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
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
1879 #define GEMM_DEFAULT_OFFSET_A 64
1880 #define GEMM_DEFAULT_OFFSET_B 0
1881 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
1886 #define SGEMM_DEFAULT_UNROLL_M 4
1887 #define DGEMM_DEFAULT_UNROLL_M 2
1888 #define QGEMM_DEFAULT_UNROLL_M 2
1889 #define CGEMM_DEFAULT_UNROLL_M 2
1890 #define ZGEMM_DEFAULT_UNROLL_M 1
1891 #define XGEMM_DEFAULT_UNROLL_M 1
1893 #define SGEMM_DEFAULT_UNROLL_M 8
1894 #define DGEMM_DEFAULT_UNROLL_M 4
1895 #define QGEMM_DEFAULT_UNROLL_M 2
1896 #define CGEMM_DEFAULT_UNROLL_M 4
1897 #define ZGEMM_DEFAULT_UNROLL_M 2
1898 #define XGEMM_DEFAULT_UNROLL_M 1
1901 #define SGEMM_DEFAULT_UNROLL_N 4
1902 #define DGEMM_DEFAULT_UNROLL_N 2
1903 #define QGEMM_DEFAULT_UNROLL_N 2
1904 #define CGEMM_DEFAULT_UNROLL_N 2
1905 #define ZGEMM_DEFAULT_UNROLL_N 1
1906 #define XGEMM_DEFAULT_UNROLL_N 1
1908 #define SGEMM_DEFAULT_P sgemm_p
1909 #define SGEMM_DEFAULT_R sgemm_r
1911 #define DGEMM_DEFAULT_P dgemm_p
1912 #define DGEMM_DEFAULT_R dgemm_r
1914 #define QGEMM_DEFAULT_P qgemm_p
1915 #define QGEMM_DEFAULT_R qgemm_r
1917 #define CGEMM_DEFAULT_P cgemm_p
1918 #define CGEMM_DEFAULT_R cgemm_r
1920 #define ZGEMM_DEFAULT_P zgemm_p
1921 #define ZGEMM_DEFAULT_R zgemm_r
1923 #define XGEMM_DEFAULT_P xgemm_p
1924 #define XGEMM_DEFAULT_R xgemm_r
1926 #define SGEMM_DEFAULT_Q 256
1927 #define DGEMM_DEFAULT_Q 256
1928 #define QGEMM_DEFAULT_Q 256
1929 #define CGEMM_DEFAULT_Q 256
1930 #define ZGEMM_DEFAULT_Q 256
1931 #define XGEMM_DEFAULT_Q 256
1941 #define GEMM_DEFAULT_OFFSET_A 0
1942 #define GEMM_DEFAULT_OFFSET_B 128
1943 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
1945 #define SGEMM_DEFAULT_UNROLL_M 8
1946 #define SGEMM_DEFAULT_UNROLL_N 8
1947 #define DGEMM_DEFAULT_UNROLL_M 8
1948 #define DGEMM_DEFAULT_UNROLL_N 8
1949 #define QGEMM_DEFAULT_UNROLL_M 8
1950 #define QGEMM_DEFAULT_UNROLL_N 8
1951 #define CGEMM_DEFAULT_UNROLL_M 4
1952 #define CGEMM_DEFAULT_UNROLL_N 4
1953 #define ZGEMM_DEFAULT_UNROLL_M 4
1954 #define ZGEMM_DEFAULT_UNROLL_N 4
1955 #define XGEMM_DEFAULT_UNROLL_M 4
1956 #define XGEMM_DEFAULT_UNROLL_N 4
1958 #define SGEMM_DEFAULT_P sgemm_p
1959 #define DGEMM_DEFAULT_P dgemm_p
1960 #define QGEMM_DEFAULT_P qgemm_p
1961 #define CGEMM_DEFAULT_P cgemm_p
1962 #define ZGEMM_DEFAULT_P zgemm_p
1963 #define XGEMM_DEFAULT_P xgemm_p
1965 #define SGEMM_DEFAULT_Q 1024
1966 #define DGEMM_DEFAULT_Q 1024
1967 #define QGEMM_DEFAULT_Q 1024
1968 #define CGEMM_DEFAULT_Q 1024
1969 #define ZGEMM_DEFAULT_Q 1024
1970 #define XGEMM_DEFAULT_Q 1024
1972 #define SGEMM_DEFAULT_R sgemm_r
1973 #define DGEMM_DEFAULT_R dgemm_r
1974 #define QGEMM_DEFAULT_R qgemm_r
1975 #define CGEMM_DEFAULT_R cgemm_r
1976 #define ZGEMM_DEFAULT_R zgemm_r
1977 #define XGEMM_DEFAULT_R xgemm_r
1981 #define GETRF_FACTOR 0.65
1985 #if defined(EV4) || defined(EV5) || defined(EV6)
1995 #define GEMM_DEFAULT_OFFSET_A 512
1996 #define GEMM_DEFAULT_OFFSET_B 512
1997 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
1999 #define SGEMM_DEFAULT_UNROLL_M 4
2000 #define SGEMM_DEFAULT_UNROLL_N 4
2001 #define DGEMM_DEFAULT_UNROLL_M 4
2002 #define DGEMM_DEFAULT_UNROLL_N 4
2003 #define CGEMM_DEFAULT_UNROLL_M 2
2004 #define CGEMM_DEFAULT_UNROLL_N 2
2005 #define ZGEMM_DEFAULT_UNROLL_M 2
2006 #define ZGEMM_DEFAULT_UNROLL_N 2
2011 #define SGEMM_DEFAULT_P 32
2012 #define SGEMM_DEFAULT_Q 112
2013 #define SGEMM_DEFAULT_R 256
2015 #define DGEMM_DEFAULT_P 32
2016 #define DGEMM_DEFAULT_Q 56
2017 #define DGEMM_DEFAULT_R 256
2019 #define CGEMM_DEFAULT_P 32
2020 #define CGEMM_DEFAULT_Q 64
2021 #define CGEMM_DEFAULT_R 240
2023 #define ZGEMM_DEFAULT_P 32
2024 #define ZGEMM_DEFAULT_Q 32
2025 #define ZGEMM_DEFAULT_R 240
2029 #define SGEMM_DEFAULT_P 64
2030 #define SGEMM_DEFAULT_Q 256
2032 #define DGEMM_DEFAULT_P 64
2033 #define DGEMM_DEFAULT_Q 128
2035 #define CGEMM_DEFAULT_P 64
2036 #define CGEMM_DEFAULT_Q 128
2038 #define ZGEMM_DEFAULT_P 64
2039 #define ZGEMM_DEFAULT_Q 64
2043 #define SGEMM_DEFAULT_P 256
2044 #define SGEMM_DEFAULT_Q 512
2046 #define DGEMM_DEFAULT_P 256
2047 #define DGEMM_DEFAULT_Q 256
2049 #define CGEMM_DEFAULT_P 256
2050 #define CGEMM_DEFAULT_Q 256
2052 #define ZGEMM_DEFAULT_P 128
2053 #define ZGEMM_DEFAULT_Q 256
2063 #define GEMM_DEFAULT_OFFSET_A 0
2064 #define GEMM_DEFAULT_OFFSET_B 8192
2065 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
2067 #define SGEMM_DEFAULT_UNROLL_M 16
2068 #define SGEMM_DEFAULT_UNROLL_N 4
2069 #define DGEMM_DEFAULT_UNROLL_M 4
2070 #define DGEMM_DEFAULT_UNROLL_N 4
2071 #define CGEMM_DEFAULT_UNROLL_M 8
2072 #define CGEMM_DEFAULT_UNROLL_N 2
2073 #define ZGEMM_DEFAULT_UNROLL_M 2
2074 #define ZGEMM_DEFAULT_UNROLL_N 2
2076 #define SGEMM_DEFAULT_P 128
2077 #define DGEMM_DEFAULT_P 128
2078 #define CGEMM_DEFAULT_P 128
2079 #define ZGEMM_DEFAULT_P 128
2081 #define SGEMM_DEFAULT_Q 512
2082 #define DGEMM_DEFAULT_Q 256
2083 #define CGEMM_DEFAULT_Q 256
2084 #define ZGEMM_DEFAULT_Q 128
2090 #define GEMM_DEFAULT_OFFSET_A 0
2091 #define GEMM_DEFAULT_OFFSET_B 1024
2092 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2094 #define SGEMM_DEFAULT_UNROLL_M 16
2095 #define SGEMM_DEFAULT_UNROLL_N 4
2096 #define DGEMM_DEFAULT_UNROLL_M 4
2097 #define DGEMM_DEFAULT_UNROLL_N 4
2098 #define CGEMM_DEFAULT_UNROLL_M 2
2099 #define CGEMM_DEFAULT_UNROLL_N 2
2100 #define ZGEMM_DEFAULT_UNROLL_M 2
2101 #define ZGEMM_DEFAULT_UNROLL_N 2
2103 #define SGEMM_DEFAULT_P 256
2104 #define DGEMM_DEFAULT_P 128
2105 #define CGEMM_DEFAULT_P 128
2106 #define ZGEMM_DEFAULT_P 64
2108 #define SGEMM_DEFAULT_Q 256
2109 #define DGEMM_DEFAULT_Q 256
2110 #define CGEMM_DEFAULT_Q 256
2111 #define ZGEMM_DEFAULT_Q 256
2121 #define GEMM_DEFAULT_OFFSET_A 2688
2122 #define GEMM_DEFAULT_OFFSET_B 3072
2123 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2125 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2126 #define SGEMM_DEFAULT_UNROLL_M 4
2128 #define SGEMM_DEFAULT_UNROLL_M 16
2130 #define SGEMM_DEFAULT_UNROLL_N 4
2131 #define DGEMM_DEFAULT_UNROLL_M 4
2132 #define DGEMM_DEFAULT_UNROLL_N 4
2133 #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2134 #define CGEMM_DEFAULT_UNROLL_M 2
2136 #define CGEMM_DEFAULT_UNROLL_M 8
2138 #define CGEMM_DEFAULT_UNROLL_N 2
2139 #define ZGEMM_DEFAULT_UNROLL_M 2
2140 #define ZGEMM_DEFAULT_UNROLL_N 2
2142 #if defined(OS_LINUX) || defined(OS_DARWIN) || defined(OS_FREEBSD)
2143 #if L2_SIZE == 1024976
2144 #define SGEMM_DEFAULT_P 320
2145 #define DGEMM_DEFAULT_P 256
2146 #define CGEMM_DEFAULT_P 256
2147 #define ZGEMM_DEFAULT_P 256
2149 #define SGEMM_DEFAULT_P 176
2150 #define DGEMM_DEFAULT_P 176
2151 #define CGEMM_DEFAULT_P 176
2152 #define ZGEMM_DEFAULT_P 176
2156 #define SGEMM_DEFAULT_Q 512
2157 #define DGEMM_DEFAULT_Q 256
2158 #define CGEMM_DEFAULT_Q 256
2159 #define ZGEMM_DEFAULT_Q 128
2170 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2171 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2172 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2174 #define SGEMM_DEFAULT_UNROLL_M 4
2175 #define SGEMM_DEFAULT_UNROLL_N 4
2176 #define DGEMM_DEFAULT_UNROLL_M 4
2177 #define DGEMM_DEFAULT_UNROLL_N 4
2178 #define CGEMM_DEFAULT_UNROLL_M 2
2179 #define CGEMM_DEFAULT_UNROLL_N 2
2180 #define ZGEMM_DEFAULT_UNROLL_M 2
2181 #define ZGEMM_DEFAULT_UNROLL_N 2
2183 #define SGEMM_DEFAULT_P 512
2184 #define DGEMM_DEFAULT_P 512
2185 #define CGEMM_DEFAULT_P 512
2186 #define ZGEMM_DEFAULT_P 512
2188 #define SGEMM_DEFAULT_Q 1024
2189 #define DGEMM_DEFAULT_Q 512
2190 #define CGEMM_DEFAULT_Q 512
2191 #define ZGEMM_DEFAULT_Q 256
2193 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2194 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2195 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2196 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2206 #define GEMM_DEFAULT_OFFSET_A (32 * 0)
2207 #define GEMM_DEFAULT_OFFSET_B (32 * 0)
2208 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2210 #define SGEMM_DEFAULT_UNROLL_M 8
2211 #define SGEMM_DEFAULT_UNROLL_N 4
2212 #define DGEMM_DEFAULT_UNROLL_M 8
2213 #define DGEMM_DEFAULT_UNROLL_N 4
2214 #define CGEMM_DEFAULT_UNROLL_M 4
2215 #define CGEMM_DEFAULT_UNROLL_N 2
2216 #define ZGEMM_DEFAULT_UNROLL_M 4
2217 #define ZGEMM_DEFAULT_UNROLL_N 2
2219 #define SGEMM_DEFAULT_P 128
2220 #define DGEMM_DEFAULT_P 128
2221 #define CGEMM_DEFAULT_P 128
2222 #define ZGEMM_DEFAULT_P 128
2224 #define SGEMM_DEFAULT_Q 4096
2225 #define DGEMM_DEFAULT_Q 3072
2226 #define CGEMM_DEFAULT_Q 2048
2227 #define ZGEMM_DEFAULT_Q 1024
2229 #define SGEMM_DEFAULT_Q 512
2230 #define DGEMM_DEFAULT_Q 256
2231 #define CGEMM_DEFAULT_Q 256
2232 #define ZGEMM_DEFAULT_Q 128
2240 #if defined(POWER3) || defined(POWER4) || defined(POWER5)
2241 #define GEMM_DEFAULT_OFFSET_A 0
2242 #define GEMM_DEFAULT_OFFSET_B 2048
2243 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2245 #define SGEMM_DEFAULT_UNROLL_M 4
2246 #define SGEMM_DEFAULT_UNROLL_N 4
2247 #define DGEMM_DEFAULT_UNROLL_M 4
2248 #define DGEMM_DEFAULT_UNROLL_N 4
2249 #define CGEMM_DEFAULT_UNROLL_M 2
2250 #define CGEMM_DEFAULT_UNROLL_N 2
2251 #define ZGEMM_DEFAULT_UNROLL_M 2
2252 #define ZGEMM_DEFAULT_UNROLL_N 2
2259 #define SGEMM_DEFAULT_P 256
2260 #define SGEMM_DEFAULT_Q 432
2261 #define SGEMM_DEFAULT_R 1012
2263 #define DGEMM_DEFAULT_P 256
2264 #define DGEMM_DEFAULT_Q 216
2265 #define DGEMM_DEFAULT_R 1012
2267 #define CGEMM_DEFAULT_P 256
2268 #define CGEMM_DEFAULT_Q 104
2269 #define CGEMM_DEFAULT_R 1012
2271 #define ZGEMM_DEFAULT_P 256
2272 #define ZGEMM_DEFAULT_Q 104
2273 #define ZGEMM_DEFAULT_R 1012
2277 #ifdef ALLOC_HUGETLB
2278 #define SGEMM_DEFAULT_P 184
2279 #define DGEMM_DEFAULT_P 184
2280 #define CGEMM_DEFAULT_P 184
2281 #define ZGEMM_DEFAULT_P 184
2283 #define SGEMM_DEFAULT_P 144
2284 #define DGEMM_DEFAULT_P 144
2285 #define CGEMM_DEFAULT_P 144
2286 #define ZGEMM_DEFAULT_P 144
2289 #define SGEMM_DEFAULT_Q 256
2290 #define CGEMM_DEFAULT_Q 256
2291 #define DGEMM_DEFAULT_Q 256
2292 #define ZGEMM_DEFAULT_Q 256
2296 #ifdef ALLOC_HUGETLB
2297 #define SGEMM_DEFAULT_P 512
2298 #define DGEMM_DEFAULT_P 256
2299 #define CGEMM_DEFAULT_P 256
2300 #define ZGEMM_DEFAULT_P 128
2302 #define SGEMM_DEFAULT_P 320
2303 #define DGEMM_DEFAULT_P 160
2304 #define CGEMM_DEFAULT_P 160
2305 #define ZGEMM_DEFAULT_P 80
2308 #define SGEMM_DEFAULT_Q 256
2309 #define CGEMM_DEFAULT_Q 256
2310 #define DGEMM_DEFAULT_Q 256
2311 #define ZGEMM_DEFAULT_Q 256
2323 #define GEMM_DEFAULT_OFFSET_A 384
2324 #define GEMM_DEFAULT_OFFSET_B 1024
2325 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2327 #define SGEMM_DEFAULT_UNROLL_M 4
2328 #define SGEMM_DEFAULT_UNROLL_N 4
2329 #define DGEMM_DEFAULT_UNROLL_M 4
2330 #define DGEMM_DEFAULT_UNROLL_N 4
2331 #define CGEMM_DEFAULT_UNROLL_M 2
2332 #define CGEMM_DEFAULT_UNROLL_N 4
2333 #define ZGEMM_DEFAULT_UNROLL_M 2
2334 #define ZGEMM_DEFAULT_UNROLL_N 4
2336 #define SGEMM_DEFAULT_P 992
2337 #define DGEMM_DEFAULT_P 480
2338 #define CGEMM_DEFAULT_P 488
2339 #define ZGEMM_DEFAULT_P 248
2341 #define SGEMM_DEFAULT_Q 504
2342 #define DGEMM_DEFAULT_Q 504
2343 #define CGEMM_DEFAULT_Q 400
2344 #define ZGEMM_DEFAULT_Q 400
2355 #define GEMM_DEFAULT_OFFSET_A 0
2356 #define GEMM_DEFAULT_OFFSET_B 65536
2358 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2359 #if defined(__32BIT__)
2360 #warning using BINARY32==POWER6
2361 #define SGEMM_DEFAULT_UNROLL_M 4
2362 #define SGEMM_DEFAULT_UNROLL_N 4
2363 #define DGEMM_DEFAULT_UNROLL_M 4
2364 #define DGEMM_DEFAULT_UNROLL_N 4
2365 #define CGEMM_DEFAULT_UNROLL_M 2
2366 #define CGEMM_DEFAULT_UNROLL_N 4
2367 #define ZGEMM_DEFAULT_UNROLL_M 2
2368 #define ZGEMM_DEFAULT_UNROLL_N 4
2370 #define SGEMM_DEFAULT_UNROLL_M 16
2371 #define SGEMM_DEFAULT_UNROLL_N 8
2372 #define DGEMM_DEFAULT_UNROLL_M 16
2373 #define DGEMM_DEFAULT_UNROLL_N 4
2374 #define CGEMM_DEFAULT_UNROLL_M 8
2375 #define CGEMM_DEFAULT_UNROLL_N 4
2376 #define ZGEMM_DEFAULT_UNROLL_M 8
2377 #define ZGEMM_DEFAULT_UNROLL_N 2
2379 #define SGEMM_DEFAULT_P 1280UL
2380 #define DGEMM_DEFAULT_P 640UL
2381 #define CGEMM_DEFAULT_P 640UL
2382 #define ZGEMM_DEFAULT_P 320UL
2384 #define SGEMM_DEFAULT_Q 640UL
2385 #define DGEMM_DEFAULT_Q 720UL
2386 #define CGEMM_DEFAULT_Q 640UL
2387 #define ZGEMM_DEFAULT_Q 640UL
2390 #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
2391 #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
2392 #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
2393 #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
2395 #define SGEMM_DEFAULT_R 4096
2396 #define DGEMM_DEFAULT_R 4096
2397 #define CGEMM_DEFAULT_R 4096
2398 #define ZGEMM_DEFAULT_R 4096
2409 #define GEMM_DEFAULT_OFFSET_A 0
2410 #define GEMM_DEFAULT_OFFSET_B 65536
2411 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2413 #define SWITCH_RATIO 16
2414 #define GEMM_PREFERED_SIZE 16
2416 #define SGEMM_DEFAULT_UNROLL_M 16
2417 #define SGEMM_DEFAULT_UNROLL_N 8
2418 #define DGEMM_DEFAULT_UNROLL_M 16
2419 #define DGEMM_DEFAULT_UNROLL_N 4
2420 #define CGEMM_DEFAULT_UNROLL_M 8
2421 #define CGEMM_DEFAULT_UNROLL_N 4
2422 #define ZGEMM_DEFAULT_UNROLL_M 8
2423 #define ZGEMM_DEFAULT_UNROLL_N 2
2425 #define SGEMM_DEFAULT_P 832
2426 #define DGEMM_DEFAULT_P 128
2427 #define CGEMM_DEFAULT_P 512
2428 #define ZGEMM_DEFAULT_P 256
2430 #define SGEMM_DEFAULT_Q 1026
2431 #define DGEMM_DEFAULT_Q 384
2432 #define CGEMM_DEFAULT_Q 1026
2433 #define ZGEMM_DEFAULT_Q 1026
2435 #define SGEMM_DEFAULT_R 4096
2436 #define DGEMM_DEFAULT_R 4096
2437 #define CGEMM_DEFAULT_R 4096
2438 #define ZGEMM_DEFAULT_R 4096
2444 #if defined(POWER10)
2448 #define GEMM_DEFAULT_OFFSET_A 0
2449 #define GEMM_DEFAULT_OFFSET_B 65536
2450 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
2452 #define SWITCH_RATIO 16
2453 #define GEMM_PREFERED_SIZE 16
2455 #define SGEMM_DEFAULT_UNROLL_M 16
2456 #define SGEMM_DEFAULT_UNROLL_N 8
2457 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2458 #define DGEMM_DEFAULT_UNROLL_M 16
2459 #define DGEMM_DEFAULT_UNROLL_N 4
2461 #define DGEMM_DEFAULT_UNROLL_M 8
2462 #define DGEMM_DEFAULT_UNROLL_N 8
2464 #define CGEMM_DEFAULT_UNROLL_M 8
2465 #define CGEMM_DEFAULT_UNROLL_N 4
2466 #define ZGEMM_DEFAULT_UNROLL_M 8
2467 #define ZGEMM_DEFAULT_UNROLL_N 2
2469 #define SGEMM_DEFAULT_P 512
2470 #define DGEMM_DEFAULT_P 384
2471 #define CGEMM_DEFAULT_P 512
2472 #define ZGEMM_DEFAULT_P 256
2474 #define SGEMM_DEFAULT_Q 512
2475 #define DGEMM_DEFAULT_Q 512
2476 #define CGEMM_DEFAULT_Q 384
2477 #define ZGEMM_DEFAULT_Q 384
2479 #define SGEMM_DEFAULT_R 4096
2480 #define DGEMM_DEFAULT_R 4096
2481 #define CGEMM_DEFAULT_R 4096
2482 #define ZGEMM_DEFAULT_R 4096
2486 #undef SBGEMM_DEFAULT_UNROLL_N
2487 #undef SBGEMM_DEFAULT_UNROLL_M
2488 #undef SBGEMM_DEFAULT_P
2489 #undef SBGEMM_DEFAULT_R
2490 #undef SBGEMM_DEFAULT_Q
2491 #define SBGEMM_DEFAULT_UNROLL_M 16
2492 #define SBGEMM_DEFAULT_UNROLL_N 8
2493 #define SBGEMM_DEFAULT_P 832
2494 #define SBGEMM_DEFAULT_Q 1026
2495 #define SBGEMM_DEFAULT_R 4096
2498 #if defined(SPARC) && defined(V7)
2503 #define GEMM_DEFAULT_OFFSET_A 0
2504 #define GEMM_DEFAULT_OFFSET_B 2048
2505 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2507 #define SGEMM_DEFAULT_UNROLL_M 2
2508 #define SGEMM_DEFAULT_UNROLL_N 8
2509 #define DGEMM_DEFAULT_UNROLL_M 2
2510 #define DGEMM_DEFAULT_UNROLL_N 8
2511 #define CGEMM_DEFAULT_UNROLL_M 1
2512 #define CGEMM_DEFAULT_UNROLL_N 4
2513 #define ZGEMM_DEFAULT_UNROLL_M 1
2514 #define ZGEMM_DEFAULT_UNROLL_N 4
2516 #define SGEMM_DEFAULT_P 256
2517 #define DGEMM_DEFAULT_P 256
2518 #define CGEMM_DEFAULT_P 256
2519 #define ZGEMM_DEFAULT_P 256
2521 #define SGEMM_DEFAULT_Q 512
2522 #define DGEMM_DEFAULT_Q 256
2523 #define CGEMM_DEFAULT_Q 256
2524 #define ZGEMM_DEFAULT_Q 128
2527 #define GEMM_THREAD gemm_thread_mn
2530 #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
2535 #define GEMM_DEFAULT_OFFSET_A 0
2536 #define GEMM_DEFAULT_OFFSET_B 2048
2537 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2539 #define SGEMM_DEFAULT_UNROLL_M 4
2540 #define SGEMM_DEFAULT_UNROLL_N 4
2541 #define DGEMM_DEFAULT_UNROLL_M 4
2542 #define DGEMM_DEFAULT_UNROLL_N 4
2543 #define CGEMM_DEFAULT_UNROLL_M 2
2544 #define CGEMM_DEFAULT_UNROLL_N 2
2545 #define ZGEMM_DEFAULT_UNROLL_M 2
2546 #define ZGEMM_DEFAULT_UNROLL_N 2
2548 #define SGEMM_DEFAULT_P 512
2549 #define DGEMM_DEFAULT_P 512
2550 #define CGEMM_DEFAULT_P 512
2551 #define ZGEMM_DEFAULT_P 512
2553 #define SGEMM_DEFAULT_Q 1024
2554 #define DGEMM_DEFAULT_Q 512
2555 #define CGEMM_DEFAULT_Q 512
2556 #define ZGEMM_DEFAULT_Q 256
2566 #define GEMM_DEFAULT_OFFSET_A 0
2567 #define GEMM_DEFAULT_OFFSET_B 0
2568 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2570 #define SGEMM_DEFAULT_UNROLL_M 2
2571 #define SGEMM_DEFAULT_UNROLL_N 8
2572 #define DGEMM_DEFAULT_UNROLL_M 2
2573 #define DGEMM_DEFAULT_UNROLL_N 8
2574 #define CGEMM_DEFAULT_UNROLL_M 1
2575 #define CGEMM_DEFAULT_UNROLL_N 4
2576 #define ZGEMM_DEFAULT_UNROLL_M 1
2577 #define ZGEMM_DEFAULT_UNROLL_N 4
2579 #define SGEMM_DEFAULT_P 108
2580 #define DGEMM_DEFAULT_P 112
2581 #define CGEMM_DEFAULT_P 108
2582 #define ZGEMM_DEFAULT_P 112
2584 #define SGEMM_DEFAULT_Q 288
2585 #define DGEMM_DEFAULT_Q 144
2586 #define CGEMM_DEFAULT_Q 144
2587 #define ZGEMM_DEFAULT_Q 72
2589 #define SGEMM_DEFAULT_R 2000
2590 #define DGEMM_DEFAULT_R 2000
2591 #define CGEMM_DEFAULT_R 2000
2592 #define ZGEMM_DEFAULT_R 2000
2597 #if defined(LOONGSON3R4)
2601 #define GEMM_DEFAULT_OFFSET_A 0
2602 #define GEMM_DEFAULT_OFFSET_B 0
2603 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2606 #define SGEMM_DEFAULT_UNROLL_M 8
2607 #define SGEMM_DEFAULT_UNROLL_N 8
2609 #define DGEMM_DEFAULT_UNROLL_M 8
2610 #define DGEMM_DEFAULT_UNROLL_N 4
2612 #define CGEMM_DEFAULT_UNROLL_M 8
2613 #define CGEMM_DEFAULT_UNROLL_N 4
2615 #define ZGEMM_DEFAULT_UNROLL_M 4
2616 #define ZGEMM_DEFAULT_UNROLL_N 4
2618 #define SGEMM_DEFAULT_UNROLL_M 8
2619 #define SGEMM_DEFAULT_UNROLL_N 4
2621 #define DGEMM_DEFAULT_UNROLL_M 4
2622 #define DGEMM_DEFAULT_UNROLL_N 4
2624 #define CGEMM_DEFAULT_UNROLL_M 4
2625 #define CGEMM_DEFAULT_UNROLL_N 2
2627 #define ZGEMM_DEFAULT_UNROLL_M 2
2628 #define ZGEMM_DEFAULT_UNROLL_N 2
2631 #define SGEMM_DEFAULT_P 64
2632 #define DGEMM_DEFAULT_P 44
2633 #define CGEMM_DEFAULT_P 64
2634 #define ZGEMM_DEFAULT_P 32
2636 #define SGEMM_DEFAULT_Q 192
2637 #define DGEMM_DEFAULT_Q 92
2638 #define CGEMM_DEFAULT_Q 128
2639 #define ZGEMM_DEFAULT_Q 80
2641 #define SGEMM_DEFAULT_R 640
2642 #define DGEMM_DEFAULT_R dgemm_r
2643 #define CGEMM_DEFAULT_R 640
2644 #define ZGEMM_DEFAULT_R 640
2646 #define GEMM_OFFSET_A1 0x10000
2647 #define GEMM_OFFSET_B1 0x100000
2652 #if defined(LOONGSON3R3)
2653 ////Copy from SICORTEX
2657 #define GEMM_DEFAULT_OFFSET_A 0
2658 #define GEMM_DEFAULT_OFFSET_B 0
2659 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2661 #define SGEMM_DEFAULT_UNROLL_M 8
2662 #define SGEMM_DEFAULT_UNROLL_N 4
2664 #define DGEMM_DEFAULT_UNROLL_M 4
2665 #define DGEMM_DEFAULT_UNROLL_N 4
2667 #define CGEMM_DEFAULT_UNROLL_M 4
2668 #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 64
2674 #define DGEMM_DEFAULT_P 44
2675 #define CGEMM_DEFAULT_P 64
2676 #define ZGEMM_DEFAULT_P 32
2678 #define SGEMM_DEFAULT_Q 192
2679 #define DGEMM_DEFAULT_Q 92
2680 #define CGEMM_DEFAULT_Q 128
2681 #define ZGEMM_DEFAULT_Q 80
2683 #define SGEMM_DEFAULT_R 640
2684 #define DGEMM_DEFAULT_R dgemm_r
2685 #define CGEMM_DEFAULT_R 640
2686 #define ZGEMM_DEFAULT_R 640
2688 #define GEMM_OFFSET_A1 0x10000
2689 #define GEMM_OFFSET_B1 0x100000
2694 #if defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
2698 #define GEMM_DEFAULT_OFFSET_A 0
2699 #define GEMM_DEFAULT_OFFSET_B 0
2700 #define GEMM_DEFAULT_ALIGN (BLASLONG) 0x03fffUL
2703 #define SGEMM_DEFAULT_UNROLL_M 8
2704 #define SGEMM_DEFAULT_UNROLL_N 8
2706 #define DGEMM_DEFAULT_UNROLL_M 8
2707 #define DGEMM_DEFAULT_UNROLL_N 4
2709 #define CGEMM_DEFAULT_UNROLL_M 8
2710 #define CGEMM_DEFAULT_UNROLL_N 4
2712 #define ZGEMM_DEFAULT_UNROLL_M 4
2713 #define ZGEMM_DEFAULT_UNROLL_N 4
2715 #define SGEMM_DEFAULT_UNROLL_M 2
2716 #define SGEMM_DEFAULT_UNROLL_N 2
2718 #define DGEMM_DEFAULT_UNROLL_M 2
2719 #define DGEMM_DEFAULT_UNROLL_N 2
2721 #define CGEMM_DEFAULT_UNROLL_M 2
2722 #define CGEMM_DEFAULT_UNROLL_N 2
2724 #define ZGEMM_DEFAULT_UNROLL_M 2
2725 #define ZGEMM_DEFAULT_UNROLL_N 2
2728 #define SGEMM_DEFAULT_P 128
2729 #define DGEMM_DEFAULT_P 128
2730 #define CGEMM_DEFAULT_P 96
2731 #define ZGEMM_DEFAULT_P 64
2733 #define SGEMM_DEFAULT_Q 240
2734 #define DGEMM_DEFAULT_Q 120
2735 #define CGEMM_DEFAULT_Q 120
2736 #define ZGEMM_DEFAULT_Q 120
2738 #define SGEMM_DEFAULT_R 12288
2739 #define DGEMM_DEFAULT_R 8192
2740 #define CGEMM_DEFAULT_R 4096
2741 #define ZGEMM_DEFAULT_R 4096
2746 #ifdef RISCV64_GENERIC
2747 #define GEMM_DEFAULT_OFFSET_A 0
2748 #define GEMM_DEFAULT_OFFSET_B 0
2749 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2751 #define SGEMM_DEFAULT_UNROLL_M 2
2752 #define SGEMM_DEFAULT_UNROLL_N 2
2754 #define DGEMM_DEFAULT_UNROLL_M 2
2755 #define DGEMM_DEFAULT_UNROLL_N 2
2757 #define CGEMM_DEFAULT_UNROLL_M 2
2758 #define CGEMM_DEFAULT_UNROLL_N 2
2760 #define ZGEMM_DEFAULT_UNROLL_M 2
2761 #define ZGEMM_DEFAULT_UNROLL_N 2
2763 #define SGEMM_DEFAULT_P 128
2764 #define DGEMM_DEFAULT_P 128
2765 #define CGEMM_DEFAULT_P 96
2766 #define ZGEMM_DEFAULT_P 64
2768 #define SGEMM_DEFAULT_Q 240
2769 #define DGEMM_DEFAULT_Q 120
2770 #define CGEMM_DEFAULT_Q 120
2771 #define ZGEMM_DEFAULT_Q 120
2773 #define SGEMM_DEFAULT_R 12288
2774 #define DGEMM_DEFAULT_R 8192
2775 #define CGEMM_DEFAULT_R 4096
2776 #define ZGEMM_DEFAULT_R 4096
2780 #define GEMM_DEFAULT_OFFSET_A 0
2781 #define GEMM_DEFAULT_OFFSET_B 0
2786 #define GEMM_DEFAULT_OFFSET_A 0
2787 #define GEMM_DEFAULT_OFFSET_B 0
2788 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2790 #define SGEMM_DEFAULT_UNROLL_M 16
2791 #define SGEMM_DEFAULT_UNROLL_N 4
2793 #define DGEMM_DEFAULT_UNROLL_M 8
2794 #define DGEMM_DEFAULT_UNROLL_N 4
2796 #define CGEMM_DEFAULT_UNROLL_M 2
2797 #define CGEMM_DEFAULT_UNROLL_N 2
2799 #define ZGEMM_DEFAULT_UNROLL_M 2
2800 #define ZGEMM_DEFAULT_UNROLL_N 2
2802 #define SGEMM_DEFAULT_P 160
2803 #define DGEMM_DEFAULT_P 160
2804 #define CGEMM_DEFAULT_P 96
2805 #define ZGEMM_DEFAULT_P 64
2807 #define SGEMM_DEFAULT_Q 240
2808 #define DGEMM_DEFAULT_Q 128
2809 #define CGEMM_DEFAULT_Q 120
2810 #define ZGEMM_DEFAULT_Q 120
2812 #define SGEMM_DEFAULT_R 12288
2813 #define DGEMM_DEFAULT_R 8192
2814 #define CGEMM_DEFAULT_R 4096
2815 #define ZGEMM_DEFAULT_R 4096
2819 #define GEMM_DEFAULT_OFFSET_A 0
2820 #define GEMM_DEFAULT_OFFSET_B 0
2828 #define GEMM_DEFAULT_OFFSET_A 0
2829 #define GEMM_DEFAULT_OFFSET_B 0
2830 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2832 #define SGEMM_DEFAULT_UNROLL_M 4
2833 #define SGEMM_DEFAULT_UNROLL_N 4
2835 #define DGEMM_DEFAULT_UNROLL_M 4
2836 #define DGEMM_DEFAULT_UNROLL_N 4
2838 #define CGEMM_DEFAULT_UNROLL_M 2
2839 #define CGEMM_DEFAULT_UNROLL_N 2
2841 #define ZGEMM_DEFAULT_UNROLL_M 2
2842 #define ZGEMM_DEFAULT_UNROLL_N 2
2844 #define SGEMM_DEFAULT_P 128
2845 #define DGEMM_DEFAULT_P 128
2846 #define CGEMM_DEFAULT_P 96
2847 #define ZGEMM_DEFAULT_P 64
2849 #define SGEMM_DEFAULT_Q 240
2850 #define DGEMM_DEFAULT_Q 120
2851 #define CGEMM_DEFAULT_Q 120
2852 #define ZGEMM_DEFAULT_Q 120
2854 #define SGEMM_DEFAULT_R 12288
2855 #define DGEMM_DEFAULT_R 8192
2856 #define CGEMM_DEFAULT_R 4096
2857 #define ZGEMM_DEFAULT_R 4096
2869 #define GEMM_DEFAULT_OFFSET_A 0
2870 #define GEMM_DEFAULT_OFFSET_B 0
2871 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
2873 #define SGEMM_DEFAULT_UNROLL_M 4
2874 #define SGEMM_DEFAULT_UNROLL_N 2
2876 #define DGEMM_DEFAULT_UNROLL_M 4
2877 #define DGEMM_DEFAULT_UNROLL_N 2
2879 #define CGEMM_DEFAULT_UNROLL_M 2
2880 #define CGEMM_DEFAULT_UNROLL_N 2
2882 #define ZGEMM_DEFAULT_UNROLL_M 2
2883 #define ZGEMM_DEFAULT_UNROLL_N 2
2885 #define SGEMM_DEFAULT_P 128
2886 #define DGEMM_DEFAULT_P 128
2887 #define CGEMM_DEFAULT_P 96
2888 #define ZGEMM_DEFAULT_P 64
2890 #define SGEMM_DEFAULT_Q 240
2891 #define DGEMM_DEFAULT_Q 120
2892 #define CGEMM_DEFAULT_Q 120
2893 #define ZGEMM_DEFAULT_Q 120
2895 #define SGEMM_DEFAULT_R 12288
2896 #define DGEMM_DEFAULT_R 8192
2897 #define CGEMM_DEFAULT_R 4096
2898 #define ZGEMM_DEFAULT_R 4096
2904 /* Common ARMv8 parameters */
2910 #define GEMM_DEFAULT_OFFSET_A 0
2911 #define GEMM_DEFAULT_OFFSET_B 0
2912 #define GEMM_DEFAULT_ALIGN 0x03fffUL
2916 #if defined(CORTEXA57) || \
2917 defined(CORTEXA72) || defined(CORTEXA73) || \
2918 defined(FALKOR) || defined(TSV110) || defined(EMAG8180)
2920 #define SGEMM_DEFAULT_UNROLL_M 16
2921 #define SGEMM_DEFAULT_UNROLL_N 4
2923 #define DGEMM_DEFAULT_UNROLL_M 8
2924 #define DGEMM_DEFAULT_UNROLL_N 4
2926 #define CGEMM_DEFAULT_UNROLL_M 8
2927 #define CGEMM_DEFAULT_UNROLL_N 4
2929 #define ZGEMM_DEFAULT_UNROLL_M 4
2930 #define ZGEMM_DEFAULT_UNROLL_N 4
2932 /*FIXME: this should be using the cache size, but there is currently no easy way to
2933 query that on ARM. So if getarch counted more than 8 cores we simply assume the host
2934 is a big desktop or server with abundant cache rather than a phone or embedded device */
2935 #if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180)
2936 #define SGEMM_DEFAULT_P 512
2937 #define DGEMM_DEFAULT_P 256
2938 #define CGEMM_DEFAULT_P 256
2939 #define ZGEMM_DEFAULT_P 128
2941 #define SGEMM_DEFAULT_Q 1024
2942 #define DGEMM_DEFAULT_Q 512
2943 #define CGEMM_DEFAULT_Q 512
2944 #define ZGEMM_DEFAULT_Q 512
2946 #define SGEMM_DEFAULT_P 128
2947 #define DGEMM_DEFAULT_P 160
2948 #define CGEMM_DEFAULT_P 128
2949 #define ZGEMM_DEFAULT_P 128
2951 #define SGEMM_DEFAULT_Q 352
2952 #define DGEMM_DEFAULT_Q 128
2953 #define CGEMM_DEFAULT_Q 224
2954 #define ZGEMM_DEFAULT_Q 112
2957 #define SGEMM_DEFAULT_R 4096
2958 #define DGEMM_DEFAULT_R 4096
2959 #define CGEMM_DEFAULT_R 4096
2960 #define ZGEMM_DEFAULT_R 2048
2962 #elif defined(CORTEXA53) || defined(CORTEXA55)
2964 #define SGEMM_DEFAULT_UNROLL_M 8
2965 #define SGEMM_DEFAULT_UNROLL_N 8
2967 #define DGEMM_DEFAULT_UNROLL_M 8
2968 #define DGEMM_DEFAULT_UNROLL_N 4
2970 #define CGEMM_DEFAULT_UNROLL_M 8
2971 #define CGEMM_DEFAULT_UNROLL_N 4
2973 #define ZGEMM_DEFAULT_UNROLL_M 4
2974 #define ZGEMM_DEFAULT_UNROLL_N 4
2976 #define SGEMM_DEFAULT_P 256
2977 #define DGEMM_DEFAULT_P 160
2978 #define CGEMM_DEFAULT_P 128
2979 #define ZGEMM_DEFAULT_P 128
2981 #define SGEMM_DEFAULT_Q 256
2982 #define DGEMM_DEFAULT_Q 128
2983 #define CGEMM_DEFAULT_Q 224
2984 #define ZGEMM_DEFAULT_Q 112
2986 #define SGEMM_DEFAULT_R 4096
2987 #define DGEMM_DEFAULT_R 4096
2988 #define CGEMM_DEFAULT_R 4096
2989 #define ZGEMM_DEFAULT_R 2048
2991 #elif defined(THUNDERX)
2993 #define SGEMM_DEFAULT_UNROLL_M 4
2994 #define SGEMM_DEFAULT_UNROLL_N 4
2996 #define DGEMM_DEFAULT_UNROLL_M 2
2997 #define DGEMM_DEFAULT_UNROLL_N 2
2999 #define CGEMM_DEFAULT_UNROLL_M 2
3000 #define CGEMM_DEFAULT_UNROLL_N 2
3002 #define ZGEMM_DEFAULT_UNROLL_M 2
3003 #define ZGEMM_DEFAULT_UNROLL_N 2
3005 #define SGEMM_DEFAULT_P 128
3006 #define DGEMM_DEFAULT_P 128
3007 #define CGEMM_DEFAULT_P 96
3008 #define ZGEMM_DEFAULT_P 64
3010 #define SGEMM_DEFAULT_Q 240
3011 #define DGEMM_DEFAULT_Q 120
3012 #define CGEMM_DEFAULT_Q 120
3013 #define ZGEMM_DEFAULT_Q 120
3015 #define SGEMM_DEFAULT_R 12288
3016 #define DGEMM_DEFAULT_R 8192
3017 #define CGEMM_DEFAULT_R 4096
3018 #define ZGEMM_DEFAULT_R 4096
3020 #elif defined(THUNDERX2T99)
3022 #define SGEMM_DEFAULT_UNROLL_M 16
3023 #define SGEMM_DEFAULT_UNROLL_N 4
3025 #define DGEMM_DEFAULT_UNROLL_M 8
3026 #define DGEMM_DEFAULT_UNROLL_N 4
3028 #define CGEMM_DEFAULT_UNROLL_M 8
3029 #define CGEMM_DEFAULT_UNROLL_N 4
3031 #define ZGEMM_DEFAULT_UNROLL_M 4
3032 #define ZGEMM_DEFAULT_UNROLL_N 4
3034 #define SGEMM_DEFAULT_P 128
3035 #define DGEMM_DEFAULT_P 160
3036 #define CGEMM_DEFAULT_P 128
3037 #define ZGEMM_DEFAULT_P 128
3039 #define SGEMM_DEFAULT_Q 352
3040 #define DGEMM_DEFAULT_Q 128
3041 #define CGEMM_DEFAULT_Q 224
3042 #define ZGEMM_DEFAULT_Q 112
3044 #define SGEMM_DEFAULT_R 4096
3045 #define DGEMM_DEFAULT_R 4096
3046 #define CGEMM_DEFAULT_R 4096
3047 #define ZGEMM_DEFAULT_R 4096
3049 #elif defined(THUNDERX3T110)
3051 #define SGEMM_DEFAULT_UNROLL_M 16
3052 #define SGEMM_DEFAULT_UNROLL_N 4
3054 #define DGEMM_DEFAULT_UNROLL_M 8
3055 #define DGEMM_DEFAULT_UNROLL_N 4
3057 #define CGEMM_DEFAULT_UNROLL_M 8
3058 #define CGEMM_DEFAULT_UNROLL_N 4
3060 #define ZGEMM_DEFAULT_UNROLL_M 4
3061 #define ZGEMM_DEFAULT_UNROLL_N 4
3063 #define SGEMM_DEFAULT_P 128
3064 #define DGEMM_DEFAULT_P 320
3065 #define CGEMM_DEFAULT_P 128
3066 #define ZGEMM_DEFAULT_P 128
3068 #define SGEMM_DEFAULT_Q 352
3069 #define DGEMM_DEFAULT_Q 128
3070 #define CGEMM_DEFAULT_Q 224
3071 #define ZGEMM_DEFAULT_Q 112
3073 #define SGEMM_DEFAULT_R 4096
3074 #define DGEMM_DEFAULT_R 4096
3075 #define CGEMM_DEFAULT_R 4096
3076 #define ZGEMM_DEFAULT_R 4096
3078 #elif defined(NEOVERSEN1)
3080 #define SGEMM_DEFAULT_UNROLL_M 16
3081 #define SGEMM_DEFAULT_UNROLL_N 4
3083 #define DGEMM_DEFAULT_UNROLL_M 8
3084 #define DGEMM_DEFAULT_UNROLL_N 4
3086 #define CGEMM_DEFAULT_UNROLL_M 8
3087 #define CGEMM_DEFAULT_UNROLL_N 4
3089 #define ZGEMM_DEFAULT_UNROLL_M 4
3090 #define ZGEMM_DEFAULT_UNROLL_N 4
3092 #define SGEMM_DEFAULT_P 128
3093 #define DGEMM_DEFAULT_P 160
3094 #define CGEMM_DEFAULT_P 128
3095 #define ZGEMM_DEFAULT_P 128
3097 #define SGEMM_DEFAULT_Q 352
3098 #define DGEMM_DEFAULT_Q 128
3099 #define CGEMM_DEFAULT_Q 224
3100 #define ZGEMM_DEFAULT_Q 112
3102 #define SGEMM_DEFAULT_R 4096
3103 #define DGEMM_DEFAULT_R 4096
3104 #define CGEMM_DEFAULT_R 4096
3105 #define ZGEMM_DEFAULT_R 4096
3107 #else /* Other/undetected ARMv8 cores */
3109 #define SGEMM_DEFAULT_UNROLL_M 16
3110 #define SGEMM_DEFAULT_UNROLL_N 4
3112 #define DGEMM_DEFAULT_UNROLL_M 8
3113 #define DGEMM_DEFAULT_UNROLL_N 4
3115 #define CGEMM_DEFAULT_UNROLL_M 8
3116 #define CGEMM_DEFAULT_UNROLL_N 4
3118 #define ZGEMM_DEFAULT_UNROLL_M 4
3119 #define ZGEMM_DEFAULT_UNROLL_N 4
3121 #define SGEMM_DEFAULT_P 128
3122 #define DGEMM_DEFAULT_P 160
3123 #define CGEMM_DEFAULT_P 128
3124 #define ZGEMM_DEFAULT_P 128
3126 #define SGEMM_DEFAULT_Q 352
3127 #define DGEMM_DEFAULT_Q 128
3128 #define CGEMM_DEFAULT_Q 224
3129 #define ZGEMM_DEFAULT_Q 112
3131 #define SGEMM_DEFAULT_R 4096
3132 #define DGEMM_DEFAULT_R 4096
3133 #define CGEMM_DEFAULT_R 4096
3134 #define ZGEMM_DEFAULT_R 4096
3144 #define GEMM_DEFAULT_OFFSET_A 0
3145 #define GEMM_DEFAULT_OFFSET_B 0
3146 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3148 #define SGEMM_DEFAULT_UNROLL_M 2
3149 #define SGEMM_DEFAULT_UNROLL_N 2
3151 #define DGEMM_DEFAULT_UNROLL_M 2
3152 #define DGEMM_DEFAULT_UNROLL_N 2
3154 #define CGEMM_DEFAULT_UNROLL_M 2
3155 #define CGEMM_DEFAULT_UNROLL_N 2
3157 #define ZGEMM_DEFAULT_UNROLL_M 2
3158 #define ZGEMM_DEFAULT_UNROLL_N 2
3160 #define SGEMM_DEFAULT_P 128
3161 #define DGEMM_DEFAULT_P 128
3162 #define CGEMM_DEFAULT_P 96
3163 #define ZGEMM_DEFAULT_P 64
3165 #define SGEMM_DEFAULT_Q 240
3166 #define DGEMM_DEFAULT_Q 120
3167 #define CGEMM_DEFAULT_Q 120
3168 #define ZGEMM_DEFAULT_Q 120
3170 #define SGEMM_DEFAULT_R 12288
3171 #define DGEMM_DEFAULT_R 8192
3172 #define CGEMM_DEFAULT_R 4096
3173 #define ZGEMM_DEFAULT_R 4096
3185 #define GEMM_DEFAULT_OFFSET_A 0
3186 #define GEMM_DEFAULT_OFFSET_B 0
3187 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3189 #define SGEMM_DEFAULT_UNROLL_M 4
3190 #define SGEMM_DEFAULT_UNROLL_N 4
3192 #define DGEMM_DEFAULT_UNROLL_M 4
3193 #define DGEMM_DEFAULT_UNROLL_N 4
3195 #define CGEMM_DEFAULT_UNROLL_M 2
3196 #define CGEMM_DEFAULT_UNROLL_N 2
3198 #define ZGEMM_DEFAULT_UNROLL_M 2
3199 #define ZGEMM_DEFAULT_UNROLL_N 2
3201 #define SGEMM_DEFAULT_P 128
3202 #define DGEMM_DEFAULT_P 128
3203 #define CGEMM_DEFAULT_P 96
3204 #define ZGEMM_DEFAULT_P 64
3206 #define SGEMM_DEFAULT_Q 240
3207 #define DGEMM_DEFAULT_Q 120
3208 #define CGEMM_DEFAULT_Q 120
3209 #define ZGEMM_DEFAULT_Q 120
3211 #define SGEMM_DEFAULT_R 12288
3212 #define DGEMM_DEFAULT_R 8192
3213 #define CGEMM_DEFAULT_R 4096
3214 #define ZGEMM_DEFAULT_R 4096
3226 #define GEMM_DEFAULT_OFFSET_A 0
3227 #define GEMM_DEFAULT_OFFSET_B 0
3228 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3230 #define SGEMM_DEFAULT_UNROLL_M 4
3231 #define SGEMM_DEFAULT_UNROLL_N 4
3233 #define DGEMM_DEFAULT_UNROLL_M 4
3234 #define DGEMM_DEFAULT_UNROLL_N 4
3236 #define CGEMM_DEFAULT_UNROLL_M 2
3237 #define CGEMM_DEFAULT_UNROLL_N 2
3239 #define ZGEMM_DEFAULT_UNROLL_M 2
3240 #define ZGEMM_DEFAULT_UNROLL_N 2
3242 #define SGEMM_DEFAULT_P 128
3243 #define DGEMM_DEFAULT_P 128
3244 #define CGEMM_DEFAULT_P 96
3245 #define ZGEMM_DEFAULT_P 64
3247 #define SGEMM_DEFAULT_Q 240
3248 #define DGEMM_DEFAULT_Q 120
3249 #define CGEMM_DEFAULT_Q 120
3250 #define ZGEMM_DEFAULT_Q 120
3252 #define SGEMM_DEFAULT_R 12288
3253 #define DGEMM_DEFAULT_R 8192
3254 #define CGEMM_DEFAULT_R 4096
3255 #define ZGEMM_DEFAULT_R 4096
3263 #if defined(ZARCH_GENERIC)
3267 #define GEMM_DEFAULT_OFFSET_A 0
3268 #define GEMM_DEFAULT_OFFSET_B 0
3269 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3271 #define SGEMM_DEFAULT_UNROLL_M 2
3272 #define SGEMM_DEFAULT_UNROLL_N 2
3274 #define DGEMM_DEFAULT_UNROLL_M 2
3275 #define DGEMM_DEFAULT_UNROLL_N 2
3277 #define CGEMM_DEFAULT_UNROLL_M 2
3278 #define CGEMM_DEFAULT_UNROLL_N 2
3280 #define ZGEMM_DEFAULT_UNROLL_M 2
3281 #define ZGEMM_DEFAULT_UNROLL_N 2
3283 #define SGEMM_DEFAULT_P 128
3284 #define DGEMM_DEFAULT_P 128
3285 #define CGEMM_DEFAULT_P 96
3286 #define ZGEMM_DEFAULT_P 64
3288 #define SGEMM_DEFAULT_Q 240
3289 #define DGEMM_DEFAULT_Q 120
3290 #define CGEMM_DEFAULT_Q 120
3291 #define ZGEMM_DEFAULT_Q 120
3293 #define SGEMM_DEFAULT_R 12288
3294 #define DGEMM_DEFAULT_R 8192
3295 #define CGEMM_DEFAULT_R 4096
3296 #define ZGEMM_DEFAULT_R 4096
3306 #define GEMM_DEFAULT_OFFSET_A 0
3307 #define GEMM_DEFAULT_OFFSET_B 0
3308 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL
3310 #define SGEMM_DEFAULT_UNROLL_M 8
3311 #define SGEMM_DEFAULT_UNROLL_N 4
3313 #define DGEMM_DEFAULT_UNROLL_M 8
3314 #define DGEMM_DEFAULT_UNROLL_N 4
3316 #define CGEMM_DEFAULT_UNROLL_M 4
3317 #define CGEMM_DEFAULT_UNROLL_N 4
3319 #define ZGEMM_DEFAULT_UNROLL_M 4
3320 #define ZGEMM_DEFAULT_UNROLL_N 4
3322 #define SGEMM_DEFAULT_P 456
3323 #define DGEMM_DEFAULT_P 320
3324 #define CGEMM_DEFAULT_P 480
3325 #define ZGEMM_DEFAULT_P 224
3327 #define SGEMM_DEFAULT_Q 488
3328 #define DGEMM_DEFAULT_Q 384
3329 #define CGEMM_DEFAULT_Q 128
3330 #define ZGEMM_DEFAULT_Q 352
3332 #define SGEMM_DEFAULT_R 8192
3333 #define DGEMM_DEFAULT_R 4096
3334 #define CGEMM_DEFAULT_R 4096
3335 #define ZGEMM_DEFAULT_R 2048
3346 #define GEMM_DEFAULT_OFFSET_A 0
3347 #define GEMM_DEFAULT_OFFSET_B 0
3348 #define GEMM_DEFAULT_ALIGN 0x03fffUL
3350 #define SGEMM_DEFAULT_UNROLL_M 16
3351 #define SGEMM_DEFAULT_UNROLL_N 4
3353 #define DGEMM_DEFAULT_UNROLL_M 8
3354 #define DGEMM_DEFAULT_UNROLL_N 4
3356 #define CGEMM_DEFAULT_UNROLL_M 4
3357 #define CGEMM_DEFAULT_UNROLL_N 4
3359 #define ZGEMM_DEFAULT_UNROLL_M 4
3360 #define ZGEMM_DEFAULT_UNROLL_N 4
3362 #define SGEMM_DEFAULT_P 480
3363 #define DGEMM_DEFAULT_P 320
3364 #define CGEMM_DEFAULT_P 480
3365 #define ZGEMM_DEFAULT_P 224
3367 #define SGEMM_DEFAULT_Q 512
3368 #define DGEMM_DEFAULT_Q 384
3369 #define CGEMM_DEFAULT_Q 128
3370 #define ZGEMM_DEFAULT_Q 352
3372 #define SGEMM_DEFAULT_R 8192
3373 #define DGEMM_DEFAULT_R 4096
3374 #define CGEMM_DEFAULT_R 4096
3375 #define ZGEMM_DEFAULT_R 2048
3388 #define GEMM_DEFAULT_OFFSET_A 0
3389 #define GEMM_DEFAULT_OFFSET_B 0
3390 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x0ffffUL
3392 #define SGEMM_DEFAULT_UNROLL_N 2
3393 #define DGEMM_DEFAULT_UNROLL_N 2
3394 #define QGEMM_DEFAULT_UNROLL_N 2
3395 #define CGEMM_DEFAULT_UNROLL_N 2
3396 #define ZGEMM_DEFAULT_UNROLL_N 2
3397 #define XGEMM_DEFAULT_UNROLL_N 1
3400 #define SGEMM_DEFAULT_UNROLL_M 2
3401 #define DGEMM_DEFAULT_UNROLL_M 2
3402 #define QGEMM_DEFAULT_UNROLL_M 2
3403 #define CGEMM_DEFAULT_UNROLL_M 2
3404 #define ZGEMM_DEFAULT_UNROLL_M 2
3405 #define XGEMM_DEFAULT_UNROLL_M 1
3407 #define SGEMM_DEFAULT_UNROLL_M 2
3408 #define DGEMM_DEFAULT_UNROLL_M 2
3409 #define QGEMM_DEFAULT_UNROLL_M 2
3410 #define CGEMM_DEFAULT_UNROLL_M 2
3411 #define ZGEMM_DEFAULT_UNROLL_M 2
3412 #define XGEMM_DEFAULT_UNROLL_M 1
3415 #define SGEMM_DEFAULT_P sgemm_p
3416 #define DGEMM_DEFAULT_P dgemm_p
3417 #define QGEMM_DEFAULT_P qgemm_p
3418 #define CGEMM_DEFAULT_P cgemm_p
3419 #define ZGEMM_DEFAULT_P zgemm_p
3420 #define XGEMM_DEFAULT_P xgemm_p
3422 #define SGEMM_DEFAULT_R sgemm_r
3423 #define DGEMM_DEFAULT_R dgemm_r
3424 #define QGEMM_DEFAULT_R qgemm_r
3425 #define CGEMM_DEFAULT_R cgemm_r
3426 #define ZGEMM_DEFAULT_R zgemm_r
3427 #define XGEMM_DEFAULT_R xgemm_r
3429 #define SGEMM_DEFAULT_Q 128
3430 #define DGEMM_DEFAULT_Q 128
3431 #define QGEMM_DEFAULT_Q 128
3432 #define CGEMM_DEFAULT_Q 128
3433 #define ZGEMM_DEFAULT_Q 128
3434 #define XGEMM_DEFAULT_Q 128
3440 #ifndef QGEMM_DEFAULT_UNROLL_M
3441 #define QGEMM_DEFAULT_UNROLL_M 2
3444 #ifndef QGEMM_DEFAULT_UNROLL_N
3445 #define QGEMM_DEFAULT_UNROLL_N 2
3448 #ifndef XGEMM_DEFAULT_UNROLL_M
3449 #define XGEMM_DEFAULT_UNROLL_M 2
3452 #ifndef XGEMM_DEFAULT_UNROLL_N
3453 #define XGEMM_DEFAULT_UNROLL_N 2
3457 #define SHUFPD_0 shufps $0x44,
3458 #define SHUFPD_1 shufps $0x4e,
3459 #define SHUFPD_2 shufps $0xe4,
3460 #define SHUFPD_3 shufps $0xee,
3464 #define SHUFPD_0 shufpd $0,
3468 #define SHUFPD_1 shufpd $1,
3472 #define SHUFPD_2 shufpd $2,
3476 #define SHUFPD_3 shufpd $3,
3480 #define SHUFPS_39 shufps $0x39,