1 ;*****************************************************************************
2 ;* deblock-a.asm: x86 deblocking
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2012 x264 project
6 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
7 ;* Fiona Glaser <fiona@x264.com>
8 ;* Oskar Arvidsson <oskar@irock.se>
10 ;* This program is free software; you can redistribute it and/or modify
11 ;* it under the terms of the GNU General Public License as published by
12 ;* the Free Software Foundation; either version 2 of the License, or
13 ;* (at your option) any later version.
15 ;* This program is distributed in the hope that it will be useful,
16 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
17 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 ;* GNU General Public License for more details.
20 ;* You should have received a copy of the GNU General Public License
21 ;* along with this program; if not, write to the Free Software
22 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
24 ;* This program is also available under a commercial proprietary license.
25 ;* For more information, contact us at licensing@x264.com.
26 ;*****************************************************************************
29 %include "x86util.asm"
33 transpose_shuf: db 0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15
47 ; out: %4 = |%1-%2|-%3
56 ; out: %4 = |%1-%2|<%3
62 psubw %5, %3 ; |%1-%2|-%3
63 pcmpgtw %4, %5 ; 0 > |%1-%2|-%3
87 ; in: %1=p1, %2=p0, %3=q0, %4=q1
88 ; %5=alpha, %6=beta, %7-%9=tmp
91 ABS_SUB %2, %3, %5, %8, %7 ; |p0-q0| - alpha
92 ABS_SUB %1, %2, %6, %9, %7 ; |p1-p0| - beta
94 ABS_SUB %3, %4, %6, %9, %7 ; |q1-q0| - beta
100 ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp
101 ; out: %1=p0', m2=q0'
102 %macro DEBLOCK_P0_Q0 7
111 mova %6, [pw_pixel_max]
120 ; in: %1=x2, %2=x1, %3=p0, %4=q0 %5=mask&tc, %6=tmp
122 pavgw %6, %3, %4 ; (p0+q0+1)>>1
132 %macro LUMA_DEBLOCK_ONE 3
133 DIFF_LT m5, %1, bm, m4, m6
140 LUMA_Q1 m5, %2, m1, m2, m4, m6
143 %macro LUMA_H_STORE 2
155 movhps [%1+r1*2-4], m2
157 movhps [%1+r1*4-4], m3
161 %macro DEBLOCK_LUMA 0
162 ;-----------------------------------------------------------------------------
163 ; void deblock_v_luma( uint16_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
164 ;-----------------------------------------------------------------------------
165 cglobal deblock_v_luma, 5,5,8
166 %assign pad 5*mmsize+12-(stack_offset&15)
168 %define ms1 [rsp+mmsize]
169 %define ms2 [rsp+mmsize*2]
170 %define am [rsp+mmsize*3]
171 %define bm [rsp+mmsize*4]
174 LOAD_AB m4, m5, r2, r3
188 LOAD_MASK m0, m1, m2, m3, am, bm, m7, m4, m6
193 LUMA_DEBLOCK_ONE m1, m0, ms1
197 LUMA_DEBLOCK_ONE m2, m3, ms2
207 DEBLOCK_P0_Q0 m1, m2, m0, m3, m5, m7, m6
219 cglobal deblock_h_luma, 5,6,8
220 %assign pad 7*mmsize+12-(stack_offset&15)
222 %define ms1 [rsp+mmsize]
223 %define ms2 [rsp+mmsize*2]
224 %define p1m [rsp+mmsize*3]
225 %define p2m [rsp+mmsize*4]
226 %define am [rsp+mmsize*5]
227 %define bm [rsp+mmsize*6]
230 LOAD_AB m4, m5, r2, r3
243 movq m2, [r0-8] ; y q2 q1 q0
250 TRANSPOSE4x4W 2, 5, 0, 1, 4
253 TRANSPOSE4x4W 2, 3, 6, 7, 4
255 movu m5, [r0-8] ; y q2 q1 q0 p0 p1 p2 x
259 TRANSPOSE4x4W 5, 0, 2, 3, 6
266 TRANSPOSE4x4W 4, 1, 3, 7, 6
271 SBUTTERFLY qdq, 0, 1, 7
272 SBUTTERFLY qdq, 2, 3, 7
276 LOAD_MASK m0, m1, m2, m3, am, bm, m7, m4, m6
280 LUMA_DEBLOCK_ONE m1, m0, ms1
284 LUMA_DEBLOCK_ONE m2, m3, ms2
294 DEBLOCK_P0_Q0 m1, m2, m0, m3, m5, m7, m6
297 TRANSPOSE4x4W 0, 1, 2, 3, 4
301 lea r0, [r0+r1*(mmsize/2)]
302 lea r2, [r2+r1*(mmsize/2)]
310 ; in: m0=p1, m1=p0, m2=q0, m3=q1, m8=p2, m9=q2
311 ; m12=alpha, m13=beta
312 ; out: m0=p1', m3=q1', m1=p0', m2=q0'
313 ; clobbers: m4, m5, m6, m7, m10, m11, m14
314 %macro DEBLOCK_LUMA_INTER_SSE2 0
315 LOAD_MASK m0, m1, m2, m3, m12, m13, m7, m4, m6
317 DIFF_LT m8, m1, m13, m10, m4
318 DIFF_LT m9, m2, m13, m11, m4
328 LUMA_Q1 m8, m0, m1, m2, m5, m4
332 LUMA_Q1 m9, m3, m1, m2, m5, m4
340 DEBLOCK_P0_Q0 m1, m2, m0, m3, m4, m5, m6
346 %macro DEBLOCK_LUMA_64 0
347 cglobal deblock_v_luma, 5,5,15
358 LOAD_AB m12, m13, r2, r3
371 DEBLOCK_LUMA_INTER_SSE2
383 cglobal deblock_h_luma, 5,7,15
385 LOAD_AB m12, m13, r2, r3
393 movu m8, [r0-8] ; y q2 q1 q0 p0 p1 p2 x
402 TRANSPOSE4x4W 8, 0, 2, 9, 10
403 TRANSPOSE4x4W 5, 1, 3, 7, 10
406 SBUTTERFLY qdq, 0, 1, 10
407 SBUTTERFLY qdq, 2, 3, 10
410 DEBLOCK_LUMA_INTER_SSE2
412 TRANSPOSE4x4W 0, 1, 2, 3, 4
436 ; in: t0-t2: tmp registers
437 ; %1=p0 %2=p1 %3=p2 %4=p3 %5=q0 %6=q1 %7=mask0
438 ; %8=mask1p %9=2 %10=p0' %11=p1' %12=p2'
439 %macro LUMA_INTRA_P012 12 ; p0..p3 in memory
454 paddw t0, %9 ; (p2 + p1 + p0 + q0 + 2)
455 paddw t2, t0 ; (2*p3 + 3*p2 + p1 + p0 + q0 + 4)
474 psrlw t1, 2 ; (2*p1 + p0 + q1 + 2)/4
475 psrlw t0, 3 ; (p2 + 2*p1 + 2*p0 + 2*q0 + q1 + 4)>>3
487 %macro LUMA_INTRA_INIT 1
488 %xdefine pad %1*mmsize+((gprsize*3) % mmsize)-(stack_offset&15)
495 CAT_XDEFINE t, i, [rsp+mmsize*(i-4)]
502 ; in: %1-%3=tmp, %4=p2, %5=q2
503 %macro LUMA_INTRA_INTER 5
504 LOAD_AB t0, t1, r2d, r3d
506 LOAD_MASK m0, m1, m2, m3, %1, t1, t0, t2, t3
515 paddw t3, [pw_2] ; alpha/4+2
516 DIFF_LT m1, m2, t3, t2, t0 ; t2 = |p0-q0| < alpha/4+2
520 DIFF_LT t3, m2, t1, t2, t0 ; t2 = |q2-q0| < beta
524 DIFF_LT t3, m1, t1, t2, t0 ; t2 = |p2-p0| < beta
529 %macro LUMA_H_INTRA_LOAD 0
535 TRANSPOSE4x4W 4, 5, 0, 1, 2
543 TRANSPOSE4x4W 2, 3, 4, 5, 6
555 TRANSPOSE8x8W 4, 5, 0, 1, 2, 3, 6, 7, t4, t5
563 ; in: %1=q3 %2=q2' %3=q1' %4=q0' %5=p0' %6=p1' %7=p2' %8=p3 %9=tmp
564 %macro LUMA_H_INTRA_STORE 9
566 TRANSPOSE4x4W %1, %2, %3, %4, %9
569 movq [r0+r1*2-8], m%3
572 TRANSPOSE4x4W %5, %6, %7, %1, %9
578 TRANSPOSE2x4x4W %1, %2, %3, %4, %9
581 movq [r0+r1*2-8], m%3
584 movhps [r4+r1-8], m%2
585 movhps [r4+r1*2-8], m%3
586 movhps [r4+r5-8], m%4
592 TRANSPOSE2x4x4W %5, %6, %7, %1, %9
599 movhps [r4+r1*2], m%7
605 ;-----------------------------------------------------------------------------
606 ; void deblock_v_luma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
607 ;-----------------------------------------------------------------------------
608 %macro DEBLOCK_LUMA_INTRA_64 0
609 cglobal deblock_v_luma_intra, 4,7,16
623 lea r5, [r1*3] ; 3*stride
625 add r4, r0 ; pix-4*stride
628 LOAD_AB aa, bb, r2d, r3d
637 LOAD_MASK p1, p0, q0, q1, aa, bb, m3, t0, t1
640 paddw t2, m0 ; alpha/4+2
641 DIFF_LT p0, q0, t2, m6, t0 ; m6 = |p0-q0| < alpha/4+2
642 DIFF_LT p2, p0, bb, t1, t0 ; m7 = |p2-p0| < beta
643 DIFF_LT q2, q0, bb, m7, t0 ; t1 = |q2-q0| < beta
647 LUMA_INTRA_P012 p0, p1, p2, [r4], q0, q1, m3, m6, m0, [r4+r5], [r4+2*r1], [r4+r1]
648 LUMA_INTRA_P012 q0, q1, q2, [r0+r5], p0, p1, m3, m7, m0, [r0], [r0+r1], [r0+2*r1]
655 ;-----------------------------------------------------------------------------
656 ; void deblock_h_luma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
657 ;-----------------------------------------------------------------------------
658 cglobal deblock_h_luma_intra, 4,7,16
671 %assign pad 24-(stack_offset&15)
675 lea r5, [r1*3] ; 3*stride
676 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
719 DEBLOCK_LUMA_INTRA_64
723 %macro DEBLOCK_LUMA_INTRA 0
724 ;-----------------------------------------------------------------------------
725 ; void deblock_v_luma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
726 ;-----------------------------------------------------------------------------
727 cglobal deblock_v_luma_intra, 4,7,8
735 mova m0, [r4+r1*2] ; p1
736 mova m1, [r4+r5] ; p0
738 mova m3, [r0+r1] ; q1
739 LUMA_INTRA_INTER t4, t5, t6, [r4+r1], [r0+r1*2]
740 LUMA_INTRA_P012 m1, m0, t3, [r4], m2, m3, t5, t4, [pw_2], [r4+r5], [r4+2*r1], [r4+r1]
741 mova t3, [r0+r1*2] ; q2
742 LUMA_INTRA_P012 m2, m3, t3, [r0+r5], m1, m0, t5, t6, [pw_2], [r0], [r0+r1], [r0+2*r1]
750 ;-----------------------------------------------------------------------------
751 ; void deblock_h_luma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
752 ;-----------------------------------------------------------------------------
753 cglobal deblock_h_luma_intra, 4,7,8
760 lea r5, [r1*3] ; 3*stride
761 add r4, r0 ; pix+4*stride
766 LUMA_INTRA_INTER t8, t9, t10, t5, t6
768 LUMA_INTRA_P012 m1, m0, t3, t4, m2, m3, t9, t8, [pw_2], t8, t5, t11
770 LUMA_INTRA_P012 m2, m3, t3, t7, m1, m0, t9, t10, [pw_2], m4, t6, m5
778 LUMA_H_INTRA_STORE 2, 0, 1, 3, 4, 6, 5, t7, 7
780 lea r0, [r0+r1*(mmsize/2)]
784 lea r4, [r4+r1*(mmsize/2)]
803 %endif ; HIGH_BIT_DEPTH
805 %if HIGH_BIT_DEPTH == 0
806 ; expands to [base],...,[base+7*stride]
807 %define PASS8ROWS(base, base3, stride, stride3) \
808 [base], [base+stride], [base+stride*2], [base3], \
809 [base3+stride], [base3+stride*2], [base3+stride3], [base3+stride*4]
811 %define PASS8ROWS(base, base3, stride, stride3, offset) \
812 PASS8ROWS(base+offset, base3+offset, stride, stride3)
814 ; in: 8 rows of 4 bytes in %4..%11
815 ; out: 4 rows of 8 bytes in m0..m3
816 %macro TRANSPOSE4x8_LOAD 11
843 ; in: 4 rows of 8 bytes in m0..m3
844 ; out: 8 rows of 4 bytes in %1..%8
845 %macro TRANSPOSE8x4B_STORE 8
874 %macro TRANSPOSE4x8B_LOAD 8
875 TRANSPOSE4x8_LOAD bw, wd, dq, %1, %2, %3, %4, %5, %6, %7, %8
878 %macro TRANSPOSE4x8W_LOAD 8
880 TRANSPOSE4x8_LOAD wd, dq, qdq, %1, %2, %3, %4, %5, %6, %7, %8
887 TRANSPOSE4x4W 0, 1, 2, 3, 4
891 %macro TRANSPOSE8x2W_STORE 8
924 ; in: 8 rows of 8 (only the middle 6 pels are used) in %1..%8
925 ; out: 6 rows of 8 in [%9+0*16] .. [%9+5*16]
926 %macro TRANSPOSE6x8_MEM 9
935 SBUTTERFLY bw, 0, 1, 7
936 SBUTTERFLY bw, 2, 3, 7
937 SBUTTERFLY bw, 4, 5, 7
939 SBUTTERFLY3 bw, m6, %8, m7
940 SBUTTERFLY wd, 0, 2, 3
941 SBUTTERFLY wd, 4, 6, 3
944 SBUTTERFLY3 wd, m1, [%9+0x10], m3
945 SBUTTERFLY wd, 5, 7, 0
946 SBUTTERFLY dq, 1, 5, 0
947 SBUTTERFLY dq, 2, 6, 0
957 ; in: 8 rows of 8 in %1..%8
958 ; out: 8 rows of 8 in %9..%16
959 %macro TRANSPOSE8x8_MEM 16
968 SBUTTERFLY bw, 0, 1, 7
969 SBUTTERFLY bw, 2, 3, 7
970 SBUTTERFLY bw, 4, 5, 7
971 SBUTTERFLY3 bw, m6, %8, m7
973 SBUTTERFLY wd, 0, 2, 5
974 SBUTTERFLY wd, 4, 6, 5
975 SBUTTERFLY wd, 1, 3, 5
978 SBUTTERFLY wd, 6, 7, 5
979 SBUTTERFLY dq, 0, 4, 5
980 SBUTTERFLY dq, 1, 6, 5
985 SBUTTERFLY3 dq, m2, %11, m0
986 SBUTTERFLY dq, 3, 7, 4
994 ; out: %4 = |%1-%2|>%3
1010 ; out: %4 = |%1-%2|>%3
1027 ; in: m0=p1 m1=p0 m2=q0 m3=q1 %1=alpha-1 %2=beta-1
1028 ; out: m5=beta-1, m7=mask, %3=alpha-1
1030 %macro LOAD_MASK 2-3
1035 packuswb m4, m4 ; 16x alpha-1
1036 packuswb m5, m5 ; 16x beta-1
1040 DIFF_GT m1, m2, m4, m7, m6 ; |p0-q0| > alpha-1
1041 DIFF_GT m0, m1, m5, m4, m6 ; |p1-p0| > beta-1
1043 DIFF_GT m3, m2, m5, m4, m6 ; |q1-q0| > beta-1
1049 ; in: m0=p1 m1=p0 m2=q0 m3=q1 m7=(tc&mask)
1050 ; out: m1=p0' m2=q0'
1052 %macro DEBLOCK_P0_Q0 0
1053 pxor m5, m1, m2 ; p0^q0
1054 pand m5, [pb_1] ; (p0^q0)&1
1057 pavgb m3, m0 ; (p1 - q1 + 256)>>1
1058 pavgb m3, [pb_3] ; (((p1 - q1 + 256)>>1)+4)>>1 = 64+2+(p1-q1)>>2
1060 pavgb m4, m2 ; (q0 - p0 + 256)>>1
1062 paddusb m3, m4 ; d+128+33
1075 ; %1=p1 %2=q2 %3=[q2] %4=[q1] %5=tc0 %6=tmp
1076 ; out: [q1] = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
1077 ; clobbers: q2, tmp, tc0
1080 pavgb %2, %6 ; avg(p2,avg(p0,q0))
1082 pand %6, [pb_1] ; (p2^avg(p0,q0))&1
1083 psubusb %2, %6 ; (p2+((p0+q0+1)>>1))>>1
1092 ;-----------------------------------------------------------------------------
1093 ; void deblock_v_luma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1094 ;-----------------------------------------------------------------------------
1095 %macro DEBLOCK_LUMA 0
1096 cglobal deblock_v_luma, 5,5,10
1102 add r4, r0 ; pix-3*stride
1104 mova m0, [r4+r1] ; p1
1105 mova m1, [r4+2*r1] ; p0
1107 mova m3, [r0+r1] ; q1
1111 punpcklbw m8, m8 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
1117 movdqa m3, [r4] ; p2
1118 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
1122 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
1124 movdqa m4, [r0+2*r1] ; q2
1125 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
1130 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m8, m6
1137 ;-----------------------------------------------------------------------------
1138 ; void deblock_h_luma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1139 ;-----------------------------------------------------------------------------
1141 cglobal deblock_h_luma, 5,9
1147 %define pix_tmp rsp+0x30
1153 ; transpose 6x16 -> tmp space
1154 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r1, r8), pix_tmp
1157 TRANSPOSE6x8_MEM PASS8ROWS(r6, r5, r1, r8), pix_tmp+8
1160 ; alpha, beta, tc0 are still in r2d, r3d, r4
1161 ; don't backup r6, r5, r7, r8 because deblock_v_luma_sse2 doesn't use them
1163 lea r0, [pix_tmp+0x30]
1170 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
1173 movq m0, [pix_tmp+0x18]
1174 movq m1, [pix_tmp+0x28]
1175 movq m2, [pix_tmp+0x38]
1176 movq m3, [pix_tmp+0x48]
1177 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
1183 movq m0, [pix_tmp+0x10]
1184 movq m1, [pix_tmp+0x20]
1185 movq m2, [pix_tmp+0x30]
1186 movq m3, [pix_tmp+0x40]
1187 TRANSPOSE8x4B_STORE PASS8ROWS(r6, r5, r7, r8)
1204 %macro DEBLOCK_LUMA 2
1205 ;-----------------------------------------------------------------------------
1206 ; void deblock_v8_luma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1207 ;-----------------------------------------------------------------------------
1208 cglobal deblock_%1_luma, 5,5,8,2*%2
1213 add r4, r0 ; pix-3*stride
1215 mova m0, [r4+r1] ; p1
1216 mova m1, [r4+2*r1] ; p0
1218 mova m3, [r0+r1] ; q1
1224 punpcklbw m4, m4 ; tc = 4x tc0[3], 4x tc0[2], 4x tc0[1], 4x tc0[0]
1225 mova [esp+%2], m4 ; tc
1229 mova [esp], m4 ; mask
1232 DIFF_GT2 m1, m3, m5, m6, m7 ; |p2-p0| > beta-1
1234 pand m4, [esp+%2] ; tc
1237 LUMA_Q1 m0, m3, [r4], [r4+r1], m6, m4
1239 mova m4, [r0+2*r1] ; q2
1240 DIFF_GT2 m2, m4, m5, m6, m3 ; |q2-q0| > beta-1
1241 mova m5, [esp] ; mask
1243 mova m5, [esp+%2] ; tc
1247 LUMA_Q1 m3, m4, [r0+2*r1], [r0+r1], m5, m6
1254 ;-----------------------------------------------------------------------------
1255 ; void deblock_h_luma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1256 ;-----------------------------------------------------------------------------
1258 cglobal deblock_h_luma, 0,5,8,0x60+HAVE_ALIGNED_STACK*12
1264 %define pix_tmp esp+12*HAVE_ALIGNED_STACK
1266 ; transpose 6x16 -> tmp space
1267 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp
1270 TRANSPOSE6x8_MEM PASS8ROWS(r0, r1, r3, r4), pix_tmp+8
1273 lea r0, [pix_tmp+0x30]
1279 call deblock_%1_luma
1281 add dword [esp ], 8 ; pix_tmp+0x38
1282 add dword [esp+16], 2 ; tc0+2
1283 call deblock_%1_luma
1287 ; transpose 16x4 -> original space (only the middle 4 rows were changed by the filter)
1292 movq m0, [pix_tmp+0x10]
1293 movq m1, [pix_tmp+0x20]
1294 movq m2, [pix_tmp+0x30]
1295 movq m3, [pix_tmp+0x40]
1296 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
1300 movq m0, [pix_tmp+0x18]
1301 movq m1, [pix_tmp+0x28]
1302 movq m2, [pix_tmp+0x38]
1303 movq m3, [pix_tmp+0x48]
1304 TRANSPOSE8x4B_STORE PASS8ROWS(r0, r1, r3, r4)
1307 %endmacro ; DEBLOCK_LUMA
1320 %macro LUMA_INTRA_P012 4 ; p0..p3 in memory
1330 pavgb t0, t1 ; ((p2+p1+1)/2 + (p0+q0+1)/2 + 1)/2
1348 psubb t0, t2 ; p1' = (p2+p1+p0+q0+2)/4;
1360 psubb t3, t2 ; p2+2*p1+2*p0+2*q0+q1
1364 pavgb t1, t5 ; (((p2+q1)/2 + p1+1)/2 + (p0+q0+1)/2 + 1)/2
1369 psubb t1, t3 ; p0'a = (p2+2*p1+2*p0+2*q0+q1+4)/8
1375 pavgb t2, p1 ; p0'b = (2*p1+p0+q0+2)/4
1383 mova %1, t1 ; store p0
1388 pavgb t1, t0 ; (p3+p2+1)/2 + (p2+p1+p0+q0+2)/4
1390 paddb t2, t4 ; 2*p3+3*p2+p1+p0+q0
1395 psubb t1, t2 ; p2' = (2*p3+3*p2+p1+p0+q0+4)/8
1403 mova %2, t0 ; store p1
1404 mova %3, t1 ; store p2
1407 %macro LUMA_INTRA_SWAP_PQ 0
1413 %define mask1p mask1q
1416 %macro DEBLOCK_LUMA_INTRA 1
1432 %define mask1q [rsp-24]
1436 %define spill(x) [esp+16*x]
1438 %define q2 [r0+2*r1]
1441 %define mask0 spill(2)
1442 %define mask1p spill(3)
1443 %define mask1q spill(4)
1444 %define mpb_0 [pb_0]
1445 %define mpb_1 [pb_1]
1448 ;-----------------------------------------------------------------------------
1449 ; void deblock_v_luma_intra( uint8_t *pix, intptr_t stride, int alpha, int beta )
1450 ;-----------------------------------------------------------------------------
1451 cglobal deblock_%1_luma_intra, 4,6,16,ARCH_X86_64*0x50-0x50
1453 lea r5, [r1*3] ; 3*stride
1459 add r4, r0 ; pix-4*stride
1467 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
1468 SWAP 7, 12 ; m12=mask0
1470 pavgb t5, mpb_1 ; alpha/4+1
1472 movdqa q2, [r0+2*r1]
1473 DIFF_GT2 p0, q0, t5, t0, t3 ; t0 = |p0-q0| > alpha/4+1
1474 DIFF_GT2 p0, p2, m5, t2, t5 ; mask1 = |p2-p0| > beta-1
1475 DIFF_GT2 q0, q2, m5, t4, t5 ; t4 = |q2-q0| > beta-1
1482 LOAD_MASK r2d, r3d, t5 ; m5=beta-1, t5=alpha-1, m7=mask0
1486 pavgb m4, [pb_1] ; alpha/4+1
1487 DIFF_GT2 p0, q0, m4, m6, m7 ; m6 = |p0-q0| > alpha/4+1
1489 DIFF_GT2 p0, p2, m5, m4, m7 ; m4 = |p2-p0| > beta-1
1492 DIFF_GT2 q0, q2, m5, m4, m7 ; m4 = |q2-q0| > beta-1
1496 LUMA_INTRA_P012 [r4+r5], [r4+2*r1], [r4+r1], [r4]
1498 LUMA_INTRA_P012 [r0], [r0+r1], [r0+2*r1], [r0+r5]
1504 ;-----------------------------------------------------------------------------
1505 ; void deblock_h_luma_intra( uint8_t *pix, intptr_t stride, int alpha, int beta )
1506 ;-----------------------------------------------------------------------------
1507 cglobal deblock_h_luma_intra, 4,9
1514 ; transpose 8x16 -> tmp space
1515 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r1, r8), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
1518 TRANSPOSE8x8_MEM PASS8ROWS(r6, r5, r1, r8), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
1521 lea r0, [pix_tmp+0x40]
1523 call deblock_v_luma_intra
1525 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
1527 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
1532 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r6, r5, r7, r8)
1536 cglobal deblock_h_luma_intra, 2,4,8,0x80
1542 ; transpose 8x16 -> tmp space
1543 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30)
1546 TRANSPOSE8x8_MEM PASS8ROWS(r0, r2, r1, r3), PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30)
1548 lea r0, [pix_tmp+0x40]
1553 call deblock_%1_luma_intra
1555 add dword [rsp], 8 ; pix_tmp+8
1556 call deblock_%1_luma_intra
1565 ; transpose 16x6 -> original space (but we can't write only 6 pixels, so really 16x8)
1566 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp, pix_tmp+0x30, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
1569 TRANSPOSE8x8_MEM PASS8ROWS(pix_tmp+8, pix_tmp+0x38, 0x10, 0x30), PASS8ROWS(r0, r2, r1, r3)
1571 %endif ; ARCH_X86_64
1572 %endmacro ; DEBLOCK_LUMA_INTRA
1575 DEBLOCK_LUMA_INTRA v
1577 DEBLOCK_LUMA_INTRA v
1578 %if ARCH_X86_64 == 0
1580 DEBLOCK_LUMA_INTRA v8
1582 %endif ; !HIGH_BIT_DEPTH
1585 ; in: %1=p0, %2=q0, %3=p1, %4=q1, %5=mask, %6=tmp, %7=tmp
1586 ; out: %1=p0', %2=q0'
1587 %macro CHROMA_DEBLOCK_P0_Q0_INTRA 7
1607 %macro CHROMA_H_LOAD 0-1
1608 movq m0, [r0-8] ; p1 p1 p0 p0
1609 movq m2, [r0] ; q0 q0 q1 q1
1615 punpckldq m0, m5 ; p1
1616 punpckhdq m1, m5 ; p0
1617 punpckldq m2, m7 ; q0
1618 punpckhdq m3, m7 ; q1
1620 movq m4, [r0+r1*2-8]
1624 punpckldq m0, m5 ; p1 ... p0 ...
1625 punpckldq m2, m7 ; q0 ... q1 ...
1628 punpckhqdq m1, m0, m4 ; p0
1629 punpcklqdq m0, m4 ; p1
1630 punpckhqdq m3, m2, m6 ; q1
1631 punpcklqdq m2, m6 ; q0
1635 %macro CHROMA_V_LOAD 1
1637 mova m1, [r0+r1] ; p0
1639 mova m3, [%1+r1] ; q1
1642 ; clobbers: m1, m2, m3
1643 %macro CHROMA_H_STORE 0-1
1644 SBUTTERFLY dq, 1, 2, 3
1650 movq [r0+r1*2-4], m2
1651 movhps [r0+r1-4], m1
1652 movhps [r0+%1-4], m2
1656 %macro CHROMA_V_STORE 0
1661 %macro DEBLOCK_CHROMA 0
1662 cglobal deblock_inter_body
1663 LOAD_AB m4, m5, r2, r3
1664 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1669 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
1672 ;-----------------------------------------------------------------------------
1673 ; void deblock_v_chroma( uint16_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1674 ;-----------------------------------------------------------------------------
1675 cglobal deblock_v_chroma, 5,7,8
1683 call deblock_inter_body
1692 ;-----------------------------------------------------------------------------
1693 ; void deblock_h_chroma( uint16_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1694 ;-----------------------------------------------------------------------------
1695 cglobal deblock_h_chroma, 5,7,8
1703 call deblock_inter_body
1705 lea r0, [r0+r1*(mmsize/4)]
1712 cglobal deblock_intra_body
1713 LOAD_AB m4, m5, r2, r3
1714 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1715 CHROMA_DEBLOCK_P0_Q0_INTRA m1, m2, m0, m3, m7, m5, m6
1718 ;-----------------------------------------------------------------------------
1719 ; void deblock_v_chroma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
1720 ;-----------------------------------------------------------------------------
1721 cglobal deblock_v_chroma_intra, 4,6,8
1731 call deblock_intra_body
1739 ;-----------------------------------------------------------------------------
1740 ; void deblock_h_chroma_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
1741 ;-----------------------------------------------------------------------------
1742 cglobal deblock_h_chroma_intra, 4,6,8
1750 call deblock_intra_body
1752 lea r0, [r0+r1*(mmsize/4)]
1757 ;-----------------------------------------------------------------------------
1758 ; void deblock_h_chroma_intra_mbaff( uint16_t *pix, intptr_t stride, int alpha, int beta )
1759 ;-----------------------------------------------------------------------------
1760 cglobal deblock_h_chroma_intra_mbaff, 4,6,8
1769 LOAD_AB m4, m5, r2, r3
1770 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1771 CHROMA_DEBLOCK_P0_Q0_INTRA m1, m2, m0, m3, m7, m5, m6
1774 lea r0, [r0+r1*(mmsize/4)]
1780 ;-----------------------------------------------------------------------------
1781 ; void deblock_h_chroma_mbaff( uint16_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1782 ;-----------------------------------------------------------------------------
1783 cglobal deblock_h_chroma_mbaff, 5,7,8
1791 LOAD_AB m4, m5, r2, r3
1792 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1798 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
1801 lea r0, [r0+r1*(mmsize/4)]
1808 ;-----------------------------------------------------------------------------
1809 ; void deblock_h_chroma_422_intra( uint16_t *pix, intptr_t stride, int alpha, int beta )
1810 ;-----------------------------------------------------------------------------
1811 cglobal deblock_h_chroma_422_intra, 4,6,8
1819 call deblock_intra_body
1821 lea r0, [r0+r1*(mmsize/4)]
1826 ;-----------------------------------------------------------------------------
1827 ; void deblock_h_chroma_422( uint16_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1828 ;-----------------------------------------------------------------------------
1829 cglobal deblock_h_chroma_422, 5,7,8
1835 LOAD_AB m4, m5, r2m, r3
1836 LOAD_MASK m0, m1, m2, m3, m4, m5, m7, m6, m4
1843 DEBLOCK_P0_Q0 m1, m2, m0, m3, m7, m5, m6
1845 lea r0, [r0+r1*(mmsize/4)]
1851 add r4, r2 ; increment once every 2 iterations
1856 %endmacro ; DEBLOCK_CHROMA
1858 %if ARCH_X86_64 == 0
1866 %endif ; HIGH_BIT_DEPTH
1868 %if HIGH_BIT_DEPTH == 0
1869 %macro CHROMA_V_START 0
1881 %macro CHROMA_H_START 0
1890 %macro CHROMA_V_LOOP 1
1902 %macro CHROMA_H_LOOP 1
1917 %macro DEBLOCK_CHROMA 0
1918 cglobal chroma_inter_body
1927 ;-----------------------------------------------------------------------------
1928 ; void deblock_v_chroma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1929 ;-----------------------------------------------------------------------------
1930 cglobal deblock_v_chroma, 5,6,8
1936 call chroma_inter_body
1942 ;-----------------------------------------------------------------------------
1943 ; void deblock_h_chroma( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1944 ;-----------------------------------------------------------------------------
1945 cglobal deblock_h_chroma, 5,7,8
1951 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
1952 call chroma_inter_body
1953 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
1956 %endmacro ; DEBLOCK_CHROMA
1962 %if ARCH_X86_64 == 0
1967 ;-----------------------------------------------------------------------------
1968 ; void deblock_h_chroma_mbaff( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
1969 ;-----------------------------------------------------------------------------
1970 %macro DEBLOCK_H_CHROMA_420_MBAFF 0
1971 cglobal deblock_h_chroma_mbaff, 5,7,8
1978 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
1984 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
1989 DEBLOCK_H_CHROMA_420_MBAFF
1990 %if ARCH_X86_64 == 0
1992 DEBLOCK_H_CHROMA_420_MBAFF
1995 %macro DEBLOCK_H_CHROMA_422 0
1996 cglobal deblock_h_chroma_422, 5,8,8
2000 %define cntr dword r0m
2005 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
2013 pshufw m6, m6, q0000
2017 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
2018 lea r0, [r0+r1*(mmsize/2)]
2019 lea t5, [t5+r1*(mmsize/2)]
2027 DEBLOCK_H_CHROMA_422
2029 DEBLOCK_H_CHROMA_422
2031 DEBLOCK_H_CHROMA_422
2033 ; in: %1=p0 %2=p1 %3=q1
2034 ; out: p0 = (p0 + q1 + 2*p1 + 2) >> 2
2035 %macro CHROMA_INTRA_P0 3
2037 pand m4, [pb_1] ; m4 = (p0^q1)&1
2040 pavgb %1, %2 ; dst = avg(p1, avg(p0,q1) - ((p0^q1)&1))
2046 %macro DEBLOCK_CHROMA_INTRA_BODY 0
2047 cglobal chroma_intra_body
2051 CHROMA_INTRA_P0 m1, m0, m3
2052 CHROMA_INTRA_P0 m2, m3, m0
2062 %macro DEBLOCK_CHROMA_INTRA 0
2063 ;-----------------------------------------------------------------------------
2064 ; void deblock_v_chroma_intra( uint8_t *pix, intptr_t stride, int alpha, int beta )
2065 ;-----------------------------------------------------------------------------
2066 cglobal deblock_v_chroma_intra, 4,5,8
2072 call chroma_intra_body
2078 ;-----------------------------------------------------------------------------
2079 ; void deblock_h_chroma_intra( uint8_t *pix, intptr_t stride, int alpha, int beta )
2080 ;-----------------------------------------------------------------------------
2081 cglobal deblock_h_chroma_intra, 4,6,8
2087 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
2088 call chroma_intra_body
2089 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
2093 cglobal deblock_h_chroma_422_intra, 4,7,8
2097 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
2098 call chroma_intra_body
2099 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
2100 lea r0, [r0+r1*(mmsize/2)]
2101 lea t5, [t5+r1*(mmsize/2)]
2105 %endmacro ; DEBLOCK_CHROMA_INTRA
2108 DEBLOCK_CHROMA_INTRA_BODY
2109 DEBLOCK_CHROMA_INTRA
2111 DEBLOCK_CHROMA_INTRA_BODY
2112 DEBLOCK_CHROMA_INTRA
2114 DEBLOCK_CHROMA_INTRA_BODY
2115 %if ARCH_X86_64 == 0
2116 DEBLOCK_CHROMA_INTRA
2119 ;-----------------------------------------------------------------------------
2120 ; void deblock_h_chroma_intra_mbaff( uint8_t *pix, intptr_t stride, int alpha, int beta )
2121 ;-----------------------------------------------------------------------------
2123 cglobal deblock_h_chroma_intra_mbaff, 4,6,8
2125 TRANSPOSE4x8W_LOAD PASS8ROWS(t5, r0, r1, t6)
2126 call chroma_intra_body
2127 TRANSPOSE8x2W_STORE PASS8ROWS(t5, r0, r1, t6, 2)
2129 %endif ; !HIGH_BIT_DEPTH
2133 ;-----------------------------------------------------------------------------
2134 ; static void deblock_strength( uint8_t nnz[48], int8_t ref[2][40], int16_t mv[2][40][2],
2135 ; uint8_t bs[2][4][4], int mvy_limit, int bframe )
2136 ;-----------------------------------------------------------------------------
2138 %define scan8start (4+1*8)
2139 %define nnz r0+scan8start
2140 %define ref r1+scan8start
2141 %define mv r2+scan8start*4
2145 %macro LOAD_BYTES_MMX 1
2150 punpckldq m2, [%1+8*1-1]
2151 punpckldq m0, [%1+8*1]
2152 punpckldq m3, [%1+8*3-1]
2153 punpckldq m1, [%1+8*3]
2156 %macro DEBLOCK_STRENGTH_REFS_MMX 0
2167 punpckldq m2, m0 ; row -1, row 0
2168 punpckldq m3, m1 ; row 1, row 2
2177 %macro DEBLOCK_STRENGTH_MVS_MMX 2
2190 %macro DEBLOCK_STRENGTH_NNZ_MMX 1
2197 pminub m4, m6 ; mv ? 1 : 0
2199 paddb m2, m2 ; nnz ? 2 : 0
2205 %macro LOAD_BYTES_XMM 1
2206 movu m2, [%1-4] ; FIXME could be aligned if we changed nnz's allocation
2209 shufps m2, m1, q3131 ; cur nnz, all rows
2211 shufps m0, m1, q3131 ; left neighbors
2213 movd m3, [%1-8] ; could be palignr if nnz was aligned
2214 por m1, m3 ; top neighbors
2218 cglobal deblock_strength, 6,6
2219 ; Prepare mv comparison register
2232 DEBLOCK_STRENGTH_REFS_MMX
2235 DEBLOCK_STRENGTH_MVS_MMX bs0, 4
2236 DEBLOCK_STRENGTH_MVS_MMX bs1, 4*8
2249 DEBLOCK_STRENGTH_NNZ_MMX bs0
2250 ; Transpose column output
2251 SBUTTERFLY bw, 2, 3, 4
2252 SBUTTERFLY bw, 2, 3, 4
2257 punpckldq m2, m0 ; row -1, row 0
2258 punpckldq m3, m1 ; row 1, row 2
2259 DEBLOCK_STRENGTH_NNZ_MMX bs1
2264 %macro DEBLOCK_STRENGTH_XMM 0
2265 cglobal deblock_strength, 6,6,7
2266 ; Prepare mv comparison register
2286 palignr m3, m0, [mv+4*8*0-16], 12
2287 palignr m2, m1, [mv+4*8*1-16], 12
2294 palignr m3, m2, [mv+4*8*2-16], 12
2296 palignr m3, m1, [mv+4*8*3-16], 12
2300 movu m0, [mv-4+4*8*0]
2301 movu m1, [mv-4+4*8*1]
2302 movu m2, [mv-4+4*8*2]
2303 movu m3, [mv-4+4*8*3]
2304 psubw m0, [mv+4*8*0]
2305 psubw m1, [mv+4*8*1]
2306 psubw m2, [mv+4*8*2]
2307 psubw m3, [mv+4*8*3]
2318 mova m0, [mv+4*8*-1]
2319 mova m1, [mv+4*8* 0]
2320 mova m2, [mv+4*8* 1]
2321 mova m3, [mv+4*8* 2]
2325 psubw m3, [mv+4*8* 3]
2346 pminub m4, m6 ; mv ? 1 : 0
2348 paddb m0, m0 ; nnz ? 2 : 0
2353 pshufb m4, [transpose_shuf]
2366 DEBLOCK_STRENGTH_XMM
2368 DEBLOCK_STRENGTH_XMM
2370 DEBLOCK_STRENGTH_XMM