Optimize RGB16 blitV functions with NEON for ARM platform.

src/opts/SkBlitMask_opts_arm_neon.cpp

Index: src/opts/SkBlitMask_opts_arm_neon.cpp
diff --git a/src/opts/SkBlitMask_opts_arm_neon.cpp b/src/opts/SkBlitMask_opts_arm_neon.cpp
index aa59a92caae3437ec5948ff6d28869a4d47806d7..3361a5d19dd777191700bbe8f4f6111ddb7c1d7a 100644
--- a/src/opts/SkBlitMask_opts_arm_neon.cpp
+++ b/src/opts/SkBlitMask_opts_arm_neon.cpp
@@ -258,3 +258,77 @@ void SkBlitLCD16Row_neon(SkPMColor dst[], const uint16_t src[],
         dst[i] = SkBlendLCD16(colA, colR, colG, colB, dst[i], src[i]);
     }
 }
+
+#define LOAD_LANE_16(reg, n) \
+    reg = vld1q_lane_u16(device, reg, n); \
+    device = (uint16_t*)((char*)device + deviceRB);
+
+#define STORE_LANE_16(reg, n) \
+    vst1_lane_u16(dst, reg, n); \
+    dst = (uint16_t*)((char*)dst + deviceRB);
+
+void SkRGB16BlitterBlitV_neon(uint16_t* device,
+                              int height,
+                              size_t deviceRB,
+                              unsigned scale,
+                              uint32_t src32) {
+    if (height >= 8)
+    {
+        uint16_t* dst = device;
+
+        // prepare constants
+        uint16x8_t vdev = vdupq_n_u16(0);
+        uint16x8_t vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE);
+        uint16x8_t vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE);
+        uint32x4_t vsrc32 = vdupq_n_u32(src32);
+        uint32x4_t vscale5 = vdupq_n_u32((uint32_t)scale);
+
+        while (height >= 8){
+            LOAD_LANE_16(vdev, 0)
+            LOAD_LANE_16(vdev, 1)
+            LOAD_LANE_16(vdev, 2)
+            LOAD_LANE_16(vdev, 3)
+            LOAD_LANE_16(vdev, 4)
+            LOAD_LANE_16(vdev, 5)
+            LOAD_LANE_16(vdev, 6)
+            LOAD_LANE_16(vdev, 7)
+
+            // Expand_rgb_16
+            uint16x8x2_t vdst = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16));
+            uint32x4_t vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst.val[0]), vscale5);
+            uint32x4_t vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst.val[1]), vscale5);
+
+            // Compact_rgb_16
+            vdst32_lo = vaddq_u32(vdst32_lo, vsrc32);
+            vdst32_hi = vaddq_u32(vdst32_hi, vsrc32);
+            vdst32_lo = vshrq_n_u32(vdst32_lo, 5);
+            vdst32_hi = vshrq_n_u32(vdst32_hi, 5);
+
+            uint16x4_t vtmp_lo = vmovn_u32(vdst32_lo) & vget_low_u16(vmaskq_ng16);
+            uint16x4_t vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vget_low_u16(vmaskq_g16);
+            uint16x4_t vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi);
+            vtmp_lo = vmovn_u32(vdst32_hi) & vget_low_u16(vmaskq_ng16);
+            vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vget_low_u16(vmaskq_g16);
+            uint16x4_t vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi);
+
+            STORE_LANE_16(vdst16_lo, 0)
+            STORE_LANE_16(vdst16_lo, 1)
+            STORE_LANE_16(vdst16_lo, 2)
+            STORE_LANE_16(vdst16_lo, 3)
+            STORE_LANE_16(vdst16_hi, 0)
+            STORE_LANE_16(vdst16_hi, 1)
+            STORE_LANE_16(vdst16_hi, 2)
+            STORE_LANE_16(vdst16_hi, 3)
+            height -= 8;
+        }
+    }
+    while (height != 0){
+        uint32_t dst32 = SkExpand_rgb_16(*device) * scale;
+        *device = SkCompact_rgb_16((src32 + dst32) >> 5);
+        device = (uint16_t*)((char*)device + deviceRB);
+        height--;
+    }
+}
+
+#undef LOAD_LANE_16
+#undef STORE_LANE_16