Do loads and math in parallel in SkColorXform_opts

src/opts/SkColorXform_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 SkColorXform_opts_DEFINED
 #define SkColorXform_opts_DEFINED
 
@@ skipping 168 matching lines @@
 
 static Sk4f clamp_0_to_255(const Sk4f& x) {
     // The order of the arguments is important here.  We want to make sure that NaN
     // clamps to zero.  Note that max(NaN, 0) = 0, while max(0, NaN) = NaN.
     return Sk4f::Min(Sk4f::Max(x, 0.0f), 255.0f);
 }
 
 template <const float (&linear_from_curve)[256], Sk4f (*linear_to_curve)(const Sk4f&)>
 static void color_xform_RGB1(uint32_t* dst, const uint32_t* src, int len,
                              const float matrix[16]) {
-    // Load transformation matrix.
-    auto rXgXbX = Sk4f::Load(matrix + 0),
+
+    Sk4f reds, greens, blues;
+    auto load_next_4 = [&reds, &greens, &blues, &src, &len] {
+        reds   = Sk4f{linear_from_curve[(src[0] >>  0) & 0xFF],
+                      linear_from_curve[(src[1] >>  0) & 0xFF],
+                      linear_from_curve[(src[2] >>  0) & 0xFF],
+                      linear_from_curve[(src[3] >>  0) & 0xFF]};
+        greens = Sk4f{linear_from_curve[(src[0] >>  8) & 0xFF],
+                      linear_from_curve[(src[1] >>  8) & 0xFF],
+                      linear_from_curve[(src[2] >>  8) & 0xFF],
+                      linear_from_curve[(src[3] >>  8) & 0xFF]};
+        blues  = Sk4f{linear_from_curve[(src[0] >> 16) & 0xFF],
+                      linear_from_curve[(src[1] >> 16) & 0xFF],
+                      linear_from_curve[(src[2] >> 16) & 0xFF],
+                      linear_from_curve[(src[3] >> 16) & 0xFF]};
+        src += 4;
+        len -= 4;
+    };
+
+    Sk4f dstReds, dstGreens, dstBlues;
+    Sk4f rXgXbX = Sk4f::Load(matrix + 0),
          rYgYbY = Sk4f::Load(matrix + 4),
          rZgZbZ = Sk4f::Load(matrix + 8);
+    auto gamut_transform_4 = [&reds, &greens, &blues, &dstReds, &dstGreens, &dstBlues, &rXgXbX,
+                              &rYgYbY, &rZgZbZ] {
+        dstReds   = rXgXbX[0]*reds + rYgYbY[0]*greens + rZgZbZ[0]*blues;
+        dstGreens = rXgXbX[1]*reds + rYgYbY[1]*greens + rZgZbZ[1]*blues;
+        dstBlues  = rXgXbX[2]*reds + rYgYbY[2]*greens + rZgZbZ[2]*blues;
+    };
 
-    while (len >= 4) {
-        // Convert to linear.  The look-up table has perfect accuracy.
-        auto reds   = Sk4f{linear_from_curve[(src[0] >>  0) & 0xFF],
-                           linear_from_curve[(src[1] >>  0) & 0xFF],
-                           linear_from_curve[(src[2] >>  0) & 0xFF],
-                           linear_from_curve[(src[3] >>  0) & 0xFF]};
-        auto greens = Sk4f{linear_from_curve[(src[0] >>  8) & 0xFF],
-                           linear_from_curve[(src[1] >>  8) & 0xFF],
-                           linear_from_curve[(src[2] >>  8) & 0xFF],
-                           linear_from_curve[(src[3] >>  8) & 0xFF]};
-        auto blues  = Sk4f{linear_from_curve[(src[0] >> 16) & 0xFF],
-                           linear_from_curve[(src[1] >> 16) & 0xFF],
-                           linear_from_curve[(src[2] >> 16) & 0xFF],
-                           linear_from_curve[(src[3] >> 16) & 0xFF]};
-
-        // Apply the transformation matrix to dst gamut.
-        auto dstReds   = rXgXbX[0]*reds + rYgYbY[0]*greens + rZgZbZ[0]*blues,
-             dstGreens = rXgXbX[1]*reds + rYgYbY[1]*greens + rZgZbZ[1]*blues,
-             dstBlues  = rXgXbX[2]*reds + rYgYbY[2]*greens + rZgZbZ[2]*blues;
-
-        // Convert to dst gamma.
+    auto gamma_transform_4 = [&dstReds, &dstGreens, &dstBlues] {
         dstReds   = linear_to_curve(dstReds);
         dstGreens = linear_to_curve(dstGreens);
         dstBlues  = linear_to_curve(dstBlues);
 
-        // Clamp floats to byte range.
         dstReds   = clamp_0_to_255(dstReds);
         dstGreens = clamp_0_to_255(dstGreens);
         dstBlues  = clamp_0_to_255(dstBlues);
+    };
 
-        // Convert to bytes and store to memory.
-        auto rgba = (Sk4i{(int)0xFF000000}          )
-                  | (SkNx_cast<int>(dstReds)        )
-                  | (SkNx_cast<int>(dstGreens) <<  8)
-                  | (SkNx_cast<int>(dstBlues)  << 16);
+    Sk4i rgba;
+    auto store_4 = [&dstReds, &dstGreens, &dstBlues, &rgba, &dst] {
+        rgba = (Sk4i{(int)0xFF000000}          )
+             | (SkNx_cast<int>(dstReds)        )
+             | (SkNx_cast<int>(dstGreens) <<  8)
+             | (SkNx_cast<int>(dstBlues)  << 16);
         rgba.store(dst);
+        dst += 4;
+    };
 
-        dst += 4;
-        src += 4;
-        len -= 4;
+    if (len >= 4) {
+        load_next_4();
     }
 
+    while (len >= 4) {
+        gamut_transform_4();
+        load_next_4();
+        gamma_transform_4();

[mtklein, 2016/06/22 19:33:34]

Why not fuse gamma_transform_4() into store_4()?  They're always called one
after the other, and the fused linear->stored-gamma would mirror load_next_4()'s
stored-gamma->linear.

I'll also admit part of my request here is that I find it hard to read
gamut_transform_4() followed by gamma_transform_4().  I know they're different,
but every time I read it they look like duplicated code.

[msarett, 2016/06/22 19:44:43]

Yup, looks good this way.

+        store_4();
+    }
+
+    gamut_transform_4();
+    gamma_transform_4();
+    store_4();
+
     while (len > 0) {
         // Splat r,g,b across a register each.
         auto r = Sk4f{linear_from_curve[(*src >>  0) & 0xFF]},
              g = Sk4f{linear_from_curve[(*src >>  8) & 0xFF]},
              b = Sk4f{linear_from_curve[(*src >> 16) & 0xFF]};
 
         // Apply transformation matrix to dst gamut.
         auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
 
         // Convert to dst gamma.
@@ skipping 30 matching lines @@
 }
 
 static void color_xform_RGB1_2dot2_to_srgb(uint32_t* dst, const uint32_t* src, int len,
                                            const float matrix[16]) {
     color_xform_RGB1<linear_from_2dot2, linear_to_srgb>(dst, src, len, matrix);
 }
 
 }  // namespace SK_OPTS_NS
 
 #endif // SkColorXform_opts_DEFINED