-
Notifications
You must be signed in to change notification settings - Fork 259
/
Copy pathlooprestoration_avx512.asm
2122 lines (2109 loc) · 67.1 KB
/
looprestoration_avx512.asm
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
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
; Copyright © 2021, VideoLAN and dav1d authors
; Copyright © 2021, Two Orioles, LLC
; All rights reserved.
;
; Redistribution and use in source and binary forms, with or without
; modification, are permitted provided that the following conditions are met:
;
; 1. Redistributions of source code must retain the above copyright notice, this
; list of conditions and the following disclaimer.
;
; 2. Redistributions in binary form must reproduce the above copyright notice,
; this list of conditions and the following disclaimer in the documentation
; and/or other materials provided with the distribution.
;
; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
; WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
; DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
; ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
; (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
; ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
; (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
; SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
%include "config.asm"
%include "ext/x86/x86inc.asm"
%if ARCH_X86_64
SECTION_RODATA 32
wiener_shufA: db 1, 2, 7, 6, 3, 4, 9, 8, 5, 6, 11, 10, 7, 8, 13, 12
wiener_shufB: db 2, 3, 8, 7, 4, 5, 10, 9, 6, 7, 12, 11, 8, 9, 14, 13
wiener_shufC: db 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11
wiener_shufD: db 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
wiener_perm32: db 1, 9, 3, 11, 5, 13, 7, 15, 33, 41, 35, 43, 37, 45, 39, 47
db 17, 25, 19, 27, 21, 29, 23, 31, 49, 57, 51, 59, 53, 61, 55, 63
sgr_shuf: db 128, 1, -1, 2,132, 3, -1, 4,136, 5, -1, 6,140, 7, -1, 8
db 129, 9, -1, 10,133, 11, -1, 12,137, -1, -1, -1,141, -1, 0,128
sgr_mix_perm: db 1, 3, 5, 7, 17, 19, 21, 23, 33, 35, 37, 39, 49, 51, 53, 55
r_ext_mask: times 68 db -1
times 4 db 0
wiener_x_shuf: db 0, 2, -1, 0
wiener_x_add: db 0, 1,127, 0
pw_61448: times 2 dw 61448
pw_164_455: dw 164, 455
pd_m16380: dd -16380
pd_m4096: dd -4096
pd_m25 dd -25
pd_m9: dd -9
pd_34816: dd 34816
pd_8421376: dd 8421376
cextern sgr_x_by_x
SECTION .text
DECLARE_REG_TMP 8, 7, 9, 11, 12, 13, 14 ; ring buffer pointers
INIT_ZMM avx512icl
cglobal wiener_filter7_8bpc, 4, 15, 20, -384*12-16, dst, stride, left, lpf, \
w, h, edge, flt
mov fltq, r6mp
mov wd, wm
movifnidn hd, hm
mov edged, r7m
vbroadcasti32x4 m6, [wiener_shufA]
vbroadcasti32x4 m7, [wiener_shufB]
mov r10d, 0xfffe
vbroadcasti32x4 m8, [wiener_shufC]
vbroadcasti32x4 m9, [wiener_shufD]
kmovw k1, r10d
vpbroadcastd m0, [wiener_x_shuf]
vpbroadcastd m1, [wiener_x_add]
mov r10, 0xaaaaaaaaaaaaaaaa
vpbroadcastd m11, [fltq+ 0]
vpbroadcastd m12, [fltq+ 4]
kmovq k2, r10
vpbroadcastd m10, [pd_m16380]
packsswb m11, m11 ; x0 x1 x0 x1
vpbroadcastd m14, [fltq+16]
pshufb m12, m0
vpbroadcastd m15, [fltq+20]
paddb m12, m1 ; x2 x3+1 x2 127
vpbroadcastd m13, [pd_8421376]
psllw m14, 5 ; y0 y1
psllw m15, 5 ; y2 y3
cmp wd, 32 ; the minimum lr unit size for chroma in 4:2:0 is 32
jle .w32 ; pixels, so we need a special case for small widths
lea t1, [rsp+wq*2+16]
add lpfq, wq
add dstq, wq
neg wq
test edgeb, 4 ; LR_HAVE_TOP
jz .no_top
call .h_top
add lpfq, strideq
mov t6, t1
mov t5, t1
add t1, 384*2
call .h_top
lea r10, [lpfq+strideq*4]
mov lpfq, dstq
mov t4, t1
add t1, 384*2
add r10, strideq
mov [rsp], r10 ; below
call .h
mov t3, t1
mov t2, t1
dec hd
jz .v1
add lpfq, strideq
add t1, 384*2
call .h
mov t2, t1
dec hd
jz .v2
add lpfq, strideq
add t1, 384*2
call .h
dec hd
jz .v3
.main:
lea t0, [t1+384*2]
.main_loop:
call .hv
dec hd
jnz .main_loop
test edgeb, 8 ; LR_HAVE_BOTTOM
jz .v3
mov lpfq, [rsp]
call .hv_bottom
add lpfq, strideq
call .hv_bottom
.v1:
call .v
RET
.no_top:
lea r10, [lpfq+strideq*4]
mov lpfq, dstq
lea r10, [r10+strideq*2]
mov [rsp], r10
call .h
mov t6, t1
mov t5, t1
mov t4, t1
mov t3, t1
mov t2, t1
dec hd
jz .v1
add lpfq, strideq
add t1, 384*2
call .h
mov t2, t1
dec hd
jz .v2
add lpfq, strideq
add t1, 384*2
call .h
dec hd
jz .v3
lea t0, [t1+384*2]
call .hv
dec hd
jz .v3
add t0, 384*8
call .hv
dec hd
jnz .main
.v3:
call .v
.v2:
call .v
jmp .v1
.h:
mov r10, wq
test edgeb, 1 ; LR_HAVE_LEFT
jz .h_extend_left
movd xm16, [leftq]
vmovdqu32 m16{k1}, [lpfq+r10-4]
add leftq, 4
jmp .h_main
.h_extend_left:
vpbroadcastb xm16, [lpfq+r10] ; the masked load ensures that no exception
vmovdqu32 m16{k1}, [lpfq+r10-4] ; gets raised from accessing invalid memory
jmp .h_main
.h_top:
mov r10, wq
test edgeb, 1 ; LR_HAVE_LEFT
jz .h_extend_left
.h_loop:
movu m16, [lpfq+r10-4]
.h_main:
movu m17, [lpfq+r10+4]
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .h_have_right
cmp r10d, -66
jl .h_have_right
push r0
lea r0, [r_ext_mask+65]
vpbroadcastb m0, [lpfq-1]
vpternlogd m16, m0, [r0+r10+0], 0xe4 ; c ? a : b
vpternlogd m17, m0, [r0+r10+8], 0xe4
pop r0
.h_have_right:
pshufb m4, m16, m6
mova m0, m10
vpdpbusd m0, m4, m11
pshufb m4, m16, m7
mova m2, m10
vpdpbusd m2, m4, m11
pshufb m4, m17, m6
mova m1, m10
vpdpbusd m1, m4, m11
pshufb m4, m17, m7
mova m3, m10
vpdpbusd m3, m4, m11
pshufb m4, m16, m8
vpdpbusd m0, m4, m12
pshufb m16, m9
vpdpbusd m2, m16, m12
pshufb m4, m17, m8
vpdpbusd m1, m4, m12
pshufb m17, m9
vpdpbusd m3, m17, m12
packssdw m0, m2
packssdw m1, m3
psraw m0, 3
psraw m1, 3
mova [t1+r10*2+ 0], m0
mova [t1+r10*2+64], m1
add r10, 64
jl .h_loop
ret
ALIGN function_align
.hv:
add lpfq, strideq
mov r10, wq
test edgeb, 1 ; LR_HAVE_LEFT
jz .hv_extend_left
movd xm16, [leftq]
vmovdqu32 m16{k1}, [lpfq+r10-4]
add leftq, 4
jmp .hv_main
.hv_extend_left:
vpbroadcastb xm16, [lpfq+r10]
vmovdqu32 m16{k1}, [lpfq+r10-4]
jmp .hv_main
.hv_bottom:
mov r10, wq
test edgeb, 1 ; LR_HAVE_LEFT
jz .hv_extend_left
.hv_loop:
movu m16, [lpfq+r10-4]
.hv_main:
movu m17, [lpfq+r10+4]
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .hv_have_right
cmp r10d, -66
jl .hv_have_right
push r0
lea r0, [r_ext_mask+65]
vpbroadcastb m0, [lpfq-1]
vpternlogd m16, m0, [r0+r10+0], 0xe4 ; c ? a : b
vpternlogd m17, m0, [r0+r10+8], 0xe4
pop r0
.hv_have_right:
pshufb m4, m16, m6
mova m0, m10
vpdpbusd m0, m4, m11
pshufb m4, m16, m7
mova m2, m10
vpdpbusd m2, m4, m11
pshufb m4, m17, m6
mova m1, m10
vpdpbusd m1, m4, m11
pshufb m4, m17, m7
mova m3, m10
vpdpbusd m3, m4, m11
pshufb m4, m16, m8
vpdpbusd m0, m4, m12
pshufb m16, m9
vpdpbusd m2, m16, m12
pshufb m4, m17, m8
vpdpbusd m1, m4, m12
pshufb m17, m9
vpdpbusd m3, m17, m12
packssdw m0, m2
packssdw m1, m3
psraw m0, 3
psraw m1, 3
mova m16, [t4+r10*2]
paddw m16, [t2+r10*2]
mova m3, [t3+r10*2]
mova m17, [t4+r10*2+64]
paddw m17, [t2+r10*2+64]
mova m5, [t3+r10*2+64]
punpcklwd m4, m16, m3
mova m2, m13
vpdpwssd m2, m4, m15
punpcklwd m18, m17, m5
mova m4, m13
vpdpwssd m4, m18, m15
punpckhwd m16, m3
mova m3, m13
vpdpwssd m3, m16, m15
punpckhwd m17, m5
mova m5, m13
vpdpwssd m5, m17, m15
mova m17, [t5+r10*2]
paddw m17, [t1+r10*2]
paddw m16, m0, [t6+r10*2]
mova m19, [t5+r10*2+64]
paddw m19, [t1+r10*2+64]
paddw m18, m1, [t6+r10*2+64]
mova [t0+r10*2+ 0], m0
mova [t0+r10*2+64], m1
punpcklwd m0, m16, m17
vpdpwssd m2, m0, m14
punpcklwd m1, m18, m19
vpdpwssd m4, m1, m14
punpckhwd m16, m17
vpdpwssd m3, m16, m14
punpckhwd m18, m19
vpdpwssd m5, m18, m14
packuswb m2, m4
psrlw m2, 8
vpackuswb m2{k2}, m3, m5
movu [dstq+r10], m2 ; We don't have a separate 5-tap version so the 7-tap
add r10, 64 ; function is used for chroma as well, and in some
jl .hv_loop ; esoteric edge cases chroma dst pointers may only
mov t6, t5 ; have a 32-byte alignment despite having a width
mov t5, t4 ; larger than 32, so use an unaligned store here.
mov t4, t3
mov t3, t2
mov t2, t1
mov t1, t0
mov t0, t6
add dstq, strideq
ret
.v:
mov r10, wq
.v_loop:
mova m4, [t4+r10*2+ 0]
paddw m4, [t2+r10*2+ 0]
mova m1, [t3+r10*2+ 0]
mova m5, [t4+r10*2+64]
paddw m5, [t2+r10*2+64]
mova m3, [t3+r10*2+64]
punpcklwd m6, m4, m1
mova m0, m13
vpdpwssd m0, m6, m15
punpcklwd m6, m5, m3
mova m2, m13
vpdpwssd m2, m6, m15
punpckhwd m4, m1
mova m1, m13
vpdpwssd m1, m4, m15
punpckhwd m5, m3
mova m3, m13
vpdpwssd m3, m5, m15
mova m5, [t1+r10*2+ 0]
paddw m4, m5, [t6+r10*2+ 0]
paddw m5, [t5+r10*2+ 0]
mova m7, [t1+r10*2+64]
paddw m6, m7, [t6+r10*2+64]
paddw m7, [t5+r10*2+64]
punpcklwd m8, m4, m5
vpdpwssd m0, m8, m14
punpcklwd m8, m6, m7
vpdpwssd m2, m8, m14
punpckhwd m4, m5
vpdpwssd m1, m4, m14
punpckhwd m6, m7
vpdpwssd m3, m6, m14
packuswb m0, m2
psrlw m0, 8
vpackuswb m0{k2}, m1, m3
movu [dstq+r10], m0
add r10, 64
jl .v_loop
mov t6, t5
mov t5, t4
mov t4, t3
mov t3, t2
mov t2, t1
add dstq, strideq
ret
.w32:
lea r10, [r_ext_mask+73]
mova ym18, [wiener_perm32]
lea t1, [rsp+16]
sub r10, wq
test edgeb, 4 ; LR_HAVE_TOP
jz .w32_no_top
call .w32_h_top
add lpfq, strideq
mov t6, t1
mov t5, t1
add t1, 32*2
call .w32_h_top
lea r9, [lpfq+strideq*4]
mov lpfq, dstq
mov t4, t1
add t1, 32*2
add r9, strideq
mov [rsp], r9 ; below
call .w32_h
mov t3, t1
mov t2, t1
dec hd
jz .w32_v1
add lpfq, strideq
add t1, 32*2
call .w32_h
mov t2, t1
dec hd
jz .w32_v2
add lpfq, strideq
add t1, 32*2
call .w32_h
dec hd
jz .w32_v3
.w32_main:
lea t0, [t1+32*2]
.w32_main_loop:
call .w32_hv
dec hd
jnz .w32_main_loop
test edgeb, 8 ; LR_HAVE_BOTTOM
jz .w32_v3
mov lpfq, [rsp]
call .w32_hv_bottom
add lpfq, strideq
call .w32_hv_bottom
.w32_v1:
call .w32_v
RET
.w32_no_top:
lea r9, [lpfq+strideq*4]
mov lpfq, dstq
lea r9, [r9+strideq*2]
mov [rsp], r9
call .w32_h
mov t6, t1
mov t5, t1
mov t4, t1
mov t3, t1
mov t2, t1
dec hd
jz .w32_v1
add lpfq, strideq
add t1, 32*2
call .w32_h
mov t2, t1
dec hd
jz .w32_v2
add lpfq, strideq
add t1, 32*2
call .w32_h
dec hd
jz .w32_v3
lea t0, [t1+32*2]
call .w32_hv
dec hd
jz .w32_v3
add t0, 32*8
call .w32_hv
dec hd
jnz .w32_main
.w32_v3:
call .w32_v
.w32_v2:
call .w32_v
jmp .w32_v1
.w32_h:
test edgeb, 1 ; LR_HAVE_LEFT
jz .w32_h_extend_left
movd xm16, [leftq]
vmovdqu32 ym16{k1}, [lpfq-4]
add leftq, 4
jmp .w32_h_main
.w32_h_extend_left:
vpbroadcastb xm16, [lpfq] ; the masked load ensures that no exception
vmovdqu32 ym16{k1}, [lpfq-4] ; gets raised from accessing invalid memory
jmp .w32_h_main
.w32_h_top:
test edgeb, 1 ; LR_HAVE_LEFT
jz .w32_h_extend_left
movu ym16, [lpfq-4]
.w32_h_main:
vinserti32x8 m16, [lpfq+4], 1
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .w32_h_have_right
vpbroadcastb m0, [lpfq+wq-1]
movu ym17, [r10-8]
vinserti32x8 m17, [r10+0], 1
vpternlogd m16, m0, m17, 0xe4 ; c ? a : b
.w32_h_have_right:
pshufb m2, m16, m6
mova m0, m10
vpdpbusd m0, m2, m11
pshufb m2, m16, m7
mova m1, m10
vpdpbusd m1, m2, m11
pshufb m2, m16, m8
vpdpbusd m0, m2, m12
pshufb m16, m9
vpdpbusd m1, m16, m12
packssdw m0, m1
psraw m0, 3
mova [t1], m0
ret
.w32_hv:
add lpfq, strideq
test edgeb, 1 ; LR_HAVE_LEFT
jz .w32_hv_extend_left
movd xm16, [leftq]
vmovdqu32 ym16{k1}, [lpfq-4]
add leftq, 4
jmp .w32_hv_main
.w32_hv_extend_left:
vpbroadcastb xm16, [lpfq]
vmovdqu32 ym16{k1}, [lpfq-4]
jmp .w32_hv_main
.w32_hv_bottom:
test edgeb, 1 ; LR_HAVE_LEFT
jz .w32_hv_extend_left
movu ym16, [lpfq-4]
.w32_hv_main:
vinserti32x8 m16, [lpfq+4], 1
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .w32_hv_have_right
vpbroadcastb m0, [lpfq+wq-1]
movu ym17, [r10-8]
vinserti32x8 m17, [r10+0], 1
vpternlogd m16, m0, m17, 0xe4
.w32_hv_have_right:
mova m3, [t4]
paddw m3, [t2]
mova m2, [t3]
pshufb m4, m16, m6
mova m0, m10
vpdpbusd m0, m4, m11
pshufb m4, m16, m7
mova m5, m10
vpdpbusd m5, m4, m11
punpcklwd m4, m3, m2
mova m1, m13
vpdpwssd m1, m4, m15
punpckhwd m3, m2
mova m2, m13
vpdpwssd m2, m3, m15
pshufb m4, m16, m8
vpdpbusd m0, m4, m12
pshufb m16, m9
vpdpbusd m5, m16, m12
packssdw m0, m5
psraw m0, 3
mova m4, [t5]
paddw m4, [t1]
paddw m3, m0, [t6]
mova [t0], m0
punpcklwd m0, m3, m4
vpdpwssd m1, m0, m14
punpckhwd m3, m4
vpdpwssd m2, m3, m14
packuswb m1, m2
vpermb m16, m18, m1
mova [dstq], ym16
mov t6, t5
mov t5, t4
mov t4, t3
mov t3, t2
mov t2, t1
mov t1, t0
mov t0, t6
add dstq, strideq
ret
.w32_v:
mova m2, [t4]
paddw m2, [t2]
mova m1, [t3]
mova m4, [t1]
paddw m3, m4, [t6]
paddw m4, [t5]
punpcklwd m5, m2, m1
mova m0, m13
vpdpwssd m0, m5, m15
punpckhwd m2, m1
mova m1, m13
vpdpwssd m1, m2, m15
punpcklwd m2, m3, m4
vpdpwssd m0, m2, m14
punpckhwd m3, m4
vpdpwssd m1, m3, m14
packuswb m0, m1
vpermb m16, m18, m0
mova [dstq], ym16
mov t6, t5
mov t5, t4
mov t4, t3
mov t3, t2
mov t2, t1
add dstq, strideq
ret
cglobal sgr_filter_5x5_8bpc, 4, 13, 23, 416*24+16, dst, stride, left, lpf, \
w, h, edge, params
mov paramsq, r6mp
mov wd, wm
mov hd, hm
mov edged, r7m
vbroadcasti32x4 m5, [sgr_shuf+1]
add lpfq, wq
vbroadcasti32x4 m6, [sgr_shuf+9]
add dstq, wq
vbroadcasti32x4 m7, [sgr_shuf+3]
lea t3, [rsp+wq*4+16+416*12]
vbroadcasti32x4 m8, [sgr_shuf+7]
pxor m4, m4
vpbroadcastd m9, [pd_m25]
vpsubd m11, m4, [paramsq+0] {1to16} ; -s0
vpbroadcastw m15, [paramsq+8] ; w0
lea t1, [rsp+wq*2+20]
vpbroadcastd m10, [pw_164_455]
neg wq
vpbroadcastd m12, [pw_61448] ; (15 << 12) + (1 << 3)
mov r10d, 0xfe
vpbroadcastd m13, [pd_m4096]
kmovb k1, r10d
vpbroadcastd m14, [pd_34816] ; (1 << 11) + (1 << 15)
mov r10, 0x3333333333333333
mova m18, [sgr_x_by_x+64*0]
kmovq k2, r10
mova m19, [sgr_x_by_x+64*1]
lea r12, [r_ext_mask+75]
mova m20, [sgr_x_by_x+64*2]
psllw m15, 4
mova m21, [sgr_x_by_x+64*3]
lea r10, [lpfq+strideq*4]
mova ym22, [sgr_shuf]
add r10, strideq
mov [rsp], r10 ; below
test edgeb, 4 ; LR_HAVE_TOP
jz .no_top
call .h_top
add lpfq, strideq
mov t2, t1
call .top_fixup
add t1, 416*6
call .h_top
lea r10, [lpfq+strideq*4]
mov lpfq, dstq
add r10, strideq
mov [rsp], r10 ; below
mov t0, t2
dec hd
jz .height1
or edged, 16
call .h
.main:
add lpfq, strideq
call .hv
call .prep_n
sub hd, 2
jl .extend_bottom
.main_loop:
add lpfq, strideq
test hd, hd
jz .odd_height
call .h
add lpfq, strideq
call .hv
call .n0
call .n1
sub hd, 2
jge .main_loop
test edgeb, 8 ; LR_HAVE_BOTTOM
jz .extend_bottom
mov lpfq, [rsp]
call .h_top
add lpfq, strideq
call .hv_bottom
.end:
call .n0
call .n1
.end2:
RET
.height1:
call .hv
call .prep_n
jmp .odd_height_end
.odd_height:
call .hv
call .n0
call .n1
.odd_height_end:
call .v
call .n0
jmp .end2
.extend_bottom:
call .v
jmp .end
.no_top:
lea r10, [lpfq+strideq*4]
mov lpfq, dstq
lea r10, [r10+strideq*2]
mov [rsp], r10
call .h
lea t2, [t1+416*6]
call .top_fixup
dec hd
jz .no_top_height1
or edged, 16
mov t0, t1
mov t1, t2
jmp .main
.no_top_height1:
call .v
call .prep_n
jmp .odd_height_end
.h: ; horizontal boxsum
lea r10, [wq-2]
test edgeb, 1 ; LR_HAVE_LEFT
jz .h_extend_left
movd xm17, [leftq]
vmovdqu32 ym17{k1}, [lpfq+wq-4]
add leftq, 4
jmp .h_main
.h_extend_left:
vpbroadcastb xm17, [lpfq+wq]
vmovdqu32 ym17{k1}, [lpfq+wq-4]
jmp .h_main
.h_top:
lea r10, [wq-2]
test edgeb, 1 ; LR_HAVE_LEFT
jz .h_extend_left
.h_loop:
movu ym17, [lpfq+r10-2]
.h_main:
vinserti32x8 m17, [lpfq+r10+6], 1
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .h_have_right
cmp r10d, -34
jl .h_have_right
vpbroadcastb m0, [lpfq-1]
movu ym16, [r12+r10-8]
vinserti32x8 m16, [r12+r10+0], 1
vpternlogd m17, m0, m16, 0xe4
.h_have_right:
pshufb m3, m17, m5
pmullw m2, m3, m3
pshufb m1, m17, m6
paddw m0, m3, m1
shufps m3, m1, q2121
paddw m0, m3
punpcklwd m16, m3, m1
punpckhwd m3, m1
punpcklwd m1, m2, m4
vpdpwssd m1, m16, m16
punpckhwd m2, m4
vpdpwssd m2, m3, m3
pshufb m16, m17, m7
paddw m0, m16
pshufb m17, m8
paddw m0, m17 ; sum
punpcklwd m3, m16, m17
vpdpwssd m1, m3, m3 ; sumsq
punpckhwd m16, m17
vpdpwssd m2, m16, m16
test edgeb, 16 ; y > 0
jz .h_loop_end
paddw m0, [t1+r10*2+416*0]
paddd m1, [t1+r10*2+416*2]
paddd m2, [t1+r10*2+416*4]
.h_loop_end:
mova [t1+r10*2+416*0], m0
mova [t1+r10*2+416*2], m1
mova [t1+r10*2+416*4], m2
add r10, 32
jl .h_loop
ret
.top_fixup:
lea r10, [wq-2]
.top_fixup_loop: ; the sums of the first row needs to be doubled
mova m0, [t1+r10*2+416*0]
mova m1, [t1+r10*2+416*2]
mova m2, [t1+r10*2+416*4]
paddw m0, m0
paddd m1, m1
paddd m2, m2
mova [t2+r10*2+416*0], m0
mova [t2+r10*2+416*2], m1
mova [t2+r10*2+416*4], m2
add r10, 32
jl .top_fixup_loop
ret
ALIGN function_align
.hv: ; horizontal boxsum + vertical boxsum + ab
lea r10, [wq-2]
test edgeb, 1 ; LR_HAVE_LEFT
jz .hv_extend_left
movd xm17, [leftq]
vmovdqu32 ym17{k1}, [lpfq+wq-4]
add leftq, 4
jmp .hv_main
.hv_extend_left:
vpbroadcastb xm17, [lpfq+wq]
vmovdqu32 ym17{k1}, [lpfq+wq-4]
jmp .hv_main
.hv_bottom:
lea r10, [wq-2]
test edgeb, 1 ; LR_HAVE_LEFT
jz .hv_extend_left
.hv_loop:
movu ym17, [lpfq+r10-2]
.hv_main:
vinserti32x8 m17, [lpfq+r10+6], 1
test edgeb, 2 ; LR_HAVE_RIGHT
jnz .hv_have_right
cmp r10d, -34
jl .hv_have_right
vpbroadcastb m0, [lpfq-1]
movu ym16, [r12+r10-8]
vinserti32x8 m16, [r12+r10+0], 1
vpternlogd m17, m0, m16, 0xe4
.hv_have_right:
pshufb m1, m17, m5
pmullw m3, m1, m1
pshufb m2, m17, m6
paddw m0, m1, m2
shufps m1, m2, q2121
paddw m0, m1
punpcklwd m16, m1, m2
punpckhwd m1, m2
punpcklwd m2, m3, m4
vpdpwssd m2, m16, m16
punpckhwd m3, m4
vpdpwssd m3, m1, m1
pshufb m16, m17, m7
paddw m0, m16
pshufb m17, m8
paddw m0, m17 ; h sum
punpcklwd m1, m16, m17
vpdpwssd m2, m1, m1 ; h sumsq
punpckhwd m16, m17
vpdpwssd m3, m16, m16
paddw m1, m0, [t1+r10*2+416*0]
paddd m16, m2, [t1+r10*2+416*2]
paddd m17, m3, [t1+r10*2+416*4]
test hd, hd
jz .hv_last_row
.hv_main2:
paddd m16, [t2+r10*2+416*2] ; hv sumsq
paddd m17, [t2+r10*2+416*4]
paddw m1, [t2+r10*2+416*0] ; hv sum
mova [t0+r10*2+416*2], m2
mova [t0+r10*2+416*4], m3
mova [t0+r10*2+416*0], m0
pmulld m16, m9 ; -a * 25
pmulld m17, m9
punpcklwd m0, m1, m4 ; b
vpdpwssd m16, m0, m0 ; -p
punpckhwd m1, m4
vpdpwssd m17, m1, m1
pmaddwd m0, m10 ; b * 164
pmaddwd m1, m10
pmulld m16, m11 ; p * s
pmulld m17, m11
vpalignr m17{k2}, m16, m16, 2
mova m16, m20
paddusw m17, m12
psraw m17, 4 ; min(z, 255) - 256
vpermt2b m16, m17, m21 ; sgr_x_by_x[128..255]
vpmovb2m k3, m17
vpermi2b m17, m18, m19 ; sgr_x_by_x[ 0..127]
vmovdqu8 m17{k3}, m16 ; x
pandn m16, m13, m17
psrld m17, 16
pmulld m0, m16
pmulld m1, m17
paddd m0, m14 ; x * b * 164 + (1 << 11) + (1 << 15)
paddd m1, m14
vpternlogd m16, m0, m13, 0xd8 ; a | (b << 12)
vpternlogd m17, m1, m13, 0xd8
mova [t3+r10*4+ 8], m16 ; The neighbor calculations requires
mova [t3+r10*4+ 24], xm17 ; 13 bits for a and 21 bits for b.
vextracti32x4 [t3+r10*4+ 56], m17, 2 ; Packing them allows for 12+20, but
mova [t3+r10*4+ 72], m17 ; that gets us most of the way.
vextracti128 [t3+r10*4+ 72], ym16, 1
vextracti32x4 [t3+r10*4+104], m16, 3
add r10, 32
jl .hv_loop
mov t2, t1
mov t1, t0
mov t0, t2
ret
.hv_last_row: ; esoteric edge case for odd heights
mova [t1+r10*2+416*0], m1
paddw m1, m0
mova [t1+r10*2+416*2], m16
paddd m16, m2
mova [t1+r10*2+416*4], m17
paddd m17, m3
jmp .hv_main2
.v: ; vertical boxsum + ab
lea r10, [wq-2]
.v_loop:
mova m2, [t1+r10*2+416*2]
paddd m16, m2, [t2+r10*2+416*2]
mova m3, [t1+r10*2+416*4]
paddd m17, m3, [t2+r10*2+416*4]
paddd m2, m2
paddd m3, m3
paddd m16, m2 ; hv sumsq
paddd m17, m3
pmulld m16, m9 ; -a * 25
pmulld m17, m9
mova m0, [t1+r10*2+416*0]
paddw m1, m0, [t2+r10*2+416*0]
paddw m0, m0
paddw m1, m0 ; hv sum
punpcklwd m0, m1, m4 ; b
vpdpwssd m16, m0, m0 ; -p
punpckhwd m1, m4
vpdpwssd m17, m1, m1
pmaddwd m0, m10 ; b * 164
pmaddwd m1, m10
pmulld m16, m11 ; p * s
pmulld m17, m11
vpalignr m17{k2}, m16, m16, 2
mova m16, m20
paddusw m17, m12
psraw m17, 4 ; min(z, 255) - 256
vpermt2b m16, m17, m21 ; sgr_x_by_x[128..255]
vpmovb2m k3, m17
vpermi2b m17, m18, m19 ; sgr_x_by_x[ 0..127]
vmovdqu8 m17{k3}, m16 ; x
pandn m16, m13, m17
psrld m17, 16
pmulld m0, m16
pmulld m1, m17
paddd m0, m14 ; x * b * 164 + (1 << 11) + (1 << 15)
paddd m1, m14
vpternlogd m16, m0, m13, 0xd8 ; a | (b << 12)
vpternlogd m17, m1, m13, 0xd8
mova [t3+r10*4+ 8], m16
mova [t3+r10*4+ 24], xm17
vextracti32x4 [t3+r10*4+ 56], m17, 2
mova [t3+r10*4+ 72], m17
vextracti128 [t3+r10*4+ 72], ym16, 1
vextracti32x4 [t3+r10*4+104], m16, 3
add r10, 32
jl .v_loop
ret
.prep_n: ; initial neighbor setup
mov r10, wq
.prep_n_loop:
movu m0, [t3+r10*4+ 4]
movu m1, [t3+r10*4+68]
paddd m2, m0, [t3+r10*4+ 0]
paddd m3, m1, [t3+r10*4+64]
paddd m2, [t3+r10*4+ 8]
paddd m3, [t3+r10*4+72]
paddd m0, m2
pslld m2, 2
paddd m1, m3
pslld m3, 2
paddd m2, m0 ; ab 565
paddd m3, m1
pandn m0, m13, m2 ; a
psrld m2, 12 ; b
pandn m1, m13, m3
psrld m3, 12
mova [t3+r10*4+416*4+ 0], m0
mova [t3+r10*4+416*8+ 0], m2
mova [t3+r10*4+416*4+64], m1
mova [t3+r10*4+416*8+64], m3
add r10, 32
jl .prep_n_loop
ret
ALIGN function_align
.n0: ; neighbor + output (even rows)
mov r10, wq
.n0_loop:
movu m16, [t3+r10*4+ 4]
movu m17, [t3+r10*4+68]
paddd m0, m16, [t3+r10*4+ 0]
paddd m1, m17, [t3+r10*4+64]
paddd m0, [t3+r10*4+ 8]
paddd m1, [t3+r10*4+72]
paddd m16, m0
pslld m0, 2
paddd m17, m1
pslld m1, 2
paddd m0, m16