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
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
1774 #undef SBGEMM_DEFAULT_UNROLL_N
1775 #undef SBGEMM_DEFAULT_UNROLL_M
1776 #undef SBGEMM_DEFAULT_P
1777 #undef SBGEMM_DEFAULT_R
1778 #undef SBGEMM_DEFAULT_Q
1779 #define SBGEMM_DEFAULT_UNROLL_N 4
1780 #define SBGEMM_DEFAULT_UNROLL_M 16
1781 #define SBGEMM_DEFAULT_P 384
1782 #define SBGEMM_DEFAULT_Q 768
1783 #define SBGEMM_DEFAULT_R sbgemm_r
1787 #define SGEMM_DEFAULT_UNROLL_M 4
1788 #define DGEMM_DEFAULT_UNROLL_M 2
1789 #define QGEMM_DEFAULT_UNROLL_M 2
1790 #define CGEMM_DEFAULT_UNROLL_M 2
1791 #define ZGEMM_DEFAULT_UNROLL_M 1
1792 #define XGEMM_DEFAULT_UNROLL_M 1
1794 #define SGEMM_DEFAULT_UNROLL_N 4
1795 #define DGEMM_DEFAULT_UNROLL_N 4
1796 #define QGEMM_DEFAULT_UNROLL_N 2
1797 #define CGEMM_DEFAULT_UNROLL_N 2
1798 #define ZGEMM_DEFAULT_UNROLL_N 2
1799 #define XGEMM_DEFAULT_UNROLL_N 1
1803 #define SGEMM_DEFAULT_UNROLL_M 16
1804 #define DGEMM_DEFAULT_UNROLL_M 16
1805 #define QGEMM_DEFAULT_UNROLL_M 2
1806 #define CGEMM_DEFAULT_UNROLL_M 8
1807 #define ZGEMM_DEFAULT_UNROLL_M 4
1808 #define XGEMM_DEFAULT_UNROLL_M 1
1810 #define SGEMM_DEFAULT_UNROLL_N 4
1811 #define DGEMM_DEFAULT_UNROLL_N 2
1812 #define QGEMM_DEFAULT_UNROLL_N 2
1813 #define CGEMM_DEFAULT_UNROLL_N 2
1814 #define ZGEMM_DEFAULT_UNROLL_N 2
1815 #define XGEMM_DEFAULT_UNROLL_N 1
1817 #define SGEMM_DEFAULT_UNROLL_MN 32
1818 #define DGEMM_DEFAULT_UNROLL_MN 32
1823 #define SGEMM_DEFAULT_P 512
1824 #define SGEMM_DEFAULT_R sgemm_r
1825 #define DGEMM_DEFAULT_P 512
1826 #define DGEMM_DEFAULT_R dgemm_r
1827 #define QGEMM_DEFAULT_P 504
1828 #define QGEMM_DEFAULT_R qgemm_r
1829 #define CGEMM_DEFAULT_P 128
1830 #define CGEMM_DEFAULT_R 1024
1831 #define ZGEMM_DEFAULT_P 512
1832 #define ZGEMM_DEFAULT_R zgemm_r
1833 #define XGEMM_DEFAULT_P 252
1834 #define XGEMM_DEFAULT_R xgemm_r
1835 #define SGEMM_DEFAULT_Q 256
1836 #define DGEMM_DEFAULT_Q 256
1837 #define QGEMM_DEFAULT_Q 128
1838 #define CGEMM_DEFAULT_Q 256
1839 #define ZGEMM_DEFAULT_Q 192
1840 #define XGEMM_DEFAULT_Q 128
1844 #define SGEMM_DEFAULT_P 640
1845 #define DGEMM_DEFAULT_P 192
1846 #define CGEMM_DEFAULT_P 384
1847 #define ZGEMM_DEFAULT_P 256
1849 #define SGEMM_DEFAULT_Q 320
1850 #define DGEMM_DEFAULT_Q 384
1851 #define CGEMM_DEFAULT_Q 192
1852 #define ZGEMM_DEFAULT_Q 128
1854 #define SGEMM_DEFAULT_R sgemm_r
1855 #define DGEMM_DEFAULT_R 8640
1856 #define CGEMM_DEFAULT_R cgemm_r
1857 #define ZGEMM_DEFAULT_R zgemm_r
1859 #define QGEMM_DEFAULT_Q 128
1860 #define QGEMM_DEFAULT_P 504
1861 #define QGEMM_DEFAULT_R qgemm_r
1862 #define XGEMM_DEFAULT_P 252
1863 #define XGEMM_DEFAULT_R xgemm_r
1864 #define XGEMM_DEFAULT_Q 128
1866 #define CGEMM3M_DEFAULT_UNROLL_N 4
1867 #define CGEMM3M_DEFAULT_UNROLL_M 8
1868 #define ZGEMM3M_DEFAULT_UNROLL_N 4
1869 #define ZGEMM3M_DEFAULT_UNROLL_M 4
1871 #define CGEMM3M_DEFAULT_P 320
1872 #define ZGEMM3M_DEFAULT_P 256
1873 #define XGEMM3M_DEFAULT_P 112
1874 #define CGEMM3M_DEFAULT_Q 320
1875 #define ZGEMM3M_DEFAULT_Q 256
1876 #define XGEMM3M_DEFAULT_Q 224
1877 #define CGEMM3M_DEFAULT_R 12288
1878 #define ZGEMM3M_DEFAULT_R 12288
1879 #define XGEMM3M_DEFAULT_R 12288
1890 #define GEMM_DEFAULT_OFFSET_A 64
1891 #define GEMM_DEFAULT_OFFSET_B 0
1892 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
1897 #define SGEMM_DEFAULT_UNROLL_M 4
1898 #define DGEMM_DEFAULT_UNROLL_M 2
1899 #define QGEMM_DEFAULT_UNROLL_M 2
1900 #define CGEMM_DEFAULT_UNROLL_M 2
1901 #define ZGEMM_DEFAULT_UNROLL_M 1
1902 #define XGEMM_DEFAULT_UNROLL_M 1
1904 #define SGEMM_DEFAULT_UNROLL_M 8
1905 #define DGEMM_DEFAULT_UNROLL_M 4
1906 #define QGEMM_DEFAULT_UNROLL_M 2
1907 #define CGEMM_DEFAULT_UNROLL_M 4
1908 #define ZGEMM_DEFAULT_UNROLL_M 2
1909 #define XGEMM_DEFAULT_UNROLL_M 1
1912 #define SGEMM_DEFAULT_UNROLL_N 4
1913 #define DGEMM_DEFAULT_UNROLL_N 2
1914 #define QGEMM_DEFAULT_UNROLL_N 2
1915 #define CGEMM_DEFAULT_UNROLL_N 2
1916 #define ZGEMM_DEFAULT_UNROLL_N 1
1917 #define XGEMM_DEFAULT_UNROLL_N 1
1919 #define SGEMM_DEFAULT_P sgemm_p
1920 #define SGEMM_DEFAULT_R sgemm_r
1922 #define DGEMM_DEFAULT_P dgemm_p
1923 #define DGEMM_DEFAULT_R dgemm_r
1925 #define QGEMM_DEFAULT_P qgemm_p
1926 #define QGEMM_DEFAULT_R qgemm_r
1928 #define CGEMM_DEFAULT_P cgemm_p
1929 #define CGEMM_DEFAULT_R cgemm_r
1931 #define ZGEMM_DEFAULT_P zgemm_p
1932 #define ZGEMM_DEFAULT_R zgemm_r
1934 #define XGEMM_DEFAULT_P xgemm_p
1935 #define XGEMM_DEFAULT_R xgemm_r
1937 #define SGEMM_DEFAULT_Q 256
1938 #define DGEMM_DEFAULT_Q 256
1939 #define QGEMM_DEFAULT_Q 256
1940 #define CGEMM_DEFAULT_Q 256
1941 #define ZGEMM_DEFAULT_Q 256
1942 #define XGEMM_DEFAULT_Q 256
1952 #define GEMM_DEFAULT_OFFSET_A 0
1953 #define GEMM_DEFAULT_OFFSET_B 128
1954 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1956 #define SGEMM_DEFAULT_UNROLL_M 8
1957 #define SGEMM_DEFAULT_UNROLL_N 8
1958 #define DGEMM_DEFAULT_UNROLL_M 8
1959 #define DGEMM_DEFAULT_UNROLL_N 8
1960 #define QGEMM_DEFAULT_UNROLL_M 8
1961 #define QGEMM_DEFAULT_UNROLL_N 8
1962 #define CGEMM_DEFAULT_UNROLL_M 4
1963 #define CGEMM_DEFAULT_UNROLL_N 4
1964 #define ZGEMM_DEFAULT_UNROLL_M 4
1965 #define ZGEMM_DEFAULT_UNROLL_N 4
1966 #define XGEMM_DEFAULT_UNROLL_M 4
1967 #define XGEMM_DEFAULT_UNROLL_N 4
1969 #define SGEMM_DEFAULT_P sgemm_p
1970 #define DGEMM_DEFAULT_P dgemm_p
1971 #define QGEMM_DEFAULT_P qgemm_p
1972 #define CGEMM_DEFAULT_P cgemm_p
1973 #define ZGEMM_DEFAULT_P zgemm_p
1974 #define XGEMM_DEFAULT_P xgemm_p
1976 #define SGEMM_DEFAULT_Q 1024
1977 #define DGEMM_DEFAULT_Q 1024
1978 #define QGEMM_DEFAULT_Q 1024
1979 #define CGEMM_DEFAULT_Q 1024
1980 #define ZGEMM_DEFAULT_Q 1024
1981 #define XGEMM_DEFAULT_Q 1024
1983 #define SGEMM_DEFAULT_R sgemm_r
1984 #define DGEMM_DEFAULT_R dgemm_r
1985 #define QGEMM_DEFAULT_R qgemm_r
1986 #define CGEMM_DEFAULT_R cgemm_r
1987 #define ZGEMM_DEFAULT_R zgemm_r
1988 #define XGEMM_DEFAULT_R xgemm_r
1992 #define GETRF_FACTOR 0.65
1996 #if defined(EV4) || defined(EV5) || defined(EV6)
2006 #define GEMM_DEFAULT_OFFSET_A 512
2007 #define GEMM_DEFAULT_OFFSET_B 512
2008 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2010 #define SGEMM_DEFAULT_UNROLL_M 4
2011 #define SGEMM_DEFAULT_UNROLL_N 4
2012 #define DGEMM_DEFAULT_UNROLL_M 4
2013 #define DGEMM_DEFAULT_UNROLL_N 4
2014 #define CGEMM_DEFAULT_UNROLL_M 2
2015 #define CGEMM_DEFAULT_UNROLL_N 2
2016 #define ZGEMM_DEFAULT_UNROLL_M 2
2017 #define ZGEMM_DEFAULT_UNROLL_N 2
2022 #define SGEMM_DEFAULT_P 32
2023 #define SGEMM_DEFAULT_Q 112
2024 #define SGEMM_DEFAULT_R 256
2026 #define DGEMM_DEFAULT_P 32
2027 #define DGEMM_DEFAULT_Q 56
2028 #define DGEMM_DEFAULT_R 256
2030 #define CGEMM_DEFAULT_P 32
2031 #define CGEMM_DEFAULT_Q 64
2032 #define CGEMM_DEFAULT_R 240
2034 #define ZGEMM_DEFAULT_P 32
2035 #define ZGEMM_DEFAULT_Q 32
2036 #define ZGEMM_DEFAULT_R 240
2040 #define SGEMM_DEFAULT_P 64
2041 #define SGEMM_DEFAULT_Q 256
2043 #define DGEMM_DEFAULT_P 64
2044 #define DGEMM_DEFAULT_Q 128
2046 #define CGEMM_DEFAULT_P 64
2047 #define CGEMM_DEFAULT_Q 128
2049 #define ZGEMM_DEFAULT_P 64
2050 #define ZGEMM_DEFAULT_Q 64
2054 #define SGEMM_DEFAULT_P 256
2055 #define SGEMM_DEFAULT_Q 512
2057 #define DGEMM_DEFAULT_P 256
2058 #define DGEMM_DEFAULT_Q 256
2060 #define CGEMM_DEFAULT_P 256
2061 #define CGEMM_DEFAULT_Q 256
2063 #define ZGEMM_DEFAULT_P 128
2064 #define ZGEMM_DEFAULT_Q 256
2074 #define GEMM_DEFAULT_OFFSET_A 0
2075 #define GEMM_DEFAULT_OFFSET_B 8192
2076 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2078 #define SGEMM_DEFAULT_UNROLL_M 16
2079 #define SGEMM_DEFAULT_UNROLL_N 4
2080 #define DGEMM_DEFAULT_UNROLL_M 4
2081 #define DGEMM_DEFAULT_UNROLL_N 4
2082 #define CGEMM_DEFAULT_UNROLL_M 8
2083 #define CGEMM_DEFAULT_UNROLL_N 2
2084 #define ZGEMM_DEFAULT_UNROLL_M 2
2085 #define ZGEMM_DEFAULT_UNROLL_N 2
2087 #define SGEMM_DEFAULT_P 128
2088 #define DGEMM_DEFAULT_P 128
2089 #define CGEMM_DEFAULT_P 128
2090 #define ZGEMM_DEFAULT_P 128
2092 #define SGEMM_DEFAULT_Q 512
2093 #define DGEMM_DEFAULT_Q 256
2094 #define CGEMM_DEFAULT_Q 256
2095 #define ZGEMM_DEFAULT_Q 128
2101 #define GEMM_DEFAULT_OFFSET_A 0
2102 #define GEMM_DEFAULT_OFFSET_B 1024
2103 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2105 #define SGEMM_DEFAULT_UNROLL_M 16
2106 #define SGEMM_DEFAULT_UNROLL_N 4
2107 #define DGEMM_DEFAULT_UNROLL_M 4
2108 #define DGEMM_DEFAULT_UNROLL_N 4
2109 #define CGEMM_DEFAULT_UNROLL_M 2
2110 #define CGEMM_DEFAULT_UNROLL_N 2
2111 #define ZGEMM_DEFAULT_UNROLL_M 2
2112 #define ZGEMM_DEFAULT_UNROLL_N 2
2114 #define SGEMM_DEFAULT_P 256
2115 #define DGEMM_DEFAULT_P 128
2116 #define CGEMM_DEFAULT_P 128
2117 #define ZGEMM_DEFAULT_P 64
2119 #define SGEMM_DEFAULT_Q 256
2120 #define DGEMM_DEFAULT_Q 256
2121 #define CGEMM_DEFAULT_Q 256
2122 #define ZGEMM_DEFAULT_Q 256
2132 #define GEMM_DEFAULT_OFFSET_A 2688
2133 #define GEMM_DEFAULT_OFFSET_B 3072
2134 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2136 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2137 #define SGEMM_DEFAULT_UNROLL_M 4
2139 #define SGEMM_DEFAULT_UNROLL_M 16
2141 #define SGEMM_DEFAULT_UNROLL_N 4
2142 #define DGEMM_DEFAULT_UNROLL_M 4
2143 #define DGEMM_DEFAULT_UNROLL_N 4
2144 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2145 #define CGEMM_DEFAULT_UNROLL_M 2
2147 #define CGEMM_DEFAULT_UNROLL_M 8
2149 #define CGEMM_DEFAULT_UNROLL_N 2
2150 #define ZGEMM_DEFAULT_UNROLL_M 2
2151 #define ZGEMM_DEFAULT_UNROLL_N 2
2153 #if defined(OS_LINUX) || defined(OS_DARWIN) || defined(OS_FREEBSD)
2154 #if L2_SIZE == 1024976
2155 #define SGEMM_DEFAULT_P 320
2156 #define DGEMM_DEFAULT_P 256
2157 #define CGEMM_DEFAULT_P 256
2158 #define ZGEMM_DEFAULT_P 256
2160 #define SGEMM_DEFAULT_P 176
2161 #define DGEMM_DEFAULT_P 176
2162 #define CGEMM_DEFAULT_P 176
2163 #define ZGEMM_DEFAULT_P 176
2167 #define SGEMM_DEFAULT_Q 512
2168 #define DGEMM_DEFAULT_Q 256
2169 #define CGEMM_DEFAULT_Q 256
2170 #define ZGEMM_DEFAULT_Q 128
2181 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2182 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2183 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2185 #define SGEMM_DEFAULT_UNROLL_M 4
2186 #define SGEMM_DEFAULT_UNROLL_N 4
2187 #define DGEMM_DEFAULT_UNROLL_M 4
2188 #define DGEMM_DEFAULT_UNROLL_N 4
2189 #define CGEMM_DEFAULT_UNROLL_M 2
2190 #define CGEMM_DEFAULT_UNROLL_N 2
2191 #define ZGEMM_DEFAULT_UNROLL_M 2
2192 #define ZGEMM_DEFAULT_UNROLL_N 2
2194 #define SGEMM_DEFAULT_P 512
2195 #define DGEMM_DEFAULT_P 512
2196 #define CGEMM_DEFAULT_P 512
2197 #define ZGEMM_DEFAULT_P 512
2199 #define SGEMM_DEFAULT_Q 1024
2200 #define DGEMM_DEFAULT_Q 512
2201 #define CGEMM_DEFAULT_Q 512
2202 #define ZGEMM_DEFAULT_Q 256
2204 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2205 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2206 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2207 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2217 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2218 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2219 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2221 #define SGEMM_DEFAULT_UNROLL_M 8
2222 #define SGEMM_DEFAULT_UNROLL_N 4
2223 #define DGEMM_DEFAULT_UNROLL_M 8
2224 #define DGEMM_DEFAULT_UNROLL_N 4
2225 #define CGEMM_DEFAULT_UNROLL_M 4
2226 #define CGEMM_DEFAULT_UNROLL_N 2
2227 #define ZGEMM_DEFAULT_UNROLL_M 4
2228 #define ZGEMM_DEFAULT_UNROLL_N 2
2230 #define SGEMM_DEFAULT_P 128
2231 #define DGEMM_DEFAULT_P 128
2232 #define CGEMM_DEFAULT_P 128
2233 #define ZGEMM_DEFAULT_P 128
2235 #define SGEMM_DEFAULT_Q 4096
2236 #define DGEMM_DEFAULT_Q 3072
2237 #define CGEMM_DEFAULT_Q 2048
2238 #define ZGEMM_DEFAULT_Q 1024
2240 #define SGEMM_DEFAULT_Q 512
2241 #define DGEMM_DEFAULT_Q 256
2242 #define CGEMM_DEFAULT_Q 256
2243 #define ZGEMM_DEFAULT_Q 128
2251 #if defined(POWER3) || defined(POWER4) || defined(POWER5)
2252 #define GEMM_DEFAULT_OFFSET_A 0
2253 #define GEMM_DEFAULT_OFFSET_B 2048
2254 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2256 #define SGEMM_DEFAULT_UNROLL_M 4
2257 #define SGEMM_DEFAULT_UNROLL_N 4
2258 #define DGEMM_DEFAULT_UNROLL_M 4
2259 #define DGEMM_DEFAULT_UNROLL_N 4
2260 #define CGEMM_DEFAULT_UNROLL_M 2
2261 #define CGEMM_DEFAULT_UNROLL_N 2
2262 #define ZGEMM_DEFAULT_UNROLL_M 2
2263 #define ZGEMM_DEFAULT_UNROLL_N 2
2270 #define SGEMM_DEFAULT_P 256
2271 #define SGEMM_DEFAULT_Q 432
2272 #define SGEMM_DEFAULT_R 1012
2274 #define DGEMM_DEFAULT_P 256
2275 #define DGEMM_DEFAULT_Q 216
2276 #define DGEMM_DEFAULT_R 1012
2278 #define CGEMM_DEFAULT_P 256
2279 #define CGEMM_DEFAULT_Q 104
2280 #define CGEMM_DEFAULT_R 1012
2282 #define ZGEMM_DEFAULT_P 256
2283 #define ZGEMM_DEFAULT_Q 104
2284 #define ZGEMM_DEFAULT_R 1012
2288 #ifdef ALLOC_HUGETLB
2289 #define SGEMM_DEFAULT_P 184
2290 #define DGEMM_DEFAULT_P 184
2291 #define CGEMM_DEFAULT_P 184
2292 #define ZGEMM_DEFAULT_P 184
2294 #define SGEMM_DEFAULT_P 144
2295 #define DGEMM_DEFAULT_P 144
2296 #define CGEMM_DEFAULT_P 144
2297 #define ZGEMM_DEFAULT_P 144
2300 #define SGEMM_DEFAULT_Q 256
2301 #define CGEMM_DEFAULT_Q 256
2302 #define DGEMM_DEFAULT_Q 256
2303 #define ZGEMM_DEFAULT_Q 256
2307 #ifdef ALLOC_HUGETLB
2308 #define SGEMM_DEFAULT_P 512
2309 #define DGEMM_DEFAULT_P 256
2310 #define CGEMM_DEFAULT_P 256
2311 #define ZGEMM_DEFAULT_P 128
2313 #define SGEMM_DEFAULT_P 320
2314 #define DGEMM_DEFAULT_P 160
2315 #define CGEMM_DEFAULT_P 160
2316 #define ZGEMM_DEFAULT_P 80
2319 #define SGEMM_DEFAULT_Q 256
2320 #define CGEMM_DEFAULT_Q 256
2321 #define DGEMM_DEFAULT_Q 256
2322 #define ZGEMM_DEFAULT_Q 256
2334 #define GEMM_DEFAULT_OFFSET_A 384
2335 #define GEMM_DEFAULT_OFFSET_B 1024
2336 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2338 #define SGEMM_DEFAULT_UNROLL_M 4
2339 #define SGEMM_DEFAULT_UNROLL_N 4
2340 #define DGEMM_DEFAULT_UNROLL_M 4
2341 #define DGEMM_DEFAULT_UNROLL_N 4
2342 #define CGEMM_DEFAULT_UNROLL_M 2
2343 #define CGEMM_DEFAULT_UNROLL_N 4
2344 #define ZGEMM_DEFAULT_UNROLL_M 2
2345 #define ZGEMM_DEFAULT_UNROLL_N 4
2347 #define SGEMM_DEFAULT_P 992
2348 #define DGEMM_DEFAULT_P 480
2349 #define CGEMM_DEFAULT_P 488
2350 #define ZGEMM_DEFAULT_P 248
2352 #define SGEMM_DEFAULT_Q 504
2353 #define DGEMM_DEFAULT_Q 504
2354 #define CGEMM_DEFAULT_Q 400
2355 #define ZGEMM_DEFAULT_Q 400
2366 #define GEMM_DEFAULT_OFFSET_A 0
2367 #define GEMM_DEFAULT_OFFSET_B 65536
2369 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2370 #if defined(__32BIT__)
2371 #warning using BINARY32==POWER6
2372 #define SGEMM_DEFAULT_UNROLL_M 4
2373 #define SGEMM_DEFAULT_UNROLL_N 4
2374 #define DGEMM_DEFAULT_UNROLL_M 4
2375 #define DGEMM_DEFAULT_UNROLL_N 4
2376 #define CGEMM_DEFAULT_UNROLL_M 2
2377 #define CGEMM_DEFAULT_UNROLL_N 4
2378 #define ZGEMM_DEFAULT_UNROLL_M 2
2379 #define ZGEMM_DEFAULT_UNROLL_N 4
2381 #define SGEMM_DEFAULT_UNROLL_M 16
2382 #define SGEMM_DEFAULT_UNROLL_N 8
2383 #define DGEMM_DEFAULT_UNROLL_M 16
2384 #define DGEMM_DEFAULT_UNROLL_N 4
2385 #define CGEMM_DEFAULT_UNROLL_M 8
2386 #define CGEMM_DEFAULT_UNROLL_N 4
2387 #define ZGEMM_DEFAULT_UNROLL_M 8
2388 #define ZGEMM_DEFAULT_UNROLL_N 2
2390 #define SGEMM_DEFAULT_P 1280UL
2391 #define DGEMM_DEFAULT_P 640UL
2392 #define CGEMM_DEFAULT_P 640UL
2393 #define ZGEMM_DEFAULT_P 320UL
2395 #define SGEMM_DEFAULT_Q 640UL
2396 #define DGEMM_DEFAULT_Q 720UL
2397 #define CGEMM_DEFAULT_Q 640UL
2398 #define ZGEMM_DEFAULT_Q 640UL
2401 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2402 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2403 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2404 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2406 #define SGEMM_DEFAULT_R 4096
2407 #define DGEMM_DEFAULT_R 4096
2408 #define CGEMM_DEFAULT_R 4096
2409 #define ZGEMM_DEFAULT_R 4096
2420 #define GEMM_DEFAULT_OFFSET_A 0
2421 #define GEMM_DEFAULT_OFFSET_B 65536
2422 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2424 #define SWITCH_RATIO 16
2425 #define GEMM_PREFERED_SIZE 16
2427 #define SGEMM_DEFAULT_UNROLL_M 16
2428 #define SGEMM_DEFAULT_UNROLL_N 8
2429 #define DGEMM_DEFAULT_UNROLL_M 16
2430 #define DGEMM_DEFAULT_UNROLL_N 4
2431 #define CGEMM_DEFAULT_UNROLL_M 8
2432 #define CGEMM_DEFAULT_UNROLL_N 4
2433 #define ZGEMM_DEFAULT_UNROLL_M 8
2434 #define ZGEMM_DEFAULT_UNROLL_N 2
2436 #define SGEMM_DEFAULT_P 832
2437 #define DGEMM_DEFAULT_P 128
2438 #define CGEMM_DEFAULT_P 512
2439 #define ZGEMM_DEFAULT_P 256
2441 #define SGEMM_DEFAULT_Q 1026
2442 #define DGEMM_DEFAULT_Q 384
2443 #define CGEMM_DEFAULT_Q 1026
2444 #define ZGEMM_DEFAULT_Q 1026
2446 #define SGEMM_DEFAULT_R 4096
2447 #define DGEMM_DEFAULT_R 4096
2448 #define CGEMM_DEFAULT_R 4096
2449 #define ZGEMM_DEFAULT_R 4096
2455 #if defined(POWER10)
2459 #define GEMM_DEFAULT_OFFSET_A 0
2460 #define GEMM_DEFAULT_OFFSET_B 65536
2461 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2463 #define SWITCH_RATIO 16
2464 #define GEMM_PREFERED_SIZE 16
2466 #define SGEMM_DEFAULT_UNROLL_M 16
2467 #define SGEMM_DEFAULT_UNROLL_N 8
2468 #define DGEMM_DEFAULT_UNROLL_M 8
2469 #define DGEMM_DEFAULT_UNROLL_N 8
2470 #define CGEMM_DEFAULT_UNROLL_M 8
2471 #define CGEMM_DEFAULT_UNROLL_N 4
2472 #define ZGEMM_DEFAULT_UNROLL_M 8
2473 #define ZGEMM_DEFAULT_UNROLL_N 2
2475 #define SGEMM_DEFAULT_P 512
2476 #define DGEMM_DEFAULT_P 384
2477 #define CGEMM_DEFAULT_P 512
2478 #define ZGEMM_DEFAULT_P 256
2480 #define SGEMM_DEFAULT_Q 512
2481 #define DGEMM_DEFAULT_Q 512
2482 #define CGEMM_DEFAULT_Q 384
2483 #define ZGEMM_DEFAULT_Q 384
2485 #define SGEMM_DEFAULT_R 4096
2486 #define DGEMM_DEFAULT_R 4096
2487 #define CGEMM_DEFAULT_R 4096
2488 #define ZGEMM_DEFAULT_R 4096
2492 #undef SBGEMM_DEFAULT_UNROLL_N
2493 #undef SBGEMM_DEFAULT_UNROLL_M
2494 #undef SBGEMM_DEFAULT_P
2495 #undef SBGEMM_DEFAULT_R
2496 #undef SBGEMM_DEFAULT_Q
2497 #define SBGEMM_DEFAULT_UNROLL_M 16
2498 #define SBGEMM_DEFAULT_UNROLL_N 8
2499 #define SBGEMM_DEFAULT_P 832
2500 #define SBGEMM_DEFAULT_Q 1026
2501 #define SBGEMM_DEFAULT_R 4096
2504 #if defined(SPARC) && defined(V7)
2509 #define GEMM_DEFAULT_OFFSET_A 0
2510 #define GEMM_DEFAULT_OFFSET_B 2048
2511 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2513 #define SGEMM_DEFAULT_UNROLL_M 2
2514 #define SGEMM_DEFAULT_UNROLL_N 8
2515 #define DGEMM_DEFAULT_UNROLL_M 2
2516 #define DGEMM_DEFAULT_UNROLL_N 8
2517 #define CGEMM_DEFAULT_UNROLL_M 1
2518 #define CGEMM_DEFAULT_UNROLL_N 4
2519 #define ZGEMM_DEFAULT_UNROLL_M 1
2520 #define ZGEMM_DEFAULT_UNROLL_N 4
2522 #define SGEMM_DEFAULT_P 256
2523 #define DGEMM_DEFAULT_P 256
2524 #define CGEMM_DEFAULT_P 256
2525 #define ZGEMM_DEFAULT_P 256
2527 #define SGEMM_DEFAULT_Q 512
2528 #define DGEMM_DEFAULT_Q 256
2529 #define CGEMM_DEFAULT_Q 256
2530 #define ZGEMM_DEFAULT_Q 128
2533 #define GEMM_THREAD gemm_thread_mn
2536 #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
2541 #define GEMM_DEFAULT_OFFSET_A 0
2542 #define GEMM_DEFAULT_OFFSET_B 2048
2543 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2545 #define SGEMM_DEFAULT_UNROLL_M 4
2546 #define SGEMM_DEFAULT_UNROLL_N 4
2547 #define DGEMM_DEFAULT_UNROLL_M 4
2548 #define DGEMM_DEFAULT_UNROLL_N 4
2549 #define CGEMM_DEFAULT_UNROLL_M 2
2550 #define CGEMM_DEFAULT_UNROLL_N 2
2551 #define ZGEMM_DEFAULT_UNROLL_M 2
2552 #define ZGEMM_DEFAULT_UNROLL_N 2
2554 #define SGEMM_DEFAULT_P 512
2555 #define DGEMM_DEFAULT_P 512
2556 #define CGEMM_DEFAULT_P 512
2557 #define ZGEMM_DEFAULT_P 512
2559 #define SGEMM_DEFAULT_Q 1024
2560 #define DGEMM_DEFAULT_Q 512
2561 #define CGEMM_DEFAULT_Q 512
2562 #define ZGEMM_DEFAULT_Q 256
2572 #define GEMM_DEFAULT_OFFSET_A 0
2573 #define GEMM_DEFAULT_OFFSET_B 0
2574 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2576 #define SGEMM_DEFAULT_UNROLL_M 2
2577 #define SGEMM_DEFAULT_UNROLL_N 8
2578 #define DGEMM_DEFAULT_UNROLL_M 2
2579 #define DGEMM_DEFAULT_UNROLL_N 8
2580 #define CGEMM_DEFAULT_UNROLL_M 1
2581 #define CGEMM_DEFAULT_UNROLL_N 4
2582 #define ZGEMM_DEFAULT_UNROLL_M 1
2583 #define ZGEMM_DEFAULT_UNROLL_N 4
2585 #define SGEMM_DEFAULT_P 108
2586 #define DGEMM_DEFAULT_P 112
2587 #define CGEMM_DEFAULT_P 108
2588 #define ZGEMM_DEFAULT_P 112
2590 #define SGEMM_DEFAULT_Q 288
2591 #define DGEMM_DEFAULT_Q 144
2592 #define CGEMM_DEFAULT_Q 144
2593 #define ZGEMM_DEFAULT_Q 72
2595 #define SGEMM_DEFAULT_R 2000
2596 #define DGEMM_DEFAULT_R 2000
2597 #define CGEMM_DEFAULT_R 2000
2598 #define ZGEMM_DEFAULT_R 2000
2603 #if defined(LOONGSON3R4)
2607 #define GEMM_DEFAULT_OFFSET_A 0
2608 #define GEMM_DEFAULT_OFFSET_B 0
2609 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2612 #define SGEMM_DEFAULT_UNROLL_M 8
2613 #define SGEMM_DEFAULT_UNROLL_N 8
2615 #define DGEMM_DEFAULT_UNROLL_M 8
2616 #define DGEMM_DEFAULT_UNROLL_N 4
2618 #define CGEMM_DEFAULT_UNROLL_M 8
2619 #define CGEMM_DEFAULT_UNROLL_N 4
2621 #define ZGEMM_DEFAULT_UNROLL_M 4
2622 #define ZGEMM_DEFAULT_UNROLL_N 4
2624 #define SGEMM_DEFAULT_UNROLL_M 8
2625 #define SGEMM_DEFAULT_UNROLL_N 4
2627 #define DGEMM_DEFAULT_UNROLL_M 4
2628 #define DGEMM_DEFAULT_UNROLL_N 4
2630 #define CGEMM_DEFAULT_UNROLL_M 4
2631 #define CGEMM_DEFAULT_UNROLL_N 2
2633 #define ZGEMM_DEFAULT_UNROLL_M 2
2634 #define ZGEMM_DEFAULT_UNROLL_N 2
2637 #define SGEMM_DEFAULT_P 64
2638 #define DGEMM_DEFAULT_P 44
2639 #define CGEMM_DEFAULT_P 64
2640 #define ZGEMM_DEFAULT_P 32
2642 #define SGEMM_DEFAULT_Q 192
2643 #define DGEMM_DEFAULT_Q 92
2644 #define CGEMM_DEFAULT_Q 128
2645 #define ZGEMM_DEFAULT_Q 80
2647 #define SGEMM_DEFAULT_R 640
2648 #define DGEMM_DEFAULT_R dgemm_r
2649 #define CGEMM_DEFAULT_R 640
2650 #define ZGEMM_DEFAULT_R 640
2652 #define GEMM_OFFSET_A1 0x10000
2653 #define GEMM_OFFSET_B1 0x100000
2658 #if defined(LOONGSON3R3)
2659 ////Copy from SICORTEX
2663 #define GEMM_DEFAULT_OFFSET_A 0
2664 #define GEMM_DEFAULT_OFFSET_B 0
2665 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2667 #define SGEMM_DEFAULT_UNROLL_M 8
2668 #define SGEMM_DEFAULT_UNROLL_N 4
2670 #define DGEMM_DEFAULT_UNROLL_M 4
2671 #define DGEMM_DEFAULT_UNROLL_N 4
2673 #define CGEMM_DEFAULT_UNROLL_M 4
2674 #define CGEMM_DEFAULT_UNROLL_N 2
2676 #define ZGEMM_DEFAULT_UNROLL_M 2
2677 #define ZGEMM_DEFAULT_UNROLL_N 2
2679 #define SGEMM_DEFAULT_P 64
2680 #define DGEMM_DEFAULT_P 44
2681 #define CGEMM_DEFAULT_P 64
2682 #define ZGEMM_DEFAULT_P 32
2684 #define SGEMM_DEFAULT_Q 192
2685 #define DGEMM_DEFAULT_Q 92
2686 #define CGEMM_DEFAULT_Q 128
2687 #define ZGEMM_DEFAULT_Q 80
2689 #define SGEMM_DEFAULT_R 640
2690 #define DGEMM_DEFAULT_R dgemm_r
2691 #define CGEMM_DEFAULT_R 640
2692 #define ZGEMM_DEFAULT_R 640
2694 #define GEMM_OFFSET_A1 0x10000
2695 #define GEMM_OFFSET_B1 0x100000
2700 #if defined (LOONGSON3R5)
2704 #define GEMM_DEFAULT_OFFSET_A 0
2705 #define GEMM_DEFAULT_OFFSET_B 0
2706 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2708 #define SGEMM_DEFAULT_UNROLL_N 8
2709 #define DGEMM_DEFAULT_UNROLL_N 8
2710 #define QGEMM_DEFAULT_UNROLL_N 2
2711 #define CGEMM_DEFAULT_UNROLL_N 4
2712 #define ZGEMM_DEFAULT_UNROLL_N 4
2713 #define XGEMM_DEFAULT_UNROLL_N 1
2715 #define SGEMM_DEFAULT_UNROLL_M 2
2716 #define DGEMM_DEFAULT_UNROLL_M 2
2717 #define QGEMM_DEFAULT_UNROLL_M 2
2718 #define CGEMM_DEFAULT_UNROLL_M 1
2719 #define ZGEMM_DEFAULT_UNROLL_M 1
2720 #define XGEMM_DEFAULT_UNROLL_M 1
2722 #define SGEMM_DEFAULT_P sgemm_p
2723 #define DGEMM_DEFAULT_P dgemm_p
2724 #define QGEMM_DEFAULT_P qgemm_p
2725 #define CGEMM_DEFAULT_P cgemm_p
2726 #define ZGEMM_DEFAULT_P zgemm_p
2727 #define XGEMM_DEFAULT_P xgemm_p
2729 #define SGEMM_DEFAULT_R sgemm_r
2730 #define DGEMM_DEFAULT_R dgemm_r
2731 #define QGEMM_DEFAULT_R qgemm_r
2732 #define CGEMM_DEFAULT_R cgemm_r
2733 #define ZGEMM_DEFAULT_R zgemm_r
2734 #define XGEMM_DEFAULT_R xgemm_r
2736 #define SGEMM_DEFAULT_Q 128
2737 #define DGEMM_DEFAULT_Q 128
2738 #define QGEMM_DEFAULT_Q 128
2739 #define CGEMM_DEFAULT_Q 128
2740 #define ZGEMM_DEFAULT_Q 128
2741 #define XGEMM_DEFAULT_Q 128
2746 #if defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
2750 #define GEMM_DEFAULT_OFFSET_A 0
2751 #define GEMM_DEFAULT_OFFSET_B 0
2752 #define GEMM_DEFAULT_ALIGN (BLASLONG) 0x03fffUL
2755 #define SGEMM_DEFAULT_UNROLL_M 8
2756 #define SGEMM_DEFAULT_UNROLL_N 8
2758 #define DGEMM_DEFAULT_UNROLL_M 8
2759 #define DGEMM_DEFAULT_UNROLL_N 4
2761 #define CGEMM_DEFAULT_UNROLL_M 8
2762 #define CGEMM_DEFAULT_UNROLL_N 4
2764 #define ZGEMM_DEFAULT_UNROLL_M 4
2765 #define ZGEMM_DEFAULT_UNROLL_N 4
2767 #define SGEMM_DEFAULT_UNROLL_M 2
2768 #define SGEMM_DEFAULT_UNROLL_N 2
2770 #define DGEMM_DEFAULT_UNROLL_M 2
2771 #define DGEMM_DEFAULT_UNROLL_N 2
2773 #define CGEMM_DEFAULT_UNROLL_M 2
2774 #define CGEMM_DEFAULT_UNROLL_N 2
2776 #define ZGEMM_DEFAULT_UNROLL_M 2
2777 #define ZGEMM_DEFAULT_UNROLL_N 2
2780 #define SGEMM_DEFAULT_P 128
2781 #define DGEMM_DEFAULT_P 128
2782 #define CGEMM_DEFAULT_P 96
2783 #define ZGEMM_DEFAULT_P 64
2785 #define SGEMM_DEFAULT_Q 240
2786 #define DGEMM_DEFAULT_Q 120
2787 #define CGEMM_DEFAULT_Q 120
2788 #define ZGEMM_DEFAULT_Q 120
2790 #define SGEMM_DEFAULT_R 12288
2791 #define DGEMM_DEFAULT_R 8192
2792 #define CGEMM_DEFAULT_R 4096
2793 #define ZGEMM_DEFAULT_R 4096
2798 #ifdef RISCV64_GENERIC
2799 #define GEMM_DEFAULT_OFFSET_A 0
2800 #define GEMM_DEFAULT_OFFSET_B 0
2801 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2803 #define SGEMM_DEFAULT_UNROLL_M 2
2804 #define SGEMM_DEFAULT_UNROLL_N 2
2806 #define DGEMM_DEFAULT_UNROLL_M 2
2807 #define DGEMM_DEFAULT_UNROLL_N 2
2809 #define CGEMM_DEFAULT_UNROLL_M 2
2810 #define CGEMM_DEFAULT_UNROLL_N 2
2812 #define ZGEMM_DEFAULT_UNROLL_M 2
2813 #define ZGEMM_DEFAULT_UNROLL_N 2
2815 #define SGEMM_DEFAULT_P 128
2816 #define DGEMM_DEFAULT_P 128
2817 #define CGEMM_DEFAULT_P 96
2818 #define ZGEMM_DEFAULT_P 64
2820 #define SGEMM_DEFAULT_Q 240
2821 #define DGEMM_DEFAULT_Q 120
2822 #define CGEMM_DEFAULT_Q 120
2823 #define ZGEMM_DEFAULT_Q 120
2825 #define SGEMM_DEFAULT_R 12288
2826 #define DGEMM_DEFAULT_R 8192
2827 #define CGEMM_DEFAULT_R 4096
2828 #define ZGEMM_DEFAULT_R 4096
2832 #define GEMM_DEFAULT_OFFSET_A 0
2833 #define GEMM_DEFAULT_OFFSET_B 0
2838 #define GEMM_DEFAULT_OFFSET_A 0
2839 #define GEMM_DEFAULT_OFFSET_B 0
2840 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2842 #define SGEMM_DEFAULT_UNROLL_M 16
2843 #define SGEMM_DEFAULT_UNROLL_N 4
2845 #define DGEMM_DEFAULT_UNROLL_M 8
2846 #define DGEMM_DEFAULT_UNROLL_N 4
2848 #define CGEMM_DEFAULT_UNROLL_M 2
2849 #define CGEMM_DEFAULT_UNROLL_N 2
2851 #define ZGEMM_DEFAULT_UNROLL_M 2
2852 #define ZGEMM_DEFAULT_UNROLL_N 2
2854 #define SGEMM_DEFAULT_P 160
2855 #define DGEMM_DEFAULT_P 160
2856 #define CGEMM_DEFAULT_P 96
2857 #define ZGEMM_DEFAULT_P 64
2859 #define SGEMM_DEFAULT_Q 240
2860 #define DGEMM_DEFAULT_Q 128
2861 #define CGEMM_DEFAULT_Q 120
2862 #define ZGEMM_DEFAULT_Q 120
2864 #define SGEMM_DEFAULT_R 12288
2865 #define DGEMM_DEFAULT_R 8192
2866 #define CGEMM_DEFAULT_R 4096
2867 #define ZGEMM_DEFAULT_R 4096
2871 #define GEMM_DEFAULT_OFFSET_A 0
2872 #define GEMM_DEFAULT_OFFSET_B 0
2880 #define GEMM_DEFAULT_OFFSET_A 0
2881 #define GEMM_DEFAULT_OFFSET_B 0
2882 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2884 #define SGEMM_DEFAULT_UNROLL_M 4
2885 #define SGEMM_DEFAULT_UNROLL_N 4
2887 #define DGEMM_DEFAULT_UNROLL_M 4
2888 #define DGEMM_DEFAULT_UNROLL_N 4
2890 #define CGEMM_DEFAULT_UNROLL_M 2
2891 #define CGEMM_DEFAULT_UNROLL_N 2
2893 #define ZGEMM_DEFAULT_UNROLL_M 2
2894 #define ZGEMM_DEFAULT_UNROLL_N 2
2896 #define SGEMM_DEFAULT_P 128
2897 #define DGEMM_DEFAULT_P 128
2898 #define CGEMM_DEFAULT_P 96
2899 #define ZGEMM_DEFAULT_P 64
2901 #define SGEMM_DEFAULT_Q 240
2902 #define DGEMM_DEFAULT_Q 120
2903 #define CGEMM_DEFAULT_Q 120
2904 #define ZGEMM_DEFAULT_Q 120
2906 #define SGEMM_DEFAULT_R 12288
2907 #define DGEMM_DEFAULT_R 8192
2908 #define CGEMM_DEFAULT_R 4096
2909 #define ZGEMM_DEFAULT_R 4096
2921 #define GEMM_DEFAULT_OFFSET_A 0
2922 #define GEMM_DEFAULT_OFFSET_B 0
2923 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2925 #define SGEMM_DEFAULT_UNROLL_M 4
2926 #define SGEMM_DEFAULT_UNROLL_N 2
2928 #define DGEMM_DEFAULT_UNROLL_M 4
2929 #define DGEMM_DEFAULT_UNROLL_N 2
2931 #define CGEMM_DEFAULT_UNROLL_M 2
2932 #define CGEMM_DEFAULT_UNROLL_N 2
2934 #define ZGEMM_DEFAULT_UNROLL_M 2
2935 #define ZGEMM_DEFAULT_UNROLL_N 2
2937 #define SGEMM_DEFAULT_P 128
2938 #define DGEMM_DEFAULT_P 128
2939 #define CGEMM_DEFAULT_P 96
2940 #define ZGEMM_DEFAULT_P 64
2942 #define SGEMM_DEFAULT_Q 240
2943 #define DGEMM_DEFAULT_Q 120
2944 #define CGEMM_DEFAULT_Q 120
2945 #define ZGEMM_DEFAULT_Q 120
2947 #define SGEMM_DEFAULT_R 12288
2948 #define DGEMM_DEFAULT_R 8192
2949 #define CGEMM_DEFAULT_R 4096
2950 #define ZGEMM_DEFAULT_R 4096
2956 /* Common ARMv8 parameters */
2962 #define GEMM_DEFAULT_OFFSET_A 0
2963 #define GEMM_DEFAULT_OFFSET_B 0
2965 /* Use explicit casting for win64 as LLP64 datamodel is used */
2966 #define GEMM_DEFAULT_ALIGN (BLASULONG)0x03fffUL
2968 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2973 #if defined(CORTEXA57) || \
2974 defined(CORTEXA72) || defined(CORTEXA73) || \
2975 defined(FALKOR) || defined(TSV110) || defined(EMAG8180) || defined(VORTEX)
2977 #define SGEMM_DEFAULT_UNROLL_M 16
2978 #define SGEMM_DEFAULT_UNROLL_N 4
2980 #define DGEMM_DEFAULT_UNROLL_M 8
2981 #define DGEMM_DEFAULT_UNROLL_N 4
2983 #define CGEMM_DEFAULT_UNROLL_M 8
2984 #define CGEMM_DEFAULT_UNROLL_N 4
2986 #define ZGEMM_DEFAULT_UNROLL_M 4
2987 #define ZGEMM_DEFAULT_UNROLL_N 4
2989 /*FIXME: this should be using the cache size, but there is currently no easy way to
2990 query that on ARM. So if getarch counted more than 8 cores we simply assume the host
2991 is a big desktop or server with abundant cache rather than a phone or embedded device */
2992 #if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180) || defined(VORTEX)
2993 #define SGEMM_DEFAULT_P 512
2994 #define DGEMM_DEFAULT_P 256
2995 #define CGEMM_DEFAULT_P 256
2996 #define ZGEMM_DEFAULT_P 128
2998 #define SGEMM_DEFAULT_Q 1024
2999 #define DGEMM_DEFAULT_Q 512
3000 #define CGEMM_DEFAULT_Q 512
3001 #define ZGEMM_DEFAULT_Q 512
3003 #define SGEMM_DEFAULT_P 128
3004 #define DGEMM_DEFAULT_P 160
3005 #define CGEMM_DEFAULT_P 128
3006 #define ZGEMM_DEFAULT_P 128
3008 #define SGEMM_DEFAULT_Q 352
3009 #define DGEMM_DEFAULT_Q 128
3010 #define CGEMM_DEFAULT_Q 224
3011 #define ZGEMM_DEFAULT_Q 112
3014 #define SGEMM_DEFAULT_R 4096
3015 #define DGEMM_DEFAULT_R 4096
3016 #define CGEMM_DEFAULT_R 4096
3017 #define ZGEMM_DEFAULT_R 2048
3019 #elif defined(CORTEXA53) || defined(CORTEXA55)
3021 #define SGEMM_DEFAULT_UNROLL_M 8
3022 #define SGEMM_DEFAULT_UNROLL_N 8
3024 #define DGEMM_DEFAULT_UNROLL_M 8
3025 #define DGEMM_DEFAULT_UNROLL_N 4
3027 #define CGEMM_DEFAULT_UNROLL_M 8
3028 #define CGEMM_DEFAULT_UNROLL_N 4
3030 #define ZGEMM_DEFAULT_UNROLL_M 4
3031 #define ZGEMM_DEFAULT_UNROLL_N 4
3033 #define SGEMM_DEFAULT_P 256
3034 #define DGEMM_DEFAULT_P 160
3035 #define CGEMM_DEFAULT_P 128
3036 #define ZGEMM_DEFAULT_P 128
3038 #define SGEMM_DEFAULT_Q 256
3039 #define DGEMM_DEFAULT_Q 128
3040 #define CGEMM_DEFAULT_Q 224
3041 #define ZGEMM_DEFAULT_Q 112
3043 #define SGEMM_DEFAULT_R 4096
3044 #define DGEMM_DEFAULT_R 4096
3045 #define CGEMM_DEFAULT_R 4096
3046 #define ZGEMM_DEFAULT_R 2048
3048 #elif defined(THUNDERX)
3050 #define SGEMM_DEFAULT_UNROLL_M 4
3051 #define SGEMM_DEFAULT_UNROLL_N 4
3053 #define DGEMM_DEFAULT_UNROLL_M 2
3054 #define DGEMM_DEFAULT_UNROLL_N 2
3056 #define CGEMM_DEFAULT_UNROLL_M 2
3057 #define CGEMM_DEFAULT_UNROLL_N 2
3059 #define ZGEMM_DEFAULT_UNROLL_M 2
3060 #define ZGEMM_DEFAULT_UNROLL_N 2
3062 #define SGEMM_DEFAULT_P 128
3063 #define DGEMM_DEFAULT_P 128
3064 #define CGEMM_DEFAULT_P 96
3065 #define ZGEMM_DEFAULT_P 64
3067 #define SGEMM_DEFAULT_Q 240
3068 #define DGEMM_DEFAULT_Q 120
3069 #define CGEMM_DEFAULT_Q 120
3070 #define ZGEMM_DEFAULT_Q 120
3072 #define SGEMM_DEFAULT_R 12288
3073 #define DGEMM_DEFAULT_R 8192
3074 #define CGEMM_DEFAULT_R 4096
3075 #define ZGEMM_DEFAULT_R 4096
3077 #elif defined(THUNDERX2T99)
3079 #define SGEMM_DEFAULT_UNROLL_M 16
3080 #define SGEMM_DEFAULT_UNROLL_N 4
3082 #define DGEMM_DEFAULT_UNROLL_M 8
3083 #define DGEMM_DEFAULT_UNROLL_N 4
3085 #define CGEMM_DEFAULT_UNROLL_M 8
3086 #define CGEMM_DEFAULT_UNROLL_N 4
3088 #define ZGEMM_DEFAULT_UNROLL_M 4
3089 #define ZGEMM_DEFAULT_UNROLL_N 4
3091 #define SGEMM_DEFAULT_P 128
3092 #define DGEMM_DEFAULT_P 160
3093 #define CGEMM_DEFAULT_P 128
3094 #define ZGEMM_DEFAULT_P 128
3096 #define SGEMM_DEFAULT_Q 352
3097 #define DGEMM_DEFAULT_Q 128
3098 #define CGEMM_DEFAULT_Q 224
3099 #define ZGEMM_DEFAULT_Q 112
3101 #define SGEMM_DEFAULT_R 4096
3102 #define DGEMM_DEFAULT_R 4096
3103 #define CGEMM_DEFAULT_R 4096
3104 #define ZGEMM_DEFAULT_R 4096
3106 #elif defined(THUNDERX3T110)
3108 #define SGEMM_DEFAULT_UNROLL_M 16
3109 #define SGEMM_DEFAULT_UNROLL_N 4
3111 #define DGEMM_DEFAULT_UNROLL_M 8
3112 #define DGEMM_DEFAULT_UNROLL_N 4
3114 #define CGEMM_DEFAULT_UNROLL_M 8
3115 #define CGEMM_DEFAULT_UNROLL_N 4
3117 #define ZGEMM_DEFAULT_UNROLL_M 4
3118 #define ZGEMM_DEFAULT_UNROLL_N 4
3120 #define SGEMM_DEFAULT_P 128
3121 #define DGEMM_DEFAULT_P 320
3122 #define CGEMM_DEFAULT_P 128
3123 #define ZGEMM_DEFAULT_P 128
3125 #define SGEMM_DEFAULT_Q 352
3126 #define DGEMM_DEFAULT_Q 128
3127 #define CGEMM_DEFAULT_Q 224
3128 #define ZGEMM_DEFAULT_Q 112
3130 #define SGEMM_DEFAULT_R 4096
3131 #define DGEMM_DEFAULT_R 4096
3132 #define CGEMM_DEFAULT_R 4096
3133 #define ZGEMM_DEFAULT_R 4096
3135 #elif defined(NEOVERSEN1)
3137 #define SGEMM_DEFAULT_UNROLL_M 16
3138 #define SGEMM_DEFAULT_UNROLL_N 4
3140 #define DGEMM_DEFAULT_UNROLL_M 8
3141 #define DGEMM_DEFAULT_UNROLL_N 4
3143 #define CGEMM_DEFAULT_UNROLL_M 8
3144 #define CGEMM_DEFAULT_UNROLL_N 4
3146 #define ZGEMM_DEFAULT_UNROLL_M 4
3147 #define ZGEMM_DEFAULT_UNROLL_N 4
3149 #define SGEMM_DEFAULT_P 128
3150 #define DGEMM_DEFAULT_P 160
3151 #define CGEMM_DEFAULT_P 128
3152 #define ZGEMM_DEFAULT_P 128
3154 #define SGEMM_DEFAULT_Q 352
3155 #define DGEMM_DEFAULT_Q 128
3156 #define CGEMM_DEFAULT_Q 224
3157 #define ZGEMM_DEFAULT_Q 112
3159 #define SGEMM_DEFAULT_R 4096
3160 #define DGEMM_DEFAULT_R 4096
3161 #define CGEMM_DEFAULT_R 4096
3162 #define ZGEMM_DEFAULT_R 4096
3164 #else /* Other/undetected ARMv8 cores */
3166 #define SGEMM_DEFAULT_UNROLL_M 16
3167 #define SGEMM_DEFAULT_UNROLL_N 4
3169 #define DGEMM_DEFAULT_UNROLL_M 8
3170 #define DGEMM_DEFAULT_UNROLL_N 4
3172 #define CGEMM_DEFAULT_UNROLL_M 8
3173 #define CGEMM_DEFAULT_UNROLL_N 4
3175 #define ZGEMM_DEFAULT_UNROLL_M 4
3176 #define ZGEMM_DEFAULT_UNROLL_N 4
3178 #define SGEMM_DEFAULT_P 128
3179 #define DGEMM_DEFAULT_P 160
3180 #define CGEMM_DEFAULT_P 128
3181 #define ZGEMM_DEFAULT_P 128
3183 #define SGEMM_DEFAULT_Q 352
3184 #define DGEMM_DEFAULT_Q 128
3185 #define CGEMM_DEFAULT_Q 224
3186 #define ZGEMM_DEFAULT_Q 112
3188 #define SGEMM_DEFAULT_R 4096
3189 #define DGEMM_DEFAULT_R 4096
3190 #define CGEMM_DEFAULT_R 4096
3191 #define ZGEMM_DEFAULT_R 4096
3201 #define GEMM_DEFAULT_OFFSET_A 0
3202 #define GEMM_DEFAULT_OFFSET_B 0
3203 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3205 #define SGEMM_DEFAULT_UNROLL_M 2
3206 #define SGEMM_DEFAULT_UNROLL_N 2
3208 #define DGEMM_DEFAULT_UNROLL_M 2
3209 #define DGEMM_DEFAULT_UNROLL_N 2
3211 #define CGEMM_DEFAULT_UNROLL_M 2
3212 #define CGEMM_DEFAULT_UNROLL_N 2
3214 #define ZGEMM_DEFAULT_UNROLL_M 2
3215 #define ZGEMM_DEFAULT_UNROLL_N 2
3217 #define SGEMM_DEFAULT_P 128
3218 #define DGEMM_DEFAULT_P 128
3219 #define CGEMM_DEFAULT_P 96
3220 #define ZGEMM_DEFAULT_P 64
3222 #define SGEMM_DEFAULT_Q 240
3223 #define DGEMM_DEFAULT_Q 120
3224 #define CGEMM_DEFAULT_Q 120
3225 #define ZGEMM_DEFAULT_Q 120
3227 #define SGEMM_DEFAULT_R 12288
3228 #define DGEMM_DEFAULT_R 8192
3229 #define CGEMM_DEFAULT_R 4096
3230 #define ZGEMM_DEFAULT_R 4096
3242 #define GEMM_DEFAULT_OFFSET_A 0
3243 #define GEMM_DEFAULT_OFFSET_B 0
3244 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3246 #define SGEMM_DEFAULT_UNROLL_M 4
3247 #define SGEMM_DEFAULT_UNROLL_N 4
3249 #define DGEMM_DEFAULT_UNROLL_M 4
3250 #define DGEMM_DEFAULT_UNROLL_N 4
3252 #define CGEMM_DEFAULT_UNROLL_M 2
3253 #define CGEMM_DEFAULT_UNROLL_N 2
3255 #define ZGEMM_DEFAULT_UNROLL_M 2
3256 #define ZGEMM_DEFAULT_UNROLL_N 2
3258 #define SGEMM_DEFAULT_P 128
3259 #define DGEMM_DEFAULT_P 128
3260 #define CGEMM_DEFAULT_P 96
3261 #define ZGEMM_DEFAULT_P 64
3263 #define SGEMM_DEFAULT_Q 240
3264 #define DGEMM_DEFAULT_Q 120
3265 #define CGEMM_DEFAULT_Q 120
3266 #define ZGEMM_DEFAULT_Q 120
3268 #define SGEMM_DEFAULT_R 12288
3269 #define DGEMM_DEFAULT_R 8192
3270 #define CGEMM_DEFAULT_R 4096
3271 #define ZGEMM_DEFAULT_R 4096
3283 #define GEMM_DEFAULT_OFFSET_A 0
3284 #define GEMM_DEFAULT_OFFSET_B 0
3285 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3287 #define SGEMM_DEFAULT_UNROLL_M 4
3288 #define SGEMM_DEFAULT_UNROLL_N 4
3290 #define DGEMM_DEFAULT_UNROLL_M 4
3291 #define DGEMM_DEFAULT_UNROLL_N 4
3293 #define CGEMM_DEFAULT_UNROLL_M 2
3294 #define CGEMM_DEFAULT_UNROLL_N 2
3296 #define ZGEMM_DEFAULT_UNROLL_M 2
3297 #define ZGEMM_DEFAULT_UNROLL_N 2
3299 #define SGEMM_DEFAULT_P 128
3300 #define DGEMM_DEFAULT_P 128
3301 #define CGEMM_DEFAULT_P 96
3302 #define ZGEMM_DEFAULT_P 64
3304 #define SGEMM_DEFAULT_Q 240
3305 #define DGEMM_DEFAULT_Q 120
3306 #define CGEMM_DEFAULT_Q 120
3307 #define ZGEMM_DEFAULT_Q 120
3309 #define SGEMM_DEFAULT_R 12288
3310 #define DGEMM_DEFAULT_R 8192
3311 #define CGEMM_DEFAULT_R 4096
3312 #define ZGEMM_DEFAULT_R 4096
3320 #if defined(ZARCH_GENERIC)
3324 #define GEMM_DEFAULT_OFFSET_A 0
3325 #define GEMM_DEFAULT_OFFSET_B 0
3326 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3328 #define SGEMM_DEFAULT_UNROLL_M 2
3329 #define SGEMM_DEFAULT_UNROLL_N 2
3331 #define DGEMM_DEFAULT_UNROLL_M 2
3332 #define DGEMM_DEFAULT_UNROLL_N 2
3334 #define CGEMM_DEFAULT_UNROLL_M 2
3335 #define CGEMM_DEFAULT_UNROLL_N 2
3337 #define ZGEMM_DEFAULT_UNROLL_M 2
3338 #define ZGEMM_DEFAULT_UNROLL_N 2
3340 #define SGEMM_DEFAULT_P 128
3341 #define DGEMM_DEFAULT_P 128
3342 #define CGEMM_DEFAULT_P 96
3343 #define ZGEMM_DEFAULT_P 64
3345 #define SGEMM_DEFAULT_Q 240
3346 #define DGEMM_DEFAULT_Q 120
3347 #define CGEMM_DEFAULT_Q 120
3348 #define ZGEMM_DEFAULT_Q 120
3350 #define SGEMM_DEFAULT_R 12288
3351 #define DGEMM_DEFAULT_R 8192
3352 #define CGEMM_DEFAULT_R 4096
3353 #define ZGEMM_DEFAULT_R 4096
3363 #define GEMM_DEFAULT_OFFSET_A 0
3364 #define GEMM_DEFAULT_OFFSET_B 0
3365 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3367 #define SGEMM_DEFAULT_UNROLL_M 8
3368 #define SGEMM_DEFAULT_UNROLL_N 4
3370 #define DGEMM_DEFAULT_UNROLL_M 8
3371 #define DGEMM_DEFAULT_UNROLL_N 4
3373 #define CGEMM_DEFAULT_UNROLL_M 4
3374 #define CGEMM_DEFAULT_UNROLL_N 4
3376 #define ZGEMM_DEFAULT_UNROLL_M 4
3377 #define ZGEMM_DEFAULT_UNROLL_N 4
3379 #define SGEMM_DEFAULT_P 456
3380 #define DGEMM_DEFAULT_P 320
3381 #define CGEMM_DEFAULT_P 480
3382 #define ZGEMM_DEFAULT_P 224
3384 #define SGEMM_DEFAULT_Q 488
3385 #define DGEMM_DEFAULT_Q 384
3386 #define CGEMM_DEFAULT_Q 128
3387 #define ZGEMM_DEFAULT_Q 352
3389 #define SGEMM_DEFAULT_R 8192
3390 #define DGEMM_DEFAULT_R 4096
3391 #define CGEMM_DEFAULT_R 4096
3392 #define ZGEMM_DEFAULT_R 2048
3403 #define GEMM_DEFAULT_OFFSET_A 0
3404 #define GEMM_DEFAULT_OFFSET_B 0
3405 #define GEMM_DEFAULT_ALIGN 0x03fffUL
3407 #define SGEMM_DEFAULT_UNROLL_M 16
3408 #define SGEMM_DEFAULT_UNROLL_N 4
3410 #define DGEMM_DEFAULT_UNROLL_M 8
3411 #define DGEMM_DEFAULT_UNROLL_N 4
3413 #define CGEMM_DEFAULT_UNROLL_M 4
3414 #define CGEMM_DEFAULT_UNROLL_N 4
3416 #define ZGEMM_DEFAULT_UNROLL_M 4
3417 #define ZGEMM_DEFAULT_UNROLL_N 4
3419 #define SGEMM_DEFAULT_P 480
3420 #define DGEMM_DEFAULT_P 320
3421 #define CGEMM_DEFAULT_P 480
3422 #define ZGEMM_DEFAULT_P 224
3424 #define SGEMM_DEFAULT_Q 512
3425 #define DGEMM_DEFAULT_Q 384
3426 #define CGEMM_DEFAULT_Q 128
3427 #define ZGEMM_DEFAULT_Q 352
3429 #define SGEMM_DEFAULT_R 8192
3430 #define DGEMM_DEFAULT_R 4096
3431 #define CGEMM_DEFAULT_R 4096
3432 #define ZGEMM_DEFAULT_R 2048
3445 #define GEMM_DEFAULT_OFFSET_A 0
3446 #define GEMM_DEFAULT_OFFSET_B 0
3447 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
3449 #define SGEMM_DEFAULT_UNROLL_N 2
3450 #define DGEMM_DEFAULT_UNROLL_N 2
3451 #define QGEMM_DEFAULT_UNROLL_N 2
3452 #define CGEMM_DEFAULT_UNROLL_N 2
3453 #define ZGEMM_DEFAULT_UNROLL_N 2
3454 #define XGEMM_DEFAULT_UNROLL_N 1
3457 #define SGEMM_DEFAULT_UNROLL_M 2
3458 #define DGEMM_DEFAULT_UNROLL_M 2
3459 #define QGEMM_DEFAULT_UNROLL_M 2
3460 #define CGEMM_DEFAULT_UNROLL_M 2
3461 #define ZGEMM_DEFAULT_UNROLL_M 2
3462 #define XGEMM_DEFAULT_UNROLL_M 1
3464 #define SGEMM_DEFAULT_UNROLL_M 2
3465 #define DGEMM_DEFAULT_UNROLL_M 2
3466 #define QGEMM_DEFAULT_UNROLL_M 2
3467 #define CGEMM_DEFAULT_UNROLL_M 2
3468 #define ZGEMM_DEFAULT_UNROLL_M 2
3469 #define XGEMM_DEFAULT_UNROLL_M 1
3472 #define SGEMM_DEFAULT_P sgemm_p
3473 #define DGEMM_DEFAULT_P dgemm_p
3474 #define QGEMM_DEFAULT_P qgemm_p
3475 #define CGEMM_DEFAULT_P cgemm_p
3476 #define ZGEMM_DEFAULT_P zgemm_p
3477 #define XGEMM_DEFAULT_P xgemm_p
3479 #define SGEMM_DEFAULT_R sgemm_r
3480 #define DGEMM_DEFAULT_R dgemm_r
3481 #define QGEMM_DEFAULT_R qgemm_r
3482 #define CGEMM_DEFAULT_R cgemm_r
3483 #define ZGEMM_DEFAULT_R zgemm_r
3484 #define XGEMM_DEFAULT_R xgemm_r
3486 #define SGEMM_DEFAULT_Q 128
3487 #define DGEMM_DEFAULT_Q 128
3488 #define QGEMM_DEFAULT_Q 128
3489 #define CGEMM_DEFAULT_Q 128
3490 #define ZGEMM_DEFAULT_Q 128
3491 #define XGEMM_DEFAULT_Q 128
3497 #ifndef QGEMM_DEFAULT_UNROLL_M
3498 #define QGEMM_DEFAULT_UNROLL_M 2
3501 #ifndef QGEMM_DEFAULT_UNROLL_N
3502 #define QGEMM_DEFAULT_UNROLL_N 2
3505 #ifndef XGEMM_DEFAULT_UNROLL_M
3506 #define XGEMM_DEFAULT_UNROLL_M 2
3509 #ifndef XGEMM_DEFAULT_UNROLL_N
3510 #define XGEMM_DEFAULT_UNROLL_N 2
3514 #define SHUFPD_0 shufps $0x44,
3515 #define SHUFPD_1 shufps $0x4e,
3516 #define SHUFPD_2 shufps $0xe4,
3517 #define SHUFPD_3 shufps $0xee,
3521 #define SHUFPD_0 shufpd $0,
3525 #define SHUFPD_1 shufpd $1,
3529 #define SHUFPD_2 shufpd $2,
3533 #define SHUFPD_3 shufpd $3,
3537 #define SHUFPS_39 shufps $0x39,