NEON optimizations for gray -> RGBA (or BGRA) conversions

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 70 matching lines @@
                 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;
     }
 }
 
+static void gray_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++) {
+        dst[i] = (uint32_t)0xFF   << 24
+               | (uint32_t)src[i] << 16
+               | (uint32_t)src[i] <<  8
+               | (uint32_t)src[i] <<  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 152 matching lines @@
 }
 
 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);
 }
 
+static void gray_to_RGB1(uint32_t dst[], const void* vsrc, int count) {
+    const uint8_t* src = (const uint8_t*) vsrc;
+    while (count >= 16) {
+        // Load 16 pixels.
+        uint8x16_t gray = vld1q_u8(src);
+
+        // Set each of the color channels.
+        uint8x16x4_t rgba;
+        rgba.val[0] = gray;
+        rgba.val[1] = gray;
+        rgba.val[2] = gray;
+        rgba.val[3] = vdupq_n_u8(0xFF);
+
+        // Store 16 pixels.
+        vst4q_u8((uint8_t*) dst, rgba);
+        src += 16;
+        dst += 16;
+        count -= 16;
+    }
+
+    if (count >= 8) {
+        // Load 8 pixels.
+        uint8x8_t gray = vld1_u8(src);
+
+        // Set each of the color channels.
+        uint8x8x4_t rgba;
+        rgba.val[0] = gray;
+        rgba.val[1] = gray;
+        rgba.val[2] = gray;
+        rgba.val[3] = vdup_n_u8(0xFF);
+
+        // Store 16 pixels.

[mtklein, 2016/02/02 19:34:59]

16 -> 8

[msarett, 2016/02/02 20:10:27]

Done.

+        vst4_u8((uint8_t*) dst, rgba);
+        src += 8;
+        dst += 8;
+        count -= 8;
+    }
+
+    gray_to_RGB1_portable(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 121 matching lines @@
 }
 
 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);
 }
 
+static void gray_to_RGB1(uint32_t dst[], const void* src, int count) {
+    gray_to_RGB1_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);
 }
 
+static void gray_to_RGB1(uint32_t dst[], const void* src, int count) {
+    gray_to_RGB1_portable(dst, src, count);
+}
+
 #endif
 
 }
 
 #endif // SkSwizzler_opts_DEFINED