|
2504
|
1 #include <inttypes.h>
|
|
|
2 #include "../config.h"
|
|
|
3 #include "rgb2rgb.h"
|
|
2506
|
4 #include "mmx.h"
|
|
2504
|
5
|
|
|
6 void rgb24to32(uint8_t *src,uint8_t *dst,uint32_t src_size)
|
|
|
7 {
|
|
2508
|
8 uint8_t *dest = dst;
|
|
2504
|
9 uint8_t *s = src;
|
|
|
10 uint8_t *end;
|
|
2510
|
11 #ifdef HAVE_MMX
|
|
|
12 const uint64_t mask32 = 0x00FFFFFF00FFFFFFULL;
|
|
|
13 uint8_t *mm_end;
|
|
|
14 #endif
|
|
2504
|
15 end = s + src_size;
|
|
2510
|
16 #ifdef HAVE_MMX
|
|
|
17 mm_end = (uint8_t*)((((unsigned long)end)/16)*16);
|
|
|
18 __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory");
|
|
|
19 while(s < mm_end)
|
|
|
20 {
|
|
|
21 __asm __volatile(
|
|
|
22 "movd %1, %%mm0\n\t"
|
|
|
23 "movd 3%1, %%mm1\n\t"
|
|
|
24 "movd 6%1, %%mm2\n\t"
|
|
|
25 "movd 9%1, %%mm3\n\t"
|
|
|
26 "punpckldq %%mm1, %%mm0\n\t"
|
|
|
27 "punpckldq %%mm3, %%mm2\n\t"
|
|
|
28 "pand %%mm7, %%mm0\n\t"
|
|
|
29 "pand %%mm7, %%mm2\n\t"
|
|
|
30 "movq %%mm0, %0\n\t"
|
|
|
31 "movq %%mm2, 8%0"
|
|
|
32 :"=m"(*dest)
|
|
|
33 :"m"(*s)
|
|
|
34 :"memory");
|
|
|
35 dest += 16;
|
|
|
36 s += 12;
|
|
|
37 }
|
|
|
38 #endif
|
|
2504
|
39 while(s < end)
|
|
|
40 {
|
|
2508
|
41 *dest++ = *s++;
|
|
|
42 *dest++ = *s++;
|
|
|
43 *dest++ = *s++;
|
|
|
44 *dest++ = 0;
|
|
2504
|
45 }
|
|
|
46 }
|
|
2505
|
47
|
|
2510
|
48 /* TODO: MMX optimization */
|
|
2505
|
49 void rgb32to24(uint8_t *src,uint8_t *dst,uint32_t src_size)
|
|
|
50 {
|
|
|
51 uint8_t *dest = dst;
|
|
|
52 uint8_t *s = src;
|
|
|
53 uint8_t *end;
|
|
|
54 end = s + src_size;
|
|
|
55 while(s < end)
|
|
|
56 {
|
|
|
57 *dest++ = *s++;
|
|
|
58 *dest++ = *s++;
|
|
|
59 *dest++ = *s++;
|
|
|
60 s++;
|
|
|
61 }
|
|
|
62 }
|
|
2506
|
63
|
|
|
64 /* Original by Strepto/Astral
|
|
|
65 ported to gcc & bugfixed : A'rpi */
|
|
|
66 void rgb15to16(uint8_t *src,uint8_t *dst,uint32_t src_size)
|
|
|
67 {
|
|
|
68 #ifdef HAVE_MMX
|
|
|
69 static uint64_t mask_b = 0x001F001F001F001FLL; // 00000000 00011111 xxB
|
|
|
70 static uint64_t mask_rg = 0x7FE07FE07FE07FE0LL; // 01111111 11100000 RGx
|
|
|
71 register char* s=src+src_size;
|
|
|
72 register char* d=dst+src_size;
|
|
|
73 register int offs=-src_size;
|
|
|
74 movq_m2r (mask_b, mm4);
|
|
|
75 movq_m2r (mask_rg, mm5);
|
|
|
76 while(offs<0){
|
|
|
77 movq_m2r (*(s+offs), mm0);
|
|
|
78 movq_r2r (mm0, mm1);
|
|
|
79
|
|
|
80 movq_m2r (*(s+8+offs), mm2);
|
|
|
81 movq_r2r (mm2, mm3);
|
|
|
82
|
|
|
83 pand_r2r (mm4, mm0);
|
|
|
84 pand_r2r (mm5, mm1);
|
|
|
85
|
|
|
86 psllq_i2r(1,mm1);
|
|
|
87 pand_r2r (mm4, mm2);
|
|
|
88
|
|
|
89 pand_r2r (mm5, mm3);
|
|
|
90 por_r2r (mm1, mm0);
|
|
|
91
|
|
|
92 psllq_i2r(1,mm3);
|
|
|
93 movq_r2m (mm0,*(d+offs));
|
|
|
94
|
|
|
95 por_r2r (mm3,mm2);
|
|
|
96 movq_r2m (mm2,*(d+8+offs));
|
|
|
97
|
|
|
98 offs+=16;
|
|
|
99 }
|
|
|
100 emms();
|
|
|
101 #else
|
|
|
102 uint16_t *s1=( uint16_t * )src;
|
|
|
103 uint16_t *d1=( uint16_t * )dst;
|
|
|
104 uint16_t *e=((uint8_t *)s1)+src_size;
|
|
|
105 while( s1<e ){
|
|
|
106 register int x=*( s1++ );
|
|
|
107 /* rrrrrggggggbbbbb
|
|
|
108 0rrrrrgggggbbbbb
|
|
|
109 0111 1111 1110 0000=0x7FE0
|
|
|
110 00000000000001 1111=0x001F */
|
|
|
111 *( d1++ )=( x&0x001F )|( ( x&0x7FE0 )<<1 );
|
|
|
112 }
|
|
|
113 #endif
|
|
|
114 }
|
