1 ;*****************************************************************************
2 ;* MMX/SSE2/AVX-optimized H.264 deblocking code
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2011 x264 project
6 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
7 ;* Jason Garrett-Glaser <darkshikari@gmail.com>
8 ;* Oskar Arvidsson <oskar@irock.se>
10 ;* This file is part of FFmpeg.
12 ;* FFmpeg is free software; you can redistribute it and/or
13 ;* modify it under the terms of the GNU Lesser General Public
14 ;* License as published by the Free Software Foundation; either
15 ;* version 2.1 of the License, or (at your option) any later version.
17 ;* FFmpeg is distributed in the hope that it will be useful,
18 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
19 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 ;* Lesser General Public License for more details.
22 ;* You should have received a copy of the GNU Lesser General Public
23 ;* License along with FFmpeg; if not, write to the Free Software
24 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 ;******************************************************************************
27 %include "libavutil/x86/x86inc.asm"
28 %include "libavutil/x86/x86util.asm"
32 pb_3_1: times 4 db 3, 1
41 ; expands to [base],...,[base+7*stride]
42 %define PASS8ROWS(base, base3, stride, stride3) \
43 [base], [base+stride], [base+stride*2], [base3], \
44 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
46 %define PASS8ROWS(base, base3, stride, stride3, offset) \
47 PASS8ROWS(base+offset, base3+offset, stride, stride3)
49 ; in: 8 rows of 4 bytes in %4..%11
50 ; out: 4 rows of 8 bytes in m0..m3
51 %macro TRANSPOSE4x8_LOAD 11
78 ; in: 4 rows of 8 bytes in m0..m3
79 ; out: 8 rows of 4 bytes in %1..%8
80 %macro TRANSPOSE8x4B_STORE 8
109 %macro TRANSPOSE4x8B_LOAD 8
110 TRANSPOSE4x8_LOAD bw, wd, dq, %1, %2, %3, %4, %5, %6, %7, %8
118 ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8
119 ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16]
120 %macro TRANSPOSE6x8_MEM 9
129 SBUTTERFLY bw, 0, 1, 7
130 SBUTTERFLY bw, 2, 3, 7
131 SBUTTERFLY bw, 4, 5, 7
133 SBUTTERFLY3 bw, m6, %8, m7
134 SBUTTERFLY wd, 0, 2, 3
135 SBUTTERFLY wd, 4, 6, 3
138 SBUTTERFLY3 wd, m1, [%9+0x10], m3
139 SBUTTERFLY wd, 5, 7, 0
140 SBUTTERFLY dq, 1, 5, 0
141 SBUTTERFLY dq, 2, 6, 0
151 ; in: 8 rows of 8 in %1..%8
152 ; out: 8 rows of 8 in %9..%16
153 %macro TRANSPOSE8x8_MEM 16
162 SBUTTERFLY bw, 0, 1, 7
163 SBUTTERFLY bw, 2, 3, 7
164 SBUTTERFLY bw, 4, 5, 7
165 SBUTTERFLY3 bw, m6, %8, m7
167 SBUTTERFLY wd, 0, 2, 5
168 SBUTTERFLY wd, 4, 6, 5
169 SBUTTERFLY wd, 1, 3, 5
172 SBUTTERFLY wd, 6, 7, 5
173 SBUTTERFLY dq, 0, 4, 5
174 SBUTTERFLY dq, 1, 6, 5
179 SBUTTERFLY3 dq, m2, %11, m0
180 SBUTTERFLY dq, 3, 7, 4
188 ; out: %4 = |%1-%2|>%3
204 ; out: %4 = |%1-%2|>%3
221 ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1
222 ; out: m5=beta-1, m7=mask, %3=alpha-1
229 packuswb m4, m4 ; 16x alpha-1
230 packuswb m5, m5 ; 16x beta-1
234 DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1
235 DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1
237 DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1
243 ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask)
246 %macro DEBLOCK_P0_Q0 0
248 pxor m5, m1, m2 ; p0^q0
250 pand m5, [pb_1] ; (p0^q0)&1
251 pavgb m3, m0 ; (p1 - q1 + 256)>>1
253 pavgb m3, [pb_3] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2
254 pavgb m4, m2 ; (q0 - p0 + 256)>>1
257 paddusb m3, m4 ; d+128+33
269 ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp
270 ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
271 ; clobbers: q2, tmp, tc0
274 pavgb %2, %6 ; avg(p2,avg(p0,q0))
276 pand %6, [pb_1] ; (p2^avg(p0,q0))&1
277 psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1
286 ;-----------------------------------------------------------------------------
287 ; void deblock_v_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
288 ;-----------------------------------------------------------------------------
289 %macro DEBLOCK_LUMA 0
290 cglobal deblock_v_luma_8, 5,5,10
296 add r4, r0 ; pix-3*stride
298 mova m0, [r4+r1] ; p1
299 mova m1, [r4+2*r1] ; p0
301 mova m3, [r0+r1] ; q1
305 punpcklbw m8, m8 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
312 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
316 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
318 movdqa m4, [r0+2*r1] ; q2
319 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
324 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m8, m6
331 ;-----------------------------------------------------------------------------
332 ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
333 ;-----------------------------------------------------------------------------
335 cglobal deblock_h_luma_8, 5,9
342 %define pix_tmp rsp+0x30
348 ; transpose 6x16 -> tmp space
349 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp
352 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp+8
355 ; alpha, beta, tc0 are still in r2d, r3d, r4
356 ; don't backup r6, r5, r7, r8 because deblock_v_luma_sse2 doesn't use them
357 lea r0, [pix_tmp+0x30]
362 call deblock_v_luma_8
364 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
367 movq m0, [pix_tmp+0x18]
368 movq m1, [pix_tmp+0x28]
369 movq m2, [pix_tmp+0x38]
370 movq m3, [pix_tmp+0x48]
371 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
377 movq m0, [pix_tmp+0x10]
378 movq m1, [pix_tmp+0x20]
379 movq m2, [pix_tmp+0x30]
380 movq m3, [pix_tmp+0x40]
381 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
400 %macro DEBLOCK_LUMA 2
401 ;-----------------------------------------------------------------------------
402 ; void deblock_v8_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
403 ;-----------------------------------------------------------------------------
404 cglobal deblock_%1_luma_8, 5,5
409 add r4, r0 ; pix-3*stride
410 %assign pad 2*%2+12-(stack_offset&15)
413 mova m0, [r4+r1] ; p1
414 mova m1, [r4+2*r1] ; p0
416 mova m3, [r0+r1] ; q1
423 punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
424 mova [esp+%2], m4 ; tc
428 mova [esp], m4 ; mask
430 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
432 pand m4, [esp+%2] ; tc
435 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
437 mova m4, [r0+2*r1] ; q2
438 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
439 pand m6, [esp] ; mask
440 mova m5, [esp+%2] ; tc
444 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6
452 ;-----------------------------------------------------------------------------
453 ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
454 ;-----------------------------------------------------------------------------
456 cglobal deblock_h_luma_8, 0,5
462 %assign pad 0x78-(stack_offset&15)
464 %define pix_tmp esp+12
466 ; transpose 6x16 -> tmp space
467 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp
470 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8
473 lea r0, [pix_tmp+0x30]
479 call deblock_%1_luma_8
481 add dword [esp ], 8 ; pix_tmp+0x38
482 add dword [esp+16], 2 ; tc0+2
483 call deblock_%1_luma_8
487 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
491 movq m0, [pix_tmp+0x10]
492 movq m1, [pix_tmp+0x20]
494 movq m2, [pix_tmp+0x30]
495 movq m3, [pix_tmp+0x40]
496 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
500 movq m0, [pix_tmp+0x18]
501 movq m1, [pix_tmp+0x28]
502 movq m2, [pix_tmp+0x38]
503 movq m3, [pix_tmp+0x48]
504 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
508 %endmacro ; DEBLOCK_LUMA
523 %macro LUMA_INTRA_P012 4 ; p0..p3 in memory
533 pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2
551 psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4;
563 psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1
567 pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2
572 psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8
578 pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4
586 mova %1, t1 ; store p0
591 pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4
593 paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0
598 psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8
606 mova %2, t0 ; store p1
607 mova %3, t1 ; store p2
610 %macro LUMA_INTRA_SWAP_PQ 0
616 %define mask1p mask1q
619 %macro DEBLOCK_LUMA_INTRA 1
635 %define mask1q [rsp-24]
639 %define spill(x) [esp+16*x+((stack_offset+4)&15)]
644 %define mask0 spill(2)
645 %define mask1p spill(3)
646 %define mask1q spill(4)
651 ;-----------------------------------------------------------------------------
652 ; void deblock_v_luma_intra( uint8_t *pix, int stride, int alpha, int beta )
653 ;-----------------------------------------------------------------------------
654 cglobal deblock_%1_luma_intra_8, 4,6,16
659 lea r5, [r1*3] ; 3*stride
665 add r4, r0 ; pix-4*stride
673 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
674 SWAP 7, 12 ; m12=mask0
676 pavgb t5, mpb_1 ; alpha/4+1
679 DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1
680 DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1
681 DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1
688 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
692 pavgb m4, [pb_1] ; alpha/4+1
693 DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1
695 DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1
698 DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1
702 LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4]
704 LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
713 ;-----------------------------------------------------------------------------
714 ; void deblock_h_luma_intra( uint8_t *pix, int stride, int alpha, int beta )
715 ;-----------------------------------------------------------------------------
716 cglobal deblock_h_luma_intra_8, 4,9
724 ; transpose 8x16 -> tmp space
725 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
728 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
730 lea r0, [pix_tmp+0x40]
732 call deblock_v_luma_intra_8
734 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
736 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
741 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
745 cglobal deblock_h_luma_intra_8, 2,4
749 %assign pad 0x8c-(stack_offset&15)
753 ; transpose 8x16 -> tmp space
754 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
757 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
759 lea r0, [pix_tmp+0x40]
764 call deblock_%1_luma_intra_8
766 add dword [rsp], 8 ; pix_tmp+8
767 call deblock_%1_luma_intra_8
776 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
777 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
780 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
784 %endmacro ; DEBLOCK_LUMA_INTRA
794 DEBLOCK_LUMA_INTRA v8
799 %macro CHROMA_V_START 0
807 %macro CHROMA_H_START 0
819 ;-----------------------------------------------------------------------------
820 ; void ff_deblock_v_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
821 ;-----------------------------------------------------------------------------
822 cglobal deblock_v_chroma_8, 5,6
828 call ff_chroma_inter_body_mmx2
833 ;-----------------------------------------------------------------------------
834 ; void ff_deblock_h_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
835 ;-----------------------------------------------------------------------------
836 cglobal deblock_h_chroma_8, 5,7
838 %define buf0 [rsp-24]
839 %define buf1 [rsp-16]
849 TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6)
859 TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6)
866 ff_chroma_inter_body_mmx2:
876 ; in: %1=p0 %2=p1 %3=q1
877 ; out: p0 = (p0 + q1 + 2*p1 + 2) >> 2
878 %macro CHROMA_INTRA_P0 3
881 pand m4, [pb_1] ; m4 = (p0^q1)&1
884 pavgb %1, %2 ; dst = avg(p1, avg(p0,q1) - ((p0^q1)&1))
890 ;-----------------------------------------------------------------------------
891 ; void ff_deblock_v_chroma_intra( uint8_t *pix, int stride, int alpha, int beta )
892 ;-----------------------------------------------------------------------------
893 cglobal deblock_v_chroma_intra_8, 4,5
899 call ff_chroma_intra_body_mmx2
904 ;-----------------------------------------------------------------------------
905 ; void ff_deblock_h_chroma_intra( uint8_t *pix, int stride, int alpha, int beta )
906 ;-----------------------------------------------------------------------------
907 cglobal deblock_h_chroma_intra_8, 4,6
909 TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6)
910 call ff_chroma_intra_body_mmx2
911 TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6)
915 ff_chroma_intra_body_mmx2:
919 CHROMA_INTRA_P0 m1, m0, m3
920 CHROMA_INTRA_P0 m2, m3, m0
929 ;-----------------------------------------------------------------------------
930 ; void h264_loop_filter_strength(int16_t bs[2][4][4], uint8_t nnz[40],
931 ; int8_t ref[2][40], int16_t mv[2][40][2],
932 ; int bidir, int edges, int step,
933 ; int mask_mv0, int mask_mv1, int field);
941 ;-----------------------------------------------------------------------------
942 %macro loop_filter_strength_iteration 7 ; edges, step, mask_mv,
943 ; dir, d_idx, mask_dir, bidir
951 xor b_idxd, b_idxd ; for (b_idx = 0; b_idx < edges; b_idx += step)
956 test b_idxd, dword mask_mvd
957 jnz %%.skip_loop_iter ; if (!(b_idx & mask_mv))
959 movd m2, [refq+b_idxq+d_idx+12] ; { ref0[bn] }
960 punpckldq m2, [refq+b_idxq+d_idx+52] ; { ref0[bn], ref1[bn] }
961 pshufw m0, [refq+b_idxq+12], 0x44 ; { ref0[b], ref0[b] }
962 pshufw m1, [refq+b_idxq+52], 0x44 ; { ref1[b], ref1[b] }
963 pshufw m3, m2, 0x4E ; { ref1[bn], ref0[bn] }
964 psubb m0, m2 ; { ref0[b] != ref0[bn],
965 ; ref0[b] != ref1[bn] }
966 psubb m1, m3 ; { ref1[b] != ref1[bn],
967 ; ref1[b] != ref0[bn] }
970 mova m1, [mvq+b_idxq*4+(d_idx+12)*4]
971 mova m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
974 psubw m1, [mvq+b_idxq*4+12*4]
975 psubw m2, [mvq+b_idxq*4+12*4+mmsize]
976 psubw m3, [mvq+b_idxq*4+52*4]
977 psubw m4, [mvq+b_idxq*4+52*4+mmsize]
982 psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit
987 mova m1, [mvq+b_idxq*4+(d_idx+52)*4]
988 mova m2, [mvq+b_idxq*4+(d_idx+52)*4+mmsize]
991 psubw m1, [mvq+b_idxq*4+12*4]
992 psubw m2, [mvq+b_idxq*4+12*4+mmsize]
993 psubw m3, [mvq+b_idxq*4+52*4]
994 psubw m4, [mvq+b_idxq*4+52*4+mmsize]
999 psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit
1008 movd m0, [refq+b_idxq+12]
1009 psubb m0, [refq+b_idxq+d_idx+12] ; ref[b] != ref[bn]
1011 mova m1, [mvq+b_idxq*4+12*4]
1012 mova m2, [mvq+b_idxq*4+12*4+mmsize]
1013 psubw m1, [mvq+b_idxq*4+(d_idx+12)*4]
1014 psubw m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
1017 psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit
1020 %endif ; bidir == 1/0
1023 movd m1, [nnzq+b_idxq+12]
1024 por m1, [nnzq+b_idxq+d_idx+12] ; nnz[b] || nnz[bn]
1032 movq [bsq+b_idxq+32*dir], m1
1034 add b_idxd, dword stepd
1035 cmp b_idxd, dword edgesd
1040 cglobal h264_loop_filter_strength, 9, 9, 0, bs, nnz, ref, mv, bidir, edges, \
1041 step, mask_mv0, mask_mv1, field
1042 %define b_idxq bidirq
1043 %define b_idxd bidird
1056 %define mask_mv0d mask_mv0m
1057 %define mask_mv1d mask_mv1m
1059 shl dword mask_mv1d, 3
1060 shl dword mask_mv0d, 3
1064 loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 0
1065 loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 0
1067 mova m0, [bsq+mmsize*0]
1068 mova m1, [bsq+mmsize*1]
1069 mova m2, [bsq+mmsize*2]
1070 mova m3, [bsq+mmsize*3]
1071 TRANSPOSE4x4W 0, 1, 2, 3, 4
1072 mova [bsq+mmsize*0], m0
1073 mova [bsq+mmsize*1], m1
1074 mova [bsq+mmsize*2], m2
1075 mova [bsq+mmsize*3], m3
1079 loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 1
1080 loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 1
1082 mova m0, [bsq+mmsize*0]
1083 mova m1, [bsq+mmsize*1]
1084 mova m2, [bsq+mmsize*2]
1085 mova m3, [bsq+mmsize*3]
1086 TRANSPOSE4x4W 0, 1, 2, 3, 4
1087 mova [bsq+mmsize*0], m0
1088 mova [bsq+mmsize*1], m1
1089 mova [bsq+mmsize*2], m2
1090 mova [bsq+mmsize*3], m3