1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
|
#include <ulmblas.h>
#include <stdio.h>
#include <emmintrin.h>
#include <immintrin.h>
#define MC 384
#define KC 384
#define NC 4096
#define MR 4
#define NR 4
//
// Local buffers for storing panels from A, B and C
//
static double _A[MC*KC] __attribute__ ((aligned (16)));
static double _B[KC*NC] __attribute__ ((aligned (16)));
static double _C[MR*NR] __attribute__ ((aligned (16)));
//
// Packing complete panels from A (i.e. without padding)
//
static void
pack_MRxk(int k, const double *A, int incRowA, int incColA,
double *buffer)
{
int i, j;
for (j=0; j<k; ++j) {
for (i=0; i<MR; ++i) {
buffer[i] = A[i*incRowA];
}
buffer += MR;
A += incColA;
}
}
//
// Packing panels from A with padding if required
//
static void
pack_A(int mc, int kc, const double *A, int incRowA, int incColA,
double *buffer)
{
int mp = mc / MR;
int _mr = mc % MR;
int i, j;
for (i=0; i<mp; ++i) {
pack_MRxk(kc, A, incRowA, incColA, buffer);
buffer += kc*MR;
A += MR*incRowA;
}
if (_mr>0) {
for (j=0; j<kc; ++j) {
for (i=0; i<_mr; ++i) {
buffer[i] = A[i*incRowA];
}
for (i=_mr; i<MR; ++i) {
buffer[i] = 0.0;
}
buffer += MR;
A += incColA;
}
}
}
//
// Packing complete panels from B (i.e. without padding)
//
static void
pack_kxNR(int k, const double *B, int incRowB, int incColB,
double *buffer)
{
int i, j;
for (i=0; i<k; ++i) {
for (j=0; j<NR; ++j) {
buffer[j] = B[j*incColB];
}
buffer += NR;
B += incRowB;
}
}
//
// Packing panels from B with padding if required
//
static void
pack_B(int kc, int nc, const double *B, int incRowB, int incColB,
double *buffer)
{
int np = nc / NR;
int _nr = nc % NR;
int i, j;
for (j=0; j<np; ++j) {
pack_kxNR(kc, B, incRowB, incColB, buffer);
buffer += kc*NR;
B += NR*incColB;
}
if (_nr>0) {
for (i=0; i<kc; ++i) {
for (j=0; j<_nr; ++j) {
buffer[j] = B[j*incColB];
}
for (j=_nr; j<NR; ++j) {
buffer[j] = 0.0;
}
buffer += NR;
B += incRowB;
}
}
}
//
// Micro kernel for multiplying panels from A and B.
//
static void
dgemm_micro_kernel(long kc,
double alpha, const double *A, const double *B,
double beta,
double *C, long incRowC, long incColC)
{
long kb = kc / 4;
long kl = kc % 4;
//
// Compute AB = A*B
//
__asm__ volatile
(
"movq %0, %%rsi \n\t" // kb (32 bit) stored in %rsi
"movq %1, %%rdi \n\t" // kl (32 bit) stored in %rdi
"movq %2, %%rax \n\t" // Address of A stored in %rax
"movq %3, %%rbx \n\t" // Address of B stored in %rbx
" \n\t"
"movaps (%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A)
"movaps 16(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+2)
"movaps (%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B)
" \n\t"
"xorpd %%xmm8, %%xmm8 \n\t" // ab_00_11 = _mm_setzero_pd()
"xorpd %%xmm9, %%xmm9 \n\t" // ab_20_31 = _mm_setzero_pd()
"xorpd %%xmm10, %%xmm10 \n\t" // ab_01_10 = _mm_setzero_pd()
"xorpd %%xmm11, %%xmm11 \n\t" // ab_21_30 = _mm_setzero_pd()
"xorpd %%xmm12, %%xmm12 \n\t" // ab_02_13 = _mm_setzero_pd()
"xorpd %%xmm13, %%xmm13 \n\t" // ab_22_33 = _mm_setzero_pd()
"xorpd %%xmm14, %%xmm14 \n\t" // ab_03_12 = _mm_setzero_pd()
"xorpd %%xmm15, %%xmm15 \n\t" // ab_23_32 = _mm_setzero_pd()
" \n\t"
"xorpd %%xmm3, %%xmm3 \n\t" // tmp3 = _mm_setzero_pd
"xorpd %%xmm4, %%xmm4 \n\t" // tmp4 = _mm_setzero_pd
"xorpd %%xmm5, %%xmm5 \n\t" // tmp5 = _mm_setzero_pd
"xorpd %%xmm6, %%xmm6 \n\t" // tmp6 = _mm_setzero_pd
"xorpd %%xmm7, %%xmm7 \n\t" // tmp7 = _mm_setzero_pd
"testq %%rdi, %%rdi \n\t" // if kl==0 writeback to AB
" \n\t"
" \n\t"
"testq %%rsi, %%rsi \n\t" // if kb==0 handle remaining kl
"je .DCONSIDERLEFT%= \n\t" // update iterations
" \n\t"
".DLOOP%=: \n\t" // for l = kb,..,1 do
" \n\t"
" \n\t" // 1. update
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"movaps 16(%%rbx), %%xmm3 \n\t" // tmp3 = _mm_load_pd(B+2)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
"movaps %%xmm2, %%xmm6 \n\t" // tmp6 = tmp2
"pshufd $78,%%xmm2, %%xmm4 \n\t" // tmp4 = _mm_shuffle_pd(tmp2, tmp2,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm2 \n\t" // tmp2 = _mm_mul_pd(tmp2, tmp0);
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1);
" \n\t"
" \n\t"
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
"movaps %%xmm4, %%xmm7 \n\t" // tmp7 = tmp4
"mulpd %%xmm0, %%xmm4 \n\t" // tmp4 = _mm_mul_pd(tmp4, tmp0)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm2, %%xmm8 \n\t" // ab_00_11 = _mm_add_pd(ab_00_11, tmp2)
"movaps 32(%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B+4)
"addpd %%xmm6, %%xmm9 \n\t" // ab_20_31 = _mm_add_pd(ab_20_31, tmp6)
"movaps %%xmm3, %%xmm6 \n\t" // tmp6 = tmp3
"pshufd $78,%%xmm3, %%xmm5 \n\t" // tmp5 = _mm_shuffle_pd(tmp3, tmp3,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm3 \n\t" // tmp3 = _mm_mul_pd(tmp3, tmp0)
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm4, %%xmm10 \n\t" // ab_01_10 = _mm_add_pd(ab_01_10, tmp4)
"addpd %%xmm7, %%xmm11 \n\t" // ab_21_30 = _mm_add_pd(ab_21_30, tmp7)
"movaps %%xmm5, %%xmm7 \n\t" // tmp7 = tmp5
"mulpd %%xmm0, %%xmm5 \n\t" // tmp5 = _mm_mul_pd(tmp5, tmp0)
"movaps 32(%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A+4)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
"movaps 48(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+6)
" \n\t"
" \n\t"
" \n\t"
" \n\t" // 2. update
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"movaps 48(%%rbx), %%xmm3 \n\t" // tmp3 = _mm_load_pd(B+6)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
"movaps %%xmm2, %%xmm6 \n\t" // tmp6 = tmp2
"pshufd $78,%%xmm2, %%xmm4 \n\t" // tmp4 = _mm_shuffle_pd(tmp2, tmp2,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm2 \n\t" // tmp2 = _mm_mul_pd(tmp2, tmp0);
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1);
" \n\t"
" \n\t"
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
"movaps %%xmm4, %%xmm7 \n\t" // tmp7 = tmp4
"mulpd %%xmm0, %%xmm4 \n\t" // tmp4 = _mm_mul_pd(tmp4, tmp0)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm2, %%xmm8 \n\t" // ab_00_11 = _mm_add_pd(ab_00_11, tmp2)
"movaps 64(%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B+8)
"addpd %%xmm6, %%xmm9 \n\t" // ab_20_31 = _mm_add_pd(ab_20_31, tmp6)
"movaps %%xmm3, %%xmm6 \n\t" // tmp6 = tmp3
"pshufd $78,%%xmm3, %%xmm5 \n\t" // tmp5 = _mm_shuffle_pd(tmp3, tmp3,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm3 \n\t" // tmp3 = _mm_mul_pd(tmp3, tmp0)
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm4, %%xmm10 \n\t" // ab_01_10 = _mm_add_pd(ab_01_10, tmp4)
"addpd %%xmm7, %%xmm11 \n\t" // ab_21_30 = _mm_add_pd(ab_21_30, tmp7)
"movaps %%xmm5, %%xmm7 \n\t" // tmp7 = tmp5
"mulpd %%xmm0, %%xmm5 \n\t" // tmp5 = _mm_mul_pd(tmp5, tmp0)
"movaps 64(%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A+8)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
"movaps 80(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+10)
" \n\t"
" \n\t"
" \n\t" // 3. update
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"movaps 80(%%rbx), %%xmm3 \n\t" // tmp3 = _mm_load_pd(B+10)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
"movaps %%xmm2, %%xmm6 \n\t" // tmp6 = tmp2
"pshufd $78,%%xmm2, %%xmm4 \n\t" // tmp4 = _mm_shuffle_pd(tmp2, tmp2,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm2 \n\t" // tmp2 = _mm_mul_pd(tmp2, tmp0);
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1);
" \n\t"
" \n\t"
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
"movaps %%xmm4, %%xmm7 \n\t" // tmp7 = tmp4
"mulpd %%xmm0, %%xmm4 \n\t" // tmp4 = _mm_mul_pd(tmp4, tmp0)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm2, %%xmm8 \n\t" // ab_00_11 = _mm_add_pd(ab_00_11, tmp2)
"movaps 96(%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B+12)
"addpd %%xmm6, %%xmm9 \n\t" // ab_20_31 = _mm_add_pd(ab_20_31, tmp6)
"movaps %%xmm3, %%xmm6 \n\t" // tmp6 = tmp3
"pshufd $78,%%xmm3, %%xmm5 \n\t" // tmp5 = _mm_shuffle_pd(tmp3, tmp3,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm3 \n\t" // tmp3 = _mm_mul_pd(tmp3, tmp0)
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm4, %%xmm10 \n\t" // ab_01_10 = _mm_add_pd(ab_01_10, tmp4)
"addpd %%xmm7, %%xmm11 \n\t" // ab_21_30 = _mm_add_pd(ab_21_30, tmp7)
"movaps %%xmm5, %%xmm7 \n\t" // tmp7 = tmp5
"mulpd %%xmm0, %%xmm5 \n\t" // tmp5 = _mm_mul_pd(tmp5, tmp0)
"movaps 96(%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A+12)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
"movaps 112(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+14)
" \n\t"
" \n\t"
" \n\t" // 4. update
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"movaps 112(%%rbx), %%xmm3 \n\t" // tmp3 = _mm_load_pd(B+14)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
"movaps %%xmm2, %%xmm6 \n\t" // tmp6 = tmp2
"pshufd $78,%%xmm2, %%xmm4 \n\t" // tmp4 = _mm_shuffle_pd(tmp2, tmp2,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm2 \n\t" // tmp2 = _mm_mul_pd(tmp2, tmp0);
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1);
" \n\t"
"addq $32*4, %%rax \n\t" // A += 16;
" \n\t"
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
"movaps %%xmm4, %%xmm7 \n\t" // tmp7 = tmp4
"mulpd %%xmm0, %%xmm4 \n\t" // tmp4 = _mm_mul_pd(tmp4, tmp0)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm2, %%xmm8 \n\t" // ab_00_11 = _mm_add_pd(ab_00_11, tmp2)
"movaps 128(%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B+16)
"addpd %%xmm6, %%xmm9 \n\t" // ab_20_31 = _mm_add_pd(ab_20_31, tmp6)
"movaps %%xmm3, %%xmm6 \n\t" // tmp6 = tmp3
"pshufd $78,%%xmm3, %%xmm5 \n\t" // tmp5 = _mm_shuffle_pd(tmp3, tmp3,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm3 \n\t" // tmp3 = _mm_mul_pd(tmp3, tmp0)
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1)
" \n\t"
"addq $32*4, %%rbx \n\t" // B += 16;
" \n\t"
" \n\t"
"addpd %%xmm4, %%xmm10 \n\t" // ab_01_10 = _mm_add_pd(ab_01_10, tmp4)
"addpd %%xmm7, %%xmm11 \n\t" // ab_21_30 = _mm_add_pd(ab_21_30, tmp7)
"movaps %%xmm5, %%xmm7 \n\t" // tmp7 = tmp5
"mulpd %%xmm0, %%xmm5 \n\t" // tmp5 = _mm_mul_pd(tmp5, tmp0)
"movaps (%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A+16)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
"movaps 16(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+18)
" \n\t"
" \n\t"
"decq %%rsi \n\t" // --l
"jne .DLOOP%= \n\t" // if l>= 1 go back
" \n\t"
" \n\t"
".DCONSIDERLEFT%=: \n\t"
"testq %%rdi, %%rdi \n\t" // if kl==0 writeback to AB
"je .DPOSTACCUMULATE%=\n\t"
" \n\t"
".DLOOPLEFT%=: \n\t" // for l = kl,..,1 do
" \n\t"
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"movapd 16(%%rbx), %%xmm3 \n\t" // tmp3 = _mm_load_pd(B+2)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
"movapd %%xmm2, %%xmm6 \n\t" // tmp6 = tmp2
"pshufd $78,%%xmm2, %%xmm4 \n\t" // tmp4 = _mm_shuffle_pd(tmp2, tmp2,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm2 \n\t" // tmp2 = _mm_mul_pd(tmp2, tmp0);
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1);
" \n\t"
" \n\t"
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
"movapd %%xmm4, %%xmm7 \n\t" // tmp7 = tmp4
"mulpd %%xmm0, %%xmm4 \n\t" // tmp4 = _mm_mul_pd(tmp4, tmp0)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm2, %%xmm8 \n\t" // ab_00_11 = _mm_add_pd(ab_00_11, tmp2)
"movapd 32(%%rbx), %%xmm2 \n\t" // tmp2 = _mm_load_pd(B+4)
"addpd %%xmm6, %%xmm9 \n\t" // ab_20_31 = _mm_add_pd(ab_20_31, tmp6)
"movapd %%xmm3, %%xmm6 \n\t" // tmp6 = tmp3
"pshufd $78,%%xmm3, %%xmm5 \n\t" // tmp5 = _mm_shuffle_pd(tmp3, tmp3,
" \n\t" // _MM_SHUFFLE2(0, 1))
"mulpd %%xmm0, %%xmm3 \n\t" // tmp3 = _mm_mul_pd(tmp3, tmp0)
"mulpd %%xmm1, %%xmm6 \n\t" // tmp6 = _mm_mul_pd(tmp6, tmp1)
" \n\t"
" \n\t"
"addpd %%xmm4, %%xmm10 \n\t" // ab_01_10 = _mm_add_pd(ab_01_10, tmp4)
"addpd %%xmm7, %%xmm11 \n\t" // ab_21_30 = _mm_add_pd(ab_21_30, tmp7)
"movapd %%xmm5, %%xmm7 \n\t" // tmp7 = tmp5
"mulpd %%xmm0, %%xmm5 \n\t" // tmp5 = _mm_mul_pd(tmp5, tmp0)
"movapd 32(%%rax), %%xmm0 \n\t" // tmp0 = _mm_load_pd(A+4)
"mulpd %%xmm1, %%xmm7 \n\t" // tmp7 = _mm_mul_pd(tmp7, tmp1)
"movapd 48(%%rax), %%xmm1 \n\t" // tmp1 = _mm_load_pd(A+6)
" \n\t"
" \n\t"
"addq $32, %%rax \n\t" // A += 4;
"addq $32, %%rbx \n\t" // B += 4;
" \n\t"
"decq %%rdi \n\t" // --l
"jne .DLOOPLEFT%= \n\t" // if l>= 1 go back
" \n\t"
".DPOSTACCUMULATE%=: \n\t" // Update remaining ab_*_* registers
" \n\t"
"addpd %%xmm3, %%xmm12 \n\t" // ab_02_13 = _mm_add_pd(ab_02_13, tmp3)
"addpd %%xmm6, %%xmm13 \n\t" // ab_22_33 = _mm_add_pd(ab_22_33, tmp6)
" \n\t" //
"addpd %%xmm5, %%xmm14 \n\t" // ab_03_12 = _mm_add_pd(ab_03_12, tmp5)
"addpd %%xmm7, %%xmm15 \n\t" // ab_23_32 = _mm_add_pd(ab_23_32, tmp7)
" \n\t"
//
// Update C <- beta*C + alpha*AB
//
//
"movsd %4, %%xmm0 \n\t" // load alpha
"movsd %5, %%xmm1 \n\t" // load beta
"movq %6, %%rcx \n\t" // Address of C stored in %rcx
"movq %7, %%r8 \n\t" // load incRowC
"leaq (,%%r8,8), %%r8 \n\t" // incRowC *= sizeof(double)
"movq %8, %%r9 \n\t" // load incColC
"leaq (,%%r9,8), %%r9 \n\t" // incRowC *= sizeof(double)
" \n\t"
"leaq (%%rcx,%%r9), %%r10 \n\t" // Store addr of C01 in %r10
"leaq (%%rcx,%%r8,2), %%rdx \n\t" // Store addr of C20 in %rdx
"leaq (%%rdx,%%r9), %%r11 \n\t" // Store addr of C21 in %r11
" \n\t"
"unpcklpd %%xmm0, %%xmm0 \n\t" // duplicate alpha
"unpcklpd %%xmm1, %%xmm1 \n\t" // duplicate beta
" \n\t"
" \n\t"
"movlpd (%%rcx), %%xmm3 \n\t" // load (C00,
"movhpd (%%r10,%%r8), %%xmm3 \n\t" // C11)
"mulpd %%xmm0, %%xmm8 \n\t" // scale ab_00_11 by alpha
"mulpd %%xmm1, %%xmm3 \n\t" // scale (C00, C11) by beta
"addpd %%xmm8, %%xmm3 \n\t" // add results
"movlpd %%xmm3, (%%rcx) \n\t" // write back (C00,
"movhpd %%xmm3, (%%r10,%%r8) \n\t" // C11)
" \n\t"
"movlpd (%%rdx), %%xmm4 \n\t" // load (C20,
"movhpd (%%r11,%%r8), %%xmm4 \n\t" // C31)
"mulpd %%xmm0, %%xmm9 \n\t" // scale ab_20_31 by alpha
"mulpd %%xmm1, %%xmm4 \n\t" // scale (C20, C31) by beta
"addpd %%xmm9, %%xmm4 \n\t" // add results
"movlpd %%xmm4, (%%rdx) \n\t" // write back (C20,
"movhpd %%xmm4, (%%r11,%%r8) \n\t" // C31)
" \n\t"
" \n\t"
"movlpd (%%r10), %%xmm3 \n\t" // load (C01,
"movhpd (%%rcx,%%r8), %%xmm3 \n\t" // C10)
"mulpd %%xmm0, %%xmm10\n\t" // scale ab_01_10 by alpha
"mulpd %%xmm1, %%xmm3 \n\t" // scale (C01, C10) by beta
"addpd %%xmm10, %%xmm3 \n\t" // add results
"movlpd %%xmm3, (%%r10) \n\t" // write back (C01,
"movhpd %%xmm3, (%%rcx,%%r8) \n\t" // C10)
" \n\t"
"movlpd (%%r11), %%xmm4 \n\t" // load (C21,
"movhpd (%%rdx,%%r8), %%xmm4 \n\t" // C30)
"mulpd %%xmm0, %%xmm11\n\t" // scale ab_21_30 by alpha
"mulpd %%xmm1, %%xmm4 \n\t" // scale (C21, C30) by beta
"addpd %%xmm11, %%xmm4 \n\t" // add results
"movlpd %%xmm4, (%%r11) \n\t" // write back (C21,
"movhpd %%xmm4, (%%rdx,%%r8) \n\t" // C30)
" \n\t"
" \n\t"
"leaq (%%rcx,%%r9,2), %%rcx \n\t" // Store addr of C02 in %rcx
"leaq (%%r10,%%r9,2), %%r10 \n\t" // Store addr of C03 in %r10
"leaq (%%rdx,%%r9,2), %%rdx \n\t" // Store addr of C22 in $rdx
"leaq (%%r11,%%r9,2), %%r11 \n\t" // Store addr of C23 in %r11
" \n\t"
" \n\t"
"movlpd (%%rcx), %%xmm3 \n\t" // load (C02,
"movhpd (%%r10,%%r8), %%xmm3 \n\t" // C13)
"mulpd %%xmm0, %%xmm12\n\t" // scale ab_02_13 by alpha
"mulpd %%xmm1, %%xmm3 \n\t" // scale (C02, C13) by beta
"addpd %%xmm12, %%xmm3 \n\t" // add results
"movlpd %%xmm3, (%%rcx) \n\t" // write back (C02,
"movhpd %%xmm3, (%%r10,%%r8) \n\t" // C13)
" \n\t"
"movlpd (%%rdx), %%xmm4 \n\t" // load (C22,
"movhpd (%%r11, %%r8), %%xmm4 \n\t" // C33)
"mulpd %%xmm0, %%xmm13\n\t" // scale ab_22_33 by alpha
"mulpd %%xmm1, %%xmm4 \n\t" // scale (C22, C33) by beta
"addpd %%xmm13, %%xmm4 \n\t" // add results
"movlpd %%xmm4, (%%rdx) \n\t" // write back (C22,
"movhpd %%xmm4, (%%r11,%%r8) \n\t" // C33)
" \n\t"
" \n\t"
"movlpd (%%r10), %%xmm3 \n\t" // load (C03,
"movhpd (%%rcx,%%r8), %%xmm3 \n\t" // C12)
"mulpd %%xmm0, %%xmm14\n\t" // scale ab_03_12 by alpha
"mulpd %%xmm1, %%xmm3 \n\t" // scale (C03, C12) by beta
"addpd %%xmm14, %%xmm3 \n\t" // add results
"movlpd %%xmm3, (%%r10) \n\t" // write back (C03,
"movhpd %%xmm3, (%%rcx,%%r8) \n\t" // C12)
" \n\t"
"movlpd (%%r11), %%xmm4 \n\t" // load (C23,
"movhpd (%%rdx,%%r8), %%xmm4 \n\t" // C32)
"mulpd %%xmm0, %%xmm15\n\t" // scale ab_23_32 by alpha
"mulpd %%xmm1, %%xmm4 \n\t" // scale (C23, C32) by beta
"addpd %%xmm15, %%xmm4 \n\t" // add results
"movlpd %%xmm4, (%%r11) \n\t" // write back (C23,
"movhpd %%xmm4, (%%rdx,%%r8) \n\t" // C32)
: // output
: // input
"m" (kb), // 0
"m" (kl), // 1
"m" (A), // 2
"m" (B), // 3
"m" (alpha), // 4
"m" (beta), // 5
"m" (C), // 6
"m" (incRowC), // 7
"m" (incColC) // 8
: // register clobber list
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
"xmm0", "xmm1", "xmm2", "xmm3",
"xmm4", "xmm5", "xmm6", "xmm7",
"xmm8", "xmm9", "xmm10", "xmm11",
"xmm12", "xmm13", "xmm14", "xmm15"
);
}
//
// Compute Y += alpha*X
//
static void
dgeaxpy(int m,
int n,
double alpha,
const double *X,
int incRowX,
int incColX,
double *Y,
int incRowY,
int incColY)
{
int i, j;
if (alpha!=1.0) {
for (j=0; j<n; ++j) {
for (i=0; i<m; ++i) {
Y[i*incRowY+j*incColY] += alpha*X[i*incRowX+j*incColX];
}
}
} else {
for (j=0; j<n; ++j) {
for (i=0; i<m; ++i) {
Y[i*incRowY+j*incColY] += X[i*incRowX+j*incColX];
}
}
}
}
//
// Compute X *= alpha
//
static void
dgescal(int m,
int n,
double alpha,
double *X,
int incRowX,
int incColX)
{
int i, j;
if (alpha!=0.0) {
for (j=0; j<n; ++j) {
for (i=0; i<m; ++i) {
X[i*incRowX+j*incColX] *= alpha;
}
}
} else {
for (j=0; j<n; ++j) {
for (i=0; i<m; ++i) {
X[i*incRowX+j*incColX] = 0.0;
}
}
}
}
//
// Macro Kernel for the multiplication of blocks of A and B. We assume that
// these blocks were previously packed to buffers _A and _B.
//
static void
dgemm_macro_kernel(int mc,
int nc,
int kc,
double alpha,
double beta,
double *C,
int incRowC,
int incColC)
{
int mp = (mc+MR-1) / MR;
int np = (nc+NR-1) / NR;
int _mr = mc % MR;
int _nr = nc % NR;
int mr, nr;
int i, j;
for (j=0; j<np; ++j) {
nr = (j!=np-1 || _nr==0) ? NR : _nr;
for (i=0; i<mp; ++i) {
mr = (i!=mp-1 || _mr==0) ? MR : _mr;
if (mr==MR && nr==NR) {
dgemm_micro_kernel(kc, alpha, &_A[i*kc*MR], &_B[j*kc*NR],
beta,
&C[i*MR*incRowC+j*NR*incColC],
incRowC, incColC);
} else {
dgemm_micro_kernel(kc, alpha, &_A[i*kc*MR], &_B[j*kc*NR],
0.0,
_C, 1, MR);
dgescal(mr, nr, beta,
&C[i*MR*incRowC+j*NR*incColC], incRowC, incColC);
dgeaxpy(mr, nr, 1.0, _C, 1, MR,
&C[i*MR*incRowC+j*NR*incColC], incRowC, incColC);
}
}
}
}
//
// Compute C <- beta*C + alpha*A*B
//
void
ULMBLAS(dgemm_nn)(int m,
int n,
int k,
double alpha,
const double *A,
int incRowA,
int incColA,
const double *B,
int incRowB,
int incColB,
double beta,
double *C,
int incRowC,
int incColC)
{
int mb = (m+MC-1) / MC;
int nb = (n+NC-1) / NC;
int kb = (k+KC-1) / KC;
int _mc = m % MC;
int _nc = n % NC;
int _kc = k % KC;
int mc, nc, kc;
int i, j, l;
double _beta;
if (alpha==0.0 || k==0) {
dgescal(m, n, beta, C, incRowC, incColC);
return;
}
for (j=0; j<nb; ++j) {
nc = (j!=nb-1 || _nc==0) ? NC : _nc;
for (l=0; l<kb; ++l) {
kc = (l!=kb-1 || _kc==0) ? KC : _kc;
_beta = (l==0) ? beta : 1.0;
pack_B(kc, nc,
&B[l*KC*incRowB+j*NC*incColB], incRowB, incColB,
_B);
for (i=0; i<mb; ++i) {
mc = (i!=mb-1 || _mc==0) ? MC : _mc;
pack_A(mc, kc,
&A[i*MC*incRowA+l*KC*incColA], incRowA, incColA,
_A);
dgemm_macro_kernel(mc, nc, kc, alpha, _beta,
&C[i*MC*incRowC+j*NC*incColC],
incRowC, incColC);
}
}
}
}
|