Use NEON optimizations for RGB -> RGB(FF) or BGR(FF) in SkSwizzler

src/opts/SkSwizzler_opts.h

 /*
  * Copyright 2016 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkSwizzler_opts_DEFINED
 #define SkSwizzler_opts_DEFINED
 
@@ skipping 42 matching lines @@
                 b = src[i] >> 16,
                 g = src[i] >>  8,
                 r = src[i] >>  0;
         dst[i] = (uint32_t)a << 24
                | (uint32_t)r << 16
                | (uint32_t)g <<  8
                | (uint32_t)b <<  0;
     }
 }
 
+static void RGB_to_RGB1_portable(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*)vsrc;
+    for (int i = 0; i < count; i++) {
+        uint8_t r = src[0],
+                g = src[1],
+                b = src[2];
+        src += 3;
+        dst[i] = (uint32_t)0xFF << 24
+               | (uint32_t)b    << 16
+               | (uint32_t)g    <<  8
+               | (uint32_t)r    <<  0;
+    }
+}
+
+static void RGB_to_BGR1_portable(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*)vsrc;
+    for (int i = 0; i < count; i++) {
+        uint8_t r = src[0],
+                g = src[1],
+                b = src[2];
+        src += 3;
+        dst[i] = (uint32_t)0xFF << 24
+               | (uint32_t)r    << 16
+               | (uint32_t)g    <<  8
+               | (uint32_t)b    <<  0;
+    }
+}
+
 #if defined(SK_ARM_HAS_NEON)
 
 // Rounded divide by 255, (x + 127) / 255
 static uint8x8_t div255_round(uint16x8_t x) {
     // result = (x + 127) / 255
     // result = (x + 127) / 256 + error1
     //
     // error1 = (x + 127) / (255 * 256)
     // error1 = (x + 127) / (256 * 256) + error2
     //
@@ skipping 16 matching lines @@
 // Scale a byte by another, (x * y + 127) / 255
 static uint8x8_t scale(uint8x8_t x, uint8x8_t y) {
     return div255_round(vmull_u8(x, y));
 }
 
 template <bool kSwapRB>
 static void premul_should_swapRB(uint32_t* dst, const void* vsrc, int count) {
     auto src = (const uint32_t*)vsrc;
     while (count >= 8) {
         // Load 8 pixels.
-        uint8x8x4_t bgra = vld4_u8((const uint8_t*) src);
+        uint8x8x4_t rgba = vld4_u8((const uint8_t*) src);
 
-        uint8x8_t a = bgra.val[3],
-                  b = bgra.val[2],
-                  g = bgra.val[1],
-                  r = bgra.val[0];
+        uint8x8_t a = rgba.val[3],
+                  b = rgba.val[2],
+                  g = rgba.val[1],
+                  r = rgba.val[0];
 
         // Premultiply.
         b = scale(b, a);
         g = scale(g, a);
         r = scale(r, a);
 
         // Store 8 premultiplied pixels.
         if (kSwapRB) {
-            bgra.val[2] = r;
-            bgra.val[1] = g;
-            bgra.val[0] = b;
+            rgba.val[2] = r;
+            rgba.val[1] = g;
+            rgba.val[0] = b;
         } else {
-            bgra.val[2] = b;
-            bgra.val[1] = g;
-            bgra.val[0] = r;
+            rgba.val[2] = b;
+            rgba.val[1] = g;
+            rgba.val[0] = r;
         }
-        vst4_u8((uint8_t*) dst, bgra);
+        vst4_u8((uint8_t*) dst, rgba);
         src += 8;
         dst += 8;
         count -= 8;
     }
 
     // Call portable code to finish up the tail of [0,8) pixels.
     auto proc = kSwapRB ? RGBA_to_bgrA_portable : RGBA_to_rgbA_portable;
     proc(dst, src, count);
 }
 
 static void RGBA_to_rgbA(uint32_t* dst, const void* src, int count) {
     premul_should_swapRB<false>(dst, src, count);
 }
 
 static void RGBA_to_bgrA(uint32_t* dst, const void* src, int count) {
     premul_should_swapRB<true>(dst, src, count);
 }
 
 static void RGBA_to_BGRA(uint32_t* dst, const void* vsrc, int count) {
     auto src = (const uint32_t*)vsrc;
     while (count >= 16) {
         // Load 16 pixels.
-        uint8x16x4_t bgra = vld4q_u8((const uint8_t*) src);
+        uint8x16x4_t rgba = vld4q_u8((const uint8_t*) src);
 
         // Swap r and b.
-        SkTSwap(bgra.val[0], bgra.val[2]);
+        SkTSwap(rgba.val[0], rgba.val[2]);
 
         // Store 16 pixels.
-        vst4q_u8((uint8_t*) dst, bgra);
+        vst4q_u8((uint8_t*) dst, rgba);
         src += 16;
         dst += 16;
         count -= 16;
     }
 
     if (count >= 8) {
         // Load 8 pixels.
-        uint8x8x4_t bgra = vld4_u8((const uint8_t*) src);
+        uint8x8x4_t rgba = vld4_u8((const uint8_t*) src);
 
         // Swap r and b.
-        SkTSwap(bgra.val[0], bgra.val[2]);
+        SkTSwap(rgba.val[0], rgba.val[2]);
 
         // Store 8 pixels.
-        vst4_u8((uint8_t*) dst, bgra);
+        vst4_u8((uint8_t*) dst, rgba);
         src += 8;
         dst += 8;
         count -= 8;
     }
 
     RGBA_to_BGRA_portable(dst, src, count);
 }
 
+template <bool kSwapRB>
+static void insert_alpha_should_swaprb(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*) vsrc;
+    while (count >= 16) {
+        // Load 16 pixels.
+        uint8x16x3_t rgb = vld3q_u8(src);
+
+        // Insert an opaque alpha channel and swap if needed.
+        uint8x16x4_t rgba;
+        if (kSwapRB) {
+            rgba.val[0] = rgb.val[2];
+            rgba.val[2] = rgb.val[0];
+        } else {
+            rgba.val[0] = rgb.val[0];
+            rgba.val[2] = rgb.val[2];
+        }
+        rgba.val[1] = rgb.val[1];
+        rgba.val[3] = vdupq_n_u8(0xFF);
+
+        // Store 16 pixels.
+        vst4q_u8((uint8_t*) dst, rgba);
+        src += 16*3;
+        dst += 16;
+        count -= 16;
+    }
+
+    if (count >= 8) {
+        // Load 8 pixels.
+        uint8x8x3_t rgb = vld3_u8(src);
+
+        // Insert an opaque alpha channel and swap if needed.
+        uint8x8x4_t rgba;
+        if (kSwapRB) {
+            rgba.val[0] = rgb.val[2];
+            rgba.val[2] = rgb.val[0];
+        } else {
+            rgba.val[0] = rgb.val[0];
+            rgba.val[2] = rgb.val[2];
+        }
+        rgba.val[1] = rgb.val[1];
+        rgba.val[3] = vdup_n_u8(0xFF);
+
+        // Store 8 pixels.
+        vst4_u8((uint8_t*) dst, rgba);
+        src += 8*3;
+        dst += 8;
+        count -= 8;
+    }
+
+    // Call portable code to finish up the tail of [0,8) pixels.
+    auto proc = kSwapRB ? RGB_to_BGR1_portable : RGB_to_RGB1_portable;
+    proc(dst, src, count);
+}
+
+static void RGB_to_RGB1(uint32_t dst[], const void* src, int count) {
+    insert_alpha_should_swaprb<false>(dst, src, count);
+}
+
+static void RGB_to_BGR1(uint32_t dst[], const void* src, int count) {
+    insert_alpha_should_swaprb<true>(dst, src, count);
+}
+
 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
 
 template <bool kSwapRB>
 static void premul_should_swapRB(uint32_t* dst, const void* vsrc, int count) {
     auto src = (const uint32_t*)vsrc;
 
     auto premul8 = [](__m128i* lo, __m128i* hi) {
         const __m128i zeros = _mm_setzero_si128();
         const __m128i _128 = _mm_set1_epi16(128);
         const __m128i _257 = _mm_set1_epi16(257);
@@ skipping 78 matching lines @@
         _mm_storeu_si128((__m128i*) dst, bgra);
 
         src += 4;
         dst += 4;
         count -= 4;
     }
 
     RGBA_to_BGRA_portable(dst, src, count);
 }
 
+static void RGB_to_RGB1(uint32_t dst[], const void* src, int count) {
+    RGB_to_RGB1_portable(dst, src, count);
+}
+
+static void RGB_to_BGR1(uint32_t dst[], const void* src, int count) {
+    RGB_to_BGR1_portable(dst, src, count);
+}
+
 #else
 
 static void RGBA_to_rgbA(uint32_t* dst, const void* src, int count) {
     RGBA_to_rgbA_portable(dst, src, count);
 }
 
 static void RGBA_to_bgrA(uint32_t* dst, const void* src, int count) {
     RGBA_to_bgrA_portable(dst, src, count);
 }
 
 static void RGBA_to_BGRA(uint32_t* dst, const void* src, int count) {
     RGBA_to_BGRA_portable(dst, src, count);
 }
 
+static void RGB_to_RGB1(uint32_t dst[], const void* src, int count) {
+    RGB_to_RGB1_portable(dst, src, count);
+}
+
+static void RGB_to_BGR1(uint32_t dst[], const void* src, int count) {
+    RGB_to_BGR1_portable(dst, src, count);
+}
+
 #endif
 
 }
 
 #endif // SkSwizzler_opts_DEFINED