From 3aa6eb3a66a623fba3371cfda8963a1ec9bdf4ff Mon Sep 17 00:00:00 2001
From: Hajime Hoshi <hajimehoshi@gmail.com>
Date: Sun, 17 Feb 2019 19:01:43 +0900
Subject: [PATCH] graphicsdriver: Optimize shaders for the case of
 FILTER_SCREEN

Assume a color matrix or color vertex values are not used when
the filter is FILTER_SCREEN.
---
 internal/graphicsdriver/metal/driver.go  | 63 +++++++++++++++---------
 internal/graphicsdriver/opengl/shader.go | 13 +++--
 2 files changed, 49 insertions(+), 27 deletions(-)

diff --git a/internal/graphicsdriver/metal/driver.go b/internal/graphicsdriver/metal/driver.go
index e53e7ebb4..2073996a9 100644
--- a/internal/graphicsdriver/metal/driver.go
+++ b/internal/graphicsdriver/metal/driver.go
@@ -115,10 +115,10 @@ inline float2 AdjustTexelByAddress<ADDRESS_REPEAT>(float2 p, float4 tex_region)
 }
 
 template<uint8_t filter, uint8_t address>
-struct GetColorFromTexel;
+struct ColorFromTexel;
 
 template<uint8_t address>
-struct GetColorFromTexel<FILTER_NEAREST, address> {
+struct ColorFromTexel<FILTER_NEAREST, address> {
   inline float4 Do(VertexOut v, texture2d<float> texture, constant float2& source_size, float scale) {
     float2 p = AdjustTexelByAddress<address>(v.tex, v.tex_region);
     if (v.tex_region[0] <= p.x &&
@@ -133,7 +133,7 @@ struct GetColorFromTexel<FILTER_NEAREST, address> {
 };
 
 template<uint8_t address>
-struct GetColorFromTexel<FILTER_LINEAR, address> {
+struct ColorFromTexel<FILTER_LINEAR, 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;
@@ -172,7 +172,7 @@ struct GetColorFromTexel<FILTER_LINEAR, address> {
 };
 
 template<uint8_t address>
-struct GetColorFromTexel<FILTER_SCREEN, address> {
+struct ColorFromTexel<FILTER_SCREEN, 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;
@@ -193,26 +193,41 @@ struct GetColorFromTexel<FILTER_SCREEN, address> {
 };
 
 template<bool useColorM, uint8_t filter, uint8_t address>
-float4 FragmentShaderImpl(
-    VertexOut v,
-    texture2d<float> texture,
-    constant float2& source_size,
-    constant float4x4& color_matrix_body,
-    constant float4& color_matrix_translation,
-    constant float& scale) {
-  float4 c = GetColorFromTexel<filter, address>().Do(v, texture, source_size, scale);
-  if (useColorM) {
-    c.rgb /= c.a + (1.0 - sign(c.a));
-    c = (color_matrix_body * c) + color_matrix_translation;
-    c *= v.color;
-    c.rgb *= c.a;
-  } else {
-    float4 s = v.color;
-    c *= float4(s.r, s.g, s.b, 1.0) * s.a;
+struct FragmentShaderImpl {
+  inline float4 Do(
+      VertexOut v,
+      texture2d<float> texture,
+      constant float2& source_size,
+      constant float4x4& color_matrix_body,
+      constant float4& color_matrix_translation,
+      constant float& scale) {
+    float4 c = ColorFromTexel<filter, address>().Do(v, texture, source_size, scale);
+    if (useColorM) {
+      c.rgb /= c.a + (1.0 - sign(c.a));
+      c = (color_matrix_body * c) + color_matrix_translation;
+      c *= v.color;
+      c.rgb *= c.a;
+    } else {
+      float4 s = v.color;
+      c *= float4(s.r, s.g, s.b, 1.0) * s.a;
+    }
+    c = min(c, c.a);
+    return c;
   }
-  c = min(c, c.a);
-  return c;
-}
+};
+
+template<bool useColorM, uint8_t address>
+struct FragmentShaderImpl<useColorM, FILTER_SCREEN, address> {
+  inline float4 Do(
+      VertexOut v,
+      texture2d<float> texture,
+      constant float2& source_size,
+      constant float4x4& color_matrix_body,
+      constant float4& color_matrix_translation,
+      constant float& scale) {
+    return ColorFromTexel<FILTER_SCREEN, address>().Do(v, texture, source_size, scale);
+  }
+};
 
 // Define Foo and FooCp macros to force macro replacement.
 // See "6.10.3.1 Argument substitution" in ISO/IEC 9899.
@@ -228,7 +243,7 @@ float4 FragmentShaderImpl(
       constant float4x4& color_matrix_body [[buffer(3)]], \
       constant float4& color_matrix_translation [[buffer(4)]], \
       constant float& scale [[buffer(5)]]) { \
-    return FragmentShaderImpl<useColorM, filter, address>( \
+    return FragmentShaderImpl<useColorM, filter, address>().Do( \
         v, texture, source_size, color_matrix_body, color_matrix_translation, scale); \
   }
 
diff --git a/internal/graphicsdriver/opengl/shader.go b/internal/graphicsdriver/opengl/shader.go
index a4f44692f..8d52f3c9a 100644
--- a/internal/graphicsdriver/opengl/shader.go
+++ b/internal/graphicsdriver/opengl/shader.go
@@ -197,12 +197,13 @@ void main(void) {
 
 #if defined(FILTER_NEAREST)
   pos = adjustTexelByAddress(pos, varying_tex_region);
-  color = vec4(0, 0, 0, 0);
   if (varying_tex_region[0] <= pos.x &&
       varying_tex_region[1] <= pos.y &&
       pos.x < varying_tex_region[2] &&
       pos.y < varying_tex_region[3]) {
     color = texture2D(texture, pos);
+  } else {
+    color = vec4(0, 0, 0, 0);
   }
 #endif
 
@@ -240,6 +241,7 @@ void main(void) {
 #endif
 
 #if defined(FILTER_SCREEN)
+
   highp vec2 p0 = pos - texel_size / 2.0 / scale;
   highp vec2 p1 = pos + texel_size / 2.0 / scale;
 
@@ -253,8 +255,10 @@ void main(void) {
 
   vec2 rateCenter = vec2(1.0, 1.0) - texel_size / 2.0 / scale;
   vec2 rate = clamp(((fract(p0 * source_size) - rateCenter) * scale) + rateCenter, 0.0, 1.0);
-  color = mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
-#endif
+  gl_FragColor = mix(mix(c0, c1, rate.x), mix(c2, c3, rate.x), rate.y);
+  // Assume that a color matrix and color vector values are not used with FILTER_SCREEN.
+
+#else
 
 #if defined(USE_COLOR_MATRIX)
   // Un-premultiply alpha.
@@ -273,6 +277,9 @@ void main(void) {
   color = min(color, color.a);
 
   gl_FragColor = color;
+
+#endif
+
 }
 `
 )