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 ;* Fiona Glaser <fiona@x264.com>
8 ;* Oskar Arvidsson <oskar@irock.se>
10 ;* This file is part of Libav.
12 ;* Libav 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 ;* Libav 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 Libav; 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/x86util.asm"
31 pb_A1: times 16 db 0xA1
32 pb_3_1: times 4 db 3, 1
40 ; expands to [base],...,[base+7*stride]
41 %define PASS8ROWS(base, base3, stride, stride3) \
42 [base], [base+stride], [base+stride*2], [base3], \
43 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
45 %define PASS8ROWS(base, base3, stride, stride3, offset) \
46 PASS8ROWS(base+offset, base3+offset, stride, stride3)
48 ; in: 8 rows of 4 bytes in %4..%11
49 ; out: 4 rows of 8 bytes in m0..m3
50 %macro TRANSPOSE4x8_LOAD 11
77 ; in: 4 rows of 8 bytes in m0..m3
78 ; out: 8 rows of 4 bytes in %1..%8
79 %macro TRANSPOSE8x4B_STORE 8
108 %macro TRANSPOSE4x8B_LOAD 8
109 TRANSPOSE4x8_LOAD bw, wd, dq, %1, %2, %3, %4, %5, %6, %7, %8
117 ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8
118 ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16]
119 %macro TRANSPOSE6x8_MEM 9
128 SBUTTERFLY bw, 0, 1, 7
129 SBUTTERFLY bw, 2, 3, 7
130 SBUTTERFLY bw, 4, 5, 7
132 SBUTTERFLY3 bw, m6, %8, m7
133 SBUTTERFLY wd, 0, 2, 3
134 SBUTTERFLY wd, 4, 6, 3
137 SBUTTERFLY3 wd, m1, [%9+0x10], m3
138 SBUTTERFLY wd, 5, 7, 0
139 SBUTTERFLY dq, 1, 5, 0
140 SBUTTERFLY dq, 2, 6, 0
150 ; in: 8 rows of 8 in %1..%8
151 ; out: 8 rows of 8 in %9..%16
152 %macro TRANSPOSE8x8_MEM 16
161 SBUTTERFLY bw, 0, 1, 7
162 SBUTTERFLY bw, 2, 3, 7
163 SBUTTERFLY bw, 4, 5, 7
164 SBUTTERFLY3 bw, m6, %8, m7
166 SBUTTERFLY wd, 0, 2, 5
167 SBUTTERFLY wd, 4, 6, 5
168 SBUTTERFLY wd, 1, 3, 5
171 SBUTTERFLY wd, 6, 7, 5
172 SBUTTERFLY dq, 0, 4, 5
173 SBUTTERFLY dq, 1, 6, 5
178 SBUTTERFLY3 dq, m2, %11, m0
179 SBUTTERFLY dq, 3, 7, 4
187 ; out: %4 = |%1-%2|>%3
203 ; out: %4 = |%1-%2|>%3
220 ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1
221 ; out: m5=beta-1, m7=mask, %3=alpha-1
228 packuswb m4, m4 ; 16x alpha-1
229 packuswb m5, m5 ; 16x beta-1
233 DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1
234 DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1
236 DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1
242 ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask)
245 %macro DEBLOCK_P0_Q0 0
247 pxor m5, m1, m2 ; p0^q0
249 pand m5, [pb_1] ; (p0^q0)&1
250 pavgb m3, m0 ; (p1 - q1 + 256)>>1
252 pavgb m3, [pb_3] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2
253 pavgb m4, m2 ; (q0 - p0 + 256)>>1
256 paddusb m3, m4 ; d+128+33
268 ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp
269 ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
270 ; clobbers: q2, tmp, tc0
273 pavgb %2, %6 ; avg(p2,avg(p0,q0))
275 pand %6, [pb_1] ; (p2^avg(p0,q0))&1
276 psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1
285 ;-----------------------------------------------------------------------------
286 ; void ff_deblock_v_luma(uint8_t *pix, int stride, int alpha, int beta,
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 ff_deblock_h_luma(uint8_t *pix, int stride, int alpha, int beta,
334 ;-----------------------------------------------------------------------------
336 cglobal deblock_h_luma_8, 5,9,0,0x60+16*WIN64
342 %define pix_tmp rsp+0x30 ; shadow space + r4
347 ; transpose 6x16 -> tmp space
348 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp
351 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r7, r8), pix_tmp+8
354 ; alpha, beta, tc0 are still in r2d, r3d, r4
355 ; don't backup r6, r5, r7, r8 because deblock_v_luma_sse2 doesn't use them
356 lea r0, [pix_tmp+0x30]
361 call deblock_v_luma_8
363 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
366 movq m0, [pix_tmp+0x18]
367 movq m1, [pix_tmp+0x28]
368 movq m2, [pix_tmp+0x38]
369 movq m3, [pix_tmp+0x48]
370 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
376 movq m0, [pix_tmp+0x10]
377 movq m1, [pix_tmp+0x20]
378 movq m2, [pix_tmp+0x30]
379 movq m3, [pix_tmp+0x40]
380 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
392 %macro DEBLOCK_LUMA 2
393 ;-----------------------------------------------------------------------------
394 ; void ff_deblock_v8_luma(uint8_t *pix, int stride, int alpha, int beta,
396 ;-----------------------------------------------------------------------------
397 cglobal deblock_%1_luma_8, 5,5,8,2*%2
402 add r4, r0 ; pix-3*stride
404 mova m0, [r4+r1] ; p1
405 mova m1, [r4+2*r1] ; p0
407 mova m3, [r0+r1] ; q1
414 punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
415 mova [esp+%2], m4 ; tc
419 mova [esp], m4 ; mask
421 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
423 pand m4, [esp+%2] ; tc
426 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
428 mova m4, [r0+2*r1] ; q2
429 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
430 pand m6, [esp] ; mask
431 mova m5, [esp+%2] ; tc
435 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6
442 ;-----------------------------------------------------------------------------
443 ; void ff_deblock_h_luma(uint8_t *pix, int stride, int alpha, int beta,
445 ;-----------------------------------------------------------------------------
447 cglobal deblock_h_luma_8, 0,5,8,0x60+12
453 %define pix_tmp esp+12
455 ; transpose 6x16 -> tmp space
456 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp
459 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8
462 lea r0, [pix_tmp+0x30]
468 call deblock_%1_luma_8
470 add dword [esp ], 8 ; pix_tmp+0x38
471 add dword [esp+16], 2 ; tc0+2
472 call deblock_%1_luma_8
476 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
480 movq m0, [pix_tmp+0x10]
481 movq m1, [pix_tmp+0x20]
483 movq m2, [pix_tmp+0x30]
484 movq m3, [pix_tmp+0x40]
485 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
489 movq m0, [pix_tmp+0x18]
490 movq m1, [pix_tmp+0x28]
491 movq m2, [pix_tmp+0x38]
492 movq m3, [pix_tmp+0x48]
493 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
496 %endmacro ; DEBLOCK_LUMA
509 %macro LUMA_INTRA_P012 4 ; p0..p3 in memory
519 pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2
537 psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4;
549 psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1
553 pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2
558 psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8
564 pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4
572 mova %1, t1 ; store p0
577 pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4
579 paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0
584 psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8
592 mova %2, t0 ; store p1
593 mova %3, t1 ; store p2
596 %macro LUMA_INTRA_SWAP_PQ 0
602 %define mask1p mask1q
605 %macro DEBLOCK_LUMA_INTRA 1
624 %define mask1q [rsp-24]
629 %define spill(x) [esp+16*x]
634 %define mask0 spill(2)
635 %define mask1p spill(3)
636 %define mask1q spill(4)
641 ;-----------------------------------------------------------------------------
642 ; void ff_deblock_v_luma_intra(uint8_t *pix, int stride, int alpha, int beta)
643 ;-----------------------------------------------------------------------------
645 cglobal deblock_%1_luma_intra_8, 4,6,16,0x10
647 cglobal deblock_%1_luma_intra_8, 4,6,16,ARCH_X86_64*0x50-0x50
650 lea r5, [r1*3] ; 3*stride
656 add r4, r0 ; pix-4*stride
664 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
665 SWAP 7, 12 ; m12=mask0
667 pavgb t5, mpb_1 ; alpha/4+1
670 DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1
671 DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1
672 DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1
679 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
683 pavgb m4, [pb_1] ; alpha/4+1
684 DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1
686 DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1
689 DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1
693 LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4]
695 LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
701 ;-----------------------------------------------------------------------------
702 ; void ff_deblock_h_luma_intra(uint8_t *pix, int stride, int alpha, int beta)
703 ;-----------------------------------------------------------------------------
704 cglobal deblock_h_luma_intra_8, 4,9,0,0x80
710 %define pix_tmp rsp+0x20 ; shadow space
715 ; transpose 8x16 -> tmp space
716 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
719 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r7, r8), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
721 lea r0, [pix_tmp+0x40]
723 call deblock_v_luma_intra_8
725 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
727 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
732 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
735 cglobal deblock_h_luma_intra_8, 2,4,8,0x80
741 ; transpose 8x16 -> tmp space
742 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
745 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
747 lea r0, [pix_tmp+0x40]
752 call deblock_%1_luma_intra_8
754 add dword [rsp], 8 ; pix_tmp+8
755 call deblock_%1_luma_intra_8
764 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
765 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
768 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
771 %endmacro ; DEBLOCK_LUMA_INTRA
779 DEBLOCK_LUMA_INTRA v8
784 %macro CHROMA_V_START 0
792 %macro CHROMA_H_START 0
804 ;-----------------------------------------------------------------------------
805 ; void ff_deblock_v_chroma(uint8_t *pix, int stride, int alpha, int beta,
807 ;-----------------------------------------------------------------------------
808 cglobal deblock_v_chroma_8, 5,6
814 call ff_chroma_inter_body_mmxext
819 ;-----------------------------------------------------------------------------
820 ; void ff_deblock_h_chroma(uint8_t *pix, int stride, int alpha, int beta,
822 ;-----------------------------------------------------------------------------
823 cglobal deblock_h_chroma_8, 5,7
825 ; This could use the red zone on 64 bit unix to avoid the stack pointer
826 ; readjustment, but valgrind assumes the red zone is clobbered on
827 ; function calls and returns.
836 TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6)
839 call ff_chroma_inter_body_mmxext
842 TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6)
849 ff_chroma_inter_body_mmxext:
859 ; in: %1=p0 %2=p1 %3=q1
860 ; out: p0 = (p0 + q1 + 2*p1 + 2) >> 2
861 %macro CHROMA_INTRA_P0 3
864 pand m4, [pb_1] ; m4 = (p0^q1)&1
867 pavgb %1, %2 ; dst = avg(p1, avg(p0,q1) - ((p0^q1)&1))
873 ;------------------------------------------------------------------------------
874 ; void ff_deblock_v_chroma_intra(uint8_t *pix, int stride, int alpha, int beta)
875 ;------------------------------------------------------------------------------
876 cglobal deblock_v_chroma_intra_8, 4,5
882 call ff_chroma_intra_body_mmxext
887 ;------------------------------------------------------------------------------
888 ; void ff_deblock_h_chroma_intra(uint8_t *pix, int stride, int alpha, int beta)
889 ;------------------------------------------------------------------------------
890 cglobal deblock_h_chroma_intra_8, 4,6
892 TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6)
893 call ff_chroma_intra_body_mmxext
894 TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6)
898 ff_chroma_intra_body_mmxext:
902 CHROMA_INTRA_P0 m1, m0, m3
903 CHROMA_INTRA_P0 m2, m3, m0
912 ;-----------------------------------------------------------------------------
913 ; void ff_h264_loop_filter_strength(int16_t bs[2][4][4], uint8_t nnz[40],
914 ; int8_t ref[2][40], int16_t mv[2][40][2],
915 ; int bidir, int edges, int step,
916 ; int mask_mv0, int mask_mv1, int field);
924 ;-----------------------------------------------------------------------------
925 %macro loop_filter_strength_iteration 7 ; edges, step, mask_mv,
926 ; dir, d_idx, mask_dir, bidir
934 xor b_idxd, b_idxd ; for (b_idx = 0; b_idx < edges; b_idx += step)
939 test b_idxd, dword mask_mvd
940 jnz %%.skip_loop_iter ; if (!(b_idx & mask_mv))
942 movd m2, [refq+b_idxq+d_idx+12] ; { ref0[bn] }
943 punpckldq m2, [refq+b_idxq+d_idx+52] ; { ref0[bn], ref1[bn] }
944 pshufw m0, [refq+b_idxq+12], 0x44 ; { ref0[b], ref0[b] }
945 pshufw m1, [refq+b_idxq+52], 0x44 ; { ref1[b], ref1[b] }
946 pshufw m3, m2, 0x4E ; { ref1[bn], ref0[bn] }
947 psubb m0, m2 ; { ref0[b] != ref0[bn],
948 ; ref0[b] != ref1[bn] }
949 psubb m1, m3 ; { ref1[b] != ref1[bn],
950 ; ref1[b] != ref0[bn] }
953 mova m1, [mvq+b_idxq*4+(d_idx+12)*4]
954 mova m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
957 psubw m1, [mvq+b_idxq*4+12*4]
958 psubw m2, [mvq+b_idxq*4+12*4+mmsize]
959 psubw m3, [mvq+b_idxq*4+52*4]
960 psubw m4, [mvq+b_idxq*4+52*4+mmsize]
965 psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit
970 mova m1, [mvq+b_idxq*4+(d_idx+52)*4]
971 mova m2, [mvq+b_idxq*4+(d_idx+52)*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
991 movd m0, [refq+b_idxq+12]
992 psubb m0, [refq+b_idxq+d_idx+12] ; ref[b] != ref[bn]
994 mova m1, [mvq+b_idxq*4+12*4]
995 mova m2, [mvq+b_idxq*4+12*4+mmsize]
996 psubw m1, [mvq+b_idxq*4+(d_idx+12)*4]
997 psubw m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
1000 psubusb m1, m5 ; abs(mv[b] - mv[bn]) >= limit
1003 %endif ; bidir == 1/0
1006 movd m1, [nnzq+b_idxq+12]
1007 por m1, [nnzq+b_idxq+d_idx+12] ; nnz[b] || nnz[bn]
1015 movq [bsq+b_idxq+32*dir], m1
1017 add b_idxd, dword stepd
1018 cmp b_idxd, dword edgesd
1023 cglobal h264_loop_filter_strength, 9, 9, 0, bs, nnz, ref, mv, bidir, edges, \
1024 step, mask_mv0, mask_mv1, field
1025 %define b_idxq bidirq
1026 %define b_idxd bidird
1039 %define mask_mv0d mask_mv0m
1040 %define mask_mv1d mask_mv1m
1042 shl dword mask_mv1d, 3
1043 shl dword mask_mv0d, 3
1047 loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 0
1048 loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 0
1050 mova m0, [bsq+mmsize*0]
1051 mova m1, [bsq+mmsize*1]
1052 mova m2, [bsq+mmsize*2]
1053 mova m3, [bsq+mmsize*3]
1054 TRANSPOSE4x4W 0, 1, 2, 3, 4
1055 mova [bsq+mmsize*0], m0
1056 mova [bsq+mmsize*1], m1
1057 mova [bsq+mmsize*2], m2
1058 mova [bsq+mmsize*3], m3
1062 loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8, 0, 1
1063 loop_filter_strength_iteration 32, 8, mask_mv0d, 0, -1, -1, 1
1065 mova m0, [bsq+mmsize*0]
1066 mova m1, [bsq+mmsize*1]
1067 mova m2, [bsq+mmsize*2]
1068 mova m3, [bsq+mmsize*3]
1069 TRANSPOSE4x4W 0, 1, 2, 3, 4
1070 mova [bsq+mmsize*0], m0
1071 mova [bsq+mmsize*1], m1
1072 mova [bsq+mmsize*2], m2
1073 mova [bsq+mmsize*3], m3