NEON Optimized RGBA->PMColor sampling 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
 
 #include "SkColorPriv.h"
 
 namespace SK_OPTS_NS {
 
+template <int kSampleSize>
 static void RGBA_to_rgbA_portable(uint32_t* dst, const void* vsrc, int count) {
     auto src = (const uint32_t*)vsrc;
+    int j = 0;
     for (int i = 0; i < count; i++) {
-        uint8_t a = src[i] >> 24,
-                b = src[i] >> 16,
-                g = src[i] >>  8,
-                r = src[i] >>  0;
+        uint8_t a = src[j] >> 24,
+                b = src[j] >> 16,
+                g = src[j] >>  8,
+                r = src[j] >>  0;
+        j += kSampleSize;
         b = (b*a+127)/255;
         g = (g*a+127)/255;
         r = (r*a+127)/255;
         dst[i] = (uint32_t)a << 24
                | (uint32_t)b << 16
                | (uint32_t)g <<  8
                | (uint32_t)r <<  0;
     }
 }
 
+template <int kSampleSize>
 static void RGBA_to_bgrA_portable(uint32_t* dst, const void* vsrc, int count) {
     auto src = (const uint32_t*)vsrc;
+    int j = 0;
     for (int i = 0; i < count; i++) {
-        uint8_t a = src[i] >> 24,
-                b = src[i] >> 16,
-                g = src[i] >>  8,
-                r = src[i] >>  0;
+        uint8_t a = src[j] >> 24,
+                b = src[j] >> 16,
+                g = src[j] >>  8,
+                r = src[j] >>  0;
+        j += kSampleSize;
         b = (b*a+127)/255;
         g = (g*a+127)/255;
         r = (r*a+127)/255;
         dst[i] = (uint32_t)a << 24
                | (uint32_t)r << 16
                | (uint32_t)g <<  8
                | (uint32_t)b <<  0;
     }
 }
 
@@ skipping 169 matching lines @@
             rgba.val[1] = g;
             rgba.val[0] = r;
         }
         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;
+    auto proc = kSwapRB ? RGBA_to_bgrA_portable<1> : RGBA_to_rgbA_portable<1>;
     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);
 }
@@ skipping 238 matching lines @@
 }
 
 static void inverted_CMYK_to_RGB1(uint32_t dst[], const void* src, int count) {
     inverted_cmyk_to<kRGB1>(dst, src, count);
 }
 
 static void inverted_CMYK_to_BGR1(uint32_t dst[], const void* src, int count) {
     inverted_cmyk_to<kBGR1>(dst, src, count);
 }
 
+static void load_rgba_sample2(const uint32_t* src, uint8x8_t* r, uint8x8_t* g, uint8x8_t* b,
+        uint8x8_t* a) {
+    // Load 16 pixels.
+    uint8x16x4_t rgba = vld4q_u8((const uint8_t*) src);
+
+    // Choose 8 pixels.
+    // pxpxpxpxpxpxpxpx -> pppppppp
+    *r = vmovn_u16(vreinterpretq_u16_u8(rgba.val[0]));
+    *g = vmovn_u16(vreinterpretq_u16_u8(rgba.val[1]));
+    *b = vmovn_u16(vreinterpretq_u16_u8(rgba.val[2]));
+    *a = vmovn_u16(vreinterpretq_u16_u8(rgba.val[3]));
+}
+
+template <int kSampleSize>
+static void load_rgba_sample(const uint32_t* src, uint8x8_t* r, uint8x8_t* g, uint8x8_t* b,
+        uint8x8_t* a) {
+    uint8x8_t rgba0 = vld1_u8((const uint8_t*) src);                 // rgba xxxx
+    uint8x8_t rgba1 = vld1_u8((const uint8_t*) src + kSampleSize);   // rgba xxxx
+    uint8x8_t rgba01 = vzip_u8(rgba0, rgba1).val[0];                 // rrgg bbaa
+
+    uint8x8_t rgba2 = vld1_u8((const uint8_t*) src + 2*kSampleSize); // rgba xxxx
+    uint8x8_t rgba3 = vld1_u8((const uint8_t*) src + 3*kSampleSize); // rgba xxxx
+    uint8x8_t rgba23 = vzip_u8(rgba2, rgba3).val[0];                 // rrgg bbaa
+
+    uint16x4x2_t rgba03 = vzip_u16(vreinterpret_u16_u8(rgba01),      // rrrr gggg
+                                   vreinterpret_u16_u8(rgba23));     // bbbb aaaa
+
+    uint8x8_t rgba4 = vld1_u8((const uint8_t*) src + 4*kSampleSize); // rgba xxxx
+    uint8x8_t rgba5 = vld1_u8((const uint8_t*) src + 5*kSampleSize); // rgba xxxx
+    uint8x8_t rgba45 = vzip_u8(rgba4, rgba5).val[0];                 // rrgg bbaa
+
+    uint8x8_t rgba6 = vld1_u8((const uint8_t*) src + 6*kSampleSize); // rgba xxxx
+    uint8x8_t rgba7 = vld1_u8((const uint8_t*) src + 7*kSampleSize); // rgba xxxx
+    uint8x8_t rgba67 = vzip_u8(rgba6, rgba7).val[0];                 // rrgg bbaa
+
+    uint16x4x2_t rgba47 = vzip_u16(vreinterpret_u16_u8(rgba45),      // rrrr gggg
+                                   vreinterpret_u16_u8(rgba67));     // bbbb aaaa
+
+    uint32x2x2_t rg = vzip_u32(vreinterpret_u32_u16(rgba03.val[0]),  // rrrr rrrr
+                               vreinterpret_u32_u16(rgba47.val[0])); // gggg gggg
+    uint32x2x2_t ba = vzip_u32(vreinterpret_u32_u16(rgba03.val[1]),  // bbbb bbbb
+                               vreinterpret_u32_u16(rgba47.val[1])); // aaaa aaaa
+
+    *r = vreinterpret_u8_u32(rg.val[0]);
+    *g = vreinterpret_u8_u32(rg.val[1]);
+    *b = vreinterpret_u8_u32(ba.val[0]);
+    *a = vreinterpret_u8_u32(ba.val[1]);
+}
+
+template <bool kSwapRB, int kSampleSize>
+static void premul_should_swapRB_sample(uint32_t* dst, const void* vsrc, int count) {
+    auto src = (const uint32_t*)vsrc;
+
+    // We must use 9 as the limit to be sure that we don't read past the end of our memory.
+    while (count >= 9) {
+        // Load pixels.
+        uint8x8_t r, g, b, a;
+        if (2 == kSampleSize) {
+            load_rgba_sample2(src, &r, &g, &b, &a);
+        } else {
+            load_rgba_sample<kSampleSize>(src, &r, &g, &b, &a);
+        }
+
+        // Premultiply.
+        r = scale(r, a);
+        g = scale(g, a);
+        b = scale(b, a);
+
+        // Store 8 premultiplied pixels.
+        uint8x8x4_t result;
+        if (kSwapRB) {
+            result.val[0] = b;
+            result.val[1] = g;
+            result.val[2] = r;
+            result.val[3] = a;
+        } else {
+            result.val[0] = r;
+            result.val[1] = g;
+            result.val[2] = b;
+            result.val[3] = a;
+        }
+        vst4_u8((uint8_t*) dst, result);
+        src += 8*kSampleSize;
+        dst += 8;
+        count -= 8;
+    }
+
+    auto proc = kSwapRB ? RGBA_to_bgrA_portable<kSampleSize> : RGBA_to_rgbA_portable<kSampleSize>;
+    proc(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample2(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<false, 2>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample2(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<true, 2>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample4(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<false, 4>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample4(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<true, 4>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample8(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<false, 8>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample8(uint32_t dst[], const void* src, int count) {
+    premul_should_swapRB_sample<true, 8>(dst, src, count);
+}
+
 #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
 
 // Scale a byte by another.
 // Inputs are stored in 16-bit lanes, but are not larger than 8-bits.
 static __m128i scale(__m128i x, __m128i y) {
     const __m128i _128 = _mm_set1_epi16(128);
     const __m128i _257 = _mm_set1_epi16(257);
 
     // (x+127)/255 == ((x+128)*257)>>16 for 0 <= x <= 255*255.
     return _mm_mulhi_epu16(_mm_add_epi16(_mm_mullo_epi16(x, y), _128), _257);
@@ skipping 57 matching lines @@
         premul8(&lo, &hi);
 
         _mm_storeu_si128((__m128i*) dst, lo);
 
         src += 4;
         dst += 4;
         count -= 4;
     }
 
     // Call portable code to finish up the tail of [0,4) pixels.
-    auto proc = kSwapRB ? RGBA_to_bgrA_portable : RGBA_to_rgbA_portable;
+    auto proc = kSwapRB ? RGBA_to_bgrA_portable<1> : RGBA_to_rgbA_portable<1>;
     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);
 }
@@ skipping 209 matching lines @@
 }
 
 static void inverted_CMYK_to_RGB1(uint32_t dst[], const void* src, int count) {
     inverted_cmyk_to<kRGB1>(dst, src, count);
 }
 
 static void inverted_CMYK_to_BGR1(uint32_t dst[], const void* src, int count) {
     inverted_cmyk_to<kBGR1>(dst, src, count);
 }
 
+static void RGBA_to_rgbA_sample2(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<2>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample2(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<2>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample4(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<4>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample4(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<4>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample8(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<8>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample8(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<8>(dst, src, count);
+}
+
 #else
 
 static void RGBA_to_rgbA(uint32_t* dst, const void* src, int count) {
-    RGBA_to_rgbA_portable(dst, src, count);
+    RGBA_to_rgbA_portable<1>(dst, src, count);
 }
 
 static void RGBA_to_bgrA(uint32_t* dst, const void* src, int count) {
-    RGBA_to_bgrA_portable(dst, src, count);
+    RGBA_to_bgrA_portable<1>(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);
 }
 
@@ skipping 14 matching lines @@
 }
 
 static void inverted_CMYK_to_RGB1(uint32_t dst[], const void* src, int count) {
     inverted_CMYK_to_RGB1_portable(dst, src, count);
 }
 
 static void inverted_CMYK_to_BGR1(uint32_t dst[], const void* src, int count) {
     inverted_CMYK_to_BGR1_portable(dst, src, count);
 }
 
+static void RGBA_to_rgbA_sample2(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<2>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample2(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<2>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample4(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<4>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample4(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<4>(dst, src, count);
+}
+
+static void RGBA_to_rgbA_sample8(uint32_t dst[], const void* src, int count) {
+    RGBA_to_rgbA_portable<8>(dst, src, count);
+}
+
+static void RGBA_to_bgrA_sample8(uint32_t dst[], const void* src, int count) {
+    RGBA_to_bgrA_portable<8>(dst, src, count);
+}
+
 #endif
 
 }
 
 #endif // SkSwizzler_opts_DEFINED