ARM Skia NEON patches - 13 - S32A_Opaque

src/opts/SkBlitRow_opts_arm_neon.cpp

 /*
  * Copyright 2012 The Android Open Source Project
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkBlitRow_opts_arm.h"
 
 #include "SkBlitMask.h"
@@ skipping 408 matching lines @@
 
     static const uint8_t alpha_mask_setup[] = {3,3,3,3,7,7,7,7};
     alpha_mask = vld1_u8(alpha_mask_setup);
 
     /* do the NEON unrolled code */
 #define    UNROLL    4
     while (count >= UNROLL) {
         uint8x8_t src_raw, dst_raw, dst_final;
         uint8x8_t src_raw_2, dst_raw_2, dst_final_2;
 
+        __builtin_prefetch(src+32);
+        __builtin_prefetch(dst+32);

[djsollen, 2013/07/11 15:24:47]

can you put the comment here that this *may* be slightly slower if the count is
low, but worth it for the general case.

[kevin.petit.not.used.account, 2013/07/11 15:42:14]

Done.

+
         /* get the source */
         src_raw = vreinterpret_u8_u32(vld1_u32(src));
 #if    UNROLL > 2
         src_raw_2 = vreinterpret_u8_u32(vld1_u32(src+2));
 #endif
 
         /* get and hold the dst too */
         dst_raw = vreinterpret_u8_u32(vld1_u32(dst));
 #if    UNROLL > 2
         dst_raw_2 = vreinterpret_u8_u32(vld1_u32(dst+2));
 #endif
 
     /* 1st and 2nd bits of the unrolling */
     {
         uint8x8_t dst_cooked;
         uint16x8_t dst_wide;
         uint8x8_t alpha_narrow;
         uint16x8_t alpha_wide;
 
         /* get the alphas spread out properly */
         alpha_narrow = vtbl1_u8(src_raw, alpha_mask);
-#if 1
-        /* reflect SkAlpha255To256() semantics a+1 vs a+a>>7 */
-        /* we collapsed (255-a)+1 ... */
         alpha_wide = vsubw_u8(vdupq_n_u16(256), alpha_narrow);
-#else
-        alpha_wide = vsubw_u8(vdupq_n_u16(255), alpha_narrow);
-        alpha_wide = vaddq_u16(alpha_wide, vshrq_n_u16(alpha_wide,7));
-#endif
 
         /* spread the dest */
         dst_wide = vmovl_u8(dst_raw);
 
         /* alpha mul the dest */
         dst_wide = vmulq_u16 (dst_wide, alpha_wide);
         dst_cooked = vshrn_n_u16(dst_wide, 8);
 
         /* sum -- ignoring any byte lane overflows */
         dst_final = vadd_u8(src_raw, dst_cooked);
     }
 
 #if    UNROLL > 2
     /* the 3rd and 4th bits of our unrolling */
     {
         uint8x8_t dst_cooked;
         uint16x8_t dst_wide;
         uint8x8_t alpha_narrow;
         uint16x8_t alpha_wide;
 
         alpha_narrow = vtbl1_u8(src_raw_2, alpha_mask);
-#if 1
-        /* reflect SkAlpha255To256() semantics a+1 vs a+a>>7 */
-        /* we collapsed (255-a)+1 ... */
         alpha_wide = vsubw_u8(vdupq_n_u16(256), alpha_narrow);
-#else
-        alpha_wide = vsubw_u8(vdupq_n_u16(255), alpha_narrow);
-        alpha_wide = vaddq_u16(alpha_wide, vshrq_n_u16(alpha_wide,7));
-#endif
 
         /* spread the dest */
         dst_wide = vmovl_u8(dst_raw_2);
 
         /* alpha mul the dest */
         dst_wide = vmulq_u16 (dst_wide, alpha_wide);
         dst_cooked = vshrn_n_u16(dst_wide, 8);
 
         /* sum -- ignoring any byte lane overflows */
         dst_final_2 = vadd_u8(src_raw_2, dst_cooked);
@@ skipping 792 matching lines @@
      * case where we do not inspect the src alpha.
      */
 #if SK_A32_SHIFT == 24
     // This proc assumes the alpha value occupies bits 24-32 of each SkPMColor
     S32A_Opaque_BlitRow32_neon_src_alpha,   // S32A_Opaque,
 #else
     S32A_Opaque_BlitRow32_neon,     // S32A_Opaque,
 #endif
     S32A_Blend_BlitRow32_arm        // S32A_Blend
 };