Add capability for SkColorXform to output half floats

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
 
 #include "SkNx.h"
 #include "SkColorPriv.h"
+#include "SkHalf.h"
 #include "SkSRGB.h"
+#include "SkTemplates.h"
 
 namespace SK_OPTS_NS {
 
 static Sk4f linear_to_2dot2(const Sk4f& x) {
     // x^(29/64) is a very good approximation of the true value, x^(1/2.2).
     auto x2  = x.rsqrt(),                            // x^(-1/2)
          x32 = x2.rsqrt().rsqrt().rsqrt().rsqrt(),   // x^(-1/32)
          x64 = x32.rsqrt();                          // x^(+1/64)
 
     // 29 = 32 - 2 - 1
     return 255.0f * x2.invert() * x32 * x64.invert();
 }
 
 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);
 }
 
 enum DstGamma {
+    // 8888
     kSRGB_DstGamma,
     k2Dot2_DstGamma,
     kTable_DstGamma,
+
+    // F16
+    kLinear_DstGamma,
 };
 
 template <DstGamma kDstGamma>
-static void color_xform_RGB1(uint32_t* dst, const uint32_t* src, int len,
+static void color_xform_RGB1(void* dst, const uint32_t* src, int len,
                              const float* const srcTables[3], const float matrix[16],
                              const uint8_t* const dstTables[3]) {
     Sk4f rXgXbX = Sk4f::Load(matrix + 0),
          rYgYbY = Sk4f::Load(matrix + 4),
          rZgZbZ = Sk4f::Load(matrix + 8);
 
     if (len >= 4) {
         Sk4f reds, greens, blues;
         auto load_next_4 = [&reds, &greens, &blues, &src, &len, &srcTables] {
             reds   = Sk4f{srcTables[0][(src[0] >>  0) & 0xFF],
@@ skipping 26 matching lines @@
                         (kSRGB_DstGamma == kDstGamma) ? sk_linear_to_srgb : linear_to_2dot2;
 
                 dstReds   = linear_to_curve(dstReds);
                 dstGreens = linear_to_curve(dstGreens);
                 dstBlues  = linear_to_curve(dstBlues);
 
                 dstReds   = clamp_0_to_255(dstReds);
                 dstGreens = clamp_0_to_255(dstGreens);
                 dstBlues  = clamp_0_to_255(dstBlues);
 
+#ifdef SK_PMCOLOR_IS_RGBA
                 auto rgba = (Sk4f_round(dstReds)        )

[mtklein, 2016/07/15 16:45:41]

Might feel better to write this sort of stuff using the _SHIFT or _INDEX macros
like this:

auto rgba = (Sk4f_round(dstReds  ) << SK_R32_SHIFT)
          | (Sk4f_round(dstGreens) << SK_G32_SHIFT)
          | ...

[msarett, 2016/07/15 17:18:23]

Yes way better!

                           | (Sk4f_round(dstGreens) <<  8)
                           | (Sk4f_round(dstBlues)  << 16)
                           | (Sk4i{      0xFF       << 24});
-                rgba.store(dst);
-            } else {
+#else
+                auto rgba = (Sk4f_round(dstBlues)       )
+                          | (Sk4f_round(dstGreens) <<  8)
+                          | (Sk4f_round(dstReds)   << 16)
+                          | (Sk4i{      0xFF       << 24});
+#endif
+                rgba.store((uint32_t*) dst);
+
+                dst = SkTAddOffset<void>(dst, 4 * sizeof(uint32_t));
+            } else if (kTable_DstGamma == kDstGamma) {
                 Sk4f scaledReds   = Sk4f::Min(Sk4f::Max(1023.0f * dstReds,   0.0f), 1023.0f);
                 Sk4f scaledGreens = Sk4f::Min(Sk4f::Max(1023.0f * dstGreens, 0.0f), 1023.0f);
                 Sk4f scaledBlues  = Sk4f::Min(Sk4f::Max(1023.0f * dstBlues,  0.0f), 1023.0f);
 
-                Sk4i indicesReds   = Sk4f_round(scaledReds);
-                Sk4i indicesGreens = Sk4f_round(scaledGreens);
-                Sk4i indicesBlues  = Sk4f_round(scaledBlues);
+#ifdef SK_PMCOLOR_IS_RGBA
+                Sk4i indicesReds   = Sk4f_round(scaledReds + 0.5f);
+                Sk4i indicesGreens = Sk4f_round(scaledGreens + 0.5f);
+                Sk4i indicesBlues  = Sk4f_round(scaledBlues + 0.5f);
+#else
+                Sk4i indicesReds   = Sk4f_round(scaledBlues + 0.5f);
+                Sk4i indicesGreens = Sk4f_round(scaledGreens + 0.5f);
+                Sk4i indicesBlues  = Sk4f_round(scaledReds + 0.5f);
+#endif
 
-                dst[0] = dstTables[0][indicesReds  [0]]
-                       | dstTables[1][indicesGreens[0]] <<  8
-                       | dstTables[2][indicesBlues [0]] << 16
-                       | 0xFF                           << 24;
-                dst[1] = dstTables[0][indicesReds  [1]]
-                       | dstTables[1][indicesGreens[1]] <<  8
-                       | dstTables[2][indicesBlues [1]] << 16
-                       | 0xFF                           << 24;
-                dst[2] = dstTables[0][indicesReds  [2]]
-                       | dstTables[1][indicesGreens[2]] <<  8
-                       | dstTables[2][indicesBlues [2]] << 16
-                       | 0xFF                           << 24;
-                dst[3] = dstTables[0][indicesReds  [3]]
-                       | dstTables[1][indicesGreens[3]] <<  8
-                       | dstTables[2][indicesBlues [3]] << 16
-                       | 0xFF                           << 24;
+                uint32_t* dst32 = (uint32_t*) dst;
+                dst32[0] = dstTables[0][indicesReds  [0]]
+                         | dstTables[1][indicesGreens[0]] <<  8
+                         | dstTables[2][indicesBlues [0]] << 16
+                         | 0xFF                           << 24;
+                dst32[1] = dstTables[0][indicesReds  [1]]
+                         | dstTables[1][indicesGreens[1]] <<  8
+                         | dstTables[2][indicesBlues [1]] << 16
+                         | 0xFF                           << 24;
+                dst32[2] = dstTables[0][indicesReds  [2]]
+                         | dstTables[1][indicesGreens[2]] <<  8
+                         | dstTables[2][indicesBlues [2]] << 16
+                         | 0xFF                           << 24;
+                dst32[3] = dstTables[0][indicesReds  [3]]
+                         | dstTables[1][indicesGreens[3]] <<  8
+                         | dstTables[2][indicesBlues [3]] << 16
+                         | 0xFF                           << 24;
+
+                dst = SkTAddOffset<void>(dst, 4 * sizeof(uint32_t));
+            } else {
+                // FIXME (msarett):
+                // Can we do better here?  Should we store half floats as planar?
+                // Should we write Intel/Arm specific code?  Should we add a transpose

[mtklein, 2016/07/15 16:45:41]

I think,

Very likely,
No (Or Maybe, but let's make it not your problem),
Yes,
Maybe,
No.

[msarett, 2016/07/15 17:18:23]

Haha thanks.

+                // function to SkNx?  Should we rewrite the algorithm to be interleaved?
+                uint64_t* dst64 = (uint64_t*) dst;
+                dst64[0] = SkFloatToHalf_finite(Sk4f(dstReds[0], dstGreens[0], dstBlues[0], 1.0f));
+                dst64[1] = SkFloatToHalf_finite(Sk4f(dstReds[1], dstGreens[1], dstBlues[1], 1.0f));
+                dst64[2] = SkFloatToHalf_finite(Sk4f(dstReds[2], dstGreens[2], dstBlues[2], 1.0f));
+                dst64[3] = SkFloatToHalf_finite(Sk4f(dstReds[3], dstGreens[3], dstBlues[3], 1.0f));
+
+                dst = SkTAddOffset<void>(dst, 4 * sizeof(uint64_t));
             }
-
-            dst += 4;
         };
 
         load_next_4();
 
         while (len >= 4) {
             transform_4();
             load_next_4();
             store_4();
         }
 
         transform_4();
         store_4();
     }
 
     while (len > 0) {
         // Splat r,g,b across a register each.
         auto r = Sk4f{srcTables[0][(*src >>  0) & 0xFF]},
              g = Sk4f{srcTables[1][(*src >>  8) & 0xFF]},
              b = Sk4f{srcTables[2][(*src >> 16) & 0xFF]};
 
-        // Apply transformation matrix to dst gamut.
-        auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
+        if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) {
+#ifdef SK_PMCOLOR_IS_RGBA

[msarett, 2016/07/15 17:18:23]

This is nonsense.  Never was going to work...

+            auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
+#else
+            auto dstPixel = rXgXbX*b + rYgYbY*g + rZgZbZ*r;
+#endif
 
-        if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) {
             Sk4f (*linear_to_curve)(const Sk4f&) =
                     (kSRGB_DstGamma == kDstGamma) ? sk_linear_to_srgb : linear_to_2dot2;
 
             dstPixel = linear_to_curve(dstPixel);
 
             dstPixel = clamp_0_to_255(dstPixel);
 
             uint32_t rgba;
             SkNx_cast<uint8_t>(Sk4f_round(dstPixel)).store(&rgba);
             rgba |= 0xFF000000;
-            *dst = rgba;
-        } else {
+            *((uint32_t*) dst) = rgba;
+            dst = SkTAddOffset<void>(dst, sizeof(uint32_t));
+        } else if (kTable_DstGamma == kDstGamma) {
+#ifdef SK_PMCOLOR_IS_RGBA
+            auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
+#else
+            auto dstPixel = rXgXbX*b + rYgYbY*g + rZgZbZ*r;
+#endif
+
             Sk4f scaledPixel = Sk4f::Min(Sk4f::Max(1023.0f * dstPixel, 0.0f), 1023.0f);
 
             Sk4i indices = Sk4f_round(scaledPixel);
 
-            *dst = dstTables[0][indices[0]]
-                 | dstTables[1][indices[1]] <<  8
-                 | dstTables[2][indices[2]] << 16
-                 | 0xFF                     << 24;
+            *((uint32_t*) dst) = dstTables[0][indices[0]]
+                               | dstTables[1][indices[1]] <<  8
+                               | dstTables[2][indices[2]] << 16
+                               | 0xFF                     << 24;
+
+            dst = SkTAddOffset<void>(dst, sizeof(uint32_t));
+        } else {
+            auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b;
+
+            uint64_t rgba = SkFloatToHalf_finite(dstPixel);
+
+            // Set alpha to 1.0
+            rgba |= 0x3C00000000000000;
+            *((uint64_t*) dst) = rgba;
+            dst = SkTAddOffset<void>(dst, sizeof(uint64_t));
         }
 
-        dst += 1;
         src += 1;
         len -= 1;
     }
 }
 
 static void color_xform_RGB1_to_2dot2(uint32_t* dst, const uint32_t* src, int len,
-                                      const float* const srcTables[3], const float matrix[16]) {
+                                      const float* const srcTables[3], const float matrix[12]) {
     color_xform_RGB1<k2Dot2_DstGamma>(dst, src, len, srcTables, matrix, nullptr);
 }
 
 static void color_xform_RGB1_to_srgb(uint32_t* dst, const uint32_t* src, int len,
-                                     const float* const srcTables[3], const float matrix[16]) {
+                                     const float* const srcTables[3], const float matrix[12]) {
     color_xform_RGB1<kSRGB_DstGamma>(dst, src, len, srcTables, matrix, nullptr);
 }
 
 static void color_xform_RGB1_to_table(uint32_t* dst, const uint32_t* src, int len,
-                                      const float* const srcTables[3], const float matrix[16],
+                                      const float* const srcTables[3], const float matrix[12],
                                       const uint8_t* const dstTables[3]) {
     color_xform_RGB1<kTable_DstGamma>(dst, src, len, srcTables, matrix, dstTables);
 }
 
+static void color_xform_RGB1_to_linear(uint64_t* dst, const uint32_t* src, int len,
+                                       const float* const srcTables[3], const float matrix[12]) {
+    color_xform_RGB1<kLinear_DstGamma>(dst, src, len, srcTables, matrix, nullptr);
+}
+
 }  // namespace SK_OPTS_NS
 
 #endif // SkColorXform_opts_DEFINED