annotate x86/dct32_sse.c @ 12475:9fef0a8ddd63 libavcodec

Move mm_support() from libavcodec to libavutil, make it a public function and rename it to av_get_cpu_flags().
author stefano
date Wed, 08 Sep 2010 15:07:14 +0000
parents db9ef48dc0e4
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12100
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
1 /*
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
2 * 32 point SSE-optimized DCT transform
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
3 * Copyright (c) 2010 Vitor Sessak
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
4 *
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
5 * This file is part of FFmpeg.
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
6 *
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
11 *
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
15 * Lesser General Public License for more details.
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
16 *
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
20 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
21
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
22 #include <stdint.h>
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
23
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
24 #include "libavutil/x86_cpu.h"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
25 #include "libavutil/mem.h"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
26 #include "libavcodec/dsputil.h"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
27 #include "fft.h"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
28
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
29 DECLARE_ALIGNED(16, static const float, b1)[] = {
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
30 0.500603, 0.505471, 0.515447, 0.531043,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
31 0.553104, 0.582935, 0.622504, 0.674808,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
32 -1.169440, -0.972568, -0.839350, -0.744536,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
33 -10.190008, -3.407609, -2.057781, -1.484165,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
34 0.502419, 0.522499, 0.566944, 0.646822,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
35 0.788155, 1.060678, 1.722447, 5.101149,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
36 0.509796, 0.601345, 0.899976, 2.562916,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
37 1.000000, 1.000000, 1.306563, 0.541196,
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
38 1.000000, 0.707107, 1.000000, -0.707107
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
39 };
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
40
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
41 DECLARE_ALIGNED(16, static const int32_t, smask)[4] = {
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
42 0, 0, 0x80000000, 0x80000000
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
43 };
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
44
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
45 /* butterfly operator */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
46 #define BUTTERFLY(a,b,c,tmp) \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
47 "movaps %%" #a ", %%" #tmp " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
48 "subps %%" #b ", %%" #a " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
49 "addps %%" #tmp ", %%" #b " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
50 "mulps " #c ", %%" #a " \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
51
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
52 ///* Same as BUTTERFLY when vectors a and b overlap */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
53 #define BUTTERFLY0(val, mask, cos, tmp, shuf) \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
54 "movaps %%" #val ", %%" #tmp " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
55 "shufps " #shuf ", %%" #val ",%%" #val " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
56 "xorps %%" #mask ", %%" #tmp " \n\t" /* flip signs */ \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
57 "addps %%" #tmp ", %%" #val " \n\t" \
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
58 "mulps %%" #cos ", %%" #val " \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
59
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
60 #define BUTTERFLY2(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0x1b)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
61 #define BUTTERFLY3(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0xb1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
62
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
63 void ff_dct32_float_sse(FFTSample *out, const FFTSample *in)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
64 {
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
65 int32_t tmp1 = 0;
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
66 __asm__ volatile(
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
67 /* pass 1 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
68
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
69 "movaps (%4), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
70 "movaps 112(%4), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
71 "shufps $0x1b, %%xmm1, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
72 BUTTERFLY(xmm0, xmm1, (%2), xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
73
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
74 "movaps 64(%4), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
75 "movaps 48(%4), %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
76 "shufps $0x1b, %%xmm4, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
77 BUTTERFLY(xmm7, xmm4, 48(%2), xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
78
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
79
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
80 /* pass 2 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
81 "movaps 64(%2), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
82 BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
83 "movaps %%xmm1, 48(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
84 "movaps %%xmm4, (%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
85
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
86 /* pass 1 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
87 "movaps 16(%4), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
88 "movaps 96(%4), %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
89 "shufps $0x1b, %%xmm6, %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
90 BUTTERFLY(xmm1, xmm6, 16(%2), xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
91
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
92 "movaps 80(%4), %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
93 "movaps 32(%4), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
94 "shufps $0x1b, %%xmm5, %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
95 BUTTERFLY(xmm4, xmm5, 32(%2), xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
96
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
97 /* pass 2 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
98 BUTTERFLY(xmm0, xmm7, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
99
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
100 "movaps 80(%2), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
101 BUTTERFLY(xmm6, xmm5, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
102
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
103 BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
104
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
105 /* pass 3 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
106 "movaps 96(%2), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
107 "shufps $0x1b, %%xmm1, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
108 BUTTERFLY(xmm0, xmm1, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
109 "movaps %%xmm0, 112(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
110 "movaps %%xmm1, 96(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
111
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
112 "movaps 0(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
113 "shufps $0x1b, %%xmm5, %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
114 BUTTERFLY(xmm0, xmm5, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
115
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
116 "movaps 48(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
117 "shufps $0x1b, %%xmm6, %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
118 BUTTERFLY(xmm1, xmm6, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
119 "movaps %%xmm1, 48(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
120
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
121 "shufps $0x1b, %%xmm4, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
122 BUTTERFLY(xmm7, xmm4, %%xmm2, xmm3)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
123
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
124 /* pass 4 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
125 "movaps (%3), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
126 "movaps 112(%2), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
127
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
128 BUTTERFLY2(xmm5, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
129
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
130 BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
131 "movaps %%xmm0, 16(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
132
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
133 BUTTERFLY2(xmm6, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
134 "movaps %%xmm6, 32(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
135
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
136 "movaps 48(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
137 BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
138 "movaps %%xmm0, 48(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
139
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
140 BUTTERFLY2(xmm4, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
141
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
142 BUTTERFLY2(xmm7, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
143
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
144 "movaps 96(%1), %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
145 BUTTERFLY2(xmm6, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
146
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
147 "movaps 112(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
148 BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
149
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
150 /* pass 5 */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
151 "movaps 128(%2), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
152 "shufps $0xCC, %%xmm3,%%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
153
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
154 BUTTERFLY3(xmm5, xmm3, xmm2, xmm1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
155 "movaps %%xmm5, (%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
156
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
157 "movaps 16(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
158 BUTTERFLY3(xmm1, xmm3, xmm2, xmm5)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
159 "movaps %%xmm1, 16(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
160
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
161 BUTTERFLY3(xmm4, xmm3, xmm2, xmm5)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
162 "movaps %%xmm4, 64(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
163
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
164 BUTTERFLY3(xmm7, xmm3, xmm2, xmm5)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
165 "movaps %%xmm7, 80(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
166
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
167 "movaps 32(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
168 BUTTERFLY3(xmm5, xmm3, xmm2, xmm7)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
169 "movaps %%xmm5, 32(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
170
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
171 "movaps 48(%1), %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
172 BUTTERFLY3(xmm4, xmm3, xmm2, xmm7)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
173 "movaps %%xmm4, 48(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
174
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
175 BUTTERFLY3(xmm6, xmm3, xmm2, xmm7)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
176 "movaps %%xmm6, 96(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
177
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
178 BUTTERFLY3(xmm0, xmm3, xmm2, xmm7)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
179 "movaps %%xmm0, 112(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
180
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
181
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
182 /* pass 6, no SIMD... */
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
183 "movss 56(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
184 "movl 4(%1), %0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
185 "addss 60(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
186 "movss 72(%1), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
187 "addss %%xmm3, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
188 "movss 52(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
189 "addss %%xmm3, %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
190 "movss 24(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
191 "addss 28(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
192 "addss 76(%1), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
193 "addss %%xmm3, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
194 "addss %%xmm4, %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
195 "movss %%xmm1, 16(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
196 "movss 20(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
197 "addss %%xmm3, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
198 "movss 40(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
199 "movss %%xmm1, 48(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
200 "addss 44(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
201 "movss 20(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
202 "addss %%xmm3, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
203 "addss %%xmm2, %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
204 "addss 28(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
205 "movss %%xmm3, 40(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
206 "addss 36(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
207 "movss 8(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
208 "movss %%xmm2, 56(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
209 "addss 12(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
210 "movss %%xmm5, 8(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
211 "movss %%xmm3, 32(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
212 "movss 52(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
213 "movss 80(%1), %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
214 "movss 120(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
215 "movss %%xmm1, 80(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
216 "movss %%xmm4, 24(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
217 "addss 124(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
218 "movss 64(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
219 "addss 60(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
220 "addss %%xmm5, %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
221 "addss 116(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
222 "movl %0, 64(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
223 "addss %%xmm0, %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
224 "addss %%xmm6, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
225 "movl 12(%1), %0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
226 "movss %%xmm1, 4(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
227 "movss 88(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
228 "movl %0, 96(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
229 "addss 92(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
230 "movss 104(%1), %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
231 "movl 28(%1), %0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
232 "addss 108(%1), %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
233 "addss %%xmm4, %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
234 "addss %%xmm1, %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
235 "addss 84(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
236 "addss %%xmm5, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
237 "addss %%xmm3, %%xmm6 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
238 "addss %%xmm0, %%xmm3 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
239 "addss %%xmm7, %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
240 "addss 100(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
241 "addss %%xmm4, %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
242 "movl %0, 112(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
243 "movss %%xmm0, 28(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
244 "movss 36(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
245 "movss %%xmm7, 36(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
246 "addss %%xmm1, %%xmm4 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
247 "movss 116(%1), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
248 "addss %%xmm2, %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
249 "addss 124(%1), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
250 "movss %%xmm0, 72(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
251 "movss 44(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
252 "movss %%xmm6, 12(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
253 "movss %%xmm3, 20(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
254 "addss %%xmm0, %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
255 "movss %%xmm4, 44(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
256 "movss %%xmm2, 88(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
257 "addss 60(%1), %%xmm0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
258 "movl 60(%1), %0 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
259 "movl %0, 120(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
260 "movss %%xmm0, 104(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
261 "addss %%xmm5, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
262 "addss 68(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
263 "movss %%xmm1, 52(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
264 "movss %%xmm5, 60(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
265 "movss 68(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
266 "movss 100(%1), %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
267 "addss %%xmm7, %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
268 "addss 108(%1), %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
269 "addss %%xmm5, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
270 "movss 84(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
271 "addss 92(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
272 "addss %%xmm2, %%xmm5 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
273 "movss %%xmm1, 68(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
274 "addss %%xmm7, %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
275 "movss 76(%1), %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
276 "movss %%xmm2, 84(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
277 "movss %%xmm5, 76(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
278 "movss 108(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
279 "addss %%xmm1, %%xmm7 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
280 "addss 124(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
281 "addss %%xmm2, %%xmm1 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
282 "addss 92(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
283 "movss %%xmm1, 100(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
284 "movss %%xmm2, 108(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
285 "movss 92(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
286 "movss %%xmm7, 92(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
287 "addss 124(%1), %%xmm2 \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
288 "movss %%xmm2, 116(%1) \n\t"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
289 :"+&r"(tmp1)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
290 :"r"(out), "r"(b1), "r"(smask), "r"(in)
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
291 :"memory"
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
292 );
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
293 }
db9ef48dc0e4 Move SSE optimized 32-point DCT to its own file. Should fix breakage with YASM
vitor
parents:
diff changeset
294