5 #include "../mmx_defs.h"
7 void rgb24to32(uint8_t *src,uint8_t *dst,uint32_t src_size)
13 const uint64_t mask32 = 0x00FFFFFF00FFFFFFULL;
18 __asm __volatile(PREFETCH" %0\n\t"::"m"(*s):"memory");
19 mm_end = (uint8_t*)((((unsigned long)end)/(MMREG_SIZE*2))*(MMREG_SIZE*2));
20 __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory");
21 if(mm_end == end) mm_end -= MMREG_SIZE*2;
30 "punpckldq %%mm1, %%mm0\n\t"
31 "punpckldq %%mm3, %%mm2\n\t"
32 "pand %%mm7, %%mm0\n\t"
33 "pand %%mm7, %%mm2\n\t"
34 MOVNTQ" %%mm0, %0\n\t"
42 __asm __volatile(SFENCE:::"memory");
43 __asm __volatile(EMMS:::"memory");
54 /* TODO: MMX optimization */
55 void rgb32to24(uint8_t *src,uint8_t *dst,uint32_t src_size)
70 /* Original by Strepto/Astral
71 ported to gcc & bugfixed : A'rpi */
72 void rgb15to16(uint8_t *src,uint8_t *dst,uint32_t src_size)
75 static uint64_t mask_b = 0x001F001F001F001FLL; // 00000000 00011111 xxB
76 static uint64_t mask_rg = 0x7FE07FE07FE07FE0LL; // 01111111 11100000 RGx
77 register char* s=src+src_size;
78 register char* d=dst+src_size;
79 register int offs=-src_size;
80 movq_m2r (mask_b, mm4);
81 movq_m2r (mask_rg, mm5);
83 movq_m2r (*(s+offs), mm0);
86 movq_m2r (*(s+8+offs), mm2);
99 movq_r2m (mm0,*(d+offs));
102 movq_r2m (mm2,*(d+8+offs));
108 uint16_t *s1=( uint16_t * )src;
109 uint16_t *d1=( uint16_t * )dst;
110 uint16_t *e=((uint8_t *)s1)+src_size;
112 register int x=*( s1++ );
115 0111 1111 1110 0000=0x7FE0
116 00000000000001 1111=0x001F */
117 *( d1++ )=( x&0x001F )|( ( x&0x7FE0 )<<1 );