#include <stddef.h>
#include <inttypes.h> /* for __WORDSIZE */
-#include "asmalign.h"
-
#ifndef __WORDSIZE
// #warning You have misconfigured system and probably will lose performance!
#define __WORDSIZE MP_WORDSIZE
"movq %3, %%mm5 \n\t"
"movq %4, %%mm6 \n\t"
"movq %5, %%mm7 \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%1) \n\t"
"movd (%1), %%mm0 \n\t"
"movq %3, %%mm5 \n\t"
"movq %4, %%mm6 \n\t"
"movq %5, %%mm7 \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%1) \n\t"
"movd (%1), %%mm0 \n\t"
/* TODO: unroll this loop */
asm volatile (
"xor %%"REG_a", %%"REG_a" \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%0, %%"REG_a") \n\t"
"movq (%0, %%"REG_a"), %%mm0 \n\t"
"movq "MANGLE(mask24r)", %%mm5 \n\t"
"movq "MANGLE(mask24g)", %%mm6 \n\t"
"movq "MANGLE(mask24b)", %%mm7 \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%1, %%"REG_a") \n\t"
"movq (%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
//FIXME handle 2 lines a once (fewer prefetch, reuse some chrom, but very likely limited by mem anyway)
asm volatile(
"xor %%"REG_a", %%"REG_a" \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
PREFETCH" 32(%2, %%"REG_a") \n\t"
//FIXME handle 2 lines a once (fewer prefetch, reuse some chrom, but very likely limited by mem anyway)
asm volatile(
"xor %%"REG_a", %%"REG_a" \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
PREFETCH" 32(%2, %%"REG_a") \n\t"
"xor %%"REG_a", %%"REG_a" \n\t"
"pcmpeqw %%mm7, %%mm7 \n\t"
"psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
asm volatile(
"xor %%"REG_a", %%"REG_a" \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
"xorl %%eax, %%eax \n\t"
"pcmpeqw %%mm7, %%mm7 \n\t"
"psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%eax, 4) \n\t"
"movq (%0, %%eax, 4), %%mm0 \n\t" // UYVY UYVY(0)
asm volatile(
"xorl %%eax, %%eax \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%eax, 4) \n\t"
"movq (%0, %%eax, 4), %%mm0 \n\t" // YUYV YUYV(0)
"movq "MANGLE(w1111)", %%mm5 \n\t"
"pxor %%mm7, %%mm7 \n\t"
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_b"\n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%"REG_b") \n\t"
"movd (%0, %%"REG_b"), %%mm0 \n\t"
"pxor %%mm7, %%mm7 \n\t"
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_b"\n\t"
"add %%"REG_b", %%"REG_b" \n\t"
- ASMALIGN16
+ ASMALIGN(4)
"1: \n\t"
PREFETCH" 64(%0, %%"REG_b") \n\t"
PREFETCH" 64(%1, %%"REG_b") \n\t"