565 support for SIMD xfermodes

This uses the most basic approach possible:
  - to load an Sk4px from 565, convert to SkPMColors on the stack serially then load those SkPMColors.
  - to store an Sk4px to 565, store to SkPMColors on the stack then convert to 565 serially.

Clearly, we can optimize these loads and stores.  That's a TODO.

The code using SkPMFloat is the same idea but a little more long-term viable, as we're only operating on one pixel at a time anyway.  We could probably write 565 <-> SkPMFloat methods, but I'd rather not until it's really compelling.

The speedups are varied but similar across SSE and NEON: a few uninteresting, many 50% faster, some 2x faster, and SoftLight ~4x faster.

This will cause minor GM diffs, but I don't think any layout test changes.

BUG=skia:

Committed: https://skia.googlesource.com/skia/+/942930dcaa51f66d82cdaf46ae62efebd16c8cd0

Committed: https://skia.googlesource.com/skia/+/860dcaa2ddfdadc050af4f943a84a9d499315066

Review URL: https://codereview.chromium.org/1245673002

1 files changed, 71 insertions, 0 deletions

diff --git a/src/opts/Sk4px_NEON.h b/src/opts/Sk4px_NEON.h
index cd6dea9979..89841d927e 100644
--- a/src/opts/Sk4px_NEON.h
+++ b/src/opts/Sk4px_NEON.h
@@ -89,5 +89,76 @@ inline Sk4px Sk4px::zeroAlphas() const {
     return Sk16b(vbicq_u8(this->fVec, (uint8x16_t)vdupq_n_u32(0xFF << SK_A32_SHIFT)));
 }
 
+static inline uint8x16_t widen_to_8888(uint16x4_t v) {
+    // RGB565 format:   |R....|G.....|B....|
+    //           Bit:  16    11      5     0
+
+    // First get each pixel into its own 32-bit lane.
+    //      v ==                       rgb3 rgb2  rgb1 rgb0
+    // spread == 0000 rgb3  0000 rgb2  0000 rgb1  0000 rgb0
+    uint32x4_t spread = vmovl_u16(v);
+
+    // Get each color independently, still in 565 precison but down at bit 0.
+    auto r5 = vshrq_n_u32(spread, 11),
+         g6 = vandq_u32(vdupq_n_u32(63), vshrq_n_u32(spread, 5)),
+         b5 = vandq_u32(vdupq_n_u32(31), spread);
+
+    // Scale 565 precision up to 8-bit each, filling low 323 bits with high bits of each component.
+    auto r8 = vorrq_u32(vshlq_n_u32(r5, 3), vshrq_n_u32(r5, 2)),
+         g8 = vorrq_u32(vshlq_n_u32(g6, 2), vshrq_n_u32(g6, 4)),
+         b8 = vorrq_u32(vshlq_n_u32(b5, 3), vshrq_n_u32(b5, 2));
+
+    // Now put all the 8-bit components into SkPMColor order.
+    return (uint8x16_t)vorrq_u32(vshlq_n_u32(r8, SK_R32_SHIFT),   // TODO: one shift is zero...
+                       vorrq_u32(vshlq_n_u32(g8, SK_G32_SHIFT),
+                       vorrq_u32(vshlq_n_u32(b8, SK_B32_SHIFT),
+                                 vdupq_n_u32(0xFF << SK_A32_SHIFT))));
+}
+
+static inline uint16x4_t narrow_to_565(uint8x16_t w8x16) {
+    uint32x4_t w = (uint32x4_t)w8x16;
+
+    // Extract out top RGB 565 bits of each pixel, with no rounding.
+    auto r5 = vandq_u32(vdupq_n_u32(31), vshrq_n_u32(w, SK_R32_SHIFT + 3)),
+         g6 = vandq_u32(vdupq_n_u32(63), vshrq_n_u32(w, SK_G32_SHIFT + 2)),
+         b5 = vandq_u32(vdupq_n_u32(31), vshrq_n_u32(w, SK_B32_SHIFT + 3));
+
+    // Now put the bits in place in the low 16-bits of each 32-bit lane.
+    auto spread = vorrq_u32(vshlq_n_u32(r5, 11),
+                  vorrq_u32(vshlq_n_u32(g6,  5),
+                            b5));
+
+    // Pack the low 16-bits of our 128-bit register down into a 64-bit register.
+    // spread == 0000 rgb3  0000 rgb2  0000 rgb1  0000 rgb0
+    //      v ==                       rgb3 rgb2  rgb1 rgb0
+    auto v = vmovn_u32(spread);
+    return v;
+}
+
+
+inline Sk4px Sk4px::Load4(const SkPMColor16 src[4]) {
+    return Sk16b(widen_to_8888(vld1_u16(src)));
+}
+inline Sk4px Sk4px::Load2(const SkPMColor16 src[2]) {
+    auto src2 = ((uint32_t)src[0]      )
+              | ((uint32_t)src[1] << 16);
+    return Sk16b(widen_to_8888(vcreate_u16(src2)));
+}
+inline Sk4px Sk4px::Load1(const SkPMColor16 src[1]) {
+    return Sk16b(widen_to_8888(vcreate_u16(src[0])));
+}
+
+inline void Sk4px::store4(SkPMColor16 dst[4]) const {
+    vst1_u16(dst, narrow_to_565(this->fVec));
+}
+inline void Sk4px::store2(SkPMColor16 dst[2]) const {
+    auto v = narrow_to_565(this->fVec);
+    dst[0] = vget_lane_u16(v, 0);
+    dst[1] = vget_lane_u16(v, 1);
+}
+inline void Sk4px::store1(SkPMColor16 dst[1]) const {
+    dst[0] = vget_lane_u16(narrow_to_565(this->fVec), 0);
+}
+
 } // namespace