annotate i386/cavsdsp_mmx.c @ 7462:5cbf11f56c02 libavcodec

Picture.ref_count/ref_poc have to be stored per field (actually also per slice but thats a seperate bug) Fixes at least: CABREF3_Sand_D.264 camp_mot_fld0_full.26l CVFI2_Sony_H.jsv CVNLFI2_Sony_H.jsv
author michael
date Sat, 02 Aug 2008 23:13:27 +0000
parents f7cbb7733146
children eebc7209c47f
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
1 /*
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
2 * Chinese AVS video (AVS1-P2, JiZhun profile) decoder.
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
3 * Copyright (c) 2006 Stefan Gehrer <stefan.gehrer@gmx.de>
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
4 *
5967
15ed47af1838 Misc spelling fixes, prefer American over British English.
diego
parents: 5963
diff changeset
5 * MMX-optimized DSP functions, based on H.264 optimizations by
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
6 * Michael Niedermayer and Loren Merritt
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
7 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
8 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
9 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
10 * FFmpeg is free software; you can redistribute it and/or
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
11 * modify it under the terms of the GNU Lesser General Public
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
12 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
13 * version 2.1 of the License, or (at your option) any later version.
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
14 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
15 * FFmpeg is distributed in the hope that it will be useful,
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
18 * Lesser General Public License for more details.
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
19 *
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
20 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3524
diff changeset
21 * License along with FFmpeg; if not, write to the Free Software
5215
2b72f9bc4f06 license header consistency cosmetics
diego
parents: 5010
diff changeset
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
23 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
24
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
25 #include "libavutil/common.h"
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
26 #include "libavutil/x86_cpu.h"
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
27 #include "libavcodec/dsputil.h"
5946
55251379b5b1 make ff_p* vars extern so that they can be used in various *_mmx.c files
aurel
parents: 5215
diff changeset
28 #include "dsputil_mmx.h"
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
29
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
30 /*****************************************************************************
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
31 *
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
32 * inverse transform
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
33 *
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
34 ****************************************************************************/
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
35
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
36 static inline void cavs_idct8_1d(int16_t *block, uint64_t bias)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
37 {
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
38 asm volatile(
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
39 "movq 112(%0), %%mm4 \n\t" /* mm4 = src7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
40 "movq 16(%0), %%mm5 \n\t" /* mm5 = src1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
41 "movq 80(%0), %%mm2 \n\t" /* mm2 = src5 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
42 "movq 48(%0), %%mm7 \n\t" /* mm7 = src3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
43 "movq %%mm4, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
44 "movq %%mm5, %%mm3 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
45 "movq %%mm2, %%mm6 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
46 "movq %%mm7, %%mm1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
47
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
48 "paddw %%mm4, %%mm4 \n\t" /* mm4 = 2*src7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
49 "paddw %%mm3, %%mm3 \n\t" /* mm3 = 2*src1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
50 "paddw %%mm6, %%mm6 \n\t" /* mm6 = 2*src5 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
51 "paddw %%mm1, %%mm1 \n\t" /* mm1 = 2*src3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
52 "paddw %%mm4, %%mm0 \n\t" /* mm0 = 3*src7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
53 "paddw %%mm3, %%mm5 \n\t" /* mm5 = 3*src1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
54 "paddw %%mm6, %%mm2 \n\t" /* mm2 = 3*src5 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
55 "paddw %%mm1, %%mm7 \n\t" /* mm7 = 3*src3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
56 "psubw %%mm4, %%mm5 \n\t" /* mm5 = 3*src1 - 2*src7 = a0 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
57 "paddw %%mm6, %%mm7 \n\t" /* mm7 = 3*src3 + 2*src5 = a1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
58 "psubw %%mm2, %%mm1 \n\t" /* mm1 = 2*src3 - 3*src5 = a2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
59 "paddw %%mm0, %%mm3 \n\t" /* mm3 = 2*src1 + 3*src7 = a3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
60
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
61 "movq %%mm5, %%mm4 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
62 "movq %%mm7, %%mm6 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
63 "movq %%mm3, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
64 "movq %%mm1, %%mm2 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
65 SUMSUB_BA( %%mm7, %%mm5 ) /* mm7 = a0 + a1 mm5 = a0 - a1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
66 "paddw %%mm3, %%mm7 \n\t" /* mm7 = a0 + a1 + a3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
67 "paddw %%mm1, %%mm5 \n\t" /* mm5 = a0 - a1 + a2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
68 "paddw %%mm7, %%mm7 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
69 "paddw %%mm5, %%mm5 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
70 "paddw %%mm6, %%mm7 \n\t" /* mm7 = b4 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
71 "paddw %%mm4, %%mm5 \n\t" /* mm5 = b5 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
72
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
73 SUMSUB_BA( %%mm1, %%mm3 ) /* mm1 = a3 + a2 mm3 = a3 - a2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
74 "psubw %%mm1, %%mm4 \n\t" /* mm4 = a0 - a2 - a3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
75 "movq %%mm4, %%mm1 \n\t" /* mm1 = a0 - a2 - a3 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
76 "psubw %%mm6, %%mm3 \n\t" /* mm3 = a3 - a2 - a1 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
77 "paddw %%mm1, %%mm1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
78 "paddw %%mm3, %%mm3 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
79 "psubw %%mm2, %%mm1 \n\t" /* mm1 = b7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
80 "paddw %%mm0, %%mm3 \n\t" /* mm3 = b6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
81
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
82 "movq 32(%0), %%mm2 \n\t" /* mm2 = src2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
83 "movq 96(%0), %%mm6 \n\t" /* mm6 = src6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
84 "movq %%mm2, %%mm4 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
85 "movq %%mm6, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
86 "psllw $2, %%mm4 \n\t" /* mm4 = 4*src2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
87 "psllw $2, %%mm6 \n\t" /* mm6 = 4*src6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
88 "paddw %%mm4, %%mm2 \n\t" /* mm2 = 5*src2 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
89 "paddw %%mm6, %%mm0 \n\t" /* mm0 = 5*src6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
90 "paddw %%mm2, %%mm2 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
91 "paddw %%mm0, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
92 "psubw %%mm0, %%mm4 \n\t" /* mm4 = 4*src2 - 10*src6 = a7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
93 "paddw %%mm2, %%mm6 \n\t" /* mm6 = 4*src6 + 10*src2 = a6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
94
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
95 "movq (%0), %%mm2 \n\t" /* mm2 = src0 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
96 "movq 64(%0), %%mm0 \n\t" /* mm0 = src4 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
97 SUMSUB_BA( %%mm0, %%mm2 ) /* mm0 = src0+src4 mm2 = src0-src4 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
98 "psllw $3, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
99 "psllw $3, %%mm2 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
100 "paddw %1, %%mm0 \n\t" /* add rounding bias */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
101 "paddw %1, %%mm2 \n\t" /* add rounding bias */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
102
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
103 SUMSUB_BA( %%mm6, %%mm0 ) /* mm6 = a4 + a6 mm0 = a4 - a6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
104 SUMSUB_BA( %%mm4, %%mm2 ) /* mm4 = a5 + a7 mm2 = a5 - a7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
105 SUMSUB_BA( %%mm7, %%mm6 ) /* mm7 = dst0 mm6 = dst7 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
106 SUMSUB_BA( %%mm5, %%mm4 ) /* mm5 = dst1 mm4 = dst6 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
107 SUMSUB_BA( %%mm3, %%mm2 ) /* mm3 = dst2 mm2 = dst5 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
108 SUMSUB_BA( %%mm1, %%mm0 ) /* mm1 = dst3 mm0 = dst4 */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
109 :: "r"(block), "m"(bias)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
110 );
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
111 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
112
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
113 static void cavs_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
114 {
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
115 int i;
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
116 DECLARE_ALIGNED_8(int16_t, b2[64]);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
117
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
118 for(i=0; i<2; i++){
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
119 DECLARE_ALIGNED_8(uint64_t, tmp);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
120
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
121 cavs_idct8_1d(block+4*i, ff_pw_4);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
122
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
123 asm volatile(
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
124 "psraw $3, %%mm7 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
125 "psraw $3, %%mm6 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
126 "psraw $3, %%mm5 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
127 "psraw $3, %%mm4 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
128 "psraw $3, %%mm3 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
129 "psraw $3, %%mm2 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
130 "psraw $3, %%mm1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
131 "psraw $3, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
132 "movq %%mm7, %0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
133 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 )
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
134 "movq %%mm0, 8(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
135 "movq %%mm6, 24(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
136 "movq %%mm7, 40(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
137 "movq %%mm4, 56(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
138 "movq %0, %%mm7 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
139 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 )
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
140 "movq %%mm7, (%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
141 "movq %%mm1, 16(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
142 "movq %%mm0, 32(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
143 "movq %%mm3, 48(%1) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
144 : "=m"(tmp)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
145 : "r"(b2+32*i)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
146 : "memory"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
147 );
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
148 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
149
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
150 for(i=0; i<2; i++){
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
151 cavs_idct8_1d(b2+4*i, ff_pw_64);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
152
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
153 asm volatile(
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
154 "psraw $7, %%mm7 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
155 "psraw $7, %%mm6 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
156 "psraw $7, %%mm5 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
157 "psraw $7, %%mm4 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
158 "psraw $7, %%mm3 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
159 "psraw $7, %%mm2 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
160 "psraw $7, %%mm1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
161 "psraw $7, %%mm0 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
162 "movq %%mm7, (%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
163 "movq %%mm5, 16(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
164 "movq %%mm3, 32(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
165 "movq %%mm1, 48(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
166 "movq %%mm0, 64(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
167 "movq %%mm2, 80(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
168 "movq %%mm4, 96(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
169 "movq %%mm6, 112(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
170 :: "r"(b2+4*i)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
171 : "memory"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
172 );
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
173 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
174
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
175 add_pixels_clamped_mmx(b2, dst, stride);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
176
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
177 /* clear block */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
178 asm volatile(
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
179 "pxor %%mm7, %%mm7 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
180 "movq %%mm7, (%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
181 "movq %%mm7, 8(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
182 "movq %%mm7, 16(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
183 "movq %%mm7, 24(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
184 "movq %%mm7, 32(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
185 "movq %%mm7, 40(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
186 "movq %%mm7, 48(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
187 "movq %%mm7, 56(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
188 "movq %%mm7, 64(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
189 "movq %%mm7, 72(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
190 "movq %%mm7, 80(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
191 "movq %%mm7, 88(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
192 "movq %%mm7, 96(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
193 "movq %%mm7, 104(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
194 "movq %%mm7, 112(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
195 "movq %%mm7, 120(%0) \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
196 :: "r" (block)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
197 );
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
198 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
199
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
200 /*****************************************************************************
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
201 *
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
202 * motion compensation
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
203 *
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
204 ****************************************************************************/
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
205
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
206 /* vertical filter [-1 -2 96 42 -7 0] */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
207 #define QPEL_CAVSV1(A,B,C,D,E,F,OP) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
208 "movd (%0), "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
209 "movq "#C", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
210 "pmullw %5, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
211 "movq "#D", %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
212 "pmullw %6, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
213 "psllw $3, "#E" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
214 "psubw "#E", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
215 "psraw $3, "#E" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
216 "paddw %%mm7, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
217 "paddw "#E", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
218 "paddw "#B", "#B" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
219 "pxor %%mm7, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
220 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
221 "punpcklbw %%mm7, "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
222 "psubw "#B", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
223 "psraw $1, "#B" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
224 "psubw "#A", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
225 "paddw %4, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
226 "psraw $7, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
227 "packuswb %%mm6, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
228 OP(%%mm6, (%1), A, d) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
229 "add %3, %1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
230
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
231 /* vertical filter [ 0 -1 5 5 -1 0] */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
232 #define QPEL_CAVSV2(A,B,C,D,E,F,OP) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
233 "movd (%0), "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
234 "movq "#C", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
235 "paddw "#D", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
236 "pmullw %5, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
237 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
238 "punpcklbw %%mm7, "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
239 "psubw "#B", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
240 "psubw "#E", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
241 "paddw %4, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
242 "psraw $3, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
243 "packuswb %%mm6, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
244 OP(%%mm6, (%1), A, d) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
245 "add %3, %1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
246
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
247 /* vertical filter [ 0 -7 42 96 -2 -1] */
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
248 #define QPEL_CAVSV3(A,B,C,D,E,F,OP) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
249 "movd (%0), "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
250 "movq "#C", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
251 "pmullw %6, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
252 "movq "#D", %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
253 "pmullw %5, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
254 "psllw $3, "#B" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
255 "psubw "#B", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
256 "psraw $3, "#B" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
257 "paddw %%mm7, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
258 "paddw "#B", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
259 "paddw "#E", "#E" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
260 "pxor %%mm7, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
261 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
262 "punpcklbw %%mm7, "#F" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
263 "psubw "#E", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
264 "psraw $1, "#E" \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
265 "psubw "#F", %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
266 "paddw %4, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
267 "psraw $7, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
268 "packuswb %%mm6, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
269 OP(%%mm6, (%1), A, d) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
270 "add %3, %1 \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
271
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
272
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
273 #define QPEL_CAVSVNUM(VOP,OP,ADD,MUL1,MUL2)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
274 int w= 2;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
275 src -= 2*srcStride;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
276 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
277 while(w--){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
278 asm volatile(\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
279 "pxor %%mm7, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
280 "movd (%0), %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
281 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
282 "movd (%0), %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
283 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
284 "movd (%0), %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
285 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
286 "movd (%0), %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
287 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
288 "movd (%0), %%mm4 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
289 "add %2, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
290 "punpcklbw %%mm7, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
291 "punpcklbw %%mm7, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
292 "punpcklbw %%mm7, %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
293 "punpcklbw %%mm7, %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
294 "punpcklbw %%mm7, %%mm4 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
295 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
296 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
297 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
298 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
299 VOP(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
300 VOP(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
301 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
302 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
303 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
304 : "+a"(src), "+c"(dst)\
6755
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 6135
diff changeset
305 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ADD), "m"(MUL1), "m"(MUL2)\
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
306 : "memory"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
307 );\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
308 if(h==16){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
309 asm volatile(\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
310 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
311 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
312 VOP(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
313 VOP(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
314 VOP(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
315 VOP(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
316 VOP(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
317 VOP(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
318 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
319 : "+a"(src), "+c"(dst)\
6755
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 6135
diff changeset
320 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ADD), "m"(MUL1), "m"(MUL2)\
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
321 : "memory"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
322 );\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
323 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
324 src += 4-(h+5)*srcStride;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
325 dst += 4-h*dstStride;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
326 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
327
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
328 #define QPEL_CAVS(OPNAME, OP, MMX)\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
329 static void OPNAME ## cavs_qpel8_h_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
330 int h=8;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
331 asm volatile(\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
332 "pxor %%mm7, %%mm7 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
333 "movq %5, %%mm6 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
334 "1: \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
335 "movq (%0), %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
336 "movq 1(%0), %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
337 "movq %%mm0, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
338 "movq %%mm2, %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
339 "punpcklbw %%mm7, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
340 "punpckhbw %%mm7, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
341 "punpcklbw %%mm7, %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
342 "punpckhbw %%mm7, %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
343 "paddw %%mm2, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
344 "paddw %%mm3, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
345 "pmullw %%mm6, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
346 "pmullw %%mm6, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
347 "movq -1(%0), %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
348 "movq 2(%0), %%mm4 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
349 "movq %%mm2, %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
350 "movq %%mm4, %%mm5 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
351 "punpcklbw %%mm7, %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
352 "punpckhbw %%mm7, %%mm3 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
353 "punpcklbw %%mm7, %%mm4 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
354 "punpckhbw %%mm7, %%mm5 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
355 "paddw %%mm4, %%mm2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
356 "paddw %%mm3, %%mm5 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
357 "psubw %%mm2, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
358 "psubw %%mm5, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
359 "movq %6, %%mm5 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
360 "paddw %%mm5, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
361 "paddw %%mm5, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
362 "psraw $3, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
363 "psraw $3, %%mm1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
364 "packuswb %%mm1, %%mm0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
365 OP(%%mm0, (%1),%%mm5, q) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
366 "add %3, %0 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
367 "add %4, %1 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
368 "decl %2 \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
369 " jnz 1b \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
370 : "+a"(src), "+c"(dst), "+m"(h)\
6755
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 6135
diff changeset
371 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_4)\
3524
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
372 : "memory"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
373 );\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
374 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
375 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
376 static inline void OPNAME ## cavs_qpel8or16_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
377 QPEL_CAVSVNUM(QPEL_CAVSV1,OP,ff_pw_64,ff_pw_96,ff_pw_42) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
378 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
379 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
380 static inline void OPNAME ## cavs_qpel8or16_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
381 QPEL_CAVSVNUM(QPEL_CAVSV2,OP,ff_pw_4,ff_pw_5,ff_pw_5) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
382 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
383 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
384 static inline void OPNAME ## cavs_qpel8or16_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
385 QPEL_CAVSVNUM(QPEL_CAVSV3,OP,ff_pw_64,ff_pw_96,ff_pw_42) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
386 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
387 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
388 static void OPNAME ## cavs_qpel8_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
389 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst , src , dstStride, srcStride, 8);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
390 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
391 static void OPNAME ## cavs_qpel16_v1_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
392 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst , src , dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
393 OPNAME ## cavs_qpel8or16_v1_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
394 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
395 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
396 static void OPNAME ## cavs_qpel8_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
397 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst , src , dstStride, srcStride, 8);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
398 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
399 static void OPNAME ## cavs_qpel16_v2_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
400 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst , src , dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
401 OPNAME ## cavs_qpel8or16_v2_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
402 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
403 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
404 static void OPNAME ## cavs_qpel8_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
405 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst , src , dstStride, srcStride, 8);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
406 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
407 static void OPNAME ## cavs_qpel16_v3_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
408 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst , src , dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
409 OPNAME ## cavs_qpel8or16_v3_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
410 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
411 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
412 static void OPNAME ## cavs_qpel16_h_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
413 OPNAME ## cavs_qpel8_h_ ## MMX(dst , src , dstStride, srcStride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
414 OPNAME ## cavs_qpel8_h_ ## MMX(dst+8, src+8, dstStride, srcStride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
415 src += 8*srcStride;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
416 dst += 8*dstStride;\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
417 OPNAME ## cavs_qpel8_h_ ## MMX(dst , src , dstStride, srcStride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
418 OPNAME ## cavs_qpel8_h_ ## MMX(dst+8, src+8, dstStride, srcStride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
419 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
420
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
421 #define CAVS_MC(OPNAME, SIZE, MMX) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
422 static void ff_ ## OPNAME ## cavs_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
423 OPNAME ## cavs_qpel ## SIZE ## _h_ ## MMX(dst, src, stride, stride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
424 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
425 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
426 static void ff_ ## OPNAME ## cavs_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
427 OPNAME ## cavs_qpel ## SIZE ## _v1_ ## MMX(dst, src, stride, stride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
428 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
429 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
430 static void ff_ ## OPNAME ## cavs_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
431 OPNAME ## cavs_qpel ## SIZE ## _v2_ ## MMX(dst, src, stride, stride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
432 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
433 \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
434 static void ff_ ## OPNAME ## cavs_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
435 OPNAME ## cavs_qpel ## SIZE ## _v3_ ## MMX(dst, src, stride, stride);\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
436 }\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
437
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
438 #define PUT_OP(a,b,temp, size) "mov" #size " " #a ", " #b " \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
439 #define AVG_3DNOW_OP(a,b,temp, size) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
440 "mov" #size " " #b ", " #temp " \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
441 "pavgusb " #temp ", " #a " \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
442 "mov" #size " " #a ", " #b " \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
443 #define AVG_MMX2_OP(a,b,temp, size) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
444 "mov" #size " " #b ", " #temp " \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
445 "pavgb " #temp ", " #a " \n\t"\
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
446 "mov" #size " " #a ", " #b " \n\t"
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
447
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
448 QPEL_CAVS(put_, PUT_OP, 3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
449 QPEL_CAVS(avg_, AVG_3DNOW_OP, 3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
450 QPEL_CAVS(put_, PUT_OP, mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
451 QPEL_CAVS(avg_, AVG_MMX2_OP, mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
452
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
453 CAVS_MC(put_, 8, 3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
454 CAVS_MC(put_, 16,3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
455 CAVS_MC(avg_, 8, 3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
456 CAVS_MC(avg_, 16,3dnow)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
457 CAVS_MC(put_, 8, mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
458 CAVS_MC(put_, 16,mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
459 CAVS_MC(avg_, 8, mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
460 CAVS_MC(avg_, 16,mmx2)
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
461
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
462 void ff_put_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
463 void ff_avg_cavs_qpel8_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
464 void ff_put_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
465 void ff_avg_cavs_qpel16_mc00_mmx2(uint8_t *dst, uint8_t *src, int stride);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
466
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
467 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx) {
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
468 #define dspfunc(PFX, IDX, NUM) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
469 c->PFX ## _pixels_tab[IDX][ 0] = ff_ ## PFX ## NUM ## _mc00_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
470 c->PFX ## _pixels_tab[IDX][ 2] = ff_ ## PFX ## NUM ## _mc20_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
471 c->PFX ## _pixels_tab[IDX][ 4] = ff_ ## PFX ## NUM ## _mc01_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
472 c->PFX ## _pixels_tab[IDX][ 8] = ff_ ## PFX ## NUM ## _mc02_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
473 c->PFX ## _pixels_tab[IDX][12] = ff_ ## PFX ## NUM ## _mc03_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
474
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
475 dspfunc(put_cavs_qpel, 0, 16);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
476 dspfunc(put_cavs_qpel, 1, 8);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
477 dspfunc(avg_cavs_qpel, 0, 16);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
478 dspfunc(avg_cavs_qpel, 1, 8);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
479 #undef dspfunc
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
480 c->cavs_idct8_add = cavs_idct8_add_mmx;
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
481 }
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
482
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
483 void ff_cavsdsp_init_3dnow(DSPContext* c, AVCodecContext *avctx) {
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
484 #define dspfunc(PFX, IDX, NUM) \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
485 c->PFX ## _pixels_tab[IDX][ 0] = ff_ ## PFX ## NUM ## _mc00_mmx2; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
486 c->PFX ## _pixels_tab[IDX][ 2] = ff_ ## PFX ## NUM ## _mc20_3dnow; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
487 c->PFX ## _pixels_tab[IDX][ 4] = ff_ ## PFX ## NUM ## _mc01_3dnow; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
488 c->PFX ## _pixels_tab[IDX][ 8] = ff_ ## PFX ## NUM ## _mc02_3dnow; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
489 c->PFX ## _pixels_tab[IDX][12] = ff_ ## PFX ## NUM ## _mc03_3dnow; \
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
490
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
491 dspfunc(put_cavs_qpel, 0, 16);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
492 dspfunc(put_cavs_qpel, 1, 8);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
493 dspfunc(avg_cavs_qpel, 0, 16);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
494 dspfunc(avg_cavs_qpel, 1, 8);
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
495 #undef dspfunc
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
496 c->cavs_idct8_add = cavs_idct8_add_mmx;
419409926166 some MMX optimizations for the CAVS decoder
stefang
parents:
diff changeset
497 }