Mercurial > mplayer.hg
annotate postproc/rgb2rgb.c @ 2567:ea6158be8103
video frame reading cleanup
author | arpi |
---|---|
date | Tue, 30 Oct 2001 20:36:20 +0000 |
parents | 3d04a0991dce |
children | 794dec2fae64 |
rev | line source |
---|---|
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
1 /* |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
2 * |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
3 * rgb2rgb.c, Software RGB to RGB convertor |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
4 * Written by Nick Kurshev. |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
5 */ |
2504 | 6 #include <inttypes.h> |
7 #include "../config.h" | |
8 #include "rgb2rgb.h" | |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
9 #include "../mmx_defs.h" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
10 |
2535 | 11 #ifdef HAVE_MMX |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
12 static const uint64_t mask32 __attribute__((aligned(8))) = 0x00FFFFFF00FFFFFFULL; |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
13 static const uint64_t mask24l __attribute__((aligned(8))) = 0x0000000000FFFFFFULL; |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
14 static const uint64_t mask24h __attribute__((aligned(8))) = 0x0000FFFFFF000000ULL; |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
15 static const uint64_t mask15b __attribute__((aligned(8))) = 0x001F001F001F001FULL; /* 00000000 00011111 xxB */ |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
16 static const uint64_t mask15rg __attribute__((aligned(8))) = 0x7FE07FE07FE07FE0ULL; /* 01111111 11100000 RGx */ |
2535 | 17 #endif |
2513 | 18 |
2504 | 19 void rgb24to32(uint8_t *src,uint8_t *dst,uint32_t src_size) |
20 { | |
2508 | 21 uint8_t *dest = dst; |
2504 | 22 uint8_t *s = src; |
23 uint8_t *end; | |
2510 | 24 #ifdef HAVE_MMX |
25 uint8_t *mm_end; | |
26 #endif | |
2504 | 27 end = s + src_size; |
2510 | 28 #ifdef HAVE_MMX |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
29 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory"); |
2516 | 30 mm_end = (uint8_t*)((((unsigned long)end)/(MMREG_SIZE*2))*(MMREG_SIZE*2)); |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
31 __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory"); |
2516 | 32 if(mm_end == end) mm_end -= MMREG_SIZE*2; |
2510 | 33 while(s < mm_end) |
34 { | |
2511 | 35 __asm __volatile( |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
36 PREFETCH" 32%1\n\t" |
2510 | 37 "movd %1, %%mm0\n\t" |
38 "movd 3%1, %%mm1\n\t" | |
39 "movd 6%1, %%mm2\n\t" | |
40 "movd 9%1, %%mm3\n\t" | |
41 "punpckldq %%mm1, %%mm0\n\t" | |
42 "punpckldq %%mm3, %%mm2\n\t" | |
43 "pand %%mm7, %%mm0\n\t" | |
44 "pand %%mm7, %%mm2\n\t" | |
2511 | 45 MOVNTQ" %%mm0, %0\n\t" |
46 MOVNTQ" %%mm2, 8%0" | |
2510 | 47 :"=m"(*dest) |
48 :"m"(*s) | |
49 :"memory"); | |
50 dest += 16; | |
51 s += 12; | |
52 } | |
2513 | 53 __asm __volatile(SFENCE:::"memory"); |
2511 | 54 __asm __volatile(EMMS:::"memory"); |
2510 | 55 #endif |
2504 | 56 while(s < end) |
57 { | |
2508 | 58 *dest++ = *s++; |
59 *dest++ = *s++; | |
60 *dest++ = *s++; | |
61 *dest++ = 0; | |
2504 | 62 } |
63 } | |
2505 | 64 |
65 void rgb32to24(uint8_t *src,uint8_t *dst,uint32_t src_size) | |
66 { | |
67 uint8_t *dest = dst; | |
68 uint8_t *s = src; | |
69 uint8_t *end; | |
2517 | 70 #ifdef HAVE_MMX |
71 uint8_t *mm_end; | |
72 #endif | |
2505 | 73 end = s + src_size; |
2517 | 74 #ifdef HAVE_MMX |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
75 __asm __volatile(PREFETCH" %0"::"m"(*s):"memory"); |
2517 | 76 mm_end = (uint8_t*)((((unsigned long)end)/(MMREG_SIZE*2))*(MMREG_SIZE*2)); |
77 __asm __volatile( | |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
78 "movq %0, %%mm7\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
79 "movq %1, %%mm6" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
80 ::"m"(mask24l),"m"(mask24h):"memory"); |
2517 | 81 if(mm_end == end) mm_end -= MMREG_SIZE*2; |
82 while(s < mm_end) | |
83 { | |
84 __asm __volatile( | |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
85 PREFETCH" 32%1\n\t" |
2517 | 86 "movq %1, %%mm0\n\t" |
87 "movq 8%1, %%mm1\n\t" | |
88 "movq %%mm0, %%mm2\n\t" | |
89 "movq %%mm1, %%mm3\n\t" | |
90 "psrlq $8, %%mm2\n\t" | |
91 "psrlq $8, %%mm3\n\t" | |
92 "pand %%mm7, %%mm0\n\t" | |
93 "pand %%mm7, %%mm1\n\t" | |
94 "pand %%mm6, %%mm2\n\t" | |
95 "pand %%mm6, %%mm3\n\t" | |
96 "por %%mm2, %%mm0\n\t" | |
97 "por %%mm3, %%mm1\n\t" | |
98 MOVNTQ" %%mm0, %0\n\t" | |
99 MOVNTQ" %%mm1, 6%0" | |
100 :"=m"(*dest) | |
101 :"m"(*s) | |
102 :"memory"); | |
103 dest += 12; | |
104 s += 16; | |
105 } | |
106 __asm __volatile(SFENCE:::"memory"); | |
107 __asm __volatile(EMMS:::"memory"); | |
108 #endif | |
2505 | 109 while(s < end) |
110 { | |
111 *dest++ = *s++; | |
112 *dest++ = *s++; | |
113 *dest++ = *s++; | |
114 s++; | |
115 } | |
116 } | |
2506 | 117 |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
118 /* |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
119 Original by Strepto/Astral |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
120 ported to gcc & bugfixed : A'rpi |
2564 | 121 MMX2, 3DNOW optimization by Nick Kurshev |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
122 */ |
2506 | 123 void rgb15to16(uint8_t *src,uint8_t *dst,uint32_t src_size) |
124 { | |
125 #ifdef HAVE_MMX | |
126 register char* s=src+src_size; | |
127 register char* d=dst+src_size; | |
128 register int offs=-src_size; | |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
129 __asm __volatile(PREFETCH" %0"::"m"(*(s+offs)):"memory"); |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
130 __asm __volatile( |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
131 "movq %0, %%mm4\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
132 "movq %1, %%mm5" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
133 ::"m"(mask15b), "m"(mask15rg):"memory"); |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
134 while(offs<0) |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
135 { |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
136 __asm __volatile( |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
137 PREFETCH" 32%1\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
138 "movq %1, %%mm0\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
139 "movq 8%1, %%mm2\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
140 "movq %%mm0, %%mm1\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
141 "movq %%mm2, %%mm3\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
142 "pand %%mm4, %%mm0\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
143 "pand %%mm5, %%mm1\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
144 "pand %%mm4, %%mm2\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
145 "pand %%mm5, %%mm3\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
146 "psllq $1, %%mm1\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
147 "psllq $1, %%mm3\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
148 "por %%mm1, %%mm0\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
149 "por %%mm3, %%mm2\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
150 MOVNTQ" %%mm0, %0\n\t" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
151 MOVNTQ" %%mm2, 8%0" |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
152 :"=m"(*(d+offs)) |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
153 :"m"(*(s+offs)) |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
154 :"memory"); |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
155 offs+=16; |
2506 | 156 } |
2538
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
157 __asm __volatile(SFENCE:::"memory"); |
71320898b333
Finish mmx2, 3dnow optimiz. 15to16 should be tested. Better fix of can't compile
nick
parents:
2535
diff
changeset
|
158 __asm __volatile(EMMS:::"memory"); |
2506 | 159 #else |
160 uint16_t *s1=( uint16_t * )src; | |
161 uint16_t *d1=( uint16_t * )dst; | |
162 uint16_t *e=((uint8_t *)s1)+src_size; | |
163 while( s1<e ){ | |
164 register int x=*( s1++ ); | |
165 /* rrrrrggggggbbbbb | |
166 0rrrrrgggggbbbbb | |
167 0111 1111 1110 0000=0x7FE0 | |
168 00000000000001 1111=0x001F */ | |
169 *( d1++ )=( x&0x001F )|( ( x&0x7FE0 )<<1 ); | |
170 } | |
171 #endif | |
172 } |