1 ;*****************************************************************************
2 ;* MMX/SSE2/AVX-optimized 10-bit H.264 deblocking code
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2011 x264 project
6 ;* Authors: Oskar Arvidsson <oskar@irock.se>
7 ;* Loren Merritt <lorenm@u.washington.edu>
8 ;* Fiona Glaser <fiona@x264.com>
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/x86util.asm"
35 %define pw_pixel_max pw_1023
37 ; out: %4 = |%1-%2|-%3
46 ; out: %4 = |%1-%2|<%3
52 psubw %5, %3 ; |%1-%2|-%3
53 pcmpgtw %4, %5 ; 0 > |%1-%2|-%3
71 pshuflw %1, %1, 01010000b
72 pshufd %1, %1, 01010000b
77 ; in: %1=p1, %2=p0, %3=q0, %4=q1
78 ; %5=alpha, %6=beta, %7-%9=tmp
81 ABS_SUB %2, %3, %5, %8, %7 ; |p0-q0| - alpha
82 ABS_SUB %1, %2, %6, %9, %7 ; |p1-p0| - beta
84 ABS_SUB %3, %4, %6, %9, %7 ; |q1-q0| - beta
90 ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp
92 %macro DEBLOCK_P0_Q0 7
101 mova %6, [pw_pixel_max]
110 ; in: %1=x2, %2=x1, %3=p0, %4=q0 %5=mask&tc, %6=tmp
112 pavgw %6, %3, %4 ; (p0+q0+1)>>1
122 %macro LUMA_DEBLOCK_ONE 3
123 DIFF_LT m5, %1, bm, m4, m6
130 LUMA_Q1 m5, %2, m1, m2, m4, m6
133 %macro LUMA_H_STORE 2
145 movhps [%1+r1*2-4], m2
147 movhps [%1+r1*4-4], m3
151 %macro DEBLOCK_LUMA 0
152 ;-----------------------------------------------------------------------------
153 ; void ff_deblock_v_luma_10(uint16_t *pix, int stride, int alpha, int beta,
155 ;-----------------------------------------------------------------------------
156 cglobal deblock_v_luma_10, 5,5,8*(mmsize/16)
157 %assign pad 5*mmsize+12-(stack_offset&15)
159 %define ms1 [rsp+mmsize]
160 %define ms2 [rsp+mmsize*2]
161 %define am [rsp+mmsize*3]
162 %define bm [rsp+mmsize*4]
166 LOAD_AB m4, m5, r2d, r3d
180 LOAD_MASK m0, m1, m2, m3, am, bm, m7, m4, m6
185 LUMA_DEBLOCK_ONE m1, m0, ms1
189 LUMA_DEBLOCK_ONE m2, m3, ms2
199 DEBLOCK_P0_Q0 m1, m2, m0, m3, m5, m7, m6
211 cglobal deblock_h_luma_10, 5,6,8*(mmsize/16)
212 %assign pad 7*mmsize+12-(stack_offset&15)
214 %define ms1 [rsp+mmsize]
215 %define ms2 [rsp+mmsize*2]
216 %define p1m [rsp+mmsize*3]
217 %define p2m [rsp+mmsize*4]
218 %define am [rsp+mmsize*5]
219 %define bm [rsp+mmsize*6]
223 LOAD_AB m4, m5, r2d, r3d
236 movq m2, [r0-8] ; y q2 q1 q0
243 TRANSPOSE4x4W 2, 5, 0, 1, 4
246 TRANSPOSE4x4W 2, 3, 6, 7, 4
248 movu m5, [r0-8] ; y q2 q1 q0 p0 p1 p2 x
252 TRANSPOSE4x4W 5, 0, 2, 3, 6
259 TRANSPOSE4x4W 4, 1, 3, 7, 6
264 SBUTTERFLY qdq, 0, 1, 7
265 SBUTTERFLY qdq, 2, 3, 7
269 LOAD_MASK m0, m1, m2, m3, am, bm, m7, m4, m6
273 LUMA_DEBLOCK_ONE m1, m0, ms1
277 LUMA_DEBLOCK_ONE m2, m3, ms2
287 DEBLOCK_P0_Q0 m1, m2, m0, m3, m5, m7, m6
290 TRANSPOSE4x4W 0, 1, 2, 3, 4
294 lea r0, [r0+r1*(mmsize/2)]
295 lea r2, [r2+r1*(mmsize/2)]
303 ; in: m0=p1, m1=p0, m2=q0, m3=q1, m8=p2, m9=q2
304 ; m12=alpha, m13=beta
305 ; out: m0=p1', m3=q1', m1=p0', m2=q0'
306 ; clobbers: m4, m5, m6, m7, m10, m11, m14
307 %macro DEBLOCK_LUMA_INTER_SSE2 0
308 LOAD_MASK m0, m1, m2, m3, m12, m13, m7, m4, m6
310 DIFF_LT m8, m1, m13, m10, m4
311 DIFF_LT m9, m2, m13, m11, m4
321 LUMA_Q1 m8, m0, m1, m2, m5, m4
325 LUMA_Q1 m9, m3, m1, m2, m5, m4
333 DEBLOCK_P0_Q0 m1, m2, m0, m3, m4, m5, m6
339 %macro DEBLOCK_LUMA_64 0
340 cglobal deblock_v_luma_10, 5,5,15
352 LOAD_AB m12, m13, r2d, r3d
365 DEBLOCK_LUMA_INTER_SSE2
377 cglobal deblock_h_luma_10, 5,7,15
380 LOAD_AB m12, m13, r2d, r3d
388 movu m8, [r0-8] ; y q2 q1 q0 p0 p1 p2 x
397 TRANSPOSE4x4W 8, 0, 2, 9, 10
398 TRANSPOSE4x4W 5, 1, 3, 7, 10
401 SBUTTERFLY qdq, 0, 1, 10
402 SBUTTERFLY qdq, 2, 3, 10
405 DEBLOCK_LUMA_INTER_SSE2
407 TRANSPOSE4x4W 0, 1, 2, 3, 4
419 %if HAVE_AVX_EXTERNAL
433 ; in: t0-t2: tmp registers
434 ; %1=p0 %2=p1 %3=p2 %4=p3 %5=q0 %6=q1 %7=mask0
435 ; %8=mask1p %9=2 %10=p0' %11=p1' %12=p2'
436 %macro LUMA_INTRA_P012 12 ; p0..p3 in memory
451 paddw t0, %9 ; (p2 + p1 + p0 + q0 + 2)
452 paddw t2, t0 ; (2*p3 + 3*p2 + p1 + p0 + q0 + 4)
471 psrlw t1, 2 ; (2*p1 + p0 + q1 + 2)/4
472 psrlw t0, 3 ; (p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4)>>3
484 %macro LUMA_INTRA_INIT 1
485 %xdefine pad %1*mmsize+((gprsize*3) % mmsize)-(stack_offset&15)
492 CAT_XDEFINE t, i, [rsp+mmsize*(i-4)]
498 ; in: %1-%3=tmp, %4=p2, %5=q2
499 %macro LUMA_INTRA_INTER 5
500 LOAD_AB t0, t1, r2d, r3d
502 LOAD_MASK m0, m1, m2, m3, %1, t1, t0, t2, t3
511 paddw t3, [pw_2] ; alpha/4+2
512 DIFF_LT m1, m2, t3, t2, t0 ; t2 = |p0-q0| < alpha/4+2
516 DIFF_LT t3, m2, t1, t2, t0 ; t2 = |q2-q0| < beta
520 DIFF_LT t3, m1, t1, t2, t0 ; t2 = |p2-p0| < beta
525 %macro LUMA_H_INTRA_LOAD 0
531 TRANSPOSE4x4W 4, 5, 0, 1, 2
539 TRANSPOSE4x4W 2, 3, 4, 5, 6
551 TRANSPOSE8x8W 4, 5, 0, 1, 2, 3, 6, 7, t4, t5
559 ; in: %1=q3 %2=q2' %3=q1' %4=q0' %5=p0' %6=p1' %7=p2' %8=p3 %9=tmp
560 %macro LUMA_H_INTRA_STORE 9
562 TRANSPOSE4x4W %1, %2, %3, %4, %9
565 movq [r0+r1*2-8], m%3
568 TRANSPOSE4x4W %5, %6, %7, %1, %9
574 TRANSPOSE2x4x4W %1, %2, %3, %4, %9
577 movq [r0+r1*2-8], m%3
580 movhps [r4+r1-8], m%2
581 movhps [r4+r1*2-8], m%3
582 movhps [r4+r5-8], m%4
588 TRANSPOSE2x4x4W %5, %6, %7, %1, %9
595 movhps [r4+r1*2], m%7
601 ;-----------------------------------------------------------------------------
602 ; void ff_deblock_v_luma_intra_10(uint16_t *pix, int stride, int alpha,
604 ;-----------------------------------------------------------------------------
605 %macro DEBLOCK_LUMA_INTRA_64 0
606 cglobal deblock_v_luma_intra_10, 4,7,16
619 lea r5, [r1*3] ; 3*stride
621 add r4, r0 ; pix-4*stride
626 LOAD_AB aa, bb, r2d, r3d
635 LOAD_MASK p1, p0, q0, q1, aa, bb, m3, t0, t1
638 paddw t2, m0 ; alpha/4+2
639 DIFF_LT p0, q0, t2, m6, t0 ; m6 = |p0-q0| < alpha/4+2
640 DIFF_LT p2, p0, bb, t1, t0 ; m7 = |p2-p0| < beta
641 DIFF_LT q2, q0, bb, m7, t0 ; t1 = |q2-q0| < beta
645 LUMA_INTRA_P012 p0, p1, p2, [r4], q0, q1, m3, m6, m0, [r4+r5], [r4+2*r1], [r4+r1]
646 LUMA_INTRA_P012 q0, q1, q2, [r0+r5], p0, p1, m3, m7, m0, [r0], [r0+r1], [r0+2*r1]
653 ;-----------------------------------------------------------------------------
654 ; void ff_deblock_h_luma_intra_10(uint16_t *pix, int stride, int alpha,
656 ;-----------------------------------------------------------------------------
657 cglobal deblock_h_luma_intra_10, 4,7,16
670 %assign pad 24-(stack_offset&15)
673 lea r5, [r1*3] ; 3*stride
674 add r4, r0 ; pix+4*stride
688 TRANSPOSE8x8W 5, 8, 9, 10, 11, 12, 13, 4, 1
690 LOAD_AB m1, m2, r2d, r3d
691 LOAD_MASK q1, q0, p0, p1, m1, m2, m3, t0, t1
693 paddw m1, m0 ; alpha/4+2
694 DIFF_LT p0, q0, m1, m6, t0 ; m6 = |p0-q0| < alpha/4+2
695 DIFF_LT q2, q0, m2, t1, t0 ; t1 = |q2-q0| < beta
696 DIFF_LT p0, p2, m2, m7, t0 ; m7 = |p2-p0| < beta
702 LUMA_INTRA_P012 q0, q1, q2, q3, p0, p1, m3, m6, m0, m5, m1, q2
703 LUMA_INTRA_P012 p0, p1, p2, p3, q0, q1, m3, m7, m0, p0, m6, p2
706 LUMA_H_INTRA_STORE 7, 8, 1, 5, 11, 6, 13, 4, 14
717 DEBLOCK_LUMA_INTRA_64
718 %if HAVE_AVX_EXTERNAL
720 DEBLOCK_LUMA_INTRA_64
725 %macro DEBLOCK_LUMA_INTRA 0
726 ;-----------------------------------------------------------------------------
727 ; void ff_deblock_v_luma_intra_10(uint16_t *pix, int stride, int alpha,
729 ;-----------------------------------------------------------------------------
730 cglobal deblock_v_luma_intra_10, 4,7,8*(mmsize/16)
740 mova m0, [r4+r1*2] ; p1
741 mova m1, [r4+r5] ; p0
743 mova m3, [r0+r1] ; q1
744 LUMA_INTRA_INTER t4, t5, t6, [r4+r1], [r0+r1*2]
745 LUMA_INTRA_P012 m1, m0, t3, [r4], m2, m3, t5, t4, [pw_2], [r4+r5], [r4+2*r1], [r4+r1]
746 mova t3, [r0+r1*2] ; q2
747 LUMA_INTRA_P012 m2, m3, t3, [r0+r5], m1, m0, t5, t6, [pw_2], [r0], [r0+r1], [r0+2*r1]
755 ;-----------------------------------------------------------------------------
756 ; void ff_deblock_h_luma_intra_10(uint16_t *pix, int stride, int alpha,
758 ;-----------------------------------------------------------------------------
759 cglobal deblock_h_luma_intra_10, 4,7,8*(mmsize/16)
766 lea r5, [r1*3] ; 3*stride
767 add r4, r0 ; pix+4*stride
774 LUMA_INTRA_INTER t8, t9, t10, t5, t6
776 LUMA_INTRA_P012 m1, m0, t3, t4, m2, m3, t9, t8, [pw_2], t8, t5, t11
778 LUMA_INTRA_P012 m2, m3, t3, t7, m1, m0, t9, t10, [pw_2], m4, t6, m5
786 LUMA_H_INTRA_STORE 2, 0, 1, 3, 4, 6, 5, t7, 7
788 lea r0, [r0+r1*(mmsize/2)]
792 lea r4, [r4+r1*(mmsize/2)]
807 %if HAVE_AVX_EXTERNAL
814 ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp
815 ; out: %1=p0', %2=q0'
816 %macro CHROMA_DEBLOCK_P0_Q0_INTRA 7
834 %macro CHROMA_V_LOAD 1
836 mova m1, [r0+r1] ; p0
838 mova m3, [%1+r1] ; q1
841 %macro CHROMA_V_STORE 0
846 ; expands to [base],...,[base+7*stride]
847 %define PASS8ROWS(base, base3, stride, stride3) \
848 [base], [base+stride], [base+stride*2], [base3], \
849 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
851 ; in: 8 rows of 4 words in %4..%11
852 ; out: 4 rows of 8 words in m0..m3
853 %macro TRANSPOSE4x8W_LOAD 8
874 punpckhqdq m1, m0, m4
876 punpckhqdq m3, m2, m6
880 ; in: 4 rows of 8 words in m0..m3
881 ; out: 8 rows of 4 words in %1..%8
882 %macro TRANSPOSE8x4W_STORE 8
883 TRANSPOSE4x4W 0, 1, 2, 3, 4
894 ; %1 = base + 3*stride
895 ; %2 = 3*stride (unused on mmx)
896 ; %3, %4 = place to store p1 and q1 values
897 %macro CHROMA_H_LOAD 4
900 movq m1, [pix_q + stride_q - 4]
901 movq m2, [pix_q + 2*stride_q - 4]
903 TRANSPOSE4x4W 0, 1, 2, 3, 4
905 TRANSPOSE4x8W_LOAD PASS8ROWS(pix_q-4, %1-4, stride_q, %2)
911 ; %1 = base + 3*stride
912 ; %2 = 3*stride (unused on mmx)
913 ; %3, %4 = place to load p1 and q1 values
914 %macro CHROMA_H_STORE 4
918 TRANSPOSE4x4W 0, 1, 2, 3, 4
920 movq [pix_q + stride_q - 4], m1
921 movq [pix_q + 2*stride_q - 4], m2
924 TRANSPOSE8x4W_STORE PASS8ROWS(pix_q-4, %1-4, stride_q, %2)
928 %macro CHROMA_V_LOAD_TC 2
935 %macro DEBLOCK_CHROMA 0
936 ;-----------------------------------------------------------------------------
937 ; void ff_deblock_v_chroma_10(uint16_t *pix, int stride, int alpha, int beta,
939 ;-----------------------------------------------------------------------------
940 cglobal deblock_v_chroma_10, 5,7-(mmsize/16),8*(mmsize/16)
951 LOAD_AB m4, m5, r2d, r3d
952 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
954 CHROMA_V_LOAD_TC m6, r4
958 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
971 ;-----------------------------------------------------------------------------
972 ; void ff_deblock_v_chroma_intra_10(uint16_t *pix, int stride, int alpha,
974 ;-----------------------------------------------------------------------------
975 cglobal deblock_v_chroma_intra_10, 4,6-(mmsize/16),8*(mmsize/16)
986 LOAD_AB m4, m5, r2d, r3d
987 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
988 CHROMA_DEBLOCK_P0_Q0_INTRA m1, m2, m0, m3, m7, m5, m6
1000 ;-----------------------------------------------------------------------------
1001 ; void ff_deblock_h_chroma_10(uint16_t *pix, int stride, int alpha, int beta,
1003 ;-----------------------------------------------------------------------------
1004 cglobal deblock_h_chroma_10, 5, 7, 8, 0-2*mmsize, pix_, stride_, alpha_, beta_, tc0_
1008 lea r6, [3*stride_q]
1015 CHROMA_H_LOAD r5, r6, [rsp], [rsp + mmsize]
1016 LOAD_AB m4, m5, alpha_d, beta_d
1017 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1019 CHROMA_V_LOAD_TC m6, tc0_q
1023 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
1024 CHROMA_H_STORE r5, r6, [rsp], [rsp + mmsize]
1027 lea pix_q, [pix_q + 4*stride_q]
1028 lea r5, [r5 + 4*stride_q]
1035 ;-----------------------------------------------------------------------------
1036 ; void ff_deblock_h_chroma422_10(uint16_t *pix, int stride, int alpha, int beta,
1038 ;-----------------------------------------------------------------------------
1039 cglobal deblock_h_chroma422_10, 5, 7, 8, 0-3*mmsize, pix_, stride_, alpha_, beta_, tc0_
1049 lea r6, [3*stride_q]
1055 CHROMA_H_LOAD r5, r6, [rsp + 1*mmsize], [rsp + 2*mmsize]
1056 LOAD_AB m4, m5, alpha_d, beta_d
1057 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1059 movd m6, [rsp + r4 + 8]
1065 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
1066 CHROMA_H_STORE r5, r6, [rsp + 1*mmsize], [rsp + 2*mmsize]
1068 lea pix_q, [pix_q + (mmsize/2)*stride_q]
1069 lea r5, [r5 + (mmsize/2)*stride_q]
1076 %if ARCH_X86_64 == 0
1082 %if HAVE_AVX_EXTERNAL