changed: use textureOffset to get pixels, this increases performance by 12% but requires glsl 1.30

git-svn-id: https://xbmc.svn.sourceforge.net/svnroot/xbmc/trunk@28588 568bbfeb-2a22-0410-94d2-cc84cf5bfa90

1 files changed, 16 insertions, 23 deletions

diff --git a/system/shaders/convolution-6x6.glsl b/system/shaders/convolution-6x6.glsl
index f37dd6105a..ef9e8bc479 100644
--- a/system/shaders/convolution-6x6.glsl
+++ b/system/shaders/convolution-6x6.glsl
@@ -14,12 +14,12 @@ uniform float     m_stretch;
 #if (HAS_FLOAT_TEXTURE)
 half3 weight(float pos)
 {
-  return texture1D(kernelTex, pos).rgb;
+  return texture(kernelTex, pos).rgb;
 }
 #else
 half3 weight(float pos)
 {
-  return texture1D(kernelTex, pos).rgb * 2.0 - 1.0;
+  return texture(kernelTex, pos).rgb * 2.0 - 1.0;
 }
 #endif
 
@@ -36,20 +36,15 @@ vec2 stretch(vec2 pos)
 #endif
 }
 
-half3 pixel(float xpos, float ypos)
-{
-  return texture2D(img, vec2(xpos, ypos)).rgb;
-}
-
-half3 line (float ypos, vec3 xpos1, vec3 xpos2, half3 linetaps1, half3 linetaps2)
+half3 line (vec2 pos, const int yoffset, half3 linetaps1, half3 linetaps2)
 {
   return
-    pixel(xpos1.r, ypos) * linetaps1.r +
-    pixel(xpos1.g, ypos) * linetaps2.r +
-    pixel(xpos1.b, ypos) * linetaps1.g +
-    pixel(xpos2.r, ypos) * linetaps2.g +
-    pixel(xpos2.g, ypos) * linetaps1.b +
-    pixel(xpos2.b, ypos) * linetaps2.b; 
+    textureOffset(img, pos, ivec2(-2, yoffset)).rgb * linetaps1.r +
+    textureOffset(img, pos, ivec2(-1, yoffset)).rgb * linetaps2.r +
+    textureOffset(img, pos, ivec2( 0, yoffset)).rgb * linetaps1.g +
+    textureOffset(img, pos, ivec2( 1, yoffset)).rgb * linetaps2.g +
+    textureOffset(img, pos, ivec2( 2, yoffset)).rgb * linetaps1.b +
+    textureOffset(img, pos, ivec2( 3, yoffset)).rgb * linetaps2.b;
 }
 
 void main()
@@ -70,17 +65,15 @@ void main()
   columntaps1 /= sum;
   columntaps2 /= sum;
 
-  vec2 xystart = (-1.5 - f) * stepxy + pos;
-  vec3 xpos1 = vec3(xystart.x, xystart.x + stepxy.x, xystart.x + stepxy.x * 2.0);
-  vec3 xpos2 = vec3(xystart.x + stepxy.x * 3.0, xystart.x + stepxy.x * 4.0, xystart.x + stepxy.x * 5.0);
+  vec2 xystart = (0.5 - f) * stepxy + pos;
 
   gl_FragColor.rgb =
-   line(xystart.y                 , xpos1, xpos2, linetaps1, linetaps2) * columntaps1.r +
-   line(xystart.y + stepxy.y      , xpos1, xpos2, linetaps1, linetaps2) * columntaps2.r +
-   line(xystart.y + stepxy.y * 2.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.g +
-   line(xystart.y + stepxy.y * 3.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.g +
-   line(xystart.y + stepxy.y * 4.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.b +
-   line(xystart.y + stepxy.y * 5.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.b;
+   line(xystart, -2, linetaps1, linetaps2) * columntaps1.r +
+   line(xystart, -1, linetaps1, linetaps2) * columntaps2.r +
+   line(xystart,  0, linetaps1, linetaps2) * columntaps1.g +
+   line(xystart,  1, linetaps1, linetaps2) * columntaps2.g +
+   line(xystart,  2, linetaps1, linetaps2) * columntaps1.b +
+   line(xystart,  3, linetaps1, linetaps2) * columntaps2.b;
 
   gl_FragColor.a = gl_Color.a;
 }