Add SSE optimization of Color32A_D565

src/opts/SkBlitRow_opts_SSE4.cpp

Index: src/opts/SkBlitRow_opts_SSE4.cpp
diff --git a/src/opts/SkBlitRow_opts_SSE4.cpp b/src/opts/SkBlitRow_opts_SSE4.cpp
index fd837d54fee4d879fd3d1d25887c343d1c51279c..f4273d27b4b0bfaf81d1df0c5d86f8f25f5a8164 100644
--- a/src/opts/SkBlitRow_opts_SSE4.cpp
+++ b/src/opts/SkBlitRow_opts_SSE4.cpp
@@ -7,10 +7,13 @@ void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT, const SkPMColor* SK_REST
     sk_throw();
 }
 
+void Color32A_D565_SSE4(uint16_t dst[], SkPMColor src, int count, int x, int y) {
+    sk_throw();
+}
+
 #else
 
-#include <emmintrin.h>  // SSE2:   Most _mm_foo() in this file.
-#include <smmintrin.h>  // SSE4.1: _mm_testz_si128 and _mm_testc_si128.
+#include <smmintrin.h>  // SSE4.1 intrinsics
 
 #include "SkColorPriv.h"
 #include "SkColor_opts_SSE2.h"
@@ -63,4 +66,76 @@ void S32A_Opaque_BlitRow32_SSE4(SkPMColor* SK_RESTRICT dst,
     }
 }
 
+static inline uint16_t Color32A_D565_1x(uint16_t dst, unsigned scale, uint32_t src_expand) {
+    uint32_t dst_expand = SkExpand_rgb_16(dst) * scale;
+    return SkCompact_rgb_16((src_expand + dst_expand) >> 5);
+}
+
+void Color32A_D565_SSE4(uint16_t dst[], SkPMColor src, int count, int x, int y) {
+    SkASSERT(count > 0);
+
+    uint32_t src_expand = (SkGetPackedG32(src) << 24) |
+                          (SkGetPackedR32(src) << 13) |
+                          (SkGetPackedB32(src) << 2);
+    unsigned scale = SkAlpha255To256(0xFF - SkGetPackedA32(src)) >> 3;
+
+    // Check if we have enough pixels to run SIMD
+    if (count >= (int)(8 + (((16 - (size_t)dst) & 0x0F) >> 1))) {
+        __m128i* dst_wide;
+        const __m128i src_expand_wide = _mm_set1_epi32(src_expand);
+        const __m128i scale_wide = _mm_set1_epi32(scale);
+        const __m128i mask_green = _mm_set1_epi32(SK_R16_MASK_IN_PLACE |
+                                                  SK_B16_MASK_IN_PLACE |
+                                                 (SK_G16_MASK_IN_PLACE << 16));
+
+        // Align dst to an even 16 byte address (0-7 pixels)
+        while (((((size_t)dst) & 0x0F) != 0) && (count > 0)) {
+            *dst = Color32A_D565_1x(*dst, scale, src_expand);
+            dst += 1;
+            count--;
+        }
+
+        dst_wide = reinterpret_cast<__m128i*>(dst);
+        do {
+            // Load 8 RGB565 pixels
+            __m128i pixels = _mm_load_si128(dst_wide);
+
+            // Duplicate and mask
+            __m128i pixels_high = _mm_unpackhi_epi16(pixels, pixels);
+            pixels_high = _mm_and_si128(mask_green, pixels_high);
+            pixels = _mm_unpacklo_epi16(pixels, pixels);
+            pixels = _mm_and_si128(mask_green, pixels);
+
+            // Scale with alpha
+            pixels_high = _mm_mullo_epi32(pixels_high, scale_wide);
+            pixels = _mm_mullo_epi32(pixels, scale_wide);
+
+            // Add src_expand_wide and shift down again
+            pixels_high = _mm_add_epi32(pixels_high, src_expand_wide);
+            pixels_high = _mm_srli_epi32(pixels_high, 5);
+            pixels = _mm_add_epi32(pixels, src_expand_wide);
+            pixels = _mm_srli_epi32(pixels, 5);
+
+            // Mask
+            pixels_high = _mm_and_si128(mask_green, pixels_high);
+            pixels = _mm_and_si128(mask_green, pixels);
+
+            // Combine into RGB565 and store
+            pixels = _mm_hadd_epi16(pixels, pixels_high);
+            _mm_store_si128(dst_wide, pixels);
+            count -= 8;
+            dst_wide++;
+        } while (count >= 8);
+
+        dst = reinterpret_cast<uint16_t*>(dst_wide);
+    }
+
+    // Small loop to handle remaining pixels.
+    while (count > 0) {
+        *dst = Color32A_D565_1x(*dst, scale, src_expand);
+        dst += 1;
+        count--;
+    }
+}
+
 #endif