Optimize RGB16 blitH functions with NEON for ARM platform.

src/core/SkBlitter_RGB16.cpp

 /*
  * Copyright 2006 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.h"
 #include "SkCoreBlitters.h"
 #include "SkColorPriv.h"
 #include "SkDither.h"
 #include "SkShader.h"
 #include "SkUtils.h"
 #include "SkUtilsArm.h"
 #include "SkXfermode.h"
 
 #if SK_MIPS_HAS_DSP
 extern void blitmask_d565_opaque_mips(int width, int height, uint16_t* device,
                                       unsigned deviceRB, const uint8_t* alpha,
                                       uint32_t expanded32, unsigned maskRB);
 #endif
 
 #if SK_ARM_NEON_IS_ALWAYS && defined(SK_CPU_LENDIAN)
     #include <arm_neon.h>
 extern void SkRGB16BlitterBlitV_neon(uint16_t* device,
                                      int height,
                                      size_t deviceRB,
                                      unsigned scale,
                                      uint32_t src32);
+extern void SkRGB16BlitterBlitH_neon(uint16_t* device,
+                                     int width,
+                                     unsigned scale,
+                                     uint32_t src32);
 #else
     // if we don't have neon, then our black blitter is worth the extra code
     #define USE_BLACK_BLITTER
 #endif
 
 void sk_dither_memset16(uint16_t dst[], uint16_t value, uint16_t other,
                         int count) {
     if (count > 0) {
         // see if we need to write one short before we can cast to an 4byte ptr
         // (we do this subtract rather than (unsigned)dst so we don't get warnings
@@ skipping 291 matching lines @@
                                        ditherColor, count);
                 } else {
                     sk_memset16(device, srcColor, count);
                 }
             } else {
                 // TODO: respect fDoDither
                 unsigned scale5 = SkAlpha255To256(aa) >> 3;
                 uint32_t src32 = srcExpanded * scale5;
                 scale5 = 32 - scale5; // now we can use it on the device
                 int n = count;
+#if SK_ARM_NEON_IS_ALWAYS && defined(SK_CPU_LENDIAN)
+                SkRGB16BlitterBlitH_neon(device, n, scale5, src32);
+#else
                 do {
                     uint32_t dst32 = SkExpand_rgb_16(*device) * scale5;
                     *device++ = SkCompact_rgb_16((src32 + dst32) >> 5);
                 } while (--n != 0);
+#endif
                 goto DONE;
             }
         }
         device += count;
 
     DONE:
         // if we have no dithering, this will always fail
         if (count & ditherInt) {
             SkTSwap(ditherColor, srcColor);
         }
@@ skipping 239 matching lines @@
             return;
         }
         runs += count;
 
         unsigned aa = antialias[0];
         antialias += count;
         if (aa) {
             unsigned scale5 = SkAlpha255To256(aa) * scale >> (8 + 3);
             uint32_t src32 =  srcExpanded * scale5;
             scale5 = 32 - scale5;
+#if SK_ARM_NEON_IS_ALWAYS && defined(SK_CPU_LENDIAN)
+            SkRGB16BlitterBlitH_neon(device, count, scale5, src32);
+#else
             do {
                 uint32_t dst32 = SkExpand_rgb_16(*device) * scale5;
                 *device++ = SkCompact_rgb_16((src32 + dst32) >> 5);
             } while (--count != 0);
+#endif
             continue;
         }
         device += count;
     }
 }
 
 static inline void blend_8_pixels(U8CPU bw, uint16_t dst[], unsigned dst_scale,
                                   U16CPU srcColor) {
     if (bw & 0x80) dst[0] = srcColor + SkAlphaMulRGB16(dst[0], dst_scale);
     if (bw & 0x40) dst[1] = srcColor + SkAlphaMulRGB16(dst[1], dst_scale);
@@ skipping 445 matching lines @@
 #endif
         } else if (0xFF == SkColorGetA(color)) {
             blitter = allocator->createT<SkRGB16_Opaque_Blitter>(device, paint);
         } else {
             blitter = allocator->createT<SkRGB16_Blitter>(device, paint);
         }
     }
 
     return blitter;
 }