利用ARM v6的SIMD指令实现blend算法

ARM v6的SIMD指令相对较弱，我原先以为其不能实现负责的blend算法，可我在cario的pixman库中看到了，写的很精巧。

这段代码的问题在于，对于a=0,255时没有做专门的优化，更加通用一些。如果在做blend时，a值大量为0或255时，性能会稍弱一些。

 

fbCompositeSrc_8888x8888arm (pixman_op_t op,
    pixman_image_t * pSrc,
    pixman_image_t * pMask,
    pixman_image_t * pDst,
    int16_t      xSrc,
    int16_t      ySrc,
    int16_t      xMask,
    int16_t      yMask,
    int16_t      xDst,
    int16_t      yDst,
    uint16_t     width,
    uint16_t     height)
{
    uint32_t *dstLine, *dst;
    uint32_t *srcLine, *src;
    int dstStride, srcStride;
    uint16_t w;
    uint32_t component_half = 0x800080;
    uint32_t upper_component_mask = 0xff00ff00;
    uint32_t alpha_mask = 0xff;

    fbComposeGetStart (pDst, xDst, yDst, uint32_t, dstStride, dstLine, 1);
    fbComposeGetStart (pSrc, xSrc, ySrc, uint32_t, srcStride, srcLine, 1);

    while (height--)
    {
 dst = dstLine;
 dstLine += dstStride;
 src = srcLine;
 srcLine += srcStride;
 w = width;

//#define inner_branch
 asm volatile (
   "cmp %[w], #0\n\t"
   "beq 2f\n\t"
   "1:\n\t"
   /* load src */
   "ldr r5, [%[src]], #4\n\t"
#ifdef inner_branch
   /* We can avoid doing the multiplication in two cases: 0x0 or 0xff.
    * The 0x0 case also allows us to avoid doing an unecessary data
    * write which is more valuable so we only check for that */
   "cmp r5, #0\n\t"
   "beq 3f\n\t"

   /* = 255 - alpha */
   "sub r8, %[alpha_mask], r5, lsr #24\n\t"

   "ldr r4, [%[dest]] \n\t"

#else
   "ldr r4, [%[dest]] \n\t"

   /* = 255 - alpha */
   "sub r8, %[alpha_mask], r5, lsr #24\n\t"
#endif
   "uxtb16 r6, r4\n\t"
   "uxtb16 r7, r4, ror #8\n\t"

   /* multiply by 257 and divide by 65536 */
   "mla r6, r6, r8, %[component_half]\n\t"
   "mla r7, r7, r8, %[component_half]\n\t"

   "uxtab16 r6, r6, r6, ror #8\n\t"
   "uxtab16 r7, r7, r7, ror #8\n\t"

   /* recombine the 0xff00ff00 bytes of r6 and r7 */
   "and r7, r7, %[upper_component_mask]\n\t"
   "uxtab16 r6, r7, r6, ror #8\n\t"

   "uqadd8 r5, r6, r5\n\t"

#ifdef inner_branch
   "3:\n\t"

#endif
   "str r5, [%[dest]], #4\n\t"
   /* increment counter and jmp to top */
   "subs %[w], %[w], #1\n\t"
   "bne 1b\n\t"
   "2:\n\t"
   : [w] "+r" (w), [dest] "+r" (dst), [src] "+r" (src)
   : [component_half] "r" (component_half), [upper_component_mask] "r" (upper_component_mask),
     [alpha_mask] "r" (alpha_mask)
   : "r4", "r5", "r6", "r7", "r8", "cc", "memory"
   );
    }
}