annotate x86/vc1dsp_yasm.asm @ 12442:94b578d0af10 libavcodec

psymodel: Const correct FFPsyWindowInfo.
author alexc
date Mon, 30 Aug 2010 23:48:35 +0000
parents 846779f6b164
children 2982071047a2
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12144
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
1 ;******************************************************************************
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
2 ;* VC1 deblocking optimizations
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
3 ;* Copyright (c) 2009 David Conrad
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
4 ;*
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
5 ;* This file is part of FFmpeg.
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
6 ;*
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
7 ;* FFmpeg is free software; you can redistribute it and/or
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
8 ;* modify it under the terms of the GNU Lesser General Public
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
9 ;* License as published by the Free Software Foundation; either
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
10 ;* version 2.1 of the License, or (at your option) any later version.
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
11 ;*
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
12 ;* FFmpeg is distributed in the hope that it will be useful,
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
13 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
14 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
15 ;* Lesser General Public License for more details.
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
16 ;*
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
17 ;* You should have received a copy of the GNU Lesser General Public
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
18 ;* License along with FFmpeg; if not, write to the Free Software
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
19 ;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
20 ;******************************************************************************
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
21
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
22 %include "x86inc.asm"
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
23 %include "x86util.asm"
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
24
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
25 cextern pw_4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
26 cextern pw_5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
27
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
28 section .text
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
29
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
30 ; dst_low, dst_high (src), zero
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
31 ; zero-extends one vector from 8 to 16 bits
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
32 %macro UNPACK_8TO16 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
33 mova m%2, m%3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
34 punpckh%1 m%3, m%4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
35 punpckl%1 m%2, m%4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
36 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
37
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
38 %macro STORE_4_WORDS_MMX 6
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
39 movd %6, %5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
40 %if mmsize==16
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
41 psrldq %5, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
42 %else
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
43 psrlq %5, 32
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
44 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
45 mov %1, %6w
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
46 shr %6, 16
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
47 mov %2, %6w
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
48 movd %6, %5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
49 mov %3, %6w
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
50 shr %6, 16
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
51 mov %4, %6w
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
52 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
53
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
54 %macro STORE_4_WORDS_SSE4 6
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
55 pextrw %1, %5, %6+0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
56 pextrw %2, %5, %6+1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
57 pextrw %3, %5, %6+2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
58 pextrw %4, %5, %6+3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
59 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
60
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
61 ; in: p1 p0 q0 q1, clobbers p0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
62 ; out: p1 = (2*(p1 - q1) - 5*(p0 - q0) + 4) >> 3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
63 %macro VC1_LOOP_FILTER_A0 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
64 psubw %1, %4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
65 psubw %2, %3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
66 paddw %1, %1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
67 pmullw %2, [pw_5]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
68 psubw %1, %2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
69 paddw %1, [pw_4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
70 psraw %1, 3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
71 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
72
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
73 ; in: p0 q0 a0 a1 a2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
74 ; m0 m1 m7 m6 m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
75 ; %1: size
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
76 ; out: m0=p0' m1=q0'
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
77 %macro VC1_FILTER 1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
78 PABSW m4, m7
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
79 PABSW m3, m6
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
80 PABSW m2, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
81 mova m6, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
82 pminsw m3, m2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
83 pcmpgtw m6, m3 ; if (a2 < a0 || a1 < a0)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
84 psubw m3, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
85 pmullw m3, [pw_5] ; 5*(a3 - a0)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
86 PABSW m2, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
87 psraw m2, 3 ; abs(d/8)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
88 pxor m7, m3 ; d_sign ^= a0_sign
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
89
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
90 pxor m5, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
91 movd m3, r2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
92 %if %1 > 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
93 punpcklbw m3, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
94 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
95 punpcklbw m3, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
96 pcmpgtw m3, m4 ; if (a0 < pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
97 pand m6, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
98
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
99 mova m3, m0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
100 psubw m3, m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
101 PABSW m4, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
102 psraw m4, 1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
103 pxor m3, m7 ; d_sign ^ clip_sign
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
104 psraw m3, 15
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
105 pminsw m2, m4 ; min(d, clip)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
106 pcmpgtw m4, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
107 pand m6, m4 ; filt3 (C return value)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
108
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
109 ; each set of 4 pixels is not filtered if the 3rd is not
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
110 %if mmsize==16
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
111 pshuflw m4, m6, 0xaa
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
112 %if %1 > 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
113 pshufhw m4, m4, 0xaa
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
114 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
115 %else
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
116 pshufw m4, m6, 0xaa
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
117 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
118 pandn m3, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
119 pand m2, m6
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
120 pand m3, m2 ; d final
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
121
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
122 PSIGNW m3, m7
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
123 psubw m0, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
124 paddw m1, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
125 packuswb m0, m0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
126 packuswb m1, m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
127 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
128
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
129 ; 1st param: size of filter
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
130 ; 2nd param: mov suffix equivalent to the filter size
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
131 %macro VC1_V_LOOP_FILTER 2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
132 pxor m5, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
133 mov%2 m6, [r4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
134 mov%2 m4, [r4+r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
135 mov%2 m7, [r4+2*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
136 mov%2 m0, [r4+r3]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
137 punpcklbw m6, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
138 punpcklbw m4, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
139 punpcklbw m7, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
140 punpcklbw m0, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
141
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
142 VC1_LOOP_FILTER_A0 m6, m4, m7, m0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
143 mov%2 m1, [r0]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
144 mov%2 m2, [r0+r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
145 punpcklbw m1, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
146 punpcklbw m2, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
147 mova m4, m0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
148 VC1_LOOP_FILTER_A0 m7, m4, m1, m2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
149 mov%2 m3, [r0+2*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
150 mov%2 m4, [r0+r3]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
151 punpcklbw m3, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
152 punpcklbw m4, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
153 mova m5, m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
154 VC1_LOOP_FILTER_A0 m5, m2, m3, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
155
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
156 VC1_FILTER %1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
157 mov%2 [r4+r3], m0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
158 mov%2 [r0], m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
159 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
160
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
161 ; 1st param: size of filter
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
162 ; NOTE: UNPACK_8TO16 this number of 8 bit numbers are in half a register
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
163 ; 2nd (optional) param: temp register to use for storing words
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
164 %macro VC1_H_LOOP_FILTER 1-2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
165 %if %1 == 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
166 movq m0, [r0 -4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
167 movq m1, [r0+ r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
168 movq m2, [r0+2*r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
169 movq m3, [r0+ r3-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
170 TRANSPOSE4x4B 0, 1, 2, 3, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
171 %else
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
172 movq m0, [r0 -4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
173 movq m4, [r0+ r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
174 movq m1, [r0+2*r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
175 movq m5, [r0+ r3-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
176 movq m2, [r4 -4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
177 movq m6, [r4+ r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
178 movq m3, [r4+2*r1-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
179 movq m7, [r4+ r3-4]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
180 punpcklbw m0, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
181 punpcklbw m1, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
182 punpcklbw m2, m6
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
183 punpcklbw m3, m7
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
184 TRANSPOSE4x4W 0, 1, 2, 3, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
185 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
186 pxor m5, m5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
187
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
188 UNPACK_8TO16 bw, 6, 0, 5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
189 UNPACK_8TO16 bw, 7, 1, 5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
190 VC1_LOOP_FILTER_A0 m6, m0, m7, m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
191 UNPACK_8TO16 bw, 4, 2, 5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
192 mova m0, m1 ; m0 = p0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
193 VC1_LOOP_FILTER_A0 m7, m1, m4, m2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
194 UNPACK_8TO16 bw, 1, 3, 5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
195 mova m5, m4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
196 VC1_LOOP_FILTER_A0 m5, m2, m1, m3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
197 SWAP 1, 4 ; m1 = q0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
198
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
199 VC1_FILTER %1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
200 punpcklbw m0, m1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
201 %if %0 > 1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
202 STORE_4_WORDS_MMX [r0-1], [r0+r1-1], [r0+2*r1-1], [r0+r3-1], m0, %2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
203 %if %1 > 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
204 psrldq m0, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
205 STORE_4_WORDS_MMX [r4-1], [r4+r1-1], [r4+2*r1-1], [r4+r3-1], m0, %2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
206 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
207 %else
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
208 STORE_4_WORDS_SSE4 [r0-1], [r0+r1-1], [r0+2*r1-1], [r0+r3-1], m0, 0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
209 STORE_4_WORDS_SSE4 [r4-1], [r4+r1-1], [r4+2*r1-1], [r4+r3-1], m0, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
210 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
211 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
212
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
213
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
214 %macro START_V_FILTER 0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
215 mov r4, r0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
216 lea r3, [4*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
217 sub r4, r3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
218 lea r3, [r1+2*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
219 imul r2, 0x01010101
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
220 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
221
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
222 %macro START_H_FILTER 1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
223 lea r3, [r1+2*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
224 %if %1 > 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
225 lea r4, [r0+4*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
226 %endif
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
227 imul r2, 0x01010101
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
228 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
229
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
230 ; I dont know why the sign extension is needed...
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
231 %macro PSIGNW_SRA_MMX 2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
232 psraw %2, 15
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
233 PSIGNW_MMX %1, %2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
234 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
235
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
236
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
237 %macro VC1_LF_MMX 1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
238 INIT_MMX
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
239 cglobal vc1_v_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
240 VC1_V_LOOP_FILTER 4, d
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
241 ret
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
242
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
243 cglobal vc1_h_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
244 VC1_H_LOOP_FILTER 4, r4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
245 ret
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
246
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
247 ; void ff_vc1_v_loop_filter4_mmx2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
248 cglobal vc1_v_loop_filter4_%1, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
249 START_V_FILTER
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
250 call vc1_v_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
251 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
252
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
253 ; void ff_vc1_h_loop_filter4_mmx2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
254 cglobal vc1_h_loop_filter4_%1, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
255 START_H_FILTER 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
256 call vc1_h_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
257 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
258
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
259 ; void ff_vc1_v_loop_filter8_mmx2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
260 cglobal vc1_v_loop_filter8_%1, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
261 START_V_FILTER
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
262 call vc1_v_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
263 add r4, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
264 add r0, 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
265 call vc1_v_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
266 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
267
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
268 ; void ff_vc1_h_loop_filter8_mmx2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
269 cglobal vc1_h_loop_filter8_%1, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
270 START_H_FILTER 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
271 call vc1_h_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
272 lea r0, [r0+4*r1]
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
273 call vc1_h_loop_filter_internal_%1
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
274 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
275 %endmacro
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
276
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
277 %define PABSW PABSW_MMX
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
278 %define PSIGNW PSIGNW_SRA_MMX
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
279 VC1_LF_MMX mmx
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
280
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
281 %define PABSW PABSW_MMX2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
282 VC1_LF_MMX mmx2
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
283
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
284 INIT_XMM
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
285 ; void ff_vc1_v_loop_filter8_sse2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
286 cglobal vc1_v_loop_filter8_sse2, 3,5,8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
287 START_V_FILTER
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
288 VC1_V_LOOP_FILTER 8, q
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
289 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
290
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
291 ; void ff_vc1_h_loop_filter8_sse2(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
292 cglobal vc1_h_loop_filter8_sse2, 3,6,8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
293 START_H_FILTER 8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
294 VC1_H_LOOP_FILTER 8, r5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
295 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
296
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
297 %define PABSW PABSW_SSSE3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
298 %define PSIGNW PSIGNW_SSSE3
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
299
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
300 INIT_MMX
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
301 ; void ff_vc1_v_loop_filter4_ssse3(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
302 cglobal vc1_v_loop_filter4_ssse3, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
303 START_V_FILTER
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
304 VC1_V_LOOP_FILTER 4, d
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
305 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
306
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
307 ; void ff_vc1_h_loop_filter4_ssse3(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
308 cglobal vc1_h_loop_filter4_ssse3, 3,5,0
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
309 START_H_FILTER 4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
310 VC1_H_LOOP_FILTER 4, r4
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
311 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
312
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
313 INIT_XMM
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
314 ; void ff_vc1_v_loop_filter8_ssse3(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
315 cglobal vc1_v_loop_filter8_ssse3, 3,5,8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
316 START_V_FILTER
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
317 VC1_V_LOOP_FILTER 8, q
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
318 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
319
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
320 ; void ff_vc1_h_loop_filter8_ssse3(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
321 cglobal vc1_h_loop_filter8_ssse3, 3,6,8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
322 START_H_FILTER 8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
323 VC1_H_LOOP_FILTER 8, r5
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
324 RET
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
325
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
326 ; void ff_vc1_h_loop_filter8_sse4(uint8_t *src, int stride, int pq)
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
327 cglobal vc1_h_loop_filter8_sse4, 3,5,8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
328 START_H_FILTER 8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
329 VC1_H_LOOP_FILTER 8
846779f6b164 MMX/SSE VC1 loop filter
conrad
parents:
diff changeset
330 RET