skia: blend32_16_row for neon version

This includes blend32_16_row neon implementation
for aarch32 and aarch64.

For performance,
blend32_16_row is called in following tests in nanobench.
 - Xfermode_SrcOver
 - tablebench
 - rotated_rects_bw_alternating_transparent_and_opaque_srcover
 - rotated_rects_bw_changing_transparent_srcover
 - rotated_rects_bw_same_transparent_srcover
 - luma_colorfilter_large
 - luma_colorfilter_small
 - chart_bw

I can see perf increase in following two tests, especially. For others, looks
similar.
For each, I tried to run two times.

1) Xfermode_SrcOver
<org>
 - D/skia    ( 2000):    3M        57      17.3µs  17.4µs  17.4µs  17.7µs  1%
  █▃▂▃▂▂▂▁▃▂      565     Xfermode_SrcOver
 - D/skia    ( 1915):    3M        70      13.5µs  16.9µs  16.7µs  18.8µs  9%
  ▆█▄▅█▁▅▅▆▄      565     Xfermode_SrcOver

<new>
 - D/skia    ( 2000):    3M        8       11.6µs  11.8µs  12.1µs  14.4µs  7%
  ▃█▁▁▂▁▁▁▂▂      565     Xfermode_SrcOver
 - D/skia    ( 2004):    3M        62      10.3µs  12.9µs  13µs    15.2µs  11%
  █▅▅▆▁▅▅▅▇▃      565     Xfermode_SrcOver

2)
luma_colorfilter_large
<org>
 - D/skia    ( 2000):  159M        8       136µs   136µs   136µs   139µs   1%
  █▃▁▂▁▁▁▁▁▁      565     luma_colorfilter_large
 - D/skia    ( 1915):  158M        2       135µs   177µs   182µs   269µs   22%
  ▆▃█▁▁▃▃▃▃▃      565     luma_colorfilter_large

<new>
 - D/skia    ( 2000):  157M        5       84.2µs  85.3µs  87.5µs  110µs   9%
  █▁▂▁▁▁▁▁▁▁      565     luma_colorfilter_large
 - D/skia    ( 2004):  159M        6       84.7µs  110µs   112µs   144µs   18%
  █▄▇▁▁▄▃▄▄▆      565     luma_colorfilter_large

Review URL: https://codereview.chromium.org/847363002

4 files changed, 157 insertions, 23 deletions

diff --git a/src/core/SkBlitter_RGB16.cpp b/src/core/SkBlitter_RGB16.cpp
index d9193ecf8f..306c95b413 100644
--- a/src/core/SkBlitter_RGB16.cpp
+++ b/src/core/SkBlitter_RGB16.cpp
@@ -77,6 +77,8 @@ protected:
     uint16_t    fRawDither16;   // unscaled
     SkBool8     fDoDither;
 
+    SkBlitRow::ColorProc16 fColorProc16;
+
     // illegal
     SkRGB16_Blitter& operator=(const SkRGB16_Blitter&);
 
@@ -544,6 +546,19 @@ SkRGB16_Blitter::SkRGB16_Blitter(const SkBitmap& device, const SkPaint& paint)
     fColor16 = SkPackRGB16( SkAlphaMul(r, fScale) >> (8 - SK_R16_BITS),
                             SkAlphaMul(g, fScale) >> (8 - SK_G16_BITS),
                             SkAlphaMul(b, fScale) >> (8 - SK_B16_BITS));
+
+    // compute SkBlitRow::Procs
+    unsigned flags = 0;
+
+    if (SkGetPackedA32(fSrcColor32) < 0xFF) {
+        flags |= SkBlitRow::kSrcPixelAlpha_Flag;
+    }
+
+    if (fDoDither) {
+        flags |= SkBlitRow::kDither_Flag;
+    }
+
+    fColorProc16 = SkBlitRow::ColorFactory16(flags);
 }
 
 const SkBitmap* SkRGB16_Blitter::justAnOpaqueColor(uint32_t* value) {
@@ -554,31 +569,12 @@ const SkBitmap* SkRGB16_Blitter::justAnOpaqueColor(uint32_t* value) {
     return NULL;
 }
 
-static uint32_t pmcolor_to_expand16(SkPMColor c) {
-    unsigned r = SkGetPackedR32(c);
-    unsigned g = SkGetPackedG32(c);
-    unsigned b = SkGetPackedB32(c);
-    return (g << 24) | (r << 13) | (b << 2);
-}
-
-static inline void blend32_16_row(SkPMColor src, uint16_t dst[], int count) {
-    SkASSERT(count > 0);
-    uint32_t src_expand = pmcolor_to_expand16(src);
-    unsigned scale = SkAlpha255To256(0xFF - SkGetPackedA32(src)) >> 3;
-    do {
-        uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale;
-        *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5);
-        dst += 1;
-    } while (--count != 0);
-}
-
 void SkRGB16_Blitter::blitH(int x, int y, int width) {
     SkASSERT(width > 0);
     SkASSERT(x + width <= fDevice.width());
     uint16_t* SK_RESTRICT device = fDevice.getAddr16(x, y);
 
-    // TODO: respect fDoDither
-    blend32_16_row(fSrcColor32, device, width);
+    fColorProc16(device, fSrcColor32, width, x, y);
 }
 
 void SkRGB16_Blitter::blitAntiH(int x, int y,
@@ -681,10 +677,9 @@ void SkRGB16_Blitter::blitRect(int x, int y, int width, int height) {
     SkASSERT(x + width <= fDevice.width() && y + height <= fDevice.height());
     uint16_t* SK_RESTRICT device = fDevice.getAddr16(x, y);
     size_t    deviceRB = fDevice.rowBytes();
-    SkPMColor src32 = fSrcColor32;
 
     while (--height >= 0) {
-        blend32_16_row(src32, device, width);
+        fColorProc16(device, fSrcColor32, width, x, y);
         device = (uint16_t*)((char*)device + deviceRB);
     }
 }
diff --git a/src/opts/SkBlitRow_opts_arm.cpp b/src/opts/SkBlitRow_opts_arm.cpp
index 9a8cfc4b77..9dc9f3b593 100644
--- a/src/opts/SkBlitRow_opts_arm.cpp
+++ b/src/opts/SkBlitRow_opts_arm.cpp
@@ -364,6 +364,13 @@ static const SkBlitRow::Proc16 sk_blitrow_platform_565_procs_arm[] = {
     NULL,   // S32A_D565_Blend_Dither
 };
 
+static const SkBlitRow::ColorProc16 sk_blitrow_platform_565_colorprocs_arm[] = {
+    NULL,   // Color32_D565,
+    NULL,   // Color32A_D565,
+    NULL,   // Color32_D565_Dither,
+    NULL,   // Color32A_D565_Dither
+};
+
 static const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm[] = {
     NULL,   // S32_Opaque,
     NULL,   // S32_Blend,
@@ -378,7 +385,7 @@ SkBlitRow::Proc16 SkBlitRow::PlatformFactory565(unsigned flags) {
 }
 
 SkBlitRow::ColorProc16 SkBlitRow::PlatformColorFactory565(unsigned flags) {
-    return NULL;
+    return SK_ARM_NEON_WRAP(sk_blitrow_platform_565_colorprocs_arm)[flags];
 }
 
 SkBlitRow::Proc32 SkBlitRow::PlatformProcs32(unsigned flags) {
diff --git a/src/opts/SkBlitRow_opts_arm_neon.cpp b/src/opts/SkBlitRow_opts_arm_neon.cpp
index a1ff1722b5..8981fd7fee 100644
--- a/src/opts/SkBlitRow_opts_arm_neon.cpp
+++ b/src/opts/SkBlitRow_opts_arm_neon.cpp
@@ -465,6 +465,130 @@ void S32A_D565_Opaque_neon(uint16_t* SK_RESTRICT dst,
 }
 #endif // #ifdef SK_CPU_ARM32
 
+static uint32_t pmcolor_to_expand16(SkPMColor c) {
+    unsigned r = SkGetPackedR32(c);
+    unsigned g = SkGetPackedG32(c);
+    unsigned b = SkGetPackedB32(c);
+    return (g << 24) | (r << 13) | (b << 2);
+}
+
+void Color32A_D565_neon(uint16_t dst[], SkPMColor src, int count, int x, int y) {
+    uint32_t src_expand;
+    unsigned scale;
+    uint16x8_t vmask_blue;
+
+    if (count <= 0) return;
+    SkASSERT(((size_t)dst & 0x01) == 0);
+
+    /*
+     * This preamble code is in order to make dst aligned to 8 bytes
+     * in the next mutiple bytes read & write access.
+     */
+    src_expand = pmcolor_to_expand16(src);
+    scale = SkAlpha255To256(0xFF - SkGetPackedA32(src)) >> 3;
+
+#define DST_ALIGN 8
+
+    /*
+     * preamble_size is in byte, meantime, this blend32_16_row_neon updates 2 bytes at a time.
+     */
+    int preamble_size = (DST_ALIGN - (size_t)dst) & (DST_ALIGN - 1);
+
+    for (int i = 0; i < preamble_size; i+=2, dst++) {
+        uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale;
+        *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5);
+        if (--count == 0)
+            break;
+    }
+
+    int count16 = 0;
+    count16 = count >> 4;
+    vmask_blue = vmovq_n_u16(SK_B16_MASK);
+
+    if (count16) {
+        uint16x8_t wide_sr;
+        uint16x8_t wide_sg;
+        uint16x8_t wide_sb;
+        uint16x8_t wide_256_sa;
+
+        unsigned sr = SkGetPackedR32(src);
+        unsigned sg = SkGetPackedG32(src);
+        unsigned sb = SkGetPackedB32(src);
+        unsigned sa = SkGetPackedA32(src);
+
+        // Operation: dst_rgb = src_rgb + ((256 - src_a) >> 3) x dst_rgb
+        // sr: 8-bit based, dr: 5-bit based, with dr x ((256-sa)>>3), 5-bit left shifted,
+        //thus, for sr, do 2-bit left shift to match MSB : (8 + 2 = 5 + 5)
+        wide_sr = vshlq_n_u16(vmovl_u8(vdup_n_u8(sr)), 2); // widen and src_red shift
+
+        // sg: 8-bit based, dg: 6-bit based, with dg x ((256-sa)>>3), 5-bit left shifted,
+        //thus, for sg, do 3-bit left shift to match MSB : (8 + 3 = 6 + 5)
+        wide_sg = vshlq_n_u16(vmovl_u8(vdup_n_u8(sg)), 3); // widen and src_grn shift
+
+        // sb: 8-bit based, db: 5-bit based, with db x ((256-sa)>>3), 5-bit left shifted,
+        //thus, for sb, do 2-bit left shift to match MSB : (8 + 2 = 5 + 5)
+        wide_sb = vshlq_n_u16(vmovl_u8(vdup_n_u8(sb)), 2); // widen and src blu shift
+
+        wide_256_sa =
+            vshrq_n_u16(vsubw_u8(vdupq_n_u16(256), vdup_n_u8(sa)), 3); // (256 - sa) >> 3
+
+        while (count16-- > 0) {
+            uint16x8_t vdst1, vdst1_r, vdst1_g, vdst1_b;
+            uint16x8_t vdst2, vdst2_r, vdst2_g, vdst2_b;
+            vdst1 = vld1q_u16(dst);
+            dst += 8;
+            vdst2 = vld1q_u16(dst);
+            dst -= 8;    //to store dst again.
+
+            vdst1_g = vshlq_n_u16(vdst1, SK_R16_BITS);                 // shift green to top of lanes
+            vdst1_b = vdst1 & vmask_blue;                              // extract blue
+            vdst1_r = vshrq_n_u16(vdst1, SK_R16_SHIFT);                // extract red
+            vdst1_g = vshrq_n_u16(vdst1_g, SK_R16_BITS + SK_B16_BITS); // extract green
+
+            vdst2_g = vshlq_n_u16(vdst2, SK_R16_BITS);                 // shift green to top of lanes
+            vdst2_b = vdst2 & vmask_blue;                              // extract blue
+            vdst2_r = vshrq_n_u16(vdst2, SK_R16_SHIFT);                // extract red
+            vdst2_g = vshrq_n_u16(vdst2_g, SK_R16_BITS + SK_B16_BITS); // extract green
+
+            vdst1_r = vmlaq_u16(wide_sr, wide_256_sa, vdst1_r);        // sr + (256-sa) x dr1
+            vdst1_g = vmlaq_u16(wide_sg, wide_256_sa, vdst1_g);        // sg + (256-sa) x dg1
+            vdst1_b = vmlaq_u16(wide_sb, wide_256_sa, vdst1_b);        // sb + (256-sa) x db1
+
+            vdst2_r = vmlaq_u16(wide_sr, wide_256_sa, vdst2_r);        // sr + (256-sa) x dr2
+            vdst2_g = vmlaq_u16(wide_sg, wide_256_sa, vdst2_g);        // sg + (256-sa) x dg2
+            vdst2_b = vmlaq_u16(wide_sb, wide_256_sa, vdst2_b);        // sb + (256-sa) x db2
+
+            vdst1_r = vshrq_n_u16(vdst1_r, 5);                         // 5-bit right shift for 5-bit red
+            vdst1_g = vshrq_n_u16(vdst1_g, 5);                         // 5-bit right shift for 6-bit green
+            vdst1_b = vshrq_n_u16(vdst1_b, 5);                         // 5-bit right shift for 5-bit blue
+
+            vdst1 = vsliq_n_u16(vdst1_b, vdst1_g, SK_G16_SHIFT);       // insert green into blue
+            vdst1 = vsliq_n_u16(vdst1, vdst1_r, SK_R16_SHIFT);         // insert red into green/blue
+
+            vdst2_r = vshrq_n_u16(vdst2_r, 5);                         // 5-bit right shift for 5-bit red
+            vdst2_g = vshrq_n_u16(vdst2_g, 5);                         // 5-bit right shift for 6-bit green
+            vdst2_b = vshrq_n_u16(vdst2_b, 5);                         // 5-bit right shift for 5-bit blue
+
+            vdst2 = vsliq_n_u16(vdst2_b, vdst2_g, SK_G16_SHIFT);       // insert green into blue
+            vdst2 = vsliq_n_u16(vdst2, vdst2_r, SK_R16_SHIFT);         // insert red into green/blue
+
+            vst1q_u16(dst, vdst1);
+            dst += 8;
+            vst1q_u16(dst, vdst2);
+            dst += 8;
+        }
+    }
+
+    count &= 0xF;
+    if (count > 0) {
+        do {
+            uint32_t dst_expand = SkExpand_rgb_16(*dst) * scale;
+            *dst = SkCompact_rgb_16((src_expand + dst_expand) >> 5);
+            dst += 1;
+        } while (--count != 0);
+    }
+}
+
 static inline uint16x8_t SkDiv255Round_neon8(uint16x8_t prod) {
     prod += vdupq_n_u16(128);
     prod += vshrq_n_u16(prod, 8);
@@ -1665,6 +1789,13 @@ const SkBlitRow::Proc16 sk_blitrow_platform_565_procs_arm_neon[] = {
     NULL,   // S32A_D565_Blend_Dither
 };
 
+const SkBlitRow::ColorProc16 sk_blitrow_platform_565_colorprocs_arm_neon[] = {
+    Color32A_D565_neon,    // Color32_D565,
+    Color32A_D565_neon,    // Color32A_D565,
+    Color32A_D565_neon,    // Color32_D565_Dither,
+    Color32A_D565_neon,    // Color32A_D565_Dither
+};
+
 const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm_neon[] = {
     NULL,   // S32_Opaque,
     S32_Blend_BlitRow32_neon,        // S32_Blend,
diff --git a/src/opts/SkBlitRow_opts_arm_neon.h b/src/opts/SkBlitRow_opts_arm_neon.h
index 95825f781b..92d58ce3c4 100644
--- a/src/opts/SkBlitRow_opts_arm_neon.h
+++ b/src/opts/SkBlitRow_opts_arm_neon.h
@@ -10,6 +10,7 @@
 #include "SkBlitRow.h"
 
 extern const SkBlitRow::Proc16 sk_blitrow_platform_565_procs_arm_neon[];
+extern const SkBlitRow::ColorProc16 sk_blitrow_platform_565_colorprocs_arm_neon[];
 extern const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm_neon[];
 
 extern void Color32_arm_neon(SkPMColor* dst, const SkPMColor* src, int count,