annotate libswscale/rgb2rgb_template.c @ 22321:5f47768cbda5

Add av_ prefix to clip functions
author reimar
date Sun, 25 Feb 2007 10:27:43 +0000
parents 1f2ba24b4e47
children 2d1ad4285df4
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1 /*
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
3 * rgb2rgb.c, Software RGB to RGB convertor
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
4 * pluralize by Software PAL8 to RGB convertor
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
5 * Software YUV to YUV convertor
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
6 * Software YUV to RGB convertor
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
7 * Written by Nick Kurshev.
19703
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
8 * palette & YUV & runtime CPU stuff by Michael (michaelni@gmx.at)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
9 * lot of big-endian byteorder fixes by Alex Beregszaszi
19703
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
10 *
20094
aca9e9783f67 Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents: 19703
diff changeset
11 * This file is part of FFmpeg.
aca9e9783f67 Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents: 19703
diff changeset
12 *
aca9e9783f67 Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents: 19703
diff changeset
13 * FFmpeg is free software; you can redistribute it and/or modify
19703
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
14 * it under the terms of the GNU General Public License as published by
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
15 * the Free Software Foundation; either version 2 of the License, or
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
16 * (at your option) any later version.
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
17 *
20094
aca9e9783f67 Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents: 19703
diff changeset
18 * FFmpeg is distributed in the hope that it will be useful,
19703
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
21 * GNU General Public License for more details.
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
22 *
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
23 * You should have received a copy of the GNU General Public License
20094
aca9e9783f67 Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents: 19703
diff changeset
24 * along with FFmpeg; if not, write to the Free Software
19703
ad7f49a1ba95 Add official GPL header to make license explicit as discussed on ffmpeg-devel.
diego
parents: 19396
diff changeset
25 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
21029
1f2ba24b4e47 Clarify that some of the non-SIMD code is now LGPLed.
lucabe
parents: 20724
diff changeset
26 *
1f2ba24b4e47 Clarify that some of the non-SIMD code is now LGPLed.
lucabe
parents: 20724
diff changeset
27 * the C code (not assembly, mmx, ...) of this file can be used
1f2ba24b4e47 Clarify that some of the non-SIMD code is now LGPLed.
lucabe
parents: 20724
diff changeset
28 * under the LGPL license too
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
29 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
30
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
31 #include <stddef.h>
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
32 #include <inttypes.h> /* for __WORDSIZE */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
33
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
34 #ifndef __WORDSIZE
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
35 // #warning You have misconfigured system and probably will lose performance!
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
36 #define __WORDSIZE MP_WORDSIZE
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
37 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
38
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
39 #undef PREFETCH
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
40 #undef MOVNTQ
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
41 #undef EMMS
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
42 #undef SFENCE
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
43 #undef MMREG_SIZE
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
44 #undef PREFETCHW
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
45 #undef PAVGB
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
46
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
47 #ifdef HAVE_SSE2
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
48 #define MMREG_SIZE 16
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
49 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
50 #define MMREG_SIZE 8
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
51 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
52
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
53 #ifdef HAVE_3DNOW
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
54 #define PREFETCH "prefetch"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
55 #define PREFETCHW "prefetchw"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
56 #define PAVGB "pavgusb"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
57 #elif defined ( HAVE_MMX2 )
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
58 #define PREFETCH "prefetchnta"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
59 #define PREFETCHW "prefetcht0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
60 #define PAVGB "pavgb"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
61 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
62 #ifdef __APPLE__
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
63 #define PREFETCH "#"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
64 #define PREFETCHW "#"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
65 #else
20724
b8fe18a742ce Fix MacIntel build: "/nop" is illegal on Apple's older version of GAS
gpoirier
parents: 20094
diff changeset
66 #define PREFETCH " # nop"
b8fe18a742ce Fix MacIntel build: "/nop" is illegal on Apple's older version of GAS
gpoirier
parents: 20094
diff changeset
67 #define PREFETCHW " # nop"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
68 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
69 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
70
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
71 #ifdef HAVE_3DNOW
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
72 /* On K6 femms is faster of emms. On K7 femms is directly mapped on emms. */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
73 #define EMMS "femms"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
74 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
75 #define EMMS "emms"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
76 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
77
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
78 #ifdef HAVE_MMX2
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
79 #define MOVNTQ "movntq"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
80 #define SFENCE "sfence"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
81 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
82 #define MOVNTQ "movq"
20724
b8fe18a742ce Fix MacIntel build: "/nop" is illegal on Apple's older version of GAS
gpoirier
parents: 20094
diff changeset
83 #define SFENCE " # nop"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
84 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
85
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
86 static inline void RENAME(rgb24to32)(const uint8_t *src,uint8_t *dst,long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
87 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
88 uint8_t *dest = dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
89 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
90 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
91 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
92 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
93 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
94 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
95 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
96 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
97 mm_end = end - 23;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
98 __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
99 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
100 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
101 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
102 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
103 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
104 "punpckldq 3%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
105 "movd 6%1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
106 "punpckldq 9%1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
107 "movd 12%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
108 "punpckldq 15%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
109 "movd 18%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
110 "punpckldq 21%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
111 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
112 "pand %%mm7, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
113 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
114 "pand %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
115 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
116 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
117 MOVNTQ" %%mm2, 16%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
118 MOVNTQ" %%mm3, 24%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
119 :"=m"(*dest)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
120 :"m"(*s)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
121 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
122 dest += 32;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
123 s += 24;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
124 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
125 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
126 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
127 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
128 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
129 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
130 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
131 /* RGB24 (= R,G,B) -> RGB32 (= A,B,G,R) */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
132 *dest++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
133 *dest++ = s[2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
134 *dest++ = s[1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
135 *dest++ = s[0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
136 s+=3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
137 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
138 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
139 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
140 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
141 *dest++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
142 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
143 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
144 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
145
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
146 static inline void RENAME(rgb32to24)(const uint8_t *src,uint8_t *dst,long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
147 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
148 uint8_t *dest = dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
149 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
150 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
151 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
152 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
153 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
154 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
155 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
156 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
157 mm_end = end - 31;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
158 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
159 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
160 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
161 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
162 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
163 "movq 8%1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
164 "movq 16%1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
165 "movq 24%1, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
166 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
167 "movq %%mm1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
168 "movq %%mm4, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
169 "movq %%mm5, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
170 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
171 "psrlq $8, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
172 "psrlq $8, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
173 "psrlq $8, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
174 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
175 "pand %2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
176 "pand %2, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
177 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
178 "pand %3, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
179 "pand %3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
180 "pand %3, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
181 "pand %3, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
182 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
183 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
184 "por %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
185 "por %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
186
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
187 "movq %%mm1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
188 "movq %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
189 "psllq $48, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
190 "psllq $32, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
191 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
192 "pand %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
193 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
194 "psrlq $16, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
195 "psrlq $32, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
196 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
197 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
198 "pand %6, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
199 "por %%mm5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
200
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
201 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
202 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
203 MOVNTQ" %%mm4, 16%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
204 :"=m"(*dest)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
205 :"m"(*s),"m"(mask24l),
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
206 "m"(mask24h),"m"(mask24hh),"m"(mask24hhh),"m"(mask24hhhh)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
207 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
208 dest += 24;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
209 s += 32;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
210 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
211 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
212 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
213 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
214 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
215 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
216 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
217 /* RGB32 (= A,B,G,R) -> RGB24 (= R,G,B) */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
218 s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
219 dest[2] = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
220 dest[1] = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
221 dest[0] = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
222 dest += 3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
223 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
224 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
225 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
226 *dest++ = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
227 s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
228 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
229 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
230 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
231
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
232 /*
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
233 Original by Strepto/Astral
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
234 ported to gcc & bugfixed : A'rpi
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
235 MMX2, 3DNOW optimization by Nick Kurshev
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
236 32bit c version, and and&add trick by Michael Niedermayer
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
237 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
238 static inline void RENAME(rgb15to16)(const uint8_t *src,uint8_t *dst,long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
239 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
240 register const uint8_t* s=src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
241 register uint8_t* d=dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
242 register const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
243 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
244 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
245 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
246 __asm __volatile(PREFETCH" %0"::"m"(*s));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
247 __asm __volatile("movq %0, %%mm4"::"m"(mask15s));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
248 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
249 while(s<mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
250 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
251 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
252 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
253 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
254 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
255 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
256 "movq %%mm2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
257 "pand %%mm4, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
258 "pand %%mm4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
259 "paddw %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
260 "paddw %%mm3, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
261 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
262 MOVNTQ" %%mm2, 8%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
263 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
264 :"m"(*s)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
265 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
266 d+=16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
267 s+=16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
268 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
269 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
270 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
271 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
272 mm_end = end - 3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
273 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
274 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
275 register unsigned x= *((uint32_t *)s);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
276 *((uint32_t *)d) = (x&0x7FFF7FFF) + (x&0x7FE07FE0);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
277 d+=4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
278 s+=4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
279 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
280 if(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
281 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
282 register unsigned short x= *((uint16_t *)s);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
283 *((uint16_t *)d) = (x&0x7FFF) + (x&0x7FE0);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
284 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
285 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
286
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
287 static inline void RENAME(rgb16to15)(const uint8_t *src,uint8_t *dst,long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
288 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
289 register const uint8_t* s=src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
290 register uint8_t* d=dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
291 register const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
292 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
293 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
294 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
295 __asm __volatile(PREFETCH" %0"::"m"(*s));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
296 __asm __volatile("movq %0, %%mm7"::"m"(mask15rg));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
297 __asm __volatile("movq %0, %%mm6"::"m"(mask15b));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
298 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
299 while(s<mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
300 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
301 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
302 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
303 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
304 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
305 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
306 "movq %%mm2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
307 "psrlq $1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
308 "psrlq $1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
309 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
310 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
311 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
312 "pand %%mm6, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
313 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
314 "por %%mm3, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
315 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
316 MOVNTQ" %%mm2, 8%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
317 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
318 :"m"(*s)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
319 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
320 d+=16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
321 s+=16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
322 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
323 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
324 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
325 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
326 mm_end = end - 3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
327 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
328 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
329 register uint32_t x= *((uint32_t *)s);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
330 *((uint32_t *)d) = ((x>>1)&0x7FE07FE0) | (x&0x001F001F);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
331 s+=4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
332 d+=4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
333 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
334 if(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
335 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
336 register uint16_t x= *((uint16_t *)s);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
337 *((uint16_t *)d) = ((x>>1)&0x7FE0) | (x&0x001F);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
338 s+=2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
339 d+=2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
340 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
341 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
342
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
343 static inline void RENAME(rgb32to16)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
344 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
345 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
346 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
347 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
348 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
349 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
350 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
351 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
352 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
353 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
354 #if 1 //is faster only if multiplies are reasonable fast (FIXME figure out on which cpus this is faster, on Athlon its slightly faster)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
355 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
356 "movq %3, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
357 "movq %4, %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
358 "movq %5, %%mm7 \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
359 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
360 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
361 PREFETCH" 32(%1) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
362 "movd (%1), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
363 "movd 4(%1), %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
364 "punpckldq 8(%1), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
365 "punpckldq 12(%1), %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
366 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
367 "movq %%mm3, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
368 "pand %%mm6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
369 "pand %%mm6, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
370 "pmaddwd %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
371 "pmaddwd %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
372 "pand %%mm5, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
373 "pand %%mm5, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
374 "por %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
375 "por %%mm4, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
376 "psrld $5, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
377 "pslld $11, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
378 "por %%mm3, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
379 MOVNTQ" %%mm0, (%0) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
380 "add $16, %1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
381 "add $8, %0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
382 "cmp %2, %1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
383 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
384 : "+r" (d), "+r"(s)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
385 : "r" (mm_end), "m" (mask3216g), "m" (mask3216br), "m" (mul3216)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
386 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
387 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
388 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
389 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
390 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
391 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
392 ::"m"(red_16mask),"m"(green_16mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
393 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
394 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
395 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
396 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
397 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
398 "movd 4%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
399 "punpckldq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
400 "punpckldq 12%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
401 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
402 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
403 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
404 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
405 "psrlq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
406 "psrlq $3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
407 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
408 "pand %2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
409 "psrlq $5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
410 "psrlq $5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
411 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
412 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
413 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
414 "psrlq $8, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
415 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
416 "pand %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
417 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
418 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
419 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
420 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
421 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
422 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
423 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
424 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
425 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
426 s += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
427 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
428 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
429 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
430 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
431 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
432 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
433 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
434 register int rgb = *(uint32_t*)s; s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
435 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>8);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
436 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
437 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
438
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
439 static inline void RENAME(rgb32tobgr16)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
440 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
441 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
442 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
443 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
444 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
445 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
446 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
447 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
448 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
449 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
450 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
451 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
452 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
453 ::"m"(red_16mask),"m"(green_16mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
454 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
455 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
456 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
457 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
458 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
459 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
460 "movd 4%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
461 "punpckldq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
462 "punpckldq 12%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
463 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
464 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
465 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
466 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
467 "psllq $8, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
468 "psllq $8, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
469 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
470 "pand %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
471 "psrlq $5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
472 "psrlq $5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
473 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
474 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
475 "psrlq $19, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
476 "psrlq $19, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
477 "pand %2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
478 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
479 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
480 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
481 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
482 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
483 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
484 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
485 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
486 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
487 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
488 s += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
489 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
490 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
491 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
492 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
493 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
494 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
495 register int rgb = *(uint32_t*)s; s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
496 *d++ = ((rgb&0xF8)<<8) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>19);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
497 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
498 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
499
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
500 static inline void RENAME(rgb32to15)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
501 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
502 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
503 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
504 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
505 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
506 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
507 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
508 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
509 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
510 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
511 #if 1 //is faster only if multiplies are reasonable fast (FIXME figure out on which cpus this is faster, on Athlon its slightly faster)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
512 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
513 "movq %3, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
514 "movq %4, %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
515 "movq %5, %%mm7 \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
516 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
517 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
518 PREFETCH" 32(%1) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
519 "movd (%1), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
520 "movd 4(%1), %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
521 "punpckldq 8(%1), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
522 "punpckldq 12(%1), %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
523 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
524 "movq %%mm3, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
525 "pand %%mm6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
526 "pand %%mm6, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
527 "pmaddwd %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
528 "pmaddwd %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
529 "pand %%mm5, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
530 "pand %%mm5, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
531 "por %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
532 "por %%mm4, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
533 "psrld $6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
534 "pslld $10, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
535 "por %%mm3, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
536 MOVNTQ" %%mm0, (%0) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
537 "add $16, %1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
538 "add $8, %0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
539 "cmp %2, %1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
540 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
541 : "+r" (d), "+r"(s)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
542 : "r" (mm_end), "m" (mask3215g), "m" (mask3216br), "m" (mul3215)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
543 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
544 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
545 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
546 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
547 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
548 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
549 ::"m"(red_15mask),"m"(green_15mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
550 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
551 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
552 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
553 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
554 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
555 "movd 4%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
556 "punpckldq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
557 "punpckldq 12%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
558 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
559 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
560 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
561 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
562 "psrlq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
563 "psrlq $3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
564 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
565 "pand %2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
566 "psrlq $6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
567 "psrlq $6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
568 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
569 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
570 "psrlq $9, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
571 "psrlq $9, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
572 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
573 "pand %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
574 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
575 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
576 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
577 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
578 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
579 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
580 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
581 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
582 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
583 s += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
584 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
585 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
586 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
587 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
588 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
589 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
590 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
591 register int rgb = *(uint32_t*)s; s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
592 *d++ = ((rgb&0xFF)>>3) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>9);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
593 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
594 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
595
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
596 static inline void RENAME(rgb32tobgr15)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
597 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
598 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
599 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
600 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
601 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
602 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
603 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
604 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
605 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
606 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
607 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
608 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
609 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
610 ::"m"(red_15mask),"m"(green_15mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
611 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
612 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
613 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
614 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
615 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
616 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
617 "movd 4%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
618 "punpckldq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
619 "punpckldq 12%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
620 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
621 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
622 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
623 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
624 "psllq $7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
625 "psllq $7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
626 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
627 "pand %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
628 "psrlq $6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
629 "psrlq $6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
630 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
631 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
632 "psrlq $19, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
633 "psrlq $19, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
634 "pand %2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
635 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
636 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
637 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
638 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
639 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
640 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
641 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
642 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
643 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
644 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
645 s += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
646 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
647 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
648 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
649 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
650 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
651 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
652 register int rgb = *(uint32_t*)s; s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
653 *d++ = ((rgb&0xF8)<<7) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>19);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
654 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
655 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
656
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
657 static inline void RENAME(rgb24to16)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
658 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
659 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
660 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
661 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
662 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
663 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
664 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
665 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
666 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
667 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
668 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
669 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
670 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
671 ::"m"(red_16mask),"m"(green_16mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
672 mm_end = end - 11;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
673 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
674 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
675 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
676 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
677 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
678 "movd 3%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
679 "punpckldq 6%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
680 "punpckldq 9%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
681 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
682 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
683 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
684 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
685 "psrlq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
686 "psrlq $3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
687 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
688 "pand %2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
689 "psrlq $5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
690 "psrlq $5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
691 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
692 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
693 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
694 "psrlq $8, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
695 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
696 "pand %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
697 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
698 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
699 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
700 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
701 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
702 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
703 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
704 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
705 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
706 s += 12;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
707 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
708 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
709 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
710 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
711 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
712 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
713 const int b= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
714 const int g= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
715 const int r= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
716 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
717 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
718 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
719
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
720 static inline void RENAME(rgb24tobgr16)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
721 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
722 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
723 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
724 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
725 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
726 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
727 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
728 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
729 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
730 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
731 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
732 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
733 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
734 ::"m"(red_16mask),"m"(green_16mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
735 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
736 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
737 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
738 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
739 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
740 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
741 "movd 3%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
742 "punpckldq 6%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
743 "punpckldq 9%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
744 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
745 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
746 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
747 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
748 "psllq $8, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
749 "psllq $8, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
750 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
751 "pand %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
752 "psrlq $5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
753 "psrlq $5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
754 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
755 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
756 "psrlq $19, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
757 "psrlq $19, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
758 "pand %2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
759 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
760 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
761 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
762 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
763 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
764 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
765 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
766 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
767 :"=m"(*d):"m"(*s),"m"(blue_16mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
768 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
769 s += 12;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
770 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
771 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
772 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
773 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
774 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
775 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
776 const int r= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
777 const int g= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
778 const int b= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
779 *d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
780 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
781 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
782
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
783 static inline void RENAME(rgb24to15)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
784 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
785 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
786 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
787 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
788 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
789 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
790 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
791 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
792 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
793 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
794 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
795 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
796 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
797 ::"m"(red_15mask),"m"(green_15mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
798 mm_end = end - 11;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
799 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
800 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
801 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
802 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
803 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
804 "movd 3%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
805 "punpckldq 6%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
806 "punpckldq 9%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
807 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
808 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
809 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
810 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
811 "psrlq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
812 "psrlq $3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
813 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
814 "pand %2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
815 "psrlq $6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
816 "psrlq $6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
817 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
818 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
819 "psrlq $9, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
820 "psrlq $9, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
821 "pand %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
822 "pand %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
823 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
824 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
825 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
826 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
827 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
828 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
829 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
830 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
831 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
832 s += 12;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
833 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
834 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
835 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
836 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
837 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
838 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
839 const int b= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
840 const int g= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
841 const int r= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
842 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
843 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
844 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
845
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
846 static inline void RENAME(rgb24tobgr15)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
847 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
848 const uint8_t *s = src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
849 const uint8_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
850 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
851 const uint8_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
852 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
853 uint16_t *d = (uint16_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
854 end = s + src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
855 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
856 __asm __volatile(PREFETCH" %0"::"m"(*src):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
857 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
858 "movq %0, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
859 "movq %1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
860 ::"m"(red_15mask),"m"(green_15mask));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
861 mm_end = end - 15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
862 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
863 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
864 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
865 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
866 "movd %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
867 "movd 3%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
868 "punpckldq 6%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
869 "punpckldq 9%1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
870 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
871 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
872 "movq %%mm3, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
873 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
874 "psllq $7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
875 "psllq $7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
876 "pand %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
877 "pand %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
878 "psrlq $6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
879 "psrlq $6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
880 "pand %%mm6, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
881 "pand %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
882 "psrlq $19, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
883 "psrlq $19, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
884 "pand %2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
885 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
886 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
887 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
888 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
889 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
890 "psllq $16, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
891 "por %%mm3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
892 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
893 :"=m"(*d):"m"(*s),"m"(blue_15mask):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
894 d += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
895 s += 12;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
896 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
897 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
898 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
899 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
900 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
901 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
902 const int r= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
903 const int g= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
904 const int b= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
905 *d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
906 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
907 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
908
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
909 /*
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
910 I use here less accurate approximation by simply
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
911 left-shifting the input
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
912 value and filling the low order bits with
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
913 zeroes. This method improves png's
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
914 compression but this scheme cannot reproduce white exactly, since it does not
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
915 generate an all-ones maximum value; the net effect is to darken the
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
916 image slightly.
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
917
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
918 The better method should be "left bit replication":
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
919
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
920 4 3 2 1 0
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
921 ---------
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
922 1 1 0 1 1
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
923
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
924 7 6 5 4 3 2 1 0
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
925 ----------------
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
926 1 1 0 1 1 1 1 0
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
927 |=======| |===|
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
928 | Leftmost Bits Repeated to Fill Open Bits
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
929 |
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
930 Original Bits
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
931 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
932 static inline void RENAME(rgb15to24)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
933 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
934 const uint16_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
935 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
936 const uint16_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
937 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
938 uint8_t *d = (uint8_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
939 const uint16_t *s = (uint16_t *)src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
940 end = s + src_size/2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
941 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
942 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
943 mm_end = end - 7;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
944 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
945 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
946 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
947 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
948 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
949 "movq %1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
950 "movq %1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
951 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
952 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
953 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
954 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
955 "psrlq $2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
956 "psrlq $7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
957 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
958 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
959 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
960 "punpcklwd %5, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
961 "punpcklwd %5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
962 "punpcklwd %5, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
963 "punpckhwd %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
964 "punpckhwd %5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
965 "punpckhwd %5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
966 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
967 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
968 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
969 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
970 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
971 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
972 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
973 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
974
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
975 "movq %%mm0, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
976 "movq %%mm3, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
977
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
978 "movq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
979 "movq 8%1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
980 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
981 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
982 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
983 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
984 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
985 "psrlq $2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
986 "psrlq $7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
987 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
988 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
989 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
990 "punpcklwd %5, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
991 "punpcklwd %5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
992 "punpcklwd %5, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
993 "punpckhwd %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
994 "punpckhwd %5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
995 "punpckhwd %5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
996 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
997 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
998 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
999 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1000 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1001 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1002 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1003 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1004
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1005 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1006 :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r), "m"(mmx_null)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1007 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1008 /* Borrowed 32 to 24 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1009 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1010 "movq %%mm0, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1011 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1012 "movq %%mm6, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1013 "movq %%mm7, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1014
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1015 "movq %%mm4, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1016 "movq %%mm5, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1017 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1018 "movq %%mm1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1019
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1020 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1021 "psrlq $8, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1022 "psrlq $8, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1023 "psrlq $8, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1024 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1025 "pand %2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1026 "pand %2, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1027 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1028 "pand %3, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1029 "pand %3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1030 "pand %3, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1031 "pand %3, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1032 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1033 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1034 "por %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1035 "por %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1036
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1037 "movq %%mm1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1038 "movq %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1039 "psllq $48, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1040 "psllq $32, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1041 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1042 "pand %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1043 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1044 "psrlq $16, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1045 "psrlq $32, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1046 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1047 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1048 "pand %6, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1049 "por %%mm5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1050
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1051 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1052 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1053 MOVNTQ" %%mm4, 16%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1054
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1055 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1056 :"m"(*s),"m"(mask24l),"m"(mask24h),"m"(mask24hh),"m"(mask24hhh),"m"(mask24hhhh)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1057 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1058 d += 24;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1059 s += 8;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1060 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1061 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1062 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1063 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1064 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1065 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1066 register uint16_t bgr;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1067 bgr = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1068 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1069 *d++ = (bgr&0x3E0)>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1070 *d++ = (bgr&0x7C00)>>7;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1071 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1072 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1073
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1074 static inline void RENAME(rgb16to24)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1075 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1076 const uint16_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1077 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1078 const uint16_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1079 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1080 uint8_t *d = (uint8_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1081 const uint16_t *s = (const uint16_t *)src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1082 end = s + src_size/2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1083 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1084 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1085 mm_end = end - 7;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1086 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1087 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1088 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1089 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1090 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1091 "movq %1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1092 "movq %1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1093 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1094 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1095 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1096 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1097 "psrlq $3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1098 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1099 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1100 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1101 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1102 "punpcklwd %5, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1103 "punpcklwd %5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1104 "punpcklwd %5, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1105 "punpckhwd %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1106 "punpckhwd %5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1107 "punpckhwd %5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1108 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1109 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1110 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1111 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1112 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1113 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1114 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1115 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1116
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1117 "movq %%mm0, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1118 "movq %%mm3, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1119
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1120 "movq 8%1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1121 "movq 8%1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1122 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1123 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1124 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1125 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1126 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1127 "psrlq $3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1128 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1129 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1130 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1131 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1132 "punpcklwd %5, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1133 "punpcklwd %5, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1134 "punpcklwd %5, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1135 "punpckhwd %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1136 "punpckhwd %5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1137 "punpckhwd %5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1138 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1139 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1140 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1141 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1142 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1143 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1144 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1145 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1146 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1147 :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1148 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1149 /* Borrowed 32 to 24 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1150 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1151 "movq %%mm0, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1152 "movq %%mm3, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1153 "movq %%mm6, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1154 "movq %%mm7, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1155
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1156 "movq %%mm4, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1157 "movq %%mm5, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1158 "movq %%mm0, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1159 "movq %%mm1, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1160
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1161 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1162 "psrlq $8, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1163 "psrlq $8, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1164 "psrlq $8, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1165 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1166 "pand %2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1167 "pand %2, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1168 "pand %2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1169 "pand %3, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1170 "pand %3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1171 "pand %3, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1172 "pand %3, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1173 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1174 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1175 "por %%mm6, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1176 "por %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1177
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1178 "movq %%mm1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1179 "movq %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1180 "psllq $48, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1181 "psllq $32, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1182 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1183 "pand %5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1184 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1185 "psrlq $16, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1186 "psrlq $32, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1187 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1188 "por %%mm3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1189 "pand %6, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1190 "por %%mm5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1191
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1192 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1193 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1194 MOVNTQ" %%mm4, 16%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1195
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1196 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1197 :"m"(*s),"m"(mask24l),"m"(mask24h),"m"(mask24hh),"m"(mask24hhh),"m"(mask24hhhh)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1198 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1199 d += 24;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1200 s += 8;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1201 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1202 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1203 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1204 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1205 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1206 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1207 register uint16_t bgr;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1208 bgr = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1209 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1210 *d++ = (bgr&0x7E0)>>3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1211 *d++ = (bgr&0xF800)>>8;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1212 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1213 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1214
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1215 static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1216 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1217 const uint16_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1218 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1219 const uint16_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1220 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1221 uint8_t *d = (uint8_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1222 const uint16_t *s = (const uint16_t *)src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1223 end = s + src_size/2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1224 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1225 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1226 __asm __volatile("pxor %%mm7,%%mm7\n\t":::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1227 mm_end = end - 3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1228 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1229 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1230 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1231 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1232 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1233 "movq %1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1234 "movq %1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1235 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1236 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1237 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1238 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1239 "psrlq $2, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1240 "psrlq $7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1241 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1242 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1243 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1244 "punpcklwd %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1245 "punpcklwd %%mm7, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1246 "punpcklwd %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1247 "punpckhwd %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1248 "punpckhwd %%mm7, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1249 "punpckhwd %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1250 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1251 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1252 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1253 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1254 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1255 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1256 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1257 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1258 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1259 MOVNTQ" %%mm3, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1260 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1261 :"m"(*s),"m"(mask15b),"m"(mask15g),"m"(mask15r)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1262 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1263 d += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1264 s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1265 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1266 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1267 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1268 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1269 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1270 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1271 #if 0 //slightly slower on athlon
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1272 int bgr= *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1273 *((uint32_t*)d)++ = ((bgr&0x1F)<<3) + ((bgr&0x3E0)<<6) + ((bgr&0x7C00)<<9);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1274 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1275 register uint16_t bgr;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1276 bgr = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1277 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1278 *d++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1279 *d++ = (bgr&0x7C00)>>7;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1280 *d++ = (bgr&0x3E0)>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1281 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1282 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1283 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1284 *d++ = (bgr&0x3E0)>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1285 *d++ = (bgr&0x7C00)>>7;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1286 *d++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1287 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1288
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1289 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1290 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1291 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1292
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1293 static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1294 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1295 const uint16_t *end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1296 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1297 const uint16_t *mm_end;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1298 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1299 uint8_t *d = (uint8_t *)dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1300 const uint16_t *s = (uint16_t *)src;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1301 end = s + src_size/2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1302 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1303 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1304 __asm __volatile("pxor %%mm7,%%mm7\n\t":::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1305 mm_end = end - 3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1306 while(s < mm_end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1307 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1308 __asm __volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1309 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1310 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1311 "movq %1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1312 "movq %1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1313 "pand %2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1314 "pand %3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1315 "pand %4, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1316 "psllq $3, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1317 "psrlq $3, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1318 "psrlq $8, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1319 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1320 "movq %%mm1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1321 "movq %%mm2, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1322 "punpcklwd %%mm7, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1323 "punpcklwd %%mm7, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1324 "punpcklwd %%mm7, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1325 "punpckhwd %%mm7, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1326 "punpckhwd %%mm7, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1327 "punpckhwd %%mm7, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1328 "psllq $8, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1329 "psllq $16, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1330 "por %%mm1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1331 "por %%mm2, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1332 "psllq $8, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1333 "psllq $16, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1334 "por %%mm4, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1335 "por %%mm5, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1336 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1337 MOVNTQ" %%mm3, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1338 :"=m"(*d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1339 :"m"(*s),"m"(mask16b),"m"(mask16g),"m"(mask16r)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1340 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1341 d += 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1342 s += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1343 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1344 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1345 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1346 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1347 while(s < end)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1348 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1349 register uint16_t bgr;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1350 bgr = *s++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1351 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1352 *d++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1353 *d++ = (bgr&0xF800)>>8;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1354 *d++ = (bgr&0x7E0)>>3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1355 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1356 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1357 *d++ = (bgr&0x1F)<<3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1358 *d++ = (bgr&0x7E0)>>3;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1359 *d++ = (bgr&0xF800)>>8;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1360 *d++ = 0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1361 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1362 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1363 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1364
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1365 static inline void RENAME(rgb32tobgr32)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1366 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1367 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1368 /* TODO: unroll this loop */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1369 asm volatile (
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1370 "xor %%"REG_a", %%"REG_a" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1371 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1372 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1373 PREFETCH" 32(%0, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1374 "movq (%0, %%"REG_a"), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1375 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1376 "movq %%mm0, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1377 "pslld $16, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1378 "psrld $16, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1379 "pand "MANGLE(mask32r)", %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1380 "pand "MANGLE(mask32g)", %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1381 "pand "MANGLE(mask32b)", %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1382 "por %%mm0, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1383 "por %%mm1, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1384 MOVNTQ" %%mm2, (%1, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1385 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1386 "cmp %2, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1387 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1388 :: "r" (src), "r"(dst), "r" (src_size-7)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1389 : "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1390 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1391
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1392 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1393 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1394 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1395 unsigned i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1396 unsigned num_pixels = src_size >> 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1397 for(i=0; i<num_pixels; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1398 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1399 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1400 dst[4*i + 1] = src[4*i + 3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1401 dst[4*i + 2] = src[4*i + 2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1402 dst[4*i + 3] = src[4*i + 1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1403 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1404 dst[4*i + 0] = src[4*i + 2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1405 dst[4*i + 1] = src[4*i + 1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1406 dst[4*i + 2] = src[4*i + 0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1407 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1408 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1409 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1410 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1411
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1412 static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, long src_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1413 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1414 unsigned i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1415 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1416 long mmx_size= 23 - src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1417 asm volatile (
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1418 "movq "MANGLE(mask24r)", %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1419 "movq "MANGLE(mask24g)", %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1420 "movq "MANGLE(mask24b)", %%mm7 \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1421 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1422 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1423 PREFETCH" 32(%1, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1424 "movq (%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1425 "movq (%1, %%"REG_a"), %%mm1 \n\t" // BGR BGR BG
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1426 "movq 2(%1, %%"REG_a"), %%mm2 \n\t" // R BGR BGR B
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1427 "psllq $16, %%mm0 \n\t" // 00 BGR BGR
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1428 "pand %%mm5, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1429 "pand %%mm6, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1430 "pand %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1431 "por %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1432 "por %%mm2, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1433 "movq 6(%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1434 MOVNTQ" %%mm1, (%2, %%"REG_a")\n\t" // RGB RGB RG
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1435 "movq 8(%1, %%"REG_a"), %%mm1 \n\t" // R BGR BGR B
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1436 "movq 10(%1, %%"REG_a"), %%mm2 \n\t" // GR BGR BGR
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1437 "pand %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1438 "pand %%mm5, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1439 "pand %%mm6, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1440 "por %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1441 "por %%mm2, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1442 "movq 14(%1, %%"REG_a"), %%mm0 \n\t" // R BGR BGR B
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1443 MOVNTQ" %%mm1, 8(%2, %%"REG_a")\n\t" // B RGB RGB R
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1444 "movq 16(%1, %%"REG_a"), %%mm1 \n\t" // GR BGR BGR
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1445 "movq 18(%1, %%"REG_a"), %%mm2 \n\t" // BGR BGR BG
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1446 "pand %%mm6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1447 "pand %%mm7, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1448 "pand %%mm5, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1449 "por %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1450 "por %%mm2, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1451 MOVNTQ" %%mm1, 16(%2, %%"REG_a")\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1452 "add $24, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1453 " js 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1454 : "+a" (mmx_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1455 : "r" (src-mmx_size), "r"(dst-mmx_size)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1456 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1457
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1458 __asm __volatile(SFENCE:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1459 __asm __volatile(EMMS:::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1460
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1461 if(mmx_size==23) return; //finihsed, was multiple of 8
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1462
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1463 src+= src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1464 dst+= src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1465 src_size= 23-mmx_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1466 src-= src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1467 dst-= src_size;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1468 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1469 for(i=0; i<src_size; i+=3)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1470 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1471 register uint8_t x;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1472 x = src[i + 2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1473 dst[i + 1] = src[i + 1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1474 dst[i + 2] = src[i + 0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1475 dst[i + 0] = x;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1476 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1477 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1478
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1479 static inline void RENAME(yuvPlanartoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1480 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1481 long lumStride, long chromStride, long dstStride, long vertLumPerChroma)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1482 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1483 long y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1484 const long chromWidth= width>>1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1485 for(y=0; y<height; y++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1486 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1487 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1488 //FIXME handle 2 lines a once (fewer prefetch, reuse some chrom, but very likely limited by mem anyway)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1489 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1490 "xor %%"REG_a", %%"REG_a" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1491 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1492 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1493 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1494 PREFETCH" 32(%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1495 PREFETCH" 32(%3, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1496 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1497 "movq %%mm0, %%mm2 \n\t" // U(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1498 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1499 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1500 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1501
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1502 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1503 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1504 "movq %%mm3, %%mm4 \n\t" // Y(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1505 "movq %%mm5, %%mm6 \n\t" // Y(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1506 "punpcklbw %%mm0, %%mm3 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1507 "punpckhbw %%mm0, %%mm4 \n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1508 "punpcklbw %%mm2, %%mm5 \n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1509 "punpckhbw %%mm2, %%mm6 \n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1510
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1511 MOVNTQ" %%mm3, (%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1512 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1513 MOVNTQ" %%mm5, 16(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1514 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1515
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1516 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1517 "cmp %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1518 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1519 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1520 : "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1521 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1522 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1523
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1524 #if defined ARCH_ALPHA && defined HAVE_MVI
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1525 #define pl2yuy2(n) \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1526 y1 = yc[n]; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1527 y2 = yc2[n]; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1528 u = uc[n]; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1529 v = vc[n]; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1530 asm("unpkbw %1, %0" : "=r"(y1) : "r"(y1)); \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1531 asm("unpkbw %1, %0" : "=r"(y2) : "r"(y2)); \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1532 asm("unpkbl %1, %0" : "=r"(u) : "r"(u)); \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1533 asm("unpkbl %1, %0" : "=r"(v) : "r"(v)); \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1534 yuv1 = (u << 8) + (v << 24); \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1535 yuv2 = yuv1 + y2; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1536 yuv1 += y1; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1537 qdst[n] = yuv1; \
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1538 qdst2[n] = yuv2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1539
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1540 int i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1541 uint64_t *qdst = (uint64_t *) dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1542 uint64_t *qdst2 = (uint64_t *) (dst + dstStride);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1543 const uint32_t *yc = (uint32_t *) ysrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1544 const uint32_t *yc2 = (uint32_t *) (ysrc + lumStride);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1545 const uint16_t *uc = (uint16_t*) usrc, *vc = (uint16_t*) vsrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1546 for(i = 0; i < chromWidth; i += 8){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1547 uint64_t y1, y2, yuv1, yuv2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1548 uint64_t u, v;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1549 /* Prefetch */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1550 asm("ldq $31,64(%0)" :: "r"(yc));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1551 asm("ldq $31,64(%0)" :: "r"(yc2));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1552 asm("ldq $31,64(%0)" :: "r"(uc));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1553 asm("ldq $31,64(%0)" :: "r"(vc));
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1554
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1555 pl2yuy2(0);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1556 pl2yuy2(1);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1557 pl2yuy2(2);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1558 pl2yuy2(3);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1559
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1560 yc += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1561 yc2 += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1562 uc += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1563 vc += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1564 qdst += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1565 qdst2 += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1566 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1567 y++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1568 ysrc += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1569 dst += dstStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1570
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1571 #elif __WORDSIZE >= 64
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1572 int i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1573 uint64_t *ldst = (uint64_t *) dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1574 const uint8_t *yc = ysrc, *uc = usrc, *vc = vsrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1575 for(i = 0; i < chromWidth; i += 2){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1576 uint64_t k, l;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1577 k = yc[0] + (uc[0] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1578 (yc[1] << 16) + (vc[0] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1579 l = yc[2] + (uc[1] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1580 (yc[3] << 16) + (vc[1] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1581 *ldst++ = k + (l << 32);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1582 yc += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1583 uc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1584 vc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1585 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1586
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1587 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1588 int i, *idst = (int32_t *) dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1589 const uint8_t *yc = ysrc, *uc = usrc, *vc = vsrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1590 for(i = 0; i < chromWidth; i++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1591 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1592 *idst++ = (yc[0] << 24)+ (uc[0] << 16) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1593 (yc[1] << 8) + (vc[0] << 0);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1594 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1595 *idst++ = yc[0] + (uc[0] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1596 (yc[1] << 16) + (vc[0] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1597 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1598 yc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1599 uc++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1600 vc++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1601 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1602 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1603 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1604 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1605 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1606 usrc += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1607 vsrc += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1608 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1609 ysrc += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1610 dst += dstStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1611 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1612 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1613 asm( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1614 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1615 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1616 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1617 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1618
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1619 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1620 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1621 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1622 * problem for anyone then tell me, and ill fix it)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1623 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1624 static inline void RENAME(yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1625 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1626 long lumStride, long chromStride, long dstStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1627 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1628 //FIXME interpolate chroma
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1629 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1630 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1631
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1632 static inline void RENAME(yuvPlanartouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1633 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1634 long lumStride, long chromStride, long dstStride, long vertLumPerChroma)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1635 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1636 long y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1637 const long chromWidth= width>>1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1638 for(y=0; y<height; y++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1639 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1640 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1641 //FIXME handle 2 lines a once (fewer prefetch, reuse some chrom, but very likely limited by mem anyway)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1642 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1643 "xor %%"REG_a", %%"REG_a" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1644 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1645 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1646 PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1647 PREFETCH" 32(%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1648 PREFETCH" 32(%3, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1649 "movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1650 "movq %%mm0, %%mm2 \n\t" // U(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1651 "movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1652 "punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1653 "punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1654
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1655 "movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1656 "movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1657 "movq %%mm0, %%mm4 \n\t" // Y(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1658 "movq %%mm2, %%mm6 \n\t" // Y(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1659 "punpcklbw %%mm3, %%mm0 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1660 "punpckhbw %%mm3, %%mm4 \n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1661 "punpcklbw %%mm5, %%mm2 \n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1662 "punpckhbw %%mm5, %%mm6 \n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1663
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1664 MOVNTQ" %%mm0, (%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1665 MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1666 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1667 MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1668
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1669 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1670 "cmp %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1671 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1672 ::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1673 : "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1674 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1675 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1676 //FIXME adapt the alpha asm code from yv12->yuy2
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1677
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1678 #if __WORDSIZE >= 64
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1679 int i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1680 uint64_t *ldst = (uint64_t *) dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1681 const uint8_t *yc = ysrc, *uc = usrc, *vc = vsrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1682 for(i = 0; i < chromWidth; i += 2){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1683 uint64_t k, l;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1684 k = uc[0] + (yc[0] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1685 (vc[0] << 16) + (yc[1] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1686 l = uc[1] + (yc[2] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1687 (vc[1] << 16) + (yc[3] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1688 *ldst++ = k + (l << 32);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1689 yc += 4;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1690 uc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1691 vc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1692 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1693
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1694 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1695 int i, *idst = (int32_t *) dst;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1696 const uint8_t *yc = ysrc, *uc = usrc, *vc = vsrc;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1697 for(i = 0; i < chromWidth; i++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1698 #ifdef WORDS_BIGENDIAN
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1699 *idst++ = (uc[0] << 24)+ (yc[0] << 16) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1700 (vc[0] << 8) + (yc[1] << 0);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1701 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1702 *idst++ = uc[0] + (yc[0] << 8) +
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1703 (vc[0] << 16) + (yc[1] << 24);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1704 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1705 yc += 2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1706 uc++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1707 vc++;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1708 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1709 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1710 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1711 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1712 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1713 usrc += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1714 vsrc += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1715 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1716 ysrc += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1717 dst += dstStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1718 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1719 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1720 asm( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1721 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1722 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1723 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1724 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1725
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1726 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1727 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1728 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1729 * problem for anyone then tell me, and ill fix it)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1730 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1731 static inline void RENAME(yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1732 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1733 long lumStride, long chromStride, long dstStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1734 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1735 //FIXME interpolate chroma
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1736 RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1737 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1738
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1739 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1740 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1741 * width should be a multiple of 16
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1742 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1743 static inline void RENAME(yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1744 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1745 long lumStride, long chromStride, long dstStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1746 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1747 RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1748 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1749
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1750 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1751 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1752 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1753 * problem for anyone then tell me, and ill fix it)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1754 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1755 static inline void RENAME(yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1756 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1757 long lumStride, long chromStride, long srcStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1758 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1759 long y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1760 const long chromWidth= width>>1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1761 for(y=0; y<height; y+=2)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1762 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1763 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1764 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1765 "xor %%"REG_a", %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1766 "pcmpeqw %%mm7, %%mm7 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1767 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1768 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1769 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1770 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1771 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1772 "movq 8(%0, %%"REG_a", 4), %%mm1\n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1773 "movq %%mm0, %%mm2 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1774 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1775 "psrlw $8, %%mm0 \n\t" // U0V0 U0V0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1776 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1777 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1778 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1779 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1780 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1781
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1782 MOVNTQ" %%mm2, (%1, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1783
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1784 "movq 16(%0, %%"REG_a", 4), %%mm1\n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1785 "movq 24(%0, %%"REG_a", 4), %%mm2\n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1786 "movq %%mm1, %%mm3 \n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1787 "movq %%mm2, %%mm4 \n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1788 "psrlw $8, %%mm1 \n\t" // U0V0 U0V0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1789 "psrlw $8, %%mm2 \n\t" // U0V0 U0V0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1790 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1791 "pand %%mm7, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1792 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1793 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1794
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1795 MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1796
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1797 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1798 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1799 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1800 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1801 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1802 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1803 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1804 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1805
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1806 MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1807 MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1808
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1809 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1810 "cmp %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1811 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1812 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1813 : "memory", "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1814 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1815
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1816 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1817 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1818
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1819 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1820 "xor %%"REG_a", %%"REG_a" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
1821 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1822 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1823 PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1824 "movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1825 "movq 8(%0, %%"REG_a", 4), %%mm1\n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1826 "movq 16(%0, %%"REG_a", 4), %%mm2\n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1827 "movq 24(%0, %%"REG_a", 4), %%mm3\n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1828 "pand %%mm7, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1829 "pand %%mm7, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1830 "pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1831 "pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1832 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1833 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1834
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1835 MOVNTQ" %%mm0, (%1, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1836 MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1837
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1838 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1839 "cmp %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1840 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1841
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1842 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1843 : "memory", "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1844 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1845 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1846 long i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1847 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1848 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1849 ydst[2*i+0] = src[4*i+0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1850 udst[i] = src[4*i+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1851 ydst[2*i+1] = src[4*i+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1852 vdst[i] = src[4*i+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1853 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1854 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1855 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1856
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1857 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1858 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1859 ydst[2*i+0] = src[4*i+0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1860 ydst[2*i+1] = src[4*i+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1861 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1862 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1863 udst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1864 vdst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1865 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1866 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1867 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1868 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1869 asm volatile( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1870 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1871 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1872 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1873 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1874
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1875 static inline void RENAME(yvu9toyv12)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1876 uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1877 long width, long height, long lumStride, long chromStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1878 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1879 /* Y Plane */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1880 memcpy(ydst, ysrc, width*height);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1881
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1882 /* XXX: implement upscaling for U,V */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1883 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1884
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1885 static inline void RENAME(planar2x)(const uint8_t *src, uint8_t *dst, long srcWidth, long srcHeight, long srcStride, long dstStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1886 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1887 long x,y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1888
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1889 dst[0]= src[0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1890
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1891 // first line
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1892 for(x=0; x<srcWidth-1; x++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1893 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1894 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1895 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1896 dst[2*srcWidth-1]= src[srcWidth-1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1897
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1898 dst+= dstStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1899
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1900 for(y=1; y<srcHeight; y++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1901 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1902 const long mmxSize= srcWidth&~15;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1903 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1904 "mov %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1905 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1906 "movq (%0, %%"REG_a"), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1907 "movq (%1, %%"REG_a"), %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1908 "movq 1(%0, %%"REG_a"), %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1909 "movq 1(%1, %%"REG_a"), %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1910 "movq -1(%0, %%"REG_a"), %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1911 "movq -1(%1, %%"REG_a"), %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1912 PAVGB" %%mm0, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1913 PAVGB" %%mm0, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1914 PAVGB" %%mm0, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1915 PAVGB" %%mm0, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1916 PAVGB" %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1917 PAVGB" %%mm1, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1918 PAVGB" %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1919 PAVGB" %%mm1, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1920 "movq %%mm5, %%mm7 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1921 "movq %%mm4, %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1922 "punpcklbw %%mm3, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1923 "punpckhbw %%mm3, %%mm7 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1924 "punpcklbw %%mm2, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1925 "punpckhbw %%mm2, %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1926 #if 1
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1927 MOVNTQ" %%mm5, (%2, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1928 MOVNTQ" %%mm7, 8(%2, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1929 MOVNTQ" %%mm4, (%3, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1930 MOVNTQ" %%mm6, 8(%3, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1931 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1932 "movq %%mm5, (%2, %%"REG_a", 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1933 "movq %%mm7, 8(%2, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1934 "movq %%mm4, (%3, %%"REG_a", 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1935 "movq %%mm6, 8(%3, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1936 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1937 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1938 " js 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1939 :: "r" (src + mmxSize ), "r" (src + srcStride + mmxSize ),
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1940 "r" (dst + mmxSize*2), "r" (dst + dstStride + mmxSize*2),
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1941 "g" (-mmxSize)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1942 : "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1943
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1944 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1945 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1946 const long mmxSize=1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1947 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1948 dst[0 ]= (3*src[0] + src[srcStride])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1949 dst[dstStride]= ( src[0] + 3*src[srcStride])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1950
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1951 for(x=mmxSize-1; x<srcWidth-1; x++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1952 dst[2*x +1]= (3*src[x+0] + src[x+srcStride+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1953 dst[2*x+dstStride+2]= ( src[x+0] + 3*src[x+srcStride+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1954 dst[2*x+dstStride+1]= ( src[x+1] + 3*src[x+srcStride ])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1955 dst[2*x +2]= (3*src[x+1] + src[x+srcStride ])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1956 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1957 dst[srcWidth*2 -1 ]= (3*src[srcWidth-1] + src[srcWidth-1 + srcStride])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1958 dst[srcWidth*2 -1 + dstStride]= ( src[srcWidth-1] + 3*src[srcWidth-1 + srcStride])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1959
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1960 dst+=dstStride*2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1961 src+=srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1962 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1963
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1964 // last line
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1965 #if 1
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1966 dst[0]= src[0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1967
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1968 for(x=0; x<srcWidth-1; x++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1969 dst[2*x+1]= (3*src[x] + src[x+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1970 dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1971 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1972 dst[2*srcWidth-1]= src[srcWidth-1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1973 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1974 for(x=0; x<srcWidth; x++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1975 dst[2*x+0]=
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1976 dst[2*x+1]= src[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1977 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1978 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1979
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1980 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1981 asm volatile( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1982 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1983 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1984 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1985 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1986
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1987 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1988 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1989 * height should be a multiple of 2 and width should be a multiple of 16 (if this is a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1990 * problem for anyone then tell me, and ill fix it)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1991 * chrominance data is only taken from every secound line others are ignored FIXME write HQ version
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1992 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1993 static inline void RENAME(uyvytoyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1994 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1995 long lumStride, long chromStride, long srcStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1996 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1997 long y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1998 const long chromWidth= width>>1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
1999 for(y=0; y<height; y+=2)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2000 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2001 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2002 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2003 "xorl %%eax, %%eax \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2004 "pcmpeqw %%mm7, %%mm7 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2005 "psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
2006 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2007 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2008 PREFETCH" 64(%0, %%eax, 4) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2009 "movq (%0, %%eax, 4), %%mm0 \n\t" // UYVY UYVY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2010 "movq 8(%0, %%eax, 4), %%mm1 \n\t" // UYVY UYVY(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2011 "movq %%mm0, %%mm2 \n\t" // UYVY UYVY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2012 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2013 "pand %%mm7, %%mm0 \n\t" // U0V0 U0V0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2014 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2015 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2016 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2017 "packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2018 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2019
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2020 MOVNTQ" %%mm2, (%1, %%eax, 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2021
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2022 "movq 16(%0, %%eax, 4), %%mm1 \n\t" // UYVY UYVY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2023 "movq 24(%0, %%eax, 4), %%mm2 \n\t" // UYVY UYVY(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2024 "movq %%mm1, %%mm3 \n\t" // UYVY UYVY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2025 "movq %%mm2, %%mm4 \n\t" // UYVY UYVY(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2026 "pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2027 "pand %%mm7, %%mm2 \n\t" // U0V0 U0V0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2028 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2029 "psrlw $8, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2030 "packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2031 "packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2032
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2033 MOVNTQ" %%mm3, 8(%1, %%eax, 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2034
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2035 "movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2036 "movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2037 "psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2038 "psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2039 "pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2040 "pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2041 "packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2042 "packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2043
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2044 MOVNTQ" %%mm0, (%3, %%eax) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2045 MOVNTQ" %%mm2, (%2, %%eax) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2046
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2047 "addl $8, %%eax \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2048 "cmpl %4, %%eax \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2049 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2050 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2051 : "memory", "%eax"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2052 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2053
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2054 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2055 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2056
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2057 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2058 "xorl %%eax, %%eax \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
2059 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2060 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2061 PREFETCH" 64(%0, %%eax, 4) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2062 "movq (%0, %%eax, 4), %%mm0 \n\t" // YUYV YUYV(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2063 "movq 8(%0, %%eax, 4), %%mm1 \n\t" // YUYV YUYV(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2064 "movq 16(%0, %%eax, 4), %%mm2 \n\t" // YUYV YUYV(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2065 "movq 24(%0, %%eax, 4), %%mm3 \n\t" // YUYV YUYV(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2066 "psrlw $8, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2067 "psrlw $8, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2068 "psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2069 "psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2070 "packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2071 "packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2072
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2073 MOVNTQ" %%mm0, (%1, %%eax, 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2074 MOVNTQ" %%mm2, 8(%1, %%eax, 2) \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2075
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2076 "addl $8, %%eax \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2077 "cmpl %4, %%eax \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2078 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2079
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2080 ::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2081 : "memory", "%eax"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2082 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2083 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2084 long i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2085 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2086 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2087 udst[i] = src[4*i+0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2088 ydst[2*i+0] = src[4*i+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2089 vdst[i] = src[4*i+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2090 ydst[2*i+1] = src[4*i+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2091 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2092 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2093 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2094
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2095 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2096 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2097 ydst[2*i+0] = src[4*i+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2098 ydst[2*i+1] = src[4*i+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2099 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2100 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2101 udst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2102 vdst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2103 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2104 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2105 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2106 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2107 asm volatile( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2108 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2109 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2110 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2111 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2112
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2113 /**
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2114 *
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2115 * height should be a multiple of 2 and width should be a multiple of 2 (if this is a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2116 * problem for anyone then tell me, and ill fix it)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2117 * chrominance data is only taken from every secound line others are ignored in the C version FIXME write HQ version
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2118 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2119 static inline void RENAME(rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2120 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2121 long lumStride, long chromStride, long srcStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2122 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2123 long y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2124 const long chromWidth= width>>1;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2125 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2126 for(y=0; y<height-2; y+=2)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2127 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2128 long i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2129 for(i=0; i<2; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2130 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2131 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2132 "mov %2, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2133 "movq "MANGLE(bgr2YCoeff)", %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2134 "movq "MANGLE(w1111)", %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2135 "pxor %%mm7, %%mm7 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2136 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d"\n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
2137 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2138 "1: \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2139 PREFETCH" 64(%0, %%"REG_d") \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2140 "movd (%0, %%"REG_d"), %%mm0 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2141 "movd 3(%0, %%"REG_d"), %%mm1 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2142 "punpcklbw %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2143 "punpcklbw %%mm7, %%mm1 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2144 "movd 6(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2145 "movd 9(%0, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2146 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2147 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2148 "pmaddwd %%mm6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2149 "pmaddwd %%mm6, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2150 "pmaddwd %%mm6, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2151 "pmaddwd %%mm6, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2152 #ifndef FAST_BGR2YV12
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2153 "psrad $8, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2154 "psrad $8, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2155 "psrad $8, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2156 "psrad $8, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2157 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2158 "packssdw %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2159 "packssdw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2160 "pmaddwd %%mm5, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2161 "pmaddwd %%mm5, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2162 "packssdw %%mm2, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2163 "psraw $7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2164
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2165 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2166 "movd 15(%0, %%"REG_d"), %%mm1 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2167 "punpcklbw %%mm7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2168 "punpcklbw %%mm7, %%mm1 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2169 "movd 18(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2170 "movd 21(%0, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2171 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2172 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2173 "pmaddwd %%mm6, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2174 "pmaddwd %%mm6, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2175 "pmaddwd %%mm6, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2176 "pmaddwd %%mm6, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2177 #ifndef FAST_BGR2YV12
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2178 "psrad $8, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2179 "psrad $8, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2180 "psrad $8, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2181 "psrad $8, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2182 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2183 "packssdw %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2184 "packssdw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2185 "pmaddwd %%mm5, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2186 "pmaddwd %%mm5, %%mm2 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2187 "add $24, %%"REG_d" \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2188 "packssdw %%mm2, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2189 "psraw $7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2190
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2191 "packuswb %%mm4, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2192 "paddusb "MANGLE(bgr2YOffset)", %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2193
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2194 MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2195 "add $8, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2196 " js 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2197 : : "r" (src+width*3), "r" (ydst+width), "g" (-width)
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2198 : "%"REG_a, "%"REG_d
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2199 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2200 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2201 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2202 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2203 src -= srcStride*2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2204 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2205 "mov %4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2206 "movq "MANGLE(w1111)", %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2207 "movq "MANGLE(bgr2UCoeff)", %%mm6 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2208 "pxor %%mm7, %%mm7 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2209 "lea (%%"REG_a", %%"REG_a", 2), %%"REG_d"\n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2210 "add %%"REG_d", %%"REG_d" \n\t"
19372
6334c14b38eb Replace asmalign.h hack by ASMALIGN cpp macros from config.h.
diego
parents: 18861
diff changeset
2211 ASMALIGN(4)
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2212 "1: \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2213 PREFETCH" 64(%0, %%"REG_d") \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2214 PREFETCH" 64(%1, %%"REG_d") \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2215 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2216 "movq (%0, %%"REG_d"), %%mm0 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2217 "movq (%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2218 "movq 6(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2219 "movq 6(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2220 PAVGB" %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2221 PAVGB" %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2222 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2223 "movq %%mm2, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2224 "psrlq $24, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2225 "psrlq $24, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2226 PAVGB" %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2227 PAVGB" %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2228 "punpcklbw %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2229 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2230 #else
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2231 "movd (%0, %%"REG_d"), %%mm0 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2232 "movd (%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2233 "movd 3(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2234 "movd 3(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2235 "punpcklbw %%mm7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2236 "punpcklbw %%mm7, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2237 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2238 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2239 "paddw %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2240 "paddw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2241 "paddw %%mm2, %%mm0 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2242 "movd 6(%0, %%"REG_d"), %%mm4 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2243 "movd 6(%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2244 "movd 9(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2245 "movd 9(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2246 "punpcklbw %%mm7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2247 "punpcklbw %%mm7, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2248 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2249 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2250 "paddw %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2251 "paddw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2252 "paddw %%mm4, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2253 "psrlw $2, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2254 "psrlw $2, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2255 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2256 "movq "MANGLE(bgr2VCoeff)", %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2257 "movq "MANGLE(bgr2VCoeff)", %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2258
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2259 "pmaddwd %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2260 "pmaddwd %%mm2, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2261 "pmaddwd %%mm6, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2262 "pmaddwd %%mm6, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2263 #ifndef FAST_BGR2YV12
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2264 "psrad $8, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2265 "psrad $8, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2266 "psrad $8, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2267 "psrad $8, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2268 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2269 "packssdw %%mm2, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2270 "packssdw %%mm3, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2271 "pmaddwd %%mm5, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2272 "pmaddwd %%mm5, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2273 "packssdw %%mm1, %%mm0 \n\t" // V1 V0 U1 U0
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2274 "psraw $7, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2275
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2276 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2277 "movq 12(%0, %%"REG_d"), %%mm4 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2278 "movq 12(%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2279 "movq 18(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2280 "movq 18(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2281 PAVGB" %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2282 PAVGB" %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2283 "movq %%mm4, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2284 "movq %%mm2, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2285 "psrlq $24, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2286 "psrlq $24, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2287 PAVGB" %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2288 PAVGB" %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2289 "punpcklbw %%mm7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2290 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2291 #else
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2292 "movd 12(%0, %%"REG_d"), %%mm4 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2293 "movd 12(%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2294 "movd 15(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2295 "movd 15(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2296 "punpcklbw %%mm7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2297 "punpcklbw %%mm7, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2298 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2299 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2300 "paddw %%mm1, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2301 "paddw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2302 "paddw %%mm2, %%mm4 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2303 "movd 18(%0, %%"REG_d"), %%mm5 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2304 "movd 18(%1, %%"REG_d"), %%mm1 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2305 "movd 21(%0, %%"REG_d"), %%mm2 \n\t"
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2306 "movd 21(%1, %%"REG_d"), %%mm3 \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2307 "punpcklbw %%mm7, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2308 "punpcklbw %%mm7, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2309 "punpcklbw %%mm7, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2310 "punpcklbw %%mm7, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2311 "paddw %%mm1, %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2312 "paddw %%mm3, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2313 "paddw %%mm5, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2314 "movq "MANGLE(w1111)", %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2315 "psrlw $2, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2316 "psrlw $2, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2317 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2318 "movq "MANGLE(bgr2VCoeff)", %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2319 "movq "MANGLE(bgr2VCoeff)", %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2320
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2321 "pmaddwd %%mm4, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2322 "pmaddwd %%mm2, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2323 "pmaddwd %%mm6, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2324 "pmaddwd %%mm6, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2325 #ifndef FAST_BGR2YV12
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2326 "psrad $8, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2327 "psrad $8, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2328 "psrad $8, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2329 "psrad $8, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2330 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2331 "packssdw %%mm2, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2332 "packssdw %%mm3, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2333 "pmaddwd %%mm5, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2334 "pmaddwd %%mm5, %%mm1 \n\t"
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2335 "add $24, %%"REG_d" \n\t"
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2336 "packssdw %%mm1, %%mm4 \n\t" // V3 V2 U3 U2
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2337 "psraw $7, %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2338
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2339 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2340 "punpckldq %%mm4, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2341 "punpckhdq %%mm4, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2342 "packsswb %%mm1, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2343 "paddb "MANGLE(bgr2UVOffset)", %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2344 "movd %%mm0, (%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2345 "punpckhdq %%mm0, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2346 "movd %%mm0, (%3, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2347 "add $4, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2348 " js 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2349 : : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth)
19396
8fe37c66d10a -fPIC support for libswscale
diego
parents: 19372
diff changeset
2350 : "%"REG_a, "%"REG_d
18861
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2351 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2352
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2353 udst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2354 vdst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2355 src += srcStride*2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2356 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2357
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2358 asm volatile( EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2359 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2360 :::"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2361 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2362 y=0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2363 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2364 for(; y<height; y+=2)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2365 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2366 long i;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2367 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2368 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2369 unsigned int b= src[6*i+0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2370 unsigned int g= src[6*i+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2371 unsigned int r= src[6*i+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2372
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2373 unsigned int Y = ((RY*r + GY*g + BY*b)>>RGB2YUV_SHIFT) + 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2374 unsigned int V = ((RV*r + GV*g + BV*b)>>RGB2YUV_SHIFT) + 128;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2375 unsigned int U = ((RU*r + GU*g + BU*b)>>RGB2YUV_SHIFT) + 128;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2376
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2377 udst[i] = U;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2378 vdst[i] = V;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2379 ydst[2*i] = Y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2380
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2381 b= src[6*i+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2382 g= src[6*i+4];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2383 r= src[6*i+5];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2384
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2385 Y = ((RY*r + GY*g + BY*b)>>RGB2YUV_SHIFT) + 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2386 ydst[2*i+1] = Y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2387 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2388 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2389 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2390
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2391 for(i=0; i<chromWidth; i++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2392 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2393 unsigned int b= src[6*i+0];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2394 unsigned int g= src[6*i+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2395 unsigned int r= src[6*i+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2396
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2397 unsigned int Y = ((RY*r + GY*g + BY*b)>>RGB2YUV_SHIFT) + 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2398
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2399 ydst[2*i] = Y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2400
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2401 b= src[6*i+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2402 g= src[6*i+4];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2403 r= src[6*i+5];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2404
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2405 Y = ((RY*r + GY*g + BY*b)>>RGB2YUV_SHIFT) + 16;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2406 ydst[2*i+1] = Y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2407 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2408 udst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2409 vdst += chromStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2410 ydst += lumStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2411 src += srcStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2412 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2413 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2414
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2415 void RENAME(interleaveBytes)(uint8_t *src1, uint8_t *src2, uint8_t *dest,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2416 long width, long height, long src1Stride,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2417 long src2Stride, long dstStride){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2418 long h;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2419
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2420 for(h=0; h < height; h++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2421 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2422 long w;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2423
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2424 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2425 #ifdef HAVE_SSE2
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2426 asm(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2427 "xor %%"REG_a", %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2428 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2429 PREFETCH" 64(%1, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2430 PREFETCH" 64(%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2431 "movdqa (%1, %%"REG_a"), %%xmm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2432 "movdqa (%1, %%"REG_a"), %%xmm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2433 "movdqa (%2, %%"REG_a"), %%xmm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2434 "punpcklbw %%xmm2, %%xmm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2435 "punpckhbw %%xmm2, %%xmm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2436 "movntdq %%xmm0, (%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2437 "movntdq %%xmm1, 16(%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2438 "add $16, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2439 "cmp %3, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2440 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2441 ::"r"(dest), "r"(src1), "r"(src2), "r" (width-15)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2442 : "memory", "%"REG_a""
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2443 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2444 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2445 asm(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2446 "xor %%"REG_a", %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2447 "1: \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2448 PREFETCH" 64(%1, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2449 PREFETCH" 64(%2, %%"REG_a") \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2450 "movq (%1, %%"REG_a"), %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2451 "movq 8(%1, %%"REG_a"), %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2452 "movq %%mm0, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2453 "movq %%mm2, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2454 "movq (%2, %%"REG_a"), %%mm4 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2455 "movq 8(%2, %%"REG_a"), %%mm5 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2456 "punpcklbw %%mm4, %%mm0 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2457 "punpckhbw %%mm4, %%mm1 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2458 "punpcklbw %%mm5, %%mm2 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2459 "punpckhbw %%mm5, %%mm3 \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2460 MOVNTQ" %%mm0, (%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2461 MOVNTQ" %%mm1, 8(%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2462 MOVNTQ" %%mm2, 16(%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2463 MOVNTQ" %%mm3, 24(%0, %%"REG_a", 2)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2464 "add $16, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2465 "cmp %3, %%"REG_a" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2466 " jb 1b \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2467 ::"r"(dest), "r"(src1), "r"(src2), "r" (width-15)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2468 : "memory", "%"REG_a
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2469 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2470 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2471 for(w= (width&(~15)); w < width; w++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2472 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2473 dest[2*w+0] = src1[w];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2474 dest[2*w+1] = src2[w];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2475 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2476 #else
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2477 for(w=0; w < width; w++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2478 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2479 dest[2*w+0] = src1[w];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2480 dest[2*w+1] = src2[w];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2481 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2482 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2483 dest += dstStride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2484 src1 += src1Stride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2485 src2 += src2Stride;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2486 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2487 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2488 asm(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2489 EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2490 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2491 ::: "memory"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2492 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2493 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2494 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2495
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2496 static inline void RENAME(vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2497 uint8_t *dst1, uint8_t *dst2,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2498 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2499 long srcStride1, long srcStride2,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2500 long dstStride1, long dstStride2)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2501 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2502 long y,x,w,h;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2503 w=width/2; h=height/2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2504 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2505 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2506 PREFETCH" %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2507 PREFETCH" %1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2508 ::"m"(*(src1+srcStride1)),"m"(*(src2+srcStride2)):"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2509 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2510 for(y=0;y<h;y++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2511 const uint8_t* s1=src1+srcStride1*(y>>1);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2512 uint8_t* d=dst1+dstStride1*y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2513 x=0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2514 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2515 for(;x<w-31;x+=32)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2516 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2517 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2518 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2519 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2520 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2521 "movq 16%1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2522 "movq 24%1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2523 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2524 "movq %%mm2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2525 "movq %%mm4, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2526 "movq %%mm6, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2527 "punpcklbw %%mm0, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2528 "punpckhbw %%mm1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2529 "punpcklbw %%mm2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2530 "punpckhbw %%mm3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2531 "punpcklbw %%mm4, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2532 "punpckhbw %%mm5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2533 "punpcklbw %%mm6, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2534 "punpckhbw %%mm7, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2535 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2536 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2537 MOVNTQ" %%mm2, 16%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2538 MOVNTQ" %%mm3, 24%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2539 MOVNTQ" %%mm4, 32%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2540 MOVNTQ" %%mm5, 40%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2541 MOVNTQ" %%mm6, 48%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2542 MOVNTQ" %%mm7, 56%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2543 :"=m"(d[2*x])
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2544 :"m"(s1[x])
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2545 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2546 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2547 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2548 for(;x<w;x++) d[2*x]=d[2*x+1]=s1[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2549 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2550 for(y=0;y<h;y++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2551 const uint8_t* s2=src2+srcStride2*(y>>1);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2552 uint8_t* d=dst2+dstStride2*y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2553 x=0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2554 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2555 for(;x<w-31;x+=32)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2556 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2557 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2558 PREFETCH" 32%1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2559 "movq %1, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2560 "movq 8%1, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2561 "movq 16%1, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2562 "movq 24%1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2563 "movq %%mm0, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2564 "movq %%mm2, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2565 "movq %%mm4, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2566 "movq %%mm6, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2567 "punpcklbw %%mm0, %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2568 "punpckhbw %%mm1, %%mm1\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2569 "punpcklbw %%mm2, %%mm2\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2570 "punpckhbw %%mm3, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2571 "punpcklbw %%mm4, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2572 "punpckhbw %%mm5, %%mm5\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2573 "punpcklbw %%mm6, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2574 "punpckhbw %%mm7, %%mm7\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2575 MOVNTQ" %%mm0, %0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2576 MOVNTQ" %%mm1, 8%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2577 MOVNTQ" %%mm2, 16%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2578 MOVNTQ" %%mm3, 24%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2579 MOVNTQ" %%mm4, 32%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2580 MOVNTQ" %%mm5, 40%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2581 MOVNTQ" %%mm6, 48%0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2582 MOVNTQ" %%mm7, 56%0"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2583 :"=m"(d[2*x])
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2584 :"m"(s2[x])
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2585 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2586 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2587 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2588 for(;x<w;x++) d[2*x]=d[2*x+1]=s2[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2589 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2590 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2591 asm(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2592 EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2593 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2594 ::: "memory"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2595 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2596 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2597 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2598
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2599 static inline void RENAME(yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2600 uint8_t *dst,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2601 long width, long height,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2602 long srcStride1, long srcStride2,
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2603 long srcStride3, long dstStride)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2604 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2605 long y,x,w,h;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2606 w=width/2; h=height;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2607 for(y=0;y<h;y++){
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2608 const uint8_t* yp=src1+srcStride1*y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2609 const uint8_t* up=src2+srcStride2*(y>>2);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2610 const uint8_t* vp=src3+srcStride3*(y>>2);
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2611 uint8_t* d=dst+dstStride*y;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2612 x=0;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2613 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2614 for(;x<w-7;x+=8)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2615 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2616 asm volatile(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2617 PREFETCH" 32(%1, %0)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2618 PREFETCH" 32(%2, %0)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2619 PREFETCH" 32(%3, %0)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2620 "movq (%1, %0, 4), %%mm0\n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2621 "movq (%2, %0), %%mm1\n\t" /* U0U1U2U3U4U5U6U7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2622 "movq (%3, %0), %%mm2\n\t" /* V0V1V2V3V4V5V6V7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2623 "movq %%mm0, %%mm3\n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2624 "movq %%mm1, %%mm4\n\t" /* U0U1U2U3U4U5U6U7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2625 "movq %%mm2, %%mm5\n\t" /* V0V1V2V3V4V5V6V7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2626 "punpcklbw %%mm1, %%mm1\n\t" /* U0U0 U1U1 U2U2 U3U3 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2627 "punpcklbw %%mm2, %%mm2\n\t" /* V0V0 V1V1 V2V2 V3V3 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2628 "punpckhbw %%mm4, %%mm4\n\t" /* U4U4 U5U5 U6U6 U7U7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2629 "punpckhbw %%mm5, %%mm5\n\t" /* V4V4 V5V5 V6V6 V7V7 */
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2630
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2631 "movq %%mm1, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2632 "punpcklbw %%mm2, %%mm1\n\t" /* U0V0 U0V0 U1V1 U1V1*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2633 "punpcklbw %%mm1, %%mm0\n\t" /* Y0U0 Y1V0 Y2U0 Y3V0*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2634 "punpckhbw %%mm1, %%mm3\n\t" /* Y4U1 Y5V1 Y6U1 Y7V1*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2635 MOVNTQ" %%mm0, (%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2636 MOVNTQ" %%mm3, 8(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2637
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2638 "punpckhbw %%mm2, %%mm6\n\t" /* U2V2 U2V2 U3V3 U3V3*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2639 "movq 8(%1, %0, 4), %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2640 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2641 "punpcklbw %%mm6, %%mm0\n\t" /* Y U2 Y V2 Y U2 Y V2*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2642 "punpckhbw %%mm6, %%mm3\n\t" /* Y U3 Y V3 Y U3 Y V3*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2643 MOVNTQ" %%mm0, 16(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2644 MOVNTQ" %%mm3, 24(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2645
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2646 "movq %%mm4, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2647 "movq 16(%1, %0, 4), %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2648 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2649 "punpcklbw %%mm5, %%mm4\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2650 "punpcklbw %%mm4, %%mm0\n\t" /* Y U4 Y V4 Y U4 Y V4*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2651 "punpckhbw %%mm4, %%mm3\n\t" /* Y U5 Y V5 Y U5 Y V5*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2652 MOVNTQ" %%mm0, 32(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2653 MOVNTQ" %%mm3, 40(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2654
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2655 "punpckhbw %%mm5, %%mm6\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2656 "movq 24(%1, %0, 4), %%mm0\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2657 "movq %%mm0, %%mm3\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2658 "punpcklbw %%mm6, %%mm0\n\t" /* Y U6 Y V6 Y U6 Y V6*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2659 "punpckhbw %%mm6, %%mm3\n\t" /* Y U7 Y V7 Y U7 Y V7*/
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2660 MOVNTQ" %%mm0, 48(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2661 MOVNTQ" %%mm3, 56(%4, %0, 8)\n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2662
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2663 : "+r" (x)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2664 : "r"(yp), "r" (up), "r"(vp), "r"(d)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2665 :"memory");
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2666 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2667 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2668 for(; x<w; x++)
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2669 {
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2670 const long x2= x<<2;
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2671 d[8*x+0]=yp[x2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2672 d[8*x+1]=up[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2673 d[8*x+2]=yp[x2+1];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2674 d[8*x+3]=vp[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2675 d[8*x+4]=yp[x2+2];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2676 d[8*x+5]=up[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2677 d[8*x+6]=yp[x2+3];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2678 d[8*x+7]=vp[x];
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2679 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2680 }
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2681 #ifdef HAVE_MMX
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2682 asm(
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2683 EMMS" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2684 SFENCE" \n\t"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2685 ::: "memory"
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2686 );
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2687 #endif
8579acff875e Move postproc ---> libswscale
lucabe
parents:
diff changeset
2688 }