-
Notifications
You must be signed in to change notification settings - Fork 25
/
cpu_tma_transpose.cpp
974 lines (914 loc) · 37 KB
/
cpu_tma_transpose.cpp
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
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
#include <stdio.h>
#if defined __ANDROID__ || defined __linux__
#include <sched.h>
#if defined __ANDROID__
#include <dlfcn.h>
#endif
#include <stdint.h>
#include <sys/syscall.h>
#include <unistd.h>
#endif
#include <errno.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/time.h>
#include "mperf/cpu_affinity.h"
#include "mperf/timer.h"
#include "mperf/tma/tma.h"
namespace {
template <typename T>
void v_trans_base(const void* src, int src_step, int height, int width,
int channel, void* dst, int dst_step) {
const T* src_ptr = (const T*)src;
T* dst_ptr = (T*)dst;
for (int j = 0; j < width; j++) {
for (int i = 0; i < height; i++) {
for (int k = 0; k < channel; k++) {
dst_ptr[i * dst_step + j * channel + k] =
src_ptr[j * src_step + i * channel + k];
}
}
}
}
/* ======================== Prefetch ======================== */
#if defined(__aarch64__) || defined(__arm__)
#ifdef __aarch64__
static inline void prefetch(const void* ptr) {
asm volatile("PRFM PLDL1KEEP, [%0]" : : "r"(ptr));
}
static inline void prefetch_l2(const void* ptr) {
asm volatile("PRFM PLDL2KEEP, [%0]" : : "r"(ptr));
}
static inline void prefetch_l3(const void* ptr) {
asm volatile("PRFM PLDL3KEEP, [%0]" : : "r"(ptr));
}
static inline void prefetch_non_temporal(const void* ptr) {
asm volatile("PRFM PLDL1STRM, [%0]" : : "r"(ptr));
}
static inline void prefetchw(const void* ptr) {
asm volatile("PRFM PSTL1KEEP, [%0]" : : "r"(ptr));
}
#define ASM_PREFETCH(address) "PRFM PLDL1KEEP, " address "\n"
#else
static inline void prefetch(const void* ptr) {
asm volatile("pld [%0]" ::"r"(ptr));
}
static inline void prefetch_l2(const void* ptr) {
asm volatile("pld [%0]" ::"r"(ptr));
}
static inline void prefetch_l3(const void* ptr) {
asm volatile("pld [%0]" ::"r"(ptr));
}
static inline void prefetch_non_temporal(const void* ptr) {
asm volatile("pld [%0]" ::"r"(ptr));
}
static inline void prefetchw(const void* ptr) {
(void)ptr;
}
#define ASM_PREFETCH(address) "PLD " address "\n"
#endif
static inline void prefetch_6x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]")
ASM_PREFETCH("[%[ptr], #64]")
ASM_PREFETCH("[%[ptr], #128]")
ASM_PREFETCH("[%[ptr], #192]")
ASM_PREFETCH("[%[ptr], #256]")
ASM_PREFETCH("[%[ptr], #320]")
:
: [ptr] "r"(ptr)
: "memory");
// clang-format on
}
static inline void prefetch_5x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]")
ASM_PREFETCH("[%[ptr], #64]")
ASM_PREFETCH("[%[ptr], #128]")
ASM_PREFETCH("[%[ptr], #192]")
ASM_PREFETCH("[%[ptr], #256]")
:
: [ptr] "r"(ptr)
: "memory");
// clang-format on
}
static inline void prefetch_4x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]")
ASM_PREFETCH("[%[ptr], #64]")
ASM_PREFETCH("[%[ptr], #128]")
ASM_PREFETCH("[%[ptr], #192]")
:
: [ptr] "r"(ptr)
: "memory");
// clang-format on
}
static inline void prefetch_3x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]")
ASM_PREFETCH("[%[ptr], #64]")
ASM_PREFETCH("[%[ptr], #128]")
:
: [ptr] "r"(ptr)
: "memory");
// clang-format on
}
static inline void prefetch_2x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]")
ASM_PREFETCH("[%[ptr], #64]")
:
: [ptr] "r"(ptr)
: "memory");
// clang-format on
}
static inline void prefetch_1x(const void* ptr) {
// clang-format off
asm volatile(ASM_PREFETCH("[%[ptr]]") : : [ptr] "r"(ptr) : "memory");
// clang-format on
}
#endif
#if defined(__SSE__)
static inline void prefetch(const void* ptr) {
asm volatile("prefetcht0 %[p]" : : [p] "m"(*(const char*)ptr));
}
static inline void prefetch_l2(const void* ptr) {
asm volatile("prefetcht1 %[p]" : : [p] "m"(*(const volatile char*)ptr));
}
static inline void prefetch_l3(const void* ptr) {
asm volatile("prefetcht2 %[p]" : : [p] "m"(*(const volatile char*)ptr));
}
static inline void prefetch_non_temporal(const void* ptr) {
asm volatile("prefetchnta %[p]" : : [p] "m"(*(const volatile char*)ptr));
}
static inline void prefetchw(const void* ptr) {
asm volatile("PREFETCHW %[p]" : : [p] "m"(*(const char*)ptr));
}
static inline void prefetch_6x(const void* ptr) {
prefetch(ptr);
prefetch((const char*)ptr + 64);
prefetch((const char*)ptr + 64 * 2);
prefetch((const char*)ptr + 64 * 3);
prefetch((const char*)ptr + 64 * 4);
prefetch((const char*)ptr + 64 * 5);
}
static inline void prefetch_5x(const void* ptr) {
prefetch(ptr);
prefetch((const char*)ptr + 64);
prefetch((const char*)ptr + 64 * 2);
prefetch((const char*)ptr + 64 * 3);
prefetch((const char*)ptr + 64 * 4);
}
static inline void prefetch_4x(const void* ptr) {
prefetch(ptr);
prefetch((const char*)ptr + 64);
prefetch((const char*)ptr + 64 * 2);
prefetch((const char*)ptr + 64 * 3);
}
static inline void prefetch_3x(const void* ptr) {
prefetch(ptr);
prefetch((const char*)ptr + 64);
prefetch((const char*)ptr + 64 * 2);
}
static inline void prefetch_2x(const void* ptr) {
prefetch(ptr);
prefetch((const char*)ptr + 64);
}
static inline void prefetch_1x(const void* ptr) {
prefetch(ptr);
}
#endif
#define PRE_L2_STEP 2
// Load & Store 32B
// INT
#if defined __x86_64__ || defined _M_X64
#ifdef __AVX2__
#include <immintrin.h>
template <bool align = false>
inline __m256i v256_load_32b(const uint32_t* ptr);
template <>
inline __m256i v256_load_32b<false>(const uint32_t* ptr) {
return _mm256_loadu_si256((const __m256i*)ptr);
}
template <>
inline __m256i v256_load_32b<true>(const uint32_t* ptr) {
return _mm256_load_si256((const __m256i*)ptr);
}
template <bool align = false>
inline void v256_store_32b(uint32_t* ptr, __m256i val);
template <>
inline void v256_store_32b<false>(uint32_t* ptr, __m256i val) {
_mm256_storeu_si256((__m256i*)ptr, val);
}
template <>
inline void v256_store_32b<true>(uint32_t* ptr, __m256i val) {
_mm256_store_si256((__m256i*)ptr, val);
}
inline void v256_store2_32b(uint32_t* p0, uint32_t* p1, __m256i val) {
_mm_store_si128((__m128i*)p0, _mm256_extractf128_si256(val, 0));
_mm_store_si128((__m128i*)p1, _mm256_extractf128_si256(val, 1));
}
// float
template <bool align = false>
inline __m256 v256_load_32b(const float* ptr);
template <>
inline __m256 v256_load_32b<false>(const float* ptr) {
return _mm256_loadu_ps(ptr);
}
template <>
inline __m256 v256_load_32b<true>(const float* ptr) {
return _mm256_load_ps(ptr);
}
template <bool align = false>
inline void v256_store_32b(float* ptr, __m256 val);
template <>
inline void v256_store_32b<false>(float* ptr, __m256 val) {
_mm256_storeu_ps(ptr, val);
}
template <>
inline void v256_store_32b<true>(float* ptr, __m256 val) {
_mm256_store_ps(ptr, val);
}
template <bool align = false>
inline void v256_store2_32b(float* p0, float* p1, __m256 val);
template <>
inline void v256_store2_32b<false>(float* p0, float* p1, __m256 val) {
_mm_storeu_ps(p0, _mm256_extractf128_ps(val, 0));
_mm_storeu_ps(p1, _mm256_extractf128_ps(val, 1));
}
template <>
inline void v256_store2_32b<true>(float* p0, float* p1, __m256 val) {
_mm_store_ps(p0, _mm256_extractf128_ps(val, 0));
_mm_store_ps(p1, _mm256_extractf128_ps(val, 1));
}
template <bool align>
inline __m256 v256_load3_32b(const float* p0, const float* p1);
template <>
inline __m256 v256_load3_32b<false>(const float* p0, const float* p1) {
return _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_loadu_ps(p0)),
_mm_loadu_ps(p1), 1);
}
template <>
inline __m256 v256_load3_32b<true>(const float* p0, const float* p1) {
return _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(p0)),
_mm_load_ps(p1), 1);
}
#endif
#include <nmmintrin.h>
template <bool align>
inline __m128i v_load_32b(const uint32_t* ptr);
template <>
inline __m128i v_load_32b<false>(const uint32_t* ptr) {
return _mm_loadu_si128((const __m128i*)ptr);
}
template <>
inline __m128i v_load_32b<true>(const uint32_t* ptr) {
return _mm_load_si128((const __m128i*)ptr);
}
template <bool align>
inline void v_store_32b(uint32_t* ptr, __m128i val);
template <>
inline void v_store_32b<false>(uint32_t* ptr, __m128i val) {
#if 1
_mm_storeu_si128((__m128i*)ptr, val);
#else
_mm_stream_si128((__m128i*)ptr, val);
#endif
}
template <>
inline void v_store_32b<true>(uint32_t* ptr, __m128i val) {
#if 1
_mm_store_si128((__m128i*)ptr, val);
#else
_mm_stream_si128((__m128i*)ptr, val);
#endif
}
template <bool align = false>
static inline void v_trans_c1_8x8_32b(const void* src, void* dst, int src_step,
int dst_step) {
// v_trans_base<uint32_t>(src, src_step, 8, 8, 1, dst, dst_step);
// return;
#ifndef __AVX2__ // else branch is faster
const float* src_ptr = (const float*)src;
float* dst_ptr = (float*)dst;
__m256 a0, a1, a2, a3, a4, a5, a6, a7, b0, b1, b2, b3, b4, b5, b6, b7;
#if 0 // avx2-algo0
a0 = v256_load_32b<align>(src_ptr + 0 * src_step);
a1 = v256_load_32b<align>(src_ptr + 1 * src_step);
a2 = v256_load_32b<align>(src_ptr + 2 * src_step);
a3 = v256_load_32b<align>(src_ptr + 3 * src_step);
a4 = v256_load_32b<align>(src_ptr + 4 * src_step);
a5 = v256_load_32b<align>(src_ptr + 5 * src_step);
a6 = v256_load_32b<align>(src_ptr + 6 * src_step);
a7 = v256_load_32b<align>(src_ptr + 7 * src_step);
#if 0 // avx2-algo0-0
b0 = _mm256_unpacklo_ps(a0, a2);
b1 = _mm256_unpacklo_ps(a1, a3);
b2 = _mm256_unpackhi_ps(a0, a2);
b3 = _mm256_unpackhi_ps(a1, a3);
b4 = _mm256_unpacklo_ps(a4, a6);
b5 = _mm256_unpacklo_ps(a5, a7);
b6 = _mm256_unpackhi_ps(a4, a6);
b7 = _mm256_unpackhi_ps(a5, a7);
a0 = _mm256_unpacklo_ps(b0, b1);
a1 = _mm256_unpackhi_ps(b0, b1);
a2 = _mm256_unpacklo_ps(b2, b3);
a3 = _mm256_unpackhi_ps(b2, b3);
a4 = _mm256_unpacklo_ps(b4, b5);
a5 = _mm256_unpackhi_ps(b4, b5);
a6 = _mm256_unpacklo_ps(b6, b7);
a7 = _mm256_unpackhi_ps(b6, b7);
#else // avx2-algo0-1
b0 = _mm256_unpacklo_ps(a0, a1);
b1 = _mm256_unpackhi_ps(a0, a1);
b2 = _mm256_unpacklo_ps(a2, a3);
b3 = _mm256_unpackhi_ps(a2, a3);
b4 = _mm256_unpacklo_ps(a4, a5);
b5 = _mm256_unpackhi_ps(a4, a5);
b6 = _mm256_unpacklo_ps(a6, a7);
b7 = _mm256_unpackhi_ps(a6, a7);
a0 = _mm256_shuffle_ps(b0, b2, _MM_SHUFFLE(1, 0, 1, 0));
a1 = _mm256_shuffle_ps(b0, b2, _MM_SHUFFLE(3, 2, 3, 2));
a2 = _mm256_shuffle_ps(b1, b3, _MM_SHUFFLE(1, 0, 1, 0));
a3 = _mm256_shuffle_ps(b1, b3, _MM_SHUFFLE(3, 2, 3, 2));
a4 = _mm256_shuffle_ps(b4, b6, _MM_SHUFFLE(1, 0, 1, 0));
a5 = _mm256_shuffle_ps(b4, b6, _MM_SHUFFLE(3, 2, 3, 2));
a6 = _mm256_shuffle_ps(b5, b7, _MM_SHUFFLE(1, 0, 1, 0));
a7 = _mm256_shuffle_ps(b5, b7, _MM_SHUFFLE(3, 2, 3, 2));
#endif
b0 = _mm256_permute2f128_ps(a0, a4, 0x20);
b1 = _mm256_permute2f128_ps(a1, a5, 0x20);
b2 = _mm256_permute2f128_ps(a2, a6, 0x20);
b3 = _mm256_permute2f128_ps(a3, a7, 0x20);
b4 = _mm256_permute2f128_ps(a0, a4, 0x31);
b5 = _mm256_permute2f128_ps(a1, a5, 0x31);
b6 = _mm256_permute2f128_ps(a2, a6, 0x31);
b7 = _mm256_permute2f128_ps(a3, a7, 0x31);
v256_store_32b<align>(dst_ptr + 0 * dst_step, b0);
v256_store_32b<align>(dst_ptr + 1 * dst_step, b1);
v256_store_32b<align>(dst_ptr + 2 * dst_step, b2);
v256_store_32b<align>(dst_ptr + 3 * dst_step, b3);
v256_store_32b<align>(dst_ptr + 4 * dst_step, b4);
v256_store_32b<align>(dst_ptr + 5 * dst_step, b5);
v256_store_32b<align>(dst_ptr + 6 * dst_step, b6);
v256_store_32b<align>(dst_ptr + 7 * dst_step, b7);
#else // avx2-algo-1 replace shuffles with blends
a0 = v256_load3_32b<align>(src_ptr + 0 * src_step + 0,
src_ptr + 4 * src_step + 0);
prefetch(src_ptr + 0 * src_step + 64);
prefetch_l2(src_ptr + 0 * src_step + 64 * PRE_L2_STEP);
prefetch(src_ptr + 4 * src_step + 64);
prefetch_l2(src_ptr + 4 * src_step + 64 * PRE_L2_STEP);
a1 = v256_load3_32b<align>(src_ptr + 1 * src_step + 0,
src_ptr + 5 * src_step + 0);
prefetch(src_ptr + 1 * src_step + 64);
prefetch_l2(src_ptr + 1 * src_step + 64 * PRE_L2_STEP);
prefetch(src_ptr + 5 * src_step + 64);
prefetch_l2(src_ptr + 5 * src_step + 64 * PRE_L2_STEP);
a2 = v256_load3_32b<align>(src_ptr + 2 * src_step + 0,
src_ptr + 6 * src_step + 0);
prefetch(src_ptr + 2 * src_step + 64);
prefetch_l2(src_ptr + 2 * src_step + 64 * PRE_L2_STEP);
prefetch(src_ptr + 6 * src_step + 64);
prefetch_l2(src_ptr + 6 * src_step + 64 * PRE_L2_STEP);
a3 = v256_load3_32b<align>(src_ptr + 3 * src_step + 0,
src_ptr + 7 * src_step + 0);
prefetch(src_ptr + 3 * src_step + 64);
prefetch_l2(src_ptr + 3 * src_step + 64 * PRE_L2_STEP);
prefetch(src_ptr + 7 * src_step + 64);
prefetch_l2(src_ptr + 7 * src_step + 64 * PRE_L2_STEP);
a4 = v256_load3_32b<align>(src_ptr + 0 * src_step + 4,
src_ptr + 4 * src_step + 4);
a5 = v256_load3_32b<align>(src_ptr + 1 * src_step + 4,
src_ptr + 5 * src_step + 4);
a6 = v256_load3_32b<align>(src_ptr + 2 * src_step + 4,
src_ptr + 6 * src_step + 4);
a7 = v256_load3_32b<align>(src_ptr + 3 * src_step + 4,
src_ptr + 7 * src_step + 4);
b0 = _mm256_unpacklo_ps(a0, a1);
b1 = _mm256_unpackhi_ps(a0, a1);
b2 = _mm256_unpacklo_ps(a2, a3);
b3 = _mm256_unpackhi_ps(a2, a3);
b4 = _mm256_unpacklo_ps(a4, a5);
b5 = _mm256_unpackhi_ps(a4, a5);
b6 = _mm256_unpacklo_ps(a6, a7);
b7 = _mm256_unpackhi_ps(a6, a7);
__m256 v;
// a0 = _mm256_shuffle_ps(b0,b2, 0x44);
// a1 = _mm256_shuffle_ps(b0,b2, 0xEE);
v = _mm256_shuffle_ps(b0, b2, 0x4E);
a0 = _mm256_blend_ps(b0, v, 0xCC);
v256_store_32b<align>(dst_ptr + 0 * dst_step, a0);
a1 = _mm256_blend_ps(b2, v, 0x33);
v256_store_32b<align>(dst_ptr + 1 * dst_step, a1);
// a2 = _mm256_shuffle_ps(b1,b3, 0x44);
// a3 = _mm256_shuffle_ps(b1,b3, 0xEE);
v = _mm256_shuffle_ps(b1, b3, 0x4E);
a2 = _mm256_blend_ps(b1, v, 0xCC);
v256_store_32b<align>(dst_ptr + 2 * dst_step, a2);
a3 = _mm256_blend_ps(b3, v, 0x33);
v256_store_32b<align>(dst_ptr + 3 * dst_step, a3);
// a4 = _mm256_shuffle_ps(b4,b6, 0x44);
// a5 = _mm256_shuffle_ps(b4,b6, 0xEE);
v = _mm256_shuffle_ps(b4, b6, 0x4E);
a4 = _mm256_blend_ps(b4, v, 0xCC);
v256_store_32b<align>(dst_ptr + 4 * dst_step, a4);
a5 = _mm256_blend_ps(b6, v, 0x33);
v256_store_32b<align>(dst_ptr + 5 * dst_step, a5);
// a6 = _mm256_shuffle_ps(b5,b7, 0x44);
// a7 = _mm256_shuffle_ps(b5,b7, 0xEE);
v = _mm256_shuffle_ps(b5, b7, 0x4E);
a6 = _mm256_blend_ps(b5, v, 0xCC);
v256_store_32b<align>(dst_ptr + 6 * dst_step, a6);
a7 = _mm256_blend_ps(b7, v, 0x33);
v256_store_32b<align>(dst_ptr + 7 * dst_step, a7);
#endif
#else
const uint32_t* src_ptr = (const uint32_t*)src;
uint32_t* dst_ptr = (uint32_t*)dst;
// prefetchw(dst_ptr + 0 * dst_step);
// prefetchw(dst_ptr + 1 * dst_step);
// prefetchw(dst_ptr + 2 * dst_step);
// prefetchw(dst_ptr + 3 * dst_step);
// prefetchw(dst_ptr + 4 * dst_step);
// prefetchw(dst_ptr + 5 * dst_step);
// prefetchw(dst_ptr + 6 * dst_step);
// prefetchw(dst_ptr + 7 * dst_step);
__m128i i00 = v_load_32b<align>(src_ptr + 0 * src_step + 0);
__m128i i01 = v_load_32b<align>(src_ptr + 0 * src_step + 4);
prefetch(src_ptr + 0 * src_step + 64);
// prefetch_l2(src_ptr + 0 * src_step + 64 * PRE_L2_STEP);
__m128i i10 = v_load_32b<align>(src_ptr + 1 * src_step + 0);
__m128i i11 = v_load_32b<align>(src_ptr + 1 * src_step + 4);
prefetch(src_ptr + 1 * src_step + 64);
// prefetch_l2(src_ptr + 1 * src_step + 64 * PRE_L2_STEP);
__m128i i20 = v_load_32b<align>(src_ptr + 2 * src_step + 0);
__m128i i21 = v_load_32b<align>(src_ptr + 2 * src_step + 4);
prefetch(src_ptr + 2 * src_step + 64);
// prefetch_l2(src_ptr + 2 * src_step + 64 * PRE_L2_STEP);
__m128i i30 = v_load_32b<align>(src_ptr + 3 * src_step + 0);
__m128i i31 = v_load_32b<align>(src_ptr + 3 * src_step + 4);
prefetch(src_ptr + 3 * src_step + 64);
// prefetch_l2(src_ptr + 3 * src_step + 64 * PRE_L2_STEP);
__m128i i40 = v_load_32b<align>(src_ptr + 4 * src_step + 0);
__m128i i41 = v_load_32b<align>(src_ptr + 4 * src_step + 4);
prefetch(src_ptr + 4 * src_step + 64);
// prefetch_l2(src_ptr + 4 * src_step + 64 * PRE_L2_STEP);
__m128i i50 = v_load_32b<align>(src_ptr + 5 * src_step + 0);
__m128i i51 = v_load_32b<align>(src_ptr + 5 * src_step + 4);
prefetch(src_ptr + 5 * src_step + 64);
// prefetch_l2(src_ptr + 5 * src_step + 64 * PRE_L2_STEP);
__m128i i60 = v_load_32b<align>(src_ptr + 6 * src_step + 0);
__m128i i61 = v_load_32b<align>(src_ptr + 6 * src_step + 4);
prefetch(src_ptr + 6 * src_step + 64);
// prefetch_l2(src_ptr + 6 * src_step + 64 * PRE_L2_STEP);
__m128i i70 = v_load_32b<align>(src_ptr + 7 * src_step + 0);
__m128i i71 = v_load_32b<align>(src_ptr + 7 * src_step + 4);
prefetch(src_ptr + 7 * src_step + 64);
// prefetch_l2(src_ptr + 7 * src_step + 64 * PRE_L2_STEP);
#if 1
// Unpack 32 bit elements from:
const __m128i a00 = _mm_unpacklo_epi32(i00, i10);
const __m128i a10 = _mm_unpacklo_epi32(i20, i30);
v_store_32b<align>(dst_ptr + 0 * dst_step + 0,
_mm_unpacklo_epi64(a00, a10));
prefetchw(dst_ptr + 0 * dst_step + 64);
v_store_32b<align>(dst_ptr + 1 * dst_step + 0,
_mm_unpackhi_epi64(a00, a10));
prefetchw(dst_ptr + 1 * dst_step + 64);
const __m128i a01 = _mm_unpacklo_epi32(i01, i11);
const __m128i a11 = _mm_unpacklo_epi32(i21, i31);
v_store_32b<align>(dst_ptr + 4 * dst_step + 0,
_mm_unpacklo_epi64(a01, a11));
prefetchw(dst_ptr + 4 * dst_step + 64);
v_store_32b<align>(dst_ptr + 5 * dst_step + 0,
_mm_unpackhi_epi64(a01, a11));
prefetchw(dst_ptr + 5 * dst_step + 64);
const __m128i a20 = _mm_unpackhi_epi32(i00, i10);
const __m128i a30 = _mm_unpackhi_epi32(i20, i30);
v_store_32b<align>(dst_ptr + 2 * dst_step + 0,
_mm_unpacklo_epi64(a20, a30));
prefetchw(dst_ptr + 2 * dst_step + 64);
v_store_32b<align>(dst_ptr + 3 * dst_step + 0,
_mm_unpackhi_epi64(a20, a30));
prefetchw(dst_ptr + 3 * dst_step + 64);
const __m128i a21 = _mm_unpackhi_epi32(i01, i11);
const __m128i a31 = _mm_unpackhi_epi32(i21, i31);
v_store_32b<align>(dst_ptr + 6 * dst_step + 0,
_mm_unpacklo_epi64(a21, a31));
prefetchw(dst_ptr + 6 * dst_step + 64);
v_store_32b<align>(dst_ptr + 7 * dst_step + 0,
_mm_unpackhi_epi64(a21, a31));
prefetchw(dst_ptr + 7 * dst_step + 64);
const __m128i a40 = _mm_unpacklo_epi32(i40, i50);
const __m128i a50 = _mm_unpacklo_epi32(i60, i70);
v_store_32b<align>(dst_ptr + 0 * dst_step + 4,
_mm_unpacklo_epi64(a40, a50));
v_store_32b<align>(dst_ptr + 1 * dst_step + 4,
_mm_unpackhi_epi64(a40, a50));
const __m128i a41 = _mm_unpacklo_epi32(i41, i51);
const __m128i a51 = _mm_unpacklo_epi32(i61, i71);
v_store_32b<align>(dst_ptr + 4 * dst_step + 4,
_mm_unpacklo_epi64(a41, a51));
v_store_32b<align>(dst_ptr + 5 * dst_step + 4,
_mm_unpackhi_epi64(a41, a51));
const __m128i a60 = _mm_unpackhi_epi32(i40, i50);
const __m128i a70 = _mm_unpackhi_epi32(i60, i70);
v_store_32b<align>(dst_ptr + 2 * dst_step + 4,
_mm_unpacklo_epi64(a60, a70));
v_store_32b<align>(dst_ptr + 3 * dst_step + 4,
_mm_unpackhi_epi64(a60, a70));
const __m128i a61 = _mm_unpackhi_epi32(i41, i51);
const __m128i a71 = _mm_unpackhi_epi32(i61, i71);
v_store_32b<align>(dst_ptr + 6 * dst_step + 4,
_mm_unpacklo_epi64(a61, a71));
v_store_32b<align>(dst_ptr + 7 * dst_step + 4,
_mm_unpackhi_epi64(a61, a71));
#else
// Unpack 32 bit elements from:
const __m128i a00 = _mm_unpacklo_epi32(i00, i10);
const __m128i a01 = _mm_unpacklo_epi32(i01, i11);
const __m128i a10 = _mm_unpacklo_epi32(i20, i30);
const __m128i a11 = _mm_unpacklo_epi32(i21, i31);
const __m128i a20 = _mm_unpackhi_epi32(i00, i10);
const __m128i a21 = _mm_unpackhi_epi32(i01, i11);
const __m128i a30 = _mm_unpackhi_epi32(i20, i30);
const __m128i a31 = _mm_unpackhi_epi32(i21, i31);
const __m128i a40 = _mm_unpacklo_epi32(i40, i50);
const __m128i a41 = _mm_unpacklo_epi32(i41, i51);
const __m128i a50 = _mm_unpacklo_epi32(i60, i70);
const __m128i a51 = _mm_unpacklo_epi32(i61, i71);
const __m128i a60 = _mm_unpackhi_epi32(i40, i50);
const __m128i a61 = _mm_unpackhi_epi32(i41, i51);
const __m128i a70 = _mm_unpackhi_epi32(i60, i70);
const __m128i a71 = _mm_unpackhi_epi32(i61, i71);
// Unpack 64 bit elements to:
v_store_32b<align>(dst_ptr + 0 * dst_step + 0,
_mm_unpacklo_epi64(a00, a10));
v_store_32b<align>(dst_ptr + 0 * dst_step + 4,
_mm_unpacklo_epi64(a40, a50));
v_store_32b<align>(dst_ptr + 1 * dst_step + 0,
_mm_unpackhi_epi64(a00, a10));
v_store_32b<align>(dst_ptr + 1 * dst_step + 4,
_mm_unpackhi_epi64(a40, a50));
v_store_32b<align>(dst_ptr + 2 * dst_step + 0,
_mm_unpacklo_epi64(a20, a30));
v_store_32b<align>(dst_ptr + 2 * dst_step + 4,
_mm_unpacklo_epi64(a60, a70));
v_store_32b<align>(dst_ptr + 3 * dst_step + 0,
_mm_unpackhi_epi64(a20, a30));
v_store_32b<align>(dst_ptr + 3 * dst_step + 4,
_mm_unpackhi_epi64(a60, a70));
v_store_32b<align>(dst_ptr + 4 * dst_step + 0,
_mm_unpacklo_epi64(a01, a11));
v_store_32b<align>(dst_ptr + 4 * dst_step + 4,
_mm_unpacklo_epi64(a41, a51));
v_store_32b<align>(dst_ptr + 5 * dst_step + 0,
_mm_unpackhi_epi64(a01, a11));
v_store_32b<align>(dst_ptr + 5 * dst_step + 4,
_mm_unpackhi_epi64(a41, a51));
v_store_32b<align>(dst_ptr + 6 * dst_step + 0,
_mm_unpacklo_epi64(a21, a31));
v_store_32b<align>(dst_ptr + 6 * dst_step + 4,
_mm_unpacklo_epi64(a61, a71));
v_store_32b<align>(dst_ptr + 7 * dst_step + 0,
_mm_unpackhi_epi64(a21, a31));
v_store_32b<align>(dst_ptr + 7 * dst_step + 4,
_mm_unpackhi_epi64(a61, a71));
#endif
#endif
}
#else
#include <arm_neon.h>
static inline uint32x4x2_t vtrnq_u64_to_u32(const uint32x4_t& a0,
const uint32x4_t& a1) {
uint32x4x2_t b0;
b0.val[0] = vcombine_u32(vget_low_u32(a0), vget_low_u32(a1));
b0.val[1] = vcombine_u32(vget_high_u32(a0), vget_high_u32(a1));
return b0;
}
template <bool align = false>
static inline void v_trans_c1_8x8_32b(const void* src, void* dst, int src_step,
int dst_step) {
const uint32_t* src_ptr = (const uint32_t*)src;
uint32_t* dst_ptr = (uint32_t*)dst;
#ifdef __aarch64__
uint32x4x2_t a0 = vld1q_u32_x2(src_ptr + 0 * src_step);
prefetch(src_ptr + 0 * src_step + 64);
prefetch_l2(src_ptr + 0 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a1 = vld1q_u32_x2(src_ptr + 1 * src_step);
prefetch(src_ptr + 1 * src_step + 64);
prefetch_l2(src_ptr + 1 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a2 = vld1q_u32_x2(src_ptr + 2 * src_step);
prefetch(src_ptr + 2 * src_step + 64);
prefetch_l2(src_ptr + 2 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a3 = vld1q_u32_x2(src_ptr + 3 * src_step);
prefetch(src_ptr + 3 * src_step + 64);
prefetch_l2(src_ptr + 3 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a4 = vld1q_u32_x2(src_ptr + 4 * src_step);
prefetch(src_ptr + 4 * src_step + 64);
prefetch_l2(src_ptr + 4 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a5 = vld1q_u32_x2(src_ptr + 5 * src_step);
prefetch(src_ptr + 5 * src_step + 64);
prefetch_l2(src_ptr + 5 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a6 = vld1q_u32_x2(src_ptr + 6 * src_step);
prefetch(src_ptr + 6 * src_step + 64);
prefetch_l2(src_ptr + 6 * src_step + 64 * PRE_L2_STEP);
uint32x4x2_t a7 = vld1q_u32_x2(src_ptr + 7 * src_step);
prefetch(src_ptr + 7 * src_step + 64);
prefetch_l2(src_ptr + 7 * src_step + 64 * PRE_L2_STEP);
// Swap 32 bit elements from:
// a0: 00 01 02 03 04 05 06 07
// a1: 10 11 12 13 14 15 16 17
// a2: 20 21 22 23 24 25 26 27
// a3: 30 31 32 33 34 35 36 37
// a4: 40 41 42 43 44 45 46 47
// a5: 50 51 52 53 54 55 56 57
// a6: 60 61 62 63 64 65 66 67
// a7: 70 71 72 73 74 75 76 77
// to:
// b0: 00 10 02 12 01 11 03 13
// b1: 20 30 22 32 21 31 23 33
// b2: 40 50 42 52 41 51 43 53
// b3: 60 70 62 72 61 71 63 73
// b4: 04 14 06 16 05 15 07 17
// b5: 24 34 26 36 25 35 27 37
// b6: 44 54 46 56 45 55 47 57
// b7: 64 74 66 76 65 75 67 77
const uint32x4x2_t b0 = vtrnq_u32(a0.val[0], a1.val[0]);
const uint32x4x2_t b1 = vtrnq_u32(a2.val[0], a3.val[0]);
const uint32x4x2_t b2 = vtrnq_u32(a4.val[0], a5.val[0]);
const uint32x4x2_t b3 = vtrnq_u32(a6.val[0], a7.val[0]);
const uint32x4x2_t b4 = vtrnq_u32(a0.val[1], a1.val[1]);
const uint32x4x2_t b5 = vtrnq_u32(a2.val[1], a3.val[1]);
const uint32x4x2_t b6 = vtrnq_u32(a4.val[1], a5.val[1]);
const uint32x4x2_t b7 = vtrnq_u32(a6.val[1], a7.val[1]);
#else
uint32x4_t a00 = vld1q_u32(src_ptr + 0 * src_step + 0);
uint32x4_t a01 = vld1q_u32(src_ptr + 0 * src_step + 4);
prefetch(src_ptr + 0 * src_step + 64);
prefetch_l2(src_ptr + 0 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a10 = vld1q_u32(src_ptr + 1 * src_step + 0);
uint32x4_t a11 = vld1q_u32(src_ptr + 1 * src_step + 4);
prefetch(src_ptr + 1 * src_step + 64);
prefetch_l2(src_ptr + 1 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a20 = vld1q_u32(src_ptr + 2 * src_step + 0);
uint32x4_t a21 = vld1q_u32(src_ptr + 2 * src_step + 4);
prefetch(src_ptr + 2 * src_step + 64);
prefetch_l2(src_ptr + 2 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a30 = vld1q_u32(src_ptr + 3 * src_step + 0);
uint32x4_t a31 = vld1q_u32(src_ptr + 3 * src_step + 4);
prefetch(src_ptr + 3 * src_step + 64);
prefetch_l2(src_ptr + 3 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a40 = vld1q_u32(src_ptr + 4 * src_step + 0);
uint32x4_t a41 = vld1q_u32(src_ptr + 4 * src_step + 4);
prefetch(src_ptr + 4 * src_step + 64);
prefetch_l2(src_ptr + 4 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a50 = vld1q_u32(src_ptr + 5 * src_step + 0);
uint32x4_t a51 = vld1q_u32(src_ptr + 5 * src_step + 4);
prefetch(src_ptr + 5 * src_step + 64);
prefetch_l2(src_ptr + 5 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a60 = vld1q_u32(src_ptr + 6 * src_step + 0);
uint32x4_t a61 = vld1q_u32(src_ptr + 6 * src_step + 4);
prefetch(src_ptr + 6 * src_step + 64);
prefetch_l2(src_ptr + 6 * src_step + 64 * PRE_L2_STEP);
uint32x4_t a70 = vld1q_u32(src_ptr + 7 * src_step + 0);
uint32x4_t a71 = vld1q_u32(src_ptr + 7 * src_step + 4);
prefetch(src_ptr + 7 * src_step + 64);
prefetch_l2(src_ptr + 7 * src_step + 64 * PRE_L2_STEP);
// Swap 32 bit elements from:
const uint32x4x2_t b0 = vtrnq_u32(a00, a10);
const uint32x4x2_t b1 = vtrnq_u32(a20, a30);
const uint32x4x2_t b2 = vtrnq_u32(a40, a50);
const uint32x4x2_t b3 = vtrnq_u32(a60, a70);
const uint32x4x2_t b4 = vtrnq_u32(a01, a11);
const uint32x4x2_t b5 = vtrnq_u32(a21, a31);
const uint32x4x2_t b6 = vtrnq_u32(a41, a51);
const uint32x4x2_t b7 = vtrnq_u32(a61, a71);
#endif
// Swap 64 bit elements to:
// c0: 00 10 20 30 02 12 22 32
// c1: 01 11 21 31 03 13 23 33
// c2: 40 50 60 70 42 52 62 72
// c3: 41 51 61 71 43 53 63 73
// c4: 04 14 24 34 06 16 26 36
// c5: 05 15 25 35 07 17 27 37
// c6: 44 54 64 74 46 56 66 76
// c7: 45 55 65 75 47 57 67 77
const uint32x4x2_t c0 = vtrnq_u64_to_u32(b0.val[0], b1.val[0]);
const uint32x4x2_t c1 = vtrnq_u64_to_u32(b0.val[1], b1.val[1]);
const uint32x4x2_t c2 = vtrnq_u64_to_u32(b2.val[0], b3.val[0]);
const uint32x4x2_t c3 = vtrnq_u64_to_u32(b2.val[1], b3.val[1]);
const uint32x4x2_t c4 = vtrnq_u64_to_u32(b4.val[0], b5.val[0]);
const uint32x4x2_t c5 = vtrnq_u64_to_u32(b4.val[1], b5.val[1]);
const uint32x4x2_t c6 = vtrnq_u64_to_u32(b6.val[0], b7.val[0]);
const uint32x4x2_t c7 = vtrnq_u64_to_u32(b6.val[1], b7.val[1]);
// Swap 128 bit elements to:
// a0: 00 10 20 30 40 50 60 70
// a1: 01 11 21 31 41 51 61 71
// a2: 02 12 22 32 42 52 62 72
// a3: 03 13 23 33 43 53 63 73
// a4: 04 14 24 34 44 54 64 74
// a5: 05 15 25 35 45 55 65 75
// a6: 06 16 26 36 46 56 66 76
// a7: 07 17 27 37 47 57 67 77
vst1q_u32(dst_ptr + 0 * dst_step + 0, c0.val[0]);
vst1q_u32(dst_ptr + 0 * dst_step + 4, c2.val[0]);
vst1q_u32(dst_ptr + 1 * dst_step + 0, c1.val[0]);
vst1q_u32(dst_ptr + 1 * dst_step + 4, c3.val[0]);
vst1q_u32(dst_ptr + 2 * dst_step + 0, c0.val[1]);
vst1q_u32(dst_ptr + 2 * dst_step + 4, c2.val[1]);
vst1q_u32(dst_ptr + 3 * dst_step + 0, c1.val[1]);
vst1q_u32(dst_ptr + 3 * dst_step + 4, c3.val[1]);
vst1q_u32(dst_ptr + 4 * dst_step + 0, c4.val[0]);
vst1q_u32(dst_ptr + 4 * dst_step + 4, c6.val[0]);
vst1q_u32(dst_ptr + 5 * dst_step + 0, c5.val[0]);
vst1q_u32(dst_ptr + 5 * dst_step + 4, c7.val[0]);
vst1q_u32(dst_ptr + 6 * dst_step + 0, c4.val[1]);
vst1q_u32(dst_ptr + 6 * dst_step + 4, c6.val[1]);
vst1q_u32(dst_ptr + 7 * dst_step + 0, c5.val[1]);
vst1q_u32(dst_ptr + 7 * dst_step + 4, c7.val[1]);
}
#endif
} // namespace
//! This func used to compare with func generated by TVM.
template <int B3H, int B3W, int B2H, int B2W, int B1H, int B1W, int KH = 8,
int KW = 8>
void transpose_32b_c1_base(const float* src, float* dst, int height, int width,
int channel, int sstride1, int dstride1,
int sstride2, int dstride2) {
int h = 0;
int w = 0;
int BHS = B3H * B2H * B1H * KH;
int BWS = B3W * B2W * B1W * KW;
if (BHS <= height && BWS <= width) {
int tile_total_hnum = height / BHS;
int tile_total_wnum = width / BWS;
// printf("out(%d,%d),L3(%d,%d),L2(%d,%d),L1(%d,%d)\n", tile_total_hnum,
// tile_total_wnum, B3H, B3W, B2H, B2W, B1H, B1W);
for (int rh = 0; rh < tile_total_hnum; rh++) {
for (int rw = 0; rw < tile_total_wnum; rw++) {
for (int b3h = 0; b3h < B3H; b3h++) {
for (int b3w = 0; b3w < B3W; b3w++) {
for (int b2h = 0; b2h < B2H; b2h++) {
for (int b2w = 0; b2w < B2W; b2w++) {
for (int b1h = 0; b1h < B1H; b1h++) {
for (int b1w = 0; b1w < B1W; b1w++) {
// clang-format off
v_trans_c1_8x8_32b<false>(
src + (rh*B3H*B2H*B1H*KH + b3h*B2H*B1H*KH + b2h*B1H*KH + b1h*KH) * sstride1 + (rw*B3W*B2W*B1W*KW + b3w*B2W*B1W*KW + b2w*B1W*KW + b1w*KW) * sstride2,
dst + (rw*B3W*B2W*B1W*KW + b3w*B2W*B1W*KW + b2w*B1W*KW + b1w*KW) * dstride1 + (rh*B3H*B2H*B1H*KH + b3h*B2H*B1H*KH + b2h*B1H*KH + b1h*KH) * dstride2,
sstride1, dstride1);
// clang-format on
}
}
}
}
}
}
}
}
h += tile_total_hnum;
w += tile_total_wnum;
}
// process bottom tail region
for (int h2 = h * BHS; h2 < height; h2++) {
for (int w2 = 0; w2 < width; w2++) {
for (int c = 0; c < channel; c++) {
dst[w2 * dstride1 + h2 * dstride2 + c] =
src[h2 * sstride1 + w2 * sstride2 + c];
}
}
}
// process right tail region
for (int h2 = 0; h2 < h * BHS; h2++) {
for (int w2 = w * BWS; w2 < width; w2++) {
for (int c = 0; c < channel; c++) {
dst[w2 * dstride1 + h2 * dstride2 + c] =
src[h2 * sstride1 + w2 * sstride2 + c];
}
}
}
}
#define ENABLE_TMA 1
#define MEMCPY 0
int main() {
int dev_id = 0;
if (set_cpu_thread_affinity_spec_core(dev_id)) {
printf("faild set thread affinity(core %d)\n", dev_id);
}
// int height = 2048;
// int width = 4096;
int height = 1 << 13;
int width = 1 << 13;
size_t img_size = height * width * sizeof(float);
size_t iter_num = 10;
printf("the iter num is %zu\n", iter_num);
float* src = new float[height * width];
float* dst = new float[height * width];
for (int idx = 0; idx < height * width; ++idx) {
src[idx] = rand() % 255;
}
volatile uint8_t res = 0;
// warmup
#if MEMCPY
memcpy(dst, src, img_size);
#else
transpose_32b_c1_base<2, 2, 1, 1, 2, 2, 8, 8>(src, dst, height, width, 1,
width, height, 1, 1);
#endif
for (int y = 0; y < height; y++) {
for (int x = 0; x < width; x++) {
dst[x * height + y] = src[y * width + x];
}
}
#if ENABLE_TMA
#if defined(__aarch64__)
mperf::tma::MPFTMA mpf_tma(mperf::MPFXPUType::A55);
// mperf::tma::MPFTMA mpf_tma(mperf::MPFXPUType::A510);
// clang-format off
mpf_tma.init({"Frontend_Bound",
"Fetch_Latency",
"ICache_Misses",
"ITLB_Misses",
"Predecode_Error",
"Fetch_Bandwidth",
"Bad_Speculation",
"Branch_Mispredicts",
"Backend_Bound",
"Memory_Bound",
"Load_Bound",
"Load_DTLB",
"Load_Cache",
"Store_Bound",
"Store_TLB",
"Store_Buffer",
"Core_Bound",
"Interlock_Bound",
"Interlock_AGU",
"Interlock_FPU",
"Core_Bound_Others",
"Retiring",
"LD_Retiring",
"ST_Retiring",
"DP_Retiring",
"ASE_Retiring",
"VFP_Retiring",
"PC_Write_Retiring",
"BR_IMMED_Retiring",
"BR_RETURN_Retiring",
"BR_INDIRECT_Retiring"});
// clang-format on
#else
mperf::tma::MPFTMA mpf_tma(mperf::MPFXPUType::HSX_SERVER);
mpf_tma.init(
{"Frontend_Bound", "Bad_Speculation", "Backend_Bound", "Retiring"});
#endif
size_t gn = mpf_tma.group_num();
size_t uncore_evt_num = mpf_tma.uncore_events_num();
printf("the gn and uncore_evt_nums %zu, %zu\n", gn, uncore_evt_num);
for (size_t i = 0; i < gn; ++i) {
mpf_tma.start(i);
#endif
for (size_t j = 0; j < iter_num; ++j) {
#if MEMCPY
memcpy(dst, src, img_size);
#else
// 1, 4, 1, 1, 4, 2, 8, 8
transpose_32b_c1_base<2, 2, 1, 1, 2, 2, 8, 8>(src, dst, height, width,
1, width, height, 1, 1);
#endif
res += dst[0];
}
#if ENABLE_TMA
mpf_tma.sample_and_stop(iter_num);
}
// FIXME(hc): when the iter_numm greater than 5, we have ran into that
// uncore event value became too big, so we call `sample(1)` very iter in
// uncore event sample to walk around it
for (size_t i = 0; i < uncore_evt_num; ++i) {
mpf_tma.start_uncore(i);
#endif
for (size_t j = 0; j < iter_num; ++j) {
#if MEMCPY
memcpy(dst, src, img_size);
#else
// 1, 4, 1, 1, 4, 2, 8, 8
transpose_32b_c1_base<2, 2, 1, 1, 2, 2, 8, 8>(src, dst, height, width,
1, width, height, 1, 1);
#endif
res += dst[0];
#if ENABLE_TMA
mpf_tma.sample(1);
#endif
}
#if ENABLE_TMA
// mpf_tma.sample_and_stop(iter_num);
}
mpf_tma.deinit();
#endif
printf("the res is %u\n", res);
delete[] src;
delete[] dst;
return 0;
}