P 512
2674 #define DGEMM_DEFAULT_P 512
2675 #define CGEMM_DEFAULT_P 512
2676 #define ZGEMM_DEFAULT_P 512
2678 #define SGEMM_DEFAULT_Q 1024
2679 #define DGEMM_DEFAULT_Q 512
2680 #define CGEMM_DEFAULT_Q 512
2681 #define ZGEMM_DEFAULT_Q 256
2691 #define GEMM_DEFAULT_OFFSET_A 0
2692 #define GEMM_DEFAULT_OFFSET_B 0
2693 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2695 #define SGEMM_DEFAULT_UNROLL_M 2
2696 #define SGEMM_DEFAULT_UNROLL_N 8
2697 #define DGEMM_DEFAULT_UNROLL_M 2
2698 #define DGEMM_DEFAULT_UNROLL_N 8
2699 #define CGEMM_DEFAULT_UNROLL_M 1
2700 #define CGEMM_DEFAULT_UNROLL_N 4
2701 #define ZGEMM_DEFAULT_UNROLL_M 1
2702 #define ZGEMM_DEFAULT_UNROLL_N 4
2704 #define SGEMM_DEFAULT_P 108
2705 #define DGEMM_DEFAULT_P 112
2706 #define CGEMM_DEFAULT_P 108
2707 #define ZGEMM_DEFAULT_P 112
2709 #define SGEMM_DEFAULT_Q 288
2710 #define DGEMM_DEFAULT_Q 144
2711 #define CGEMM_DEFAULT_Q 144
2712 #define ZGEMM_DEFAULT_Q 72
2714 #define SGEMM_DEFAULT_R 2000
2715 #define DGEMM_DEFAULT_R 2000
2716 #define CGEMM_DEFAULT_R 2000
2717 #define ZGEMM_DEFAULT_R 2000
2722 #if defined(LOONGSON3R4)
2726 #define GEMM_DEFAULT_OFFSET_A 0
2727 #define GEMM_DEFAULT_OFFSET_B 0
2728 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2731 #define SGEMM_DEFAULT_UNROLL_M 8
2732 #define SGEMM_DEFAULT_UNROLL_N 8
2734 #define DGEMM_DEFAULT_UNROLL_M 8
2735 #define DGEMM_DEFAULT_UNROLL_N 4
2737 #define CGEMM_DEFAULT_UNROLL_M 8
2738 #define CGEMM_DEFAULT_UNROLL_N 4
2740 #define ZGEMM_DEFAULT_UNROLL_M 4
2741 #define ZGEMM_DEFAULT_UNROLL_N 4
2743 #define SGEMM_DEFAULT_UNROLL_M 8
2744 #define SGEMM_DEFAULT_UNROLL_N 4
2746 #define DGEMM_DEFAULT_UNROLL_M 4
2747 #define DGEMM_DEFAULT_UNROLL_N 4
2749 #define CGEMM_DEFAULT_UNROLL_M 4
2750 #define CGEMM_DEFAULT_UNROLL_N 2
2752 #define ZGEMM_DEFAULT_UNROLL_M 2
2753 #define ZGEMM_DEFAULT_UNROLL_N 2
2756 #define SGEMM_DEFAULT_P 64
2757 #define DGEMM_DEFAULT_P 44
2758 #define CGEMM_DEFAULT_P 64
2759 #define ZGEMM_DEFAULT_P 32
2761 #define SGEMM_DEFAULT_Q 192
2762 #define DGEMM_DEFAULT_Q 92
2763 #define CGEMM_DEFAULT_Q 128
2764 #define ZGEMM_DEFAULT_Q 80
2766 #define SGEMM_DEFAULT_R 640
2767 #define DGEMM_DEFAULT_R dgemm_r
2768 #define CGEMM_DEFAULT_R 640
2769 #define ZGEMM_DEFAULT_R 640
2771 #define GEMM_OFFSET_A1 0x10000
2772 #define GEMM_OFFSET_B1 0x100000
2777 #if defined(LOONGSON3R3)
2778 ////Copy from SICORTEX
2782 #define GEMM_DEFAULT_OFFSET_A 0
2783 #define GEMM_DEFAULT_OFFSET_B 0
2784 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2786 #define SGEMM_DEFAULT_UNROLL_M 8
2787 #define SGEMM_DEFAULT_UNROLL_N 4
2789 #define DGEMM_DEFAULT_UNROLL_M 4
2790 #define DGEMM_DEFAULT_UNROLL_N 4
2792 #define CGEMM_DEFAULT_UNROLL_M 4
2793 #define CGEMM_DEFAULT_UNROLL_N 2
2795 #define ZGEMM_DEFAULT_UNROLL_M 2
2796 #define ZGEMM_DEFAULT_UNROLL_N 2
2798 #define SGEMM_DEFAULT_P 64
2799 #define DGEMM_DEFAULT_P 44
2800 #define CGEMM_DEFAULT_P 64
2801 #define ZGEMM_DEFAULT_P 32
2803 #define SGEMM_DEFAULT_Q 192
2804 #define DGEMM_DEFAULT_Q 92
2805 #define CGEMM_DEFAULT_Q 128
2806 #define ZGEMM_DEFAULT_Q 80
2808 #define SGEMM_DEFAULT_R 640
2809 #define DGEMM_DEFAULT_R dgemm_r
2810 #define CGEMM_DEFAULT_R 640
2811 #define ZGEMM_DEFAULT_R 640
2813 #define GEMM_OFFSET_A1 0x10000
2814 #define GEMM_OFFSET_B1 0x100000
2819 #if defined (LOONGSON3R5)
2823 #define GEMM_DEFAULT_OFFSET_A 0
2824 #define GEMM_DEFAULT_OFFSET_B 0
2825 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2827 #define SGEMM_DEFAULT_UNROLL_N 8
2828 #define DGEMM_DEFAULT_UNROLL_N 8
2829 #define QGEMM_DEFAULT_UNROLL_N 2
2830 #define CGEMM_DEFAULT_UNROLL_N 4
2831 #define ZGEMM_DEFAULT_UNROLL_N 4
2832 #define XGEMM_DEFAULT_UNROLL_N 1
2834 #define SGEMM_DEFAULT_UNROLL_M 2
2835 #define DGEMM_DEFAULT_UNROLL_M 2
2836 #define QGEMM_DEFAULT_UNROLL_M 2
2837 #define CGEMM_DEFAULT_UNROLL_M 1
2838 #define ZGEMM_DEFAULT_UNROLL_M 1
2839 #define XGEMM_DEFAULT_UNROLL_M 1
2841 #define SGEMM_DEFAULT_P sgemm_p
2842 #define DGEMM_DEFAULT_P dgemm_p
2843 #define QGEMM_DEFAULT_P qgemm_p
2844 #define CGEMM_DEFAULT_P cgemm_p
2845 #define ZGEMM_DEFAULT_P zgemm_p
2846 #define XGEMM_DEFAULT_P xgemm_p
2848 #define SGEMM_DEFAULT_R sgemm_r
2849 #define DGEMM_DEFAULT_R dgemm_r
2850 #define QGEMM_DEFAULT_R qgemm_r
2851 #define CGEMM_DEFAULT_R cgemm_r
2852 #define ZGEMM_DEFAULT_R zgemm_r
2853 #define XGEMM_DEFAULT_R xgemm_r
2855 #define SGEMM_DEFAULT_Q 128
2856 #define DGEMM_DEFAULT_Q 128
2857 #define QGEMM_DEFAULT_Q 128
2858 #define CGEMM_DEFAULT_Q 128
2859 #define ZGEMM_DEFAULT_Q 128
2860 #define XGEMM_DEFAULT_Q 128
2865 #if defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
2869 #define GEMM_DEFAULT_OFFSET_A 0
2870 #define GEMM_DEFAULT_OFFSET_B 0
2871 #define GEMM_DEFAULT_ALIGN (BLASLONG) 0x03fffUL
2874 #define SGEMM_DEFAULT_UNROLL_M 8
2875 #define SGEMM_DEFAULT_UNROLL_N 8
2877 #define DGEMM_DEFAULT_UNROLL_M 8
2878 #define DGEMM_DEFAULT_UNROLL_N 4
2880 #define CGEMM_DEFAULT_UNROLL_M 8
2881 #define CGEMM_DEFAULT_UNROLL_N 4
2883 #define ZGEMM_DEFAULT_UNROLL_M 4
2884 #define ZGEMM_DEFAULT_UNROLL_N 4
2886 #define SGEMM_DEFAULT_UNROLL_M 2
2887 #define SGEMM_DEFAULT_UNROLL_N 2
2889 #define DGEMM_DEFAULT_UNROLL_M 2
2890 #define DGEMM_DEFAULT_UNROLL_N 2
2892 #define CGEMM_DEFAULT_UNROLL_M 2
2893 #define CGEMM_DEFAULT_UNROLL_N 2
2895 #define ZGEMM_DEFAULT_UNROLL_M 2
2896 #define ZGEMM_DEFAULT_UNROLL_N 2
2899 #define SGEMM_DEFAULT_P 128
2900 #define DGEMM_DEFAULT_P 128
2901 #define CGEMM_DEFAULT_P 96
2902 #define ZGEMM_DEFAULT_P 64
2904 #define SGEMM_DEFAULT_Q 240
2905 #define DGEMM_DEFAULT_Q 120
2906 #define CGEMM_DEFAULT_Q 120
2907 #define ZGEMM_DEFAULT_Q 120
2909 #define SGEMM_DEFAULT_R 12288
2910 #define DGEMM_DEFAULT_R 8192
2911 #define CGEMM_DEFAULT_R 4096
2912 #define ZGEMM_DEFAULT_R 4096
2917 #ifdef RISCV64_GENERIC
2918 #define GEMM_DEFAULT_OFFSET_A 0
2919 #define GEMM_DEFAULT_OFFSET_B 0
2920 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2922 #define SGEMM_DEFAULT_UNROLL_M 2
2923 #define SGEMM_DEFAULT_UNROLL_N 2
2925 #define DGEMM_DEFAULT_UNROLL_M 2
2926 #define DGEMM_DEFAULT_UNROLL_N 2
2928 #define CGEMM_DEFAULT_UNROLL_M 2
2929 #define CGEMM_DEFAULT_UNROLL_N 2
2931 #define ZGEMM_DEFAULT_UNROLL_M 2
2932 #define ZGEMM_DEFAULT_UNROLL_N 2
2934 #define SGEMM_DEFAULT_P 128
2935 #define DGEMM_DEFAULT_P 128
2936 #define CGEMM_DEFAULT_P 96
2937 #define ZGEMM_DEFAULT_P 64
2939 #define SGEMM_DEFAULT_Q 240
2940 #define DGEMM_DEFAULT_Q 120
2941 #define CGEMM_DEFAULT_Q 120
2942 #define ZGEMM_DEFAULT_Q 120
2944 #define SGEMM_DEFAULT_R 12288
2945 #define DGEMM_DEFAULT_R 8192
2946 #define CGEMM_DEFAULT_R 4096
2947 #define ZGEMM_DEFAULT_R 4096
2951 #define GEMM_DEFAULT_OFFSET_A 0
2952 #define GEMM_DEFAULT_OFFSET_B 0
2957 #define GEMM_DEFAULT_OFFSET_A 0
2958 #define GEMM_DEFAULT_OFFSET_B 0
2959 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2961 #define SGEMM_DEFAULT_UNROLL_M 16
2962 #define SGEMM_DEFAULT_UNROLL_N 4
2964 #define DGEMM_DEFAULT_UNROLL_M 8
2965 #define DGEMM_DEFAULT_UNROLL_N 4
2967 #define CGEMM_DEFAULT_UNROLL_M 2
2968 #define CGEMM_DEFAULT_UNROLL_N 2
2970 #define ZGEMM_DEFAULT_UNROLL_M 2
2971 #define ZGEMM_DEFAULT_UNROLL_N 2
2973 #define SGEMM_DEFAULT_P 160
2974 #define DGEMM_DEFAULT_P 160
2975 #define CGEMM_DEFAULT_P 96
2976 #define ZGEMM_DEFAULT_P 64
2978 #define SGEMM_DEFAULT_Q 240
2979 #define DGEMM_DEFAULT_Q 128
2980 #define CGEMM_DEFAULT_Q 120
2981 #define ZGEMM_DEFAULT_Q 120
2983 #define SGEMM_DEFAULT_R 12288
2984 #define DGEMM_DEFAULT_R 8192
2985 #define CGEMM_DEFAULT_R 4096
2986 #define ZGEMM_DEFAULT_R 4096
2990 #define GEMM_DEFAULT_OFFSET_A 0
2991 #define GEMM_DEFAULT_OFFSET_B 0
2999 #define GEMM_DEFAULT_OFFSET_A 0
3000 #define GEMM_DEFAULT_OFFSET_B 0
3001 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3003 #define SGEMM_DEFAULT_UNROLL_M 4
3004 #define SGEMM_DEFAULT_UNROLL_N 4
3006 #define DGEMM_DEFAULT_UNROLL_M 4
3007 #define DGEMM_DEFAULT_UNROLL_N 4
3009 #define CGEMM_DEFAULT_UNROLL_M 2
3010 #define CGEMM_DEFAULT_UNROLL_N 2
3012 #define ZGEMM_DEFAULT_UNROLL_M 2
3013 #define ZGEMM_DEFAULT_UNROLL_N 2
3015 #define SGEMM_DEFAULT_P 128
3016 #define DGEMM_DEFAULT_P 128
3017 #define CGEMM_DEFAULT_P 96
3018 #define ZGEMM_DEFAULT_P 64
3020 #define SGEMM_DEFAULT_Q 240
3021 #define DGEMM_DEFAULT_Q 120
3022 #define CGEMM_DEFAULT_Q 120
3023 #define ZGEMM_DEFAULT_Q 120
3025 #define SGEMM_DEFAULT_R 12288
3026 #define DGEMM_DEFAULT_R 8192
3027 #define CGEMM_DEFAULT_R 4096
3028 #define ZGEMM_DEFAULT_R 4096
3040 #define GEMM_DEFAULT_OFFSET_A 0
3041 #define GEMM_DEFAULT_OFFSET_B 0
3042 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3044 #define SGEMM_DEFAULT_UNROLL_M 4
3045 #define SGEMM_DEFAULT_UNROLL_N 2
3047 #define DGEMM_DEFAULT_UNROLL_M 4
3048 #define DGEMM_DEFAULT_UNROLL_N 2
3050 #define CGEMM_DEFAULT_UNROLL_M 2
3051 #define CGEMM_DEFAULT_UNROLL_N 2
3053 #define ZGEMM_DEFAULT_UNROLL_M 2
3054 #define ZGEMM_DEFAULT_UNROLL_N 2
3056 #define SGEMM_DEFAULT_P 128
3057 #define DGEMM_DEFAULT_P 128
3058 #define CGEMM_DEFAULT_P 96
3059 #define ZGEMM_DEFAULT_P 64
3061 #define SGEMM_DEFAULT_Q 240
3062 #define DGEMM_DEFAULT_Q 120
3063 #define CGEMM_DEFAULT_Q 120
3064 #define ZGEMM_DEFAULT_Q 120
3066 #define SGEMM_DEFAULT_R 12288
3067 #define DGEMM_DEFAULT_R 8192
3068 #define CGEMM_DEFAULT_R 4096
3069 #define ZGEMM_DEFAULT_R 4096
3075 /* Common ARMv8 parameters */
3081 #define GEMM_DEFAULT_OFFSET_A 0
3082 #define GEMM_DEFAULT_OFFSET_B 0
3084 /* Use explicit casting for win64 as LLP64 datamodel is used */
3085 #define GEMM_DEFAULT_ALIGN (BLASULONG)0x03fffUL
3087 #define GEMM_DEFAULT_ALIGN 0x03fffUL
3092 #if defined(CORTEXA57) || \
3093 defined(CORTEXA72) || defined(CORTEXA73) || \
3094 defined(FALKOR) || defined(TSV110) || defined(EMAG8180) || defined(VORTEX)
3096 #define SGEMM_DEFAULT_UNROLL_M 16
3097 #define SGEMM_DEFAULT_UNROLL_N 4
3099 #define DGEMM_DEFAULT_UNROLL_M 8
3100 #define DGEMM_DEFAULT_UNROLL_N 4
3102 #define CGEMM_DEFAULT_UNROLL_M 8
3103 #define CGEMM_DEFAULT_UNROLL_N 4
3105 #define ZGEMM_DEFAULT_UNROLL_M 4
3106 #define ZGEMM_DEFAULT_UNROLL_N 4
3108 /*FIXME: this should be using the cache size, but there is currently no easy way to
3109 query that on ARM. So if getarch counted more than 8 cores we simply assume the host
3110 is a big desktop or server with abundant cache rather than a phone or embedded device */
3111 #if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180) || defined(VORTEX)
3112 #define SGEMM_DEFAULT_P 512
3113 #define DGEMM_DEFAULT_P 256
3114 #define CGEMM_DEFAULT_P 256
3115 #define ZGEMM_DEFAULT_P 128
3117 #define SGEMM_DEFAULT_Q 1024
3118 #define DGEMM_DEFAULT_Q 512
3119 #define CGEMM_DEFAULT_Q 512
3120 #define ZGEMM_DEFAULT_Q 512
3122 #define SGEMM_DEFAULT_P 128
3123 #define DGEMM_DEFAULT_P 160
3124 #define CGEMM_DEFAULT_P 128
3125 #define ZGEMM_DEFAULT_P 128
3127 #define SGEMM_DEFAULT_Q 352
3128 #define DGEMM_DEFAULT_Q 128
3129 #define CGEMM_DEFAULT_Q 224
3130 #define ZGEMM_DEFAULT_Q 112
3133 #define SGEMM_DEFAULT_R 4096
3134 #define DGEMM_DEFAULT_R 4096
3135 #define CGEMM_DEFAULT_R 4096
3136 #define ZGEMM_DEFAULT_R 2048
3138 #elif defined(CORTEXA53) || defined(CORTEXA55)
3140 #define SGEMM_DEFAULT_UNROLL_M 8
3141 #define SGEMM_DEFAULT_UNROLL_N 8
3143 #define DGEMM_DEFAULT_UNROLL_M 8
3144 #define DGEMM_DEFAULT_UNROLL_N 4
3146 #define CGEMM_DEFAULT_UNROLL_M 8
3147 #define CGEMM_DEFAULT_UNROLL_N 4
3149 #define ZGEMM_DEFAULT_UNROLL_M 4
3150 #define ZGEMM_DEFAULT_UNROLL_N 4
3152 #define SGEMM_DEFAULT_P 256
3153 #define DGEMM_DEFAULT_P 160
3154 #define CGEMM_DEFAULT_P 128
3155 #define ZGEMM_DEFAULT_P 128
3157 #define SGEMM_DEFAULT_Q 256
3158 #define DGEMM_DEFAULT_Q 128
3159 #define CGEMM_DEFAULT_Q 224
3160 #define ZGEMM_DEFAULT_Q 112
3162 #define SGEMM_DEFAULT_R 4096
3163 #define DGEMM_DEFAULT_R 4096
3164 #define CGEMM_DEFAULT_R 4096
3165 #define ZGEMM_DEFAULT_R 2048
3167 #elif defined(THUNDERX)
3169 #define SGEMM_DEFAULT_UNROLL_M 4
3170 #define SGEMM_DEFAULT_UNROLL_N 4
3172 #define DGEMM_DEFAULT_UNROLL_M 2
3173 #define DGEMM_DEFAULT_UNROLL_N 2
3175 #define CGEMM_DEFAULT_UNROLL_M 2
3176 #define CGEMM_DEFAULT_UNROLL_N 2
3178 #define ZGEMM_DEFAULT_UNROLL_M 2
3179 #define ZGEMM_DEFAULT_UNROLL_N 2
3181 #define SGEMM_DEFAULT_P 128
3182 #define DGEMM_DEFAULT_P 128
3183 #define CGEMM_DEFAULT_P 96
3184 #define ZGEMM_DEFAULT_P 64
3186 #define SGEMM_DEFAULT_Q 240
3187 #define DGEMM_DEFAULT_Q 120
3188 #define CGEMM_DEFAULT_Q 120
3189 #define ZGEMM_DEFAULT_Q 120
3191 #define SGEMM_DEFAULT_R 12288
3192 #define DGEMM_DEFAULT_R 8192
3193 #define CGEMM_DEFAULT_R 4096
3194 #define ZGEMM_DEFAULT_R 4096
3196 #elif defined(THUNDERX2T99)
3198 #define SGEMM_DEFAULT_UNROLL_M 16
3199 #define SGEMM_DEFAULT_UNROLL_N 4
3201 #define DGEMM_DEFAULT_UNROLL_M 8
3202 #define DGEMM_DEFAULT_UNROLL_N 4
3204 #define CGEMM_DEFAULT_UNROLL_M 8
3205 #define CGEMM_DEFAULT_UNROLL_N 4
3207 #define ZGEMM_DEFAULT_UNROLL_M 4
3208 #define ZGEMM_DEFAULT_UNROLL_N 4
3210 #define SGEMM_DEFAULT_P 128
3211 #define DGEMM_DEFAULT_P 160
3212 #define CGEMM_DEFAULT_P 128
3213 #define ZGEMM_DEFAULT_P 128
3215 #define SGEMM_DEFAULT_Q 352
3216 #define DGEMM_DEFAULT_Q 128
3217 #define CGEMM_DEFAULT_Q 224
3218 #define ZGEMM_DEFAULT_Q 112
3220 #define SGEMM_DEFAULT_R 4096
3221 #define DGEMM_DEFAULT_R 4096
3222 #define CGEMM_DEFAULT_R 4096
3223 #define ZGEMM_DEFAULT_R 4096
3225 #elif defined(THUNDERX3T110)
3227 #define SGEMM_DEFAULT_UNROLL_M 16
3228 #define SGEMM_DEFAULT_UNROLL_N 4
3230 #define DGEMM_DEFAULT_UNROLL_M 8
3231 #define DGEMM_DEFAULT_UNROLL_N 4
3233 #define CGEMM_DEFAULT_UNROLL_M 8
3234 #define CGEMM_DEFAULT_UNROLL_N 4
3236 #define ZGEMM_DEFAULT_UNROLL_M 4
3237 #define ZGEMM_DEFAULT_UNROLL_N 4
3239 #define SGEMM_DEFAULT_P 128
3240 #define DGEMM_DEFAULT_P 320
3241 #define CGEMM_DEFAULT_P 128
3242 #define ZGEMM_DEFAULT_P 128
3244 #define SGEMM_DEFAULT_Q 352
3245 #define DGEMM_DEFAULT_Q 128
3246 #define CGEMM_DEFAULT_Q 224
3247 #define ZGEMM_DEFAULT_Q 112
3249 #define SGEMM_DEFAULT_R 4096
3250 #define DGEMM_DEFAULT_R 4096
3251 #define CGEMM_DEFAULT_R 4096
3252 #define ZGEMM_DEFAULT_R 4096
3254 #elif defined(NEOVERSEN1)
3256 #define SGEMM_DEFAULT_UNROLL_M 16
3257 #define SGEMM_DEFAULT_UNROLL_N 4
3259 #define DGEMM_DEFAULT_UNROLL_M 8
3260 #define DGEMM_DEFAULT_UNROLL_N 4
3262 #define CGEMM_DEFAULT_UNROLL_M 8
3263 #define CGEMM_DEFAULT_UNROLL_N 4
3265 #define ZGEMM_DEFAULT_UNROLL_M 4
3266 #define ZGEMM_DEFAULT_UNROLL_N 4
3268 #define SGEMM_DEFAULT_P 128
3269 #define DGEMM_DEFAULT_P 160
3270 #define CGEMM_DEFAULT_P 128
3271 #define ZGEMM_DEFAULT_P 128
3273 #define SGEMM_DEFAULT_Q 352
3274 #define DGEMM_DEFAULT_Q 128
3275 #define CGEMM_DEFAULT_Q 224
3276 #define ZGEMM_DEFAULT_Q 112
3278 #define SGEMM_DEFAULT_R 4096
3279 #define DGEMM_DEFAULT_R 4096
3280 #define CGEMM_DEFAULT_R 4096
3281 #define ZGEMM_DEFAULT_R 4096
3283 #else /* Other/undetected ARMv8 cores */
3285 #define SGEMM_DEFAULT_UNROLL_M 16
3286 #define SGEMM_DEFAULT_UNROLL_N 4
3288 #define DGEMM_DEFAULT_UNROLL_M 8
3289 #define DGEMM_DEFAULT_UNROLL_N 4
3291 #define CGEMM_DEFAULT_UNROLL_M 8
3292 #define CGEMM_DEFAULT_UNROLL_N 4
3294 #define ZGEMM_DEFAULT_UNROLL_M 4
3295 #define ZGEMM_DEFAULT_UNROLL_N 4
3297 #define SGEMM_DEFAULT_P 128
3298 #define DGEMM_DEFAULT_P 160
3299 #define CGEMM_DEFAULT_P 128
3300 #define ZGEMM_DEFAULT_P 128
3302 #define SGEMM_DEFAULT_Q 352
3303 #define DGEMM_DEFAULT_Q 128
3304 #define CGEMM_DEFAULT_Q 224
3305 #define ZGEMM_DEFAULT_Q 112
3307 #define SGEMM_DEFAULT_R 4096
3308 #define DGEMM_DEFAULT_R 4096
3309 #define CGEMM_DEFAULT_R 4096
3310 #define ZGEMM_DEFAULT_R 4096
3320 #define GEMM_DEFAULT_OFFSET_A 0
3321 #define GEMM_DEFAULT_OFFSET_B 0
3322 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3324 #define SGEMM_DEFAULT_UNROLL_M 2
3325 #define SGEMM_DEFAULT_UNROLL_N 2
3327 #define DGEMM_DEFAULT_UNROLL_M 2
3328 #define DGEMM_DEFAULT_UNROLL_N 2
3330 #define CGEMM_DEFAULT_UNROLL_M 2
3331 #define CGEMM_DEFAULT_UNROLL_N 2
3333 #define ZGEMM_DEFAULT_UNROLL_M 2
3334 #define ZGEMM_DEFAULT_UNROLL_N 2
3336 #define SGEMM_DEFAULT_P 128
3337 #define DGEMM_DEFAULT_P 128
3338 #define CGEMM_DEFAULT_P 96
3339 #define ZGEMM_DEFAULT_P 64
3341 #define SGEMM_DEFAULT_Q 240
3342 #define DGEMM_DEFAULT_Q 120
3343 #define CGEMM_DEFAULT_Q 120
3344 #define ZGEMM_DEFAULT_Q 120
3346 #define SGEMM_DEFAULT_R 12288
3347 #define DGEMM_DEFAULT_R 8192
3348 #define CGEMM_DEFAULT_R 4096
3349 #define ZGEMM_DEFAULT_R 4096
3361 #define GEMM_DEFAULT_OFFSET_A 0
3362 #define GEMM_DEFAULT_OFFSET_B 0
3363 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3365 #define SGEMM_DEFAULT_UNROLL_M 4
3366 #define SGEMM_DEFAULT_UNROLL_N 4
3368 #define DGEMM_DEFAULT_UNROLL_M 4
3369 #define DGEMM_DEFAULT_UNROLL_N 4
3371 #define CGEMM_DEFAULT_UNROLL_M 2
3372 #define CGEMM_DEFAULT_UNROLL_N 2
3374 #define ZGEMM_DEFAULT_UNROLL_M 2
3375 #define ZGEMM_DEFAULT_UNROLL_N 2
3377 #define SGEMM_DEFAULT_P 128
3378 #define DGEMM_DEFAULT_P 128
3379 #define CGEMM_DEFAULT_P 96
3380 #define ZGEMM_DEFAULT_P 64
3382 #define SGEMM_DEFAULT_Q 240
3383 #define DGEMM_DEFAULT_Q 120
3384 #define CGEMM_DEFAULT_Q 120
3385 #define ZGEMM_DEFAULT_Q 120
3387 #define SGEMM_DEFAULT_R 12288
3388 #define DGEMM_DEFAULT_R 8192
3389 #define CGEMM_DEFAULT_R 4096
3390 #define ZGEMM_DEFAULT_R 4096
3402 #define GEMM_DEFAULT_OFFSET_A 0
3403 #define GEMM_DEFAULT_OFFSET_B 0
3404 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3406 #define SGEMM_DEFAULT_UNROLL_M 4
3407 #define SGEMM_DEFAULT_UNROLL_N 4
3409 #define DGEMM_DEFAULT_UNROLL_M 4
3410 #define DGEMM_DEFAULT_UNROLL_N 4
3412 #define CGEMM_DEFAULT_UNROLL_M 2
3413 #define CGEMM_DEFAULT_UNROLL_N 2
3415 #define ZGEMM_DEFAULT_UNROLL_M 2
3416 #define ZGEMM_DEFAULT_UNROLL_N 2
3418 #define SGEMM_DEFAULT_P 128
3419 #define DGEMM_DEFAULT_P 128
3420 #define CGEMM_DEFAULT_P 96
3421 #define ZGEMM_DEFAULT_P 64
3423 #define SGEMM_DEFAULT_Q 240
3424 #define DGEMM_DEFAULT_Q 120
3425 #define CGEMM_DEFAULT_Q 120
3426 #define ZGEMM_DEFAULT_Q 120
3428 #define SGEMM_DEFAULT_R 12288
3429 #define DGEMM_DEFAULT_R 8192
3430 #define CGEMM_DEFAULT_R 4096
3431 #define ZGEMM_DEFAULT_R 4096
3439 #if defined(ZARCH_GENERIC)
3443 #define GEMM_DEFAULT_OFFSET_A 0
3444 #define GEMM_DEFAULT_OFFSET_B 0
3445 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3447 #define SGEMM_DEFAULT_UNROLL_M 2
3448 #define SGEMM_DEFAULT_UNROLL_N 2
3450 #define DGEMM_DEFAULT_UNROLL_M 2
3451 #define DGEMM_DEFAULT_UNROLL_N 2
3453 #define CGEMM_DEFAULT_UNROLL_M 2
3454 #define CGEMM_DEFAULT_UNROLL_N 2
3456 #define ZGEMM_DEFAULT_UNROLL_M 2
3457 #define ZGEMM_DEFAULT_UNROLL_N 2
3459 #define SGEMM_DEFAULT_P 128
3460 #define DGEMM_DEFAULT_P 128
3461 #define CGEMM_DEFAULT_P 96
3462 #define ZGEMM_DEFAULT_P 64
3464 #define SGEMM_DEFAULT_Q 240
3465 #define DGEMM_DEFAULT_Q 120
3466 #define CGEMM_DEFAULT_Q 120
3467 #define ZGEMM_DEFAULT_Q 120
3469 #define SGEMM_DEFAULT_R 12288
3470 #define DGEMM_DEFAULT_R 8192
3471 #define CGEMM_DEFAULT_R 4096
3472 #define ZGEMM_DEFAULT_R 4096
3482 #define GEMM_DEFAULT_OFFSET_A 0
3483 #define GEMM_DEFAULT_OFFSET_B 0
3484 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3486 #define SGEMM_DEFAULT_UNROLL_M 8
3487 #define SGEMM_DEFAULT_UNROLL_N 4
3489 #define DGEMM_DEFAULT_UNROLL_M 8
3490 #define DGEMM_DEFAULT_UNROLL_N 4
3492 #define CGEMM_DEFAULT_UNROLL_M 4
3493 #define CGEMM_DEFAULT_UNROLL_N 4
3495 #define ZGEMM_DEFAULT_UNROLL_M 4
3496 #define ZGEMM_DEFAULT_UNROLL_N 4
3498 #define SGEMM_DEFAULT_P 456
3499 #define DGEMM_DEFAULT_P 320
3500 #define CGEMM_DEFAULT_P 480
3501 #define ZGEMM_DEFAULT_P 224
3503 #define SGEMM_DEFAULT_Q 488
3504 #define DGEMM_DEFAULT_Q 384
3505 #define CGEMM_DEFAULT_Q 128
3506 #define ZGEMM_DEFAULT_Q 352
3508 #define SGEMM_DEFAULT_R 8192
3509 #define DGEMM_DEFAULT_R 4096
3510 #define CGEMM_DEFAULT_R 4096
3511 #define ZGEMM_DEFAULT_R 2048
3522 #define GEMM_DEFAULT_OFFSET_A 0
3523 #define GEMM_DEFAULT_OFFSET_B 0
3524 #define GEMM_DEFAULT_ALIGN 0x03fffUL
3526 #define SGEMM_DEFAULT_UNROLL_M 16
3527 #define SGEMM_DEFAULT_UNROLL_N 4
3529 #define DGEMM_DEFAULT_UNROLL_M 8
3530 #define DGEMM_DEFAULT_UNROLL_N 4
3532 #define CGEMM_DEFAULT_UNROLL_M 4
3533 #define CGEMM_DEFAULT_UNROLL_N 4
3535 #define ZGEMM_DEFAULT_UNROLL_M 4
3536 #define ZGEMM_DEFAULT_UNROLL_N 4
3538 #define SGEMM_DEFAULT_P 480
3539 #define DGEMM_DEFAULT_P 320
3540 #define CGEMM_DEFAULT_P 480
3541 #define ZGEMM_DEFAULT_P 224
3543 #define SGEMM_DEFAULT_Q 512
3544 #define DGEMM_DEFAULT_Q 384
3545 #define CGEMM_DEFAULT_Q 128
3546 #define ZGEMM_DEFAULT_Q 352
3548 #define SGEMM_DEFAULT_R 8192
3549 #define DGEMM_DEFAULT_R 4096
3550 #define CGEMM_DEFAULT_R 4096
3551 #define ZGEMM_DEFAULT_R 2048
3564 #define GEMM_DEFAULT_OFFSET_A 0
3565 #define GEMM_DEFAULT_OFFSET_B 0
3566 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
3568 #define SGEMM_DEFAULT_UNROLL_N 2
3569 #define DGEMM_DEFAULT_UNROLL_N 2
3570 #define QGEMM_DEFAULT_UNROLL_N 2
3571 #define CGEMM_DEFAULT_UNROLL_N 2
3572 #define ZGEMM_DEFAULT_UNROLL_N 2
3573 #define XGEMM_DEFAULT_UNROLL_N 1
3576 #define SGEMM_DEFAULT_UNROLL_M 2
3577 #define DGEMM_DEFAULT_UNROLL_M 2
3578 #define QGEMM_DEFAULT_UNROLL_M 2
3579 #define CGEMM_DEFAULT_UNROLL_M 2
3580 #define ZGEMM_DEFAULT_UNROLL_M 2
3581 #define XGEMM_DEFAULT_UNROLL_M 1
3583 #define SGEMM_DEFAULT_UNROLL_M 2
3584 #define DGEMM_DEFAULT_UNROLL_M 2
3585 #define QGEMM_DEFAULT_UNROLL_M 2
3586 #define CGEMM_DEFAULT_UNROLL_M 2
3587 #define ZGEMM_DEFAULT_UNROLL_M 2
3588 #define XGEMM_DEFAULT_UNROLL_M 1
3591 #define SGEMM_DEFAULT_P sgemm_p
3592 #define DGEMM_DEFAULT_P dgemm_p
3593 #define QGEMM_DEFAULT_P qgemm_p
3594 #define CGEMM_DEFAULT_P cgemm_p
3595 #define ZGEMM_DEFAULT_P zgemm_p
3596 #define XGEMM_DEFAULT_P xgemm_p
3598 #define SGEMM_DEFAULT_R sgemm_r
3599 #define DGEMM_DEFAULT_R dgemm_r
3600 #define QGEMM_DEFAULT_R qgemm_r
3601 #define CGEMM_DEFAULT_R cgemm_r
3602 #define ZGEMM_DEFAULT_R zgemm_r
3603 #define XGEMM_DEFAULT_R xgemm_r
3605 #define SGEMM_DEFAULT_Q 128
3606 #define DGEMM_DEFAULT_Q 128
3607 #define QGEMM_DEFAULT_Q 128
3608 #define CGEMM_DEFAULT_Q 128
3609 #define ZGEMM_DEFAULT_Q 128
3610 #define XGEMM_DEFAULT_Q 128
3616 #ifndef QGEMM_DEFAULT_UNROLL_M
3617 #define QGEMM_DEFAULT_UNROLL_M 2
3620 #ifndef QGEMM_DEFAULT_UNROLL_N
3621 #define QGEMM_DEFAULT_UNROLL_N 2
3624 #ifndef XGEMM_DEFAULT_UNROLL_M
3625 #define XGEMM_DEFAULT_UNROLL_M 2
3628 #ifndef XGEMM_DEFAULT_UNROLL_N
3629 #define XGEMM_DEFAULT_UNROLL_N 2
3633 #define SHUFPD_0 shufps $0x44,
3634 #define SHUFPD_1 shufps $0x4e,
3635 #define SHUFPD_2 shufps $0xe4,
3636 #define SHUFPD_3 shufps $0xee,
3640 #define SHUFPD_0 shufpd $0,
3644 #define SHUFPD_1 shufpd $1,
3648 #define SHUFPD_2 shufpd $2,
3652 #define SHUFPD_3 shufpd $3,
3656 #define SHUFPS_39 shufps $0x39,