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 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 ;******************************************************************************
28 %include "x86util.asm"
37 ; expands to [base],...,[base+7*stride]
38 %define PASS8ROWS(base, base3, stride, stride3) \
39 [base], [base+stride], [base+stride*2], [base3], \
40 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
42 %define PASS8ROWS(base, base3, stride, stride3, offset) \
43 PASS8ROWS(base+offset, base3+offset, stride, stride3)
45 ; in: 8 rows of 4 bytes in %4..%11
46 ; out: 4 rows of 8 bytes in m0..m3
47 %macro TRANSPOSE4x8_LOAD 11
74 ; in: 4 rows of 8 bytes in m0..m3
75 ; out: 8 rows of 4 bytes in %1..%8
76 %macro TRANSPOSE8x4B_STORE 8
105 %macro TRANSPOSE4x8B_LOAD 8
106 TRANSPOSE4x8_LOAD bw, wd, dq, %1, %2, %3, %4, %5, %6, %7, %8
114 ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8
115 ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16]
116 %macro TRANSPOSE6x8_MEM 9
125 SBUTTERFLY bw, 0, 1, 7
126 SBUTTERFLY bw, 2, 3, 7
127 SBUTTERFLY bw, 4, 5, 7
129 SBUTTERFLY3 bw, m6, %8, m7
130 SBUTTERFLY wd, 0, 2, 3
131 SBUTTERFLY wd, 4, 6, 3
134 SBUTTERFLY3 wd, m1, [%9+0x10], m3
135 SBUTTERFLY wd, 5, 7, 0
136 SBUTTERFLY dq, 1, 5, 0
137 SBUTTERFLY dq, 2, 6, 0
147 ; in: 8 rows of 8 in %1..%8
148 ; out: 8 rows of 8 in %9..%16
149 %macro TRANSPOSE8x8_MEM 16
158 SBUTTERFLY bw, 0, 1, 7
159 SBUTTERFLY bw, 2, 3, 7
160 SBUTTERFLY bw, 4, 5, 7
161 SBUTTERFLY3 bw, m6, %8, m7
163 SBUTTERFLY wd, 0, 2, 5
164 SBUTTERFLY wd, 4, 6, 5
165 SBUTTERFLY wd, 1, 3, 5
168 SBUTTERFLY wd, 6, 7, 5
169 SBUTTERFLY dq, 0, 4, 5
170 SBUTTERFLY dq, 1, 6, 5
175 SBUTTERFLY3 dq, m2, %11, m0
176 SBUTTERFLY dq, 3, 7, 4
184 ; out: %4 = |%1-%2|>%3
200 ; out: %4 = |%1-%2|>%3
217 ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1
218 ; out: m5=beta-1, m7=mask, %3=alpha-1
225 packuswb m4, m4 ; 16x alpha-1
226 packuswb m5, m5 ; 16x beta-1
230 DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1
231 DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1
233 DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1
239 ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask)
242 %macro DEBLOCK_P0_Q0 0
243 pxor m5, m1, m2 ; p0^q0
244 pand m5, [pb_1] ; (p0^q0)&1
247 pavgb m3, m0 ; (p1 - q1 + 256)>>1
248 pavgb m3, [pb_3] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2
250 pavgb m4, m2 ; (q0 - p0 + 256)>>1
252 paddusb m3, m4 ; d+128+33
265 ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp
266 ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
267 ; clobbers: q2, tmp, tc0
270 pavgb %2, %6 ; avg(p2,avg(p0,q0))
272 pand %6, [pb_1] ; (p2^avg(p0,q0))&1
273 psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1
282 ;-----------------------------------------------------------------------------
283 ; void deblock_v_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
284 ;-----------------------------------------------------------------------------
285 %macro DEBLOCK_LUMA 1
286 cglobal deblock_v_luma_8_%1, 5,5,10
292 add r4, r0 ; pix-3*stride
294 mova m0, [r4+r1] ; p1
295 mova m1, [r4+2*r1] ; p0
297 mova m3, [r0+r1] ; q1
301 punpcklbw m8, m8 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
308 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
312 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
314 movdqa m4, [r0+2*r1] ; q2
315 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
320 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m8, m6
327 ;-----------------------------------------------------------------------------
328 ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
329 ;-----------------------------------------------------------------------------
331 cglobal deblock_h_luma_8_%1, 5,7
338 %define pix_tmp rsp+0x30
344 ; transpose 6x16 -> tmp space
345 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r10, r11), pix_tmp
348 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r10, r11), pix_tmp+8
351 ; alpha, beta, tc0 are still in r2d, r3d, r4
352 ; don't backup r6, r5, r10, r11 because deblock_v_luma_sse2 doesn't use them
353 lea r0, [pix_tmp+0x30]
358 call deblock_v_luma_8_%1
360 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
363 movq m0, [pix_tmp+0x18]
364 movq m1, [pix_tmp+0x28]
365 movq m2, [pix_tmp+0x38]
366 movq m3, [pix_tmp+0x48]
367 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r10, r11)
373 movq m0, [pix_tmp+0x10]
374 movq m1, [pix_tmp+0x20]
375 movq m2, [pix_tmp+0x30]
376 movq m3, [pix_tmp+0x40]
377 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r10, r11)
394 %macro DEBLOCK_LUMA 3
395 ;-----------------------------------------------------------------------------
396 ; void deblock_v8_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
397 ;-----------------------------------------------------------------------------
398 cglobal deblock_%2_luma_8_%1, 5,5
403 add r4, r0 ; pix-3*stride
404 %assign pad 2*%3+12-(stack_offset&15)
407 mova m0, [r4+r1] ; p1
408 mova m1, [r4+2*r1] ; p0
410 mova m3, [r0+r1] ; q1
416 punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
417 mova [esp+%3], m4 ; tc
421 mova [esp], m4 ; mask
424 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
426 pand m4, [esp+%3] ; tc
429 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
431 mova m4, [r0+2*r1] ; q2
432 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
433 mova m5, [esp] ; mask
435 mova m5, [esp+%3] ; tc
439 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6
447 ;-----------------------------------------------------------------------------
448 ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
449 ;-----------------------------------------------------------------------------
451 cglobal deblock_h_luma_8_%1, 0,5
457 %assign pad 0x78-(stack_offset&15)
459 %define pix_tmp esp+12
461 ; transpose 6x16 -> tmp space
462 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp
465 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8
468 lea r0, [pix_tmp+0x30]
474 call deblock_%2_luma_8_%1
476 add dword [esp ], 8 ; pix_tmp+0x38
477 add dword [esp+16], 2 ; tc0+2
478 call deblock_%2_luma_8_%1
482 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
487 movq m0, [pix_tmp+0x10]
488 movq m1, [pix_tmp+0x20]
489 movq m2, [pix_tmp+0x30]
490 movq m3, [pix_tmp+0x40]
491 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
495 movq m0, [pix_tmp+0x18]
496 movq m1, [pix_tmp+0x28]
497 movq m2, [pix_tmp+0x38]
498 movq m3, [pix_tmp+0x48]
499 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
503 %endmacro ; DEBLOCK_LUMA
506 DEBLOCK_LUMA mmxext, v8, 8
508 DEBLOCK_LUMA sse2, v, 16
510 DEBLOCK_LUMA avx, v, 16
516 %macro LUMA_INTRA_P012 4 ; p0..p3 in memory
526 pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2
544 psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4;
556 psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1
560 pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2
565 psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8
571 pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4
579 mova %1, t1 ; store p0
584 pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4
586 paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0
591 psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8
599 mova %2, t0 ; store p1
600 mova %3, t1 ; store p2
603 %macro LUMA_INTRA_SWAP_PQ 0
609 %define mask1p mask1q
612 %macro DEBLOCK_LUMA_INTRA 2
628 %define mask1q [rsp-24]
632 %define spill(x) [esp+16*x+((stack_offset+4)&15)]
637 %define mask0 spill(2)
638 %define mask1p spill(3)
639 %define mask1q spill(4)
644 ;-----------------------------------------------------------------------------
645 ; void deblock_v_luma_intra( uint8_t *pix, int stride, int alpha, int beta )
646 ;-----------------------------------------------------------------------------
647 cglobal deblock_%2_luma_intra_8_%1, 4,6,16
652 lea r5, [r1*3] ; 3*stride
658 add r4, r0 ; pix-4*stride
666 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
667 SWAP 7, 12 ; m12=mask0
669 pavgb t5, mpb_1 ; alpha/4+1
672 DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1
673 DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1
674 DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1
681 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
685 pavgb m4, [pb_1] ; alpha/4+1
686 DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1
688 DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1
691 DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1
695 LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4]
697 LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
706 ;-----------------------------------------------------------------------------
707 ; void deblock_h_luma_intra( uint8_t *pix, int stride, int alpha, int beta )
708 ;-----------------------------------------------------------------------------
709 cglobal deblock_h_luma_intra_8_%1, 4,7
717 ; transpose 8x16 -> tmp space
718 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r10, r11), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
721 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r10, r11), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
723 lea r0, [pix_tmp+0x40]
725 call deblock_v_luma_intra_8_%1
727 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
729 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r6, r5, r10, r11)
734 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r6, r5, r10, r11)
738 cglobal deblock_h_luma_intra_8_%1, 2,4
742 %assign pad 0x8c-(stack_offset&15)
746 ; transpose 8x16 -> tmp space
747 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
750 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
752 lea r0, [pix_tmp+0x40]
757 call deblock_%2_luma_intra_8_%1
759 add dword [rsp], 8 ; pix_tmp+8
760 call deblock_%2_luma_intra_8_%1
769 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
770 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
773 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
777 %endmacro ; DEBLOCK_LUMA_INTRA
780 DEBLOCK_LUMA_INTRA sse2, v
782 DEBLOCK_LUMA_INTRA avx , v
785 DEBLOCK_LUMA_INTRA mmxext, v8
790 %macro CHROMA_V_START 0
798 %macro CHROMA_H_START 0
810 ;-----------------------------------------------------------------------------
811 ; void ff_deblock_v_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
812 ;-----------------------------------------------------------------------------
813 cglobal deblock_v_chroma_8_mmxext, 5,6
819 call ff_chroma_inter_body_mmxext
824 ;-----------------------------------------------------------------------------
825 ; void ff_deblock_h_chroma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
826 ;-----------------------------------------------------------------------------
827 cglobal deblock_h_chroma_8_mmxext, 5,7
829 %define buf0 [rsp-24]
830 %define buf1 [rsp-16]
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)
846 ff_chroma_inter_body_mmxext:
856 ; in: %1=p0 %2=p1 %3=q1
857 ; out: p0 = (p0 + q1 + 2*p1 + 2) >> 2
858 %macro CHROMA_INTRA_P0 3
861 pand m4, [pb_1] ; m4 = (p0^q1)&1
864 pavgb %1, %2 ; dst = avg(p1, avg(p0,q1) - ((p0^q1)&1))
870 ;-----------------------------------------------------------------------------
871 ; void ff_deblock_v_chroma_intra( uint8_t *pix, int stride, int alpha, int beta )
872 ;-----------------------------------------------------------------------------
873 cglobal deblock_v_chroma_intra_8_mmxext, 4,5
879 call ff_chroma_intra_body_mmxext
884 ;-----------------------------------------------------------------------------
885 ; void ff_deblock_h_chroma_intra( uint8_t *pix, int stride, int alpha, int beta )
886 ;-----------------------------------------------------------------------------
887 cglobal deblock_h_chroma_intra_8_mmxext, 4,6
889 TRANSPOSE4x8_LOAD bw, wd, dq, PASS8ROWS(t5, r0, r1, t6)
890 call ff_chroma_intra_body_mmxext
891 TRANSPOSE8x4B_STORE PASS8ROWS(t5, r0, r1, t6)
895 ff_chroma_intra_body_mmxext:
899 CHROMA_INTRA_P0 m1, m0, m3
900 CHROMA_INTRA_P0 m2, m3, m0