From 3ecb00f717e51039d28fabc71d3570847bb7f26c Mon Sep 17 00:00:00 2001
From: Hajime Hoshi <hajimehoshi@gmail.com>
Date: Sat, 16 Feb 2019 23:30:14 +0900
Subject: [PATCH] graphicsdriver: Optimize shaders

Simplified the case when a color matrix is not used.
---
 internal/graphicsdriver/metal/driver.go  | 13 ++++++++-----
 internal/graphicsdriver/opengl/shader.go | 10 ++++++----
 2 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/internal/graphicsdriver/metal/driver.go b/internal/graphicsdriver/metal/driver.go
index 283857cbc..e111e2a66 100644
--- a/internal/graphicsdriver/metal/driver.go
+++ b/internal/graphicsdriver/metal/driver.go
@@ -121,7 +121,6 @@ template<uint8_t address>
 struct GetColorFromTexel<FILTER_NEAREST, address> {
   inline float4 Do(VertexOut v, texture2d<float> texture, constant float2& source_size, float scale) {
     constexpr sampler texture_sampler(filter::nearest);
-    const float2 texel_size = 1 / source_size;
 
     float2 p = AdjustTexelByAddress<address>(v.tex, v.tex_region);
     if (p.x < v.tex_region[0] ||
@@ -203,13 +202,17 @@ float4 FragmentShaderImpl(
     constant float4& color_matrix_translation,
     constant float& scale) {
   float4 c = GetColorFromTexel<filter, address>().Do(v, texture, source_size, scale);
-  c.rgb /= c.a + (1.0 - sign(c.a));
   if (useColorM) {
+    c.rgb /= c.a + (1.0 - sign(c.a));
     c = (color_matrix_body * c) + color_matrix_translation;
+    c *= v.color;
+    c = clamp(c, 0.0, 1.0);
+    c.rgb *= c.a;
+  } else {
+    float4 s = v.color;
+    c *= float4(s.r, s.g, s.b, 1.0) * s.a;
+    c = clamp(c, 0.0, c.a);
   }
-  c *= v.color;
-  c = clamp(c, 0.0, 1.0);
-  c.rgb *= c.a;
   return c;
 }
 
diff --git a/internal/graphicsdriver/opengl/shader.go b/internal/graphicsdriver/opengl/shader.go
index 133e8a659..8941ad33f 100644
--- a/internal/graphicsdriver/opengl/shader.go
+++ b/internal/graphicsdriver/opengl/shader.go
@@ -256,19 +256,21 @@ void main(void) {
   color = mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
 #endif
 
+#if defined(USE_COLOR_MATRIX)
   // Un-premultiply alpha.
   // When the alpha is 0, 1.0 - sign(alpha) is 1.0, which means division does nothing.
   color.rgb /= color.a + (1.0 - sign(color.a));
-
-#if defined(USE_COLOR_MATRIX)
   // Apply the color matrix or scale.
   color = (color_matrix_body * color) + color_matrix_translation;
-#endif
-
   color *= varying_color_scale;
   color = clamp(color, 0.0, 1.0);
   // Premultiply alpha
   color.rgb *= color.a;
+#else
+  vec4 s = varying_color_scale;
+  color *= vec4(s.r, s.g, s.b, 1.0) * s.a;
+  color = clamp(color, 0.0, color.a);
+#endif
 
   gl_FragColor = color;
 }