annotate x86/h264_weight.asm @ 12485:b42e02e9bf2b libavcodec

Move AMR-NB frame unpacking code to a common file so it can be reused in the AMR-WB decoder. Patch by Marcelo Galvo Pvoa.
author vitor
date Fri, 10 Sep 2010 19:51:08 +0000
parents 2982071047a2
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
1 ;*****************************************************************************
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
2 ;* SSE2-optimized weighted prediction code
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
3 ;*****************************************************************************
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
4 ;* Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
5 ;* Copyright (C) 2010 Eli Friedman <eli.friedman@gmail.com>
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
6 ;*
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
7 ;* This file is part of FFmpeg.
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
8 ;*
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
9 ;* FFmpeg is free software; you can redistribute it and/or
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
10 ;* modify it under the terms of the GNU Lesser General Public
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
11 ;* License as published by the Free Software Foundation; either
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
12 ;* version 2.1 of the License, or (at your option) any later version.
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
13 ;*
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
14 ;* FFmpeg is distributed in the hope that it will be useful,
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
15 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
16 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
17 ;* Lesser General Public License for more details.
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
18 ;*
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
19 ;* You should have received a copy of the GNU Lesser General Public
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
20 ;* License along with FFmpeg; if not, write to the Free Software
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
21 ;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
22 ;******************************************************************************
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
23
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
24 %include "x86inc.asm"
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
25
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
26 SECTION .text
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
27
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
28 ;-----------------------------------------------------------------------------
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
29 ; biweight pred:
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
30 ;
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
31 ; void h264_biweight_16x16_sse2(uint8_t *dst, uint8_t *src, int stride,
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
32 ; int log2_denom, int weightd, int weights,
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
33 ; int offset);
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
34 ; and
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
35 ; void h264_weight_16x16_sse2(uint8_t *dst, int stride,
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
36 ; int log2_denom, int weight,
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
37 ; int offset);
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
38 ;-----------------------------------------------------------------------------
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
39
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
40 %macro WEIGHT_SETUP 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
41 add r4, r4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
42 inc r4
12457
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
43 movd m3, r3d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
44 movd m5, r4d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
45 movd m6, r2d
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
46 pslld m5, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
47 psrld m5, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
48 %if mmsize == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
49 pshuflw m3, m3, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
50 pshuflw m5, m5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
51 punpcklqdq m3, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
52 punpcklqdq m5, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
53 %else
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
54 pshufw m3, m3, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
55 pshufw m5, m5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
56 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
57 pxor m7, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
58 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
59
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
60 %macro WEIGHT_OP 2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
61 movh m0, [r0+%1]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
62 movh m1, [r0+%2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
63 punpcklbw m0, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
64 punpcklbw m1, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
65 pmullw m0, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
66 pmullw m1, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
67 paddsw m0, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
68 paddsw m1, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
69 psraw m0, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
70 psraw m1, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
71 packuswb m0, m1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
72 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
73
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
74 %macro WEIGHT_FUNC_DBL_MM 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
75 cglobal h264_weight_16x%1_mmx2, 5, 5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
76 WEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
77 mov r2, %1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
78 %if %1 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
79 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
80 WEIGHT_OP 0, 4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
81 mova [r0 ], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
82 WEIGHT_OP 8, 12
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
83 mova [r0+8], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
84 add r0, r1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
85 dec r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
86 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
87 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
88 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
89 jmp mangle(ff_h264_weight_16x16_mmx2.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
90 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
91 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
92
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
93 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
94 WEIGHT_FUNC_DBL_MM 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
95 WEIGHT_FUNC_DBL_MM 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
96
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
97 %macro WEIGHT_FUNC_MM 4
12453
35e1de8243c6 Fix typo in r25019.
darkshikari
parents: 12452
diff changeset
98 cglobal h264_weight_%1x%2_%4, 7, 7, %3
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
99 WEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
100 mov r2, %2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
101 %if %2 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
102 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
103 WEIGHT_OP 0, mmsize/2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
104 mova [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
105 add r0, r1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
106 dec r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
107 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
108 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
109 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
110 jmp mangle(ff_h264_weight_%1x16_%4.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
111 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
112 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
113
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
114 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
115 WEIGHT_FUNC_MM 8, 16, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
116 WEIGHT_FUNC_MM 8, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
117 WEIGHT_FUNC_MM 8, 4, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
118 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
119 WEIGHT_FUNC_MM 16, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
120 WEIGHT_FUNC_MM 16, 8, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
121
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
122 %macro WEIGHT_FUNC_HALF_MM 5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
123 cglobal h264_weight_%1x%2_%5, 5, 5, %4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
124 WEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
125 mov r2, %2/2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
126 lea r3, [r1*2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
127 %if %2 == mmsize
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
128 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
129 WEIGHT_OP 0, r1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
130 movh [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
131 %if mmsize == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
132 movhps [r0+r1], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
133 %else
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
134 psrlq m0, 32
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
135 movh [r0+r1], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
136 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
137 add r0, r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
138 dec r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
139 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
140 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
141 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
142 jmp mangle(ff_h264_weight_%1x%3_%5.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
143 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
144 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
145
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
146 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
147 WEIGHT_FUNC_HALF_MM 4, 8, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
148 WEIGHT_FUNC_HALF_MM 4, 4, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
149 WEIGHT_FUNC_HALF_MM 4, 2, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
150 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
151 WEIGHT_FUNC_HALF_MM 8, 16, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
152 WEIGHT_FUNC_HALF_MM 8, 8, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
153 WEIGHT_FUNC_HALF_MM 8, 4, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
154
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
155 %macro BIWEIGHT_SETUP 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
156 add r6, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
157 or r6, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
158 add r3, 1
12457
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
159 movd m3, r4d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
160 movd m4, r5d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
161 movd m5, r6d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
162 movd m6, r3d
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
163 pslld m5, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
164 psrld m5, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
165 %if mmsize == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
166 pshuflw m3, m3, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
167 pshuflw m4, m4, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
168 pshuflw m5, m5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
169 punpcklqdq m3, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
170 punpcklqdq m4, m4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
171 punpcklqdq m5, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
172 %else
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
173 pshufw m3, m3, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
174 pshufw m4, m4, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
175 pshufw m5, m5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
176 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
177 pxor m7, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
178 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
179
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
180 %macro BIWEIGHT_STEPA 3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
181 movh m%1, [r0+%3]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
182 movh m%2, [r1+%3]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
183 punpcklbw m%1, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
184 punpcklbw m%2, m7
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
185 pmullw m%1, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
186 pmullw m%2, m4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
187 paddsw m%1, m%2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
188 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
189
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
190 %macro BIWEIGHT_STEPB 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
191 paddsw m0, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
192 paddsw m1, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
193 psraw m0, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
194 psraw m1, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
195 packuswb m0, m1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
196 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
197
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
198 %macro BIWEIGHT_FUNC_DBL_MM 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
199 cglobal h264_biweight_16x%1_mmx2, 7, 7, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
200 BIWEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
201 mov r3, %1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
202 %if %1 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
203 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
204 BIWEIGHT_STEPA 0, 1, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
205 BIWEIGHT_STEPA 1, 2, 4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
206 BIWEIGHT_STEPB
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
207 mova [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
208 BIWEIGHT_STEPA 0, 1, 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
209 BIWEIGHT_STEPA 1, 2, 12
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
210 BIWEIGHT_STEPB
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
211 mova [r0+8], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
212 add r0, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
213 add r1, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
214 dec r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
215 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
216 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
217 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
218 jmp mangle(ff_h264_biweight_16x16_mmx2.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
219 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
220 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
221
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
222 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
223 BIWEIGHT_FUNC_DBL_MM 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
224 BIWEIGHT_FUNC_DBL_MM 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
225
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
226 %macro BIWEIGHT_FUNC_MM 4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
227 cglobal h264_biweight_%1x%2_%4, 7, 7, %3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
228 BIWEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
229 mov r3, %2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
230 %if %2 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
231 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
232 BIWEIGHT_STEPA 0, 1, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
233 BIWEIGHT_STEPA 1, 2, mmsize/2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
234 BIWEIGHT_STEPB
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
235 mova [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
236 add r0, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
237 add r1, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
238 dec r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
239 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
240 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
241 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
242 jmp mangle(ff_h264_biweight_%1x16_%4.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
243 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
244 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
245
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
246 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
247 BIWEIGHT_FUNC_MM 8, 16, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
248 BIWEIGHT_FUNC_MM 8, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
249 BIWEIGHT_FUNC_MM 8, 4, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
250 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
251 BIWEIGHT_FUNC_MM 16, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
252 BIWEIGHT_FUNC_MM 16, 8, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
253
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
254 %macro BIWEIGHT_FUNC_HALF_MM 5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
255 cglobal h264_biweight_%1x%2_%5, 7, 7, %4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
256 BIWEIGHT_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
257 mov r3, %2/2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
258 lea r4, [r2*2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
259 %if %2 == mmsize
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
260 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
261 BIWEIGHT_STEPA 0, 1, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
262 BIWEIGHT_STEPA 1, 2, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
263 BIWEIGHT_STEPB
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
264 movh [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
265 %if mmsize == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
266 movhps [r0+r2], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
267 %else
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
268 psrlq m0, 32
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
269 movh [r0+r2], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
270 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
271 add r0, r4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
272 add r1, r4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
273 dec r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
274 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
275 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
276 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
277 jmp mangle(ff_h264_biweight_%1x%3_%5.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
278 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
279 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
280
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
281 INIT_MMX
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
282 BIWEIGHT_FUNC_HALF_MM 4, 8, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
283 BIWEIGHT_FUNC_HALF_MM 4, 4, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
284 BIWEIGHT_FUNC_HALF_MM 4, 2, 8, 0, mmx2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
285 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
286 BIWEIGHT_FUNC_HALF_MM 8, 16, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
287 BIWEIGHT_FUNC_HALF_MM 8, 8, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
288 BIWEIGHT_FUNC_HALF_MM 8, 4, 16, 8, sse2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
289
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
290 %macro BIWEIGHT_SSSE3_SETUP 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
291 add r6, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
292 or r6, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
293 add r3, 1
12457
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
294 movd m4, r4d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
295 movd m0, r5d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
296 movd m5, r6d
2982071047a2 Use "d" suffix for general-purpose registers used with movd.
reimar
parents: 12453
diff changeset
297 movd m6, r3d
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
298 pslld m5, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
299 psrld m5, 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
300 punpcklbw m4, m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
301 pshuflw m4, m4, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
302 pshuflw m5, m5, 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
303 punpcklqdq m4, m4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
304 punpcklqdq m5, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
305 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
306
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
307 %macro BIWEIGHT_SSSE3_OP 0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
308 pmaddubsw m0, m4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
309 pmaddubsw m2, m4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
310 paddsw m0, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
311 paddsw m2, m5
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
312 psraw m0, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
313 psraw m2, m6
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
314 packuswb m0, m2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
315 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
316
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
317 %macro BIWEIGHT_SSSE3_16 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
318 cglobal h264_biweight_16x%1_ssse3, 7, 7, 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
319 BIWEIGHT_SSSE3_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
320 mov r3, %1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
321
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
322 %if %1 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
323 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
324 movh m0, [r0]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
325 movh m2, [r0+8]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
326 movh m3, [r1+8]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
327 punpcklbw m0, [r1]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
328 punpcklbw m2, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
329 BIWEIGHT_SSSE3_OP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
330 mova [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
331 add r0, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
332 add r1, r2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
333 dec r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
334 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
335 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
336 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
337 jmp mangle(ff_h264_biweight_16x16_ssse3.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
338 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
339 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
340
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
341 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
342 BIWEIGHT_SSSE3_16 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
343 BIWEIGHT_SSSE3_16 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
344
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
345 %macro BIWEIGHT_SSSE3_8 1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
346 cglobal h264_biweight_8x%1_ssse3, 7, 7, 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
347 BIWEIGHT_SSSE3_SETUP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
348 mov r3, %1/2
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
349 lea r4, [r2*2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
350
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
351 %if %1 == 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
352 .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
353 movh m0, [r0]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
354 movh m1, [r1]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
355 movh m2, [r0+r2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
356 movh m3, [r1+r2]
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
357 punpcklbw m0, m1
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
358 punpcklbw m2, m3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
359 BIWEIGHT_SSSE3_OP
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
360 movh [r0], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
361 movhps [r0+r2], m0
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
362 add r0, r4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
363 add r1, r4
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
364 dec r3
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
365 jnz .nextrow
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
366 REP_RET
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
367 %else
12452
2c28e7d75e9c Unscrew breakage after my last commit because of symbol prefixes.
rbultje
parents: 12451
diff changeset
368 jmp mangle(ff_h264_biweight_8x16_ssse3.nextrow)
12451
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
369 %endif
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
370 %endmacro
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
371
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
372 INIT_XMM
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
373 BIWEIGHT_SSSE3_8 16
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
374 BIWEIGHT_SSSE3_8 8
4c3e6ff1237e Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
diff changeset
375 BIWEIGHT_SSSE3_8 4