annotate x86/x86util.asm @ 9017:1e2c96617886 libavcodec

Change sign in ff_celp_lp_synthesis_filterf(). This makes this function useful for AMR and RA144. Patch by Kenan Gillet
author vitor
date Mon, 23 Feb 2009 22:19:16 +0000
parents ba83a0c57e9f
children c08ca946c80a
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8510
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
1 ;*****************************************************************************
8804
ba83a0c57e9f Fix wrong file name in header, noticed by David DeHaven, dave sagetv com.
diego
parents: 8510
diff changeset
2 ;* x86util.asm
8510
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
3 ;*****************************************************************************
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
4 ;* Copyright (C) 2008 Loren Merritt <lorenm@u.washington.edu>
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
5 ;*
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
6 ;* This program is free software; you can redistribute it and/or modify
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
7 ;* it under the terms of the GNU General Public License as published by
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
8 ;* the Free Software Foundation; either version 2 of the License, or
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
9 ;* (at your option) any later version.
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
10 ;*
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
11 ;* This program is distributed in the hope that it will be useful,
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
12 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
13 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
14 ;* GNU General Public License for more details.
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
15 ;*
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
16 ;* You should have received a copy of the GNU General Public License
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
17 ;* along with this program; if not, write to the Free Software
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
18 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
19 ;*****************************************************************************
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
20
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
21 %macro SBUTTERFLY 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
22 mova m%4, m%2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
23 punpckl%1 m%2, m%3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
24 punpckh%1 m%4, m%3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
25 SWAP %3, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
26 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
27
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
28 %macro TRANSPOSE4x4W 5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
29 SBUTTERFLY wd, %1, %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
30 SBUTTERFLY wd, %3, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
31 SBUTTERFLY dq, %1, %3, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
32 SBUTTERFLY dq, %2, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
33 SWAP %2, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
34 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
35
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
36 %macro TRANSPOSE2x4x4W 5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
37 SBUTTERFLY wd, %1, %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
38 SBUTTERFLY wd, %3, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
39 SBUTTERFLY dq, %1, %3, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
40 SBUTTERFLY dq, %2, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
41 SBUTTERFLY qdq, %1, %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
42 SBUTTERFLY qdq, %3, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
43 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
44
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
45 %macro TRANSPOSE4x4D 5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
46 SBUTTERFLY dq, %1, %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
47 SBUTTERFLY dq, %3, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
48 SBUTTERFLY qdq, %1, %3, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
49 SBUTTERFLY qdq, %2, %4, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
50 SWAP %2, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
51 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
52
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
53 %macro TRANSPOSE8x8W 9-11
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
54 %ifdef ARCH_X86_64
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
55 SBUTTERFLY wd, %1, %2, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
56 SBUTTERFLY wd, %3, %4, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
57 SBUTTERFLY wd, %5, %6, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
58 SBUTTERFLY wd, %7, %8, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
59 SBUTTERFLY dq, %1, %3, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
60 SBUTTERFLY dq, %2, %4, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
61 SBUTTERFLY dq, %5, %7, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
62 SBUTTERFLY dq, %6, %8, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
63 SBUTTERFLY qdq, %1, %5, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
64 SBUTTERFLY qdq, %2, %6, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
65 SBUTTERFLY qdq, %3, %7, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
66 SBUTTERFLY qdq, %4, %8, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
67 SWAP %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
68 SWAP %4, %7
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
69 %else
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
70 ; in: m0..m7, unless %11 in which case m6 is in %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
71 ; out: m0..m7, unless %11 in which case m4 is in %10
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
72 ; spills into %9 and %10
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
73 %if %0<11
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
74 movdqa %9, m%7
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
75 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
76 SBUTTERFLY wd, %1, %2, %7
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
77 movdqa %10, m%2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
78 movdqa m%7, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
79 SBUTTERFLY wd, %3, %4, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
80 SBUTTERFLY wd, %5, %6, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
81 SBUTTERFLY wd, %7, %8, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
82 SBUTTERFLY dq, %1, %3, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
83 movdqa %9, m%3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
84 movdqa m%2, %10
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
85 SBUTTERFLY dq, %2, %4, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
86 SBUTTERFLY dq, %5, %7, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
87 SBUTTERFLY dq, %6, %8, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
88 SBUTTERFLY qdq, %1, %5, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
89 SBUTTERFLY qdq, %2, %6, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
90 movdqa %10, m%2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
91 movdqa m%3, %9
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
92 SBUTTERFLY qdq, %3, %7, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
93 SBUTTERFLY qdq, %4, %8, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
94 SWAP %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
95 SWAP %4, %7
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
96 %if 0<11
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
97 movdqa m%5, %10
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
98 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
99 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
100 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
101
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
102 %macro ABS1_MMX 2 ; a, tmp
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
103 pxor %2, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
104 psubw %2, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
105 pmaxsw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
106 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
107
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
108 %macro ABS2_MMX 4 ; a, b, tmp0, tmp1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
109 pxor %3, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
110 pxor %4, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
111 psubw %3, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
112 psubw %4, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
113 pmaxsw %1, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
114 pmaxsw %2, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
115 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
116
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
117 %macro ABS1_SSSE3 2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
118 pabsw %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
119 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
120
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
121 %macro ABS2_SSSE3 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
122 pabsw %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
123 pabsw %2, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
124 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
125
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
126 %define ABS1 ABS1_MMX
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
127 %define ABS2 ABS2_MMX
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
128
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
129 %macro ABS4 6
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
130 ABS2 %1, %2, %5, %6
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
131 ABS2 %3, %4, %5, %6
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
132 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
133
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
134 %macro SPLATB_MMX 3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
135 movd %1, [%2-3] ;to avoid crossing a cacheline
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
136 punpcklbw %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
137 %if mmsize==16
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
138 pshuflw %1, %1, 0xff
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
139 punpcklqdq %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
140 %else
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
141 pshufw %1, %1, 0xff
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
142 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
143 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
144
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
145 %macro SPLATB_SSSE3 3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
146 movd %1, [%2-3]
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
147 pshufb %1, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
148 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
149
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
150 %macro PALIGNR_MMX 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
151 %ifnidn %4, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
152 mova %4, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
153 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
154 %if mmsize == 8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
155 psllq %1, (8-%3)*8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
156 psrlq %4, %3*8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
157 %else
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
158 pslldq %1, 16-%3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
159 psrldq %4, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
160 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
161 por %1, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
162 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
163
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
164 %macro PALIGNR_SSSE3 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
165 palignr %1, %2, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
166 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
167
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
168 %macro SUMSUB_BA 2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
169 paddw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
170 paddw %2, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
171 psubw %2, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
172 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
173
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
174 %macro SUMSUB_BADC 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
175 paddw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
176 paddw %3, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
177 paddw %2, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
178 paddw %4, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
179 psubw %2, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
180 psubw %4, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
181 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
182
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
183 %macro HADAMARD8_1D 8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
184 SUMSUB_BADC %1, %5, %2, %6
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
185 SUMSUB_BADC %3, %7, %4, %8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
186 SUMSUB_BADC %1, %3, %2, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
187 SUMSUB_BADC %5, %7, %6, %8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
188 SUMSUB_BADC %1, %2, %3, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
189 SUMSUB_BADC %5, %6, %7, %8
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
190 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
191
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
192 %macro SUMSUB2_AB 3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
193 mova %3, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
194 paddw %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
195 paddw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
196 psubw %3, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
197 psubw %3, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
198 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
199
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
200 %macro SUMSUBD2_AB 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
201 mova %4, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
202 mova %3, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
203 psraw %2, 1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
204 psraw %4, 1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
205 paddw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
206 psubw %4, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
207 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
208
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
209 %macro LOAD_DIFF 5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
210 %ifidn %3, none
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
211 movh %1, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
212 movh %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
213 punpcklbw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
214 punpcklbw %2, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
215 psubw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
216 %else
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
217 movh %1, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
218 punpcklbw %1, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
219 movh %2, %5
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
220 punpcklbw %2, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
221 psubw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
222 %endif
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
223 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
224
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
225 %macro LOAD_DIFF_8x4P 6-8 r0,r2 ; 4x dest, 2x temp, 2x pointer
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
226 LOAD_DIFF %1, %5, none, [%7], [%8]
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
227 LOAD_DIFF %2, %6, none, [%7+r1], [%8+r3]
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
228 LOAD_DIFF %3, %5, none, [%7+2*r1], [%8+2*r3]
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
229 LOAD_DIFF %4, %6, none, [%7+r4], [%8+r5]
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
230 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
231
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
232 %macro STORE_DIFF 4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
233 psraw %1, 6
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
234 movh %2, %4
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
235 punpcklbw %2, %3
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
236 paddsw %1, %2
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
237 packuswb %1, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
238 movh %4, %1
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
239 %endmacro
cea216e44ee3 Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
diff changeset
240