Scale horizontally to an intermediate target, then vertically to the final target. Each step is materialized by a shader pass.

The intermediate render target now has the source height and the destination width, which is reflected in the vertex coordinates and the vertex coordinates stepping parameter.

2 files changed, 100 insertions, 100 deletions

diff --git a/system/shaders/convolutionsep-4x4_d3d.fx b/system/shaders/convolutionsep-4x4_d3d.fx
index d8700f16a2..9c55bc2b36 100644
--- a/system/shaders/convolutionsep-4x4_d3d.fx
+++ b/system/shaders/convolutionsep-4x4_d3d.fx
@@ -22,7 +22,8 @@
 texture g_Texture;
 texture g_KernelTexture;
 texture g_IntermediateTexture;
-float2  g_StepXY;
+float2  g_StepXY_P0;
+float2  g_StepXY_P1;
 
 sampler RGBSampler =
   sampler_state {
@@ -82,84 +83,84 @@ half4 weight(float pos)
 #endif
 }
 
-half3 pixel(float xpos, float ypos)
+half3 pixel(sampler samp, float xpos, float ypos)
 {
-  return tex2D(RGBSampler, float2(xpos, ypos)).rgb;
+  return tex2D(samp, float2(xpos, ypos)).rgb;
 }
 
+// Code for first pass - horizontal
+
 half3 getLine(float ypos, float4 xpos, half4 linetaps)
 {
   return
-    pixel(xpos.r, ypos) * linetaps.r +
-    pixel(xpos.g, ypos) * linetaps.g +
-    pixel(xpos.b, ypos) * linetaps.b +
-    pixel(xpos.a, ypos) * linetaps.a;
+    pixel(RGBSampler, xpos.r, ypos) * linetaps.r +
+    pixel(RGBSampler, xpos.g, ypos) * linetaps.g +
+    pixel(RGBSampler, xpos.b, ypos) * linetaps.b +
+    pixel(RGBSampler, xpos.a, ypos) * linetaps.a;
 }
 
 PS_OUTPUT CONVOLUTION4x4Horiz(VS_OUTPUT In)
 {
   PS_OUTPUT OUT;
 
-  float2 pos = In.TextureUV + g_StepXY * 0.5;
-  float2 f = frac(pos / g_StepXY);
+  float2 pos = In.TextureUV + g_StepXY_P0 * 0.5;
+  float2 f = frac(pos / g_StepXY_P0);
 
-  half4 linetaps   = weight(1.0 - f.x);
-  half4 columntaps = weight(1.0 - f.y);
+  half4 linetaps = weight(1.0 - f.x);
 
   // kernel generation code made sure taps add up to 1, no need to adjust here.
 
-  float2 xystart = (-1.0 - f) * g_StepXY + In.TextureUV;
+  float2 xystart;
+  xystart.x = (-1.0 - f.x) * g_StepXY_P0.x + In.TextureUV.x;
+  xystart.y = In.TextureUV.y;
+
   float4 xpos = float4(
       xystart.x,
-      xystart.x + g_StepXY.x,
-      xystart.x + g_StepXY.x * 2.0,
-      xystart.x + g_StepXY.x * 3.0);
-
-  OUT.RGBColor.rgb =
-    getLine(xystart.y                   , xpos, linetaps) * columntaps.r +
-    getLine(xystart.y + g_StepXY.y      , xpos, linetaps) * columntaps.g +
-    getLine(xystart.y + g_StepXY.y * 2.0, xpos, linetaps) * columntaps.b +
-    getLine(xystart.y + g_StepXY.y * 3.0, xpos, linetaps) * columntaps.a;
+      xystart.x + g_StepXY_P0.x,
+      xystart.x + g_StepXY_P0.x * 2.0,
+      xystart.x + g_StepXY_P0.x * 3.0);
 
+  OUT.RGBColor.rgb = getLine(xystart.y, xpos, linetaps);
   OUT.RGBColor.a = 1.0;
+
   return OUT;
 }
 
+// Code for second pass - vertical
+
+half3 getRow(float xpos, float4 ypos, half4 columntaps)
+{
+  return
+    pixel(IntermediateSampler, xpos, ypos.r) * columntaps.r +
+    pixel(IntermediateSampler, xpos, ypos.g) * columntaps.g +
+    pixel(IntermediateSampler, xpos, ypos.b) * columntaps.b +
+    pixel(IntermediateSampler, xpos, ypos.a) * columntaps.a;
+}
+
 PS_OUTPUT CONVOLUTION4x4Vert(VS_OUTPUT In)
 {
   PS_OUTPUT OUT;
 
-  float2 pos = In.TextureUV + g_StepXY * 0.5;
-  float2 f = frac(pos / g_StepXY);
+  float2 pos = In.TextureUV + g_StepXY_P1 * 0.5;
+  float2 f = frac(pos / g_StepXY_P1);
 
-  half4 linetaps   = weight(1.0 - f.x);
   half4 columntaps = weight(1.0 - f.y);
 
   // kernel generation code made sure taps add up to 1, no need to adjust here.
 
-  float2 xystart = (-1.0 - f) * g_StepXY + In.TextureUV;
-  float4 xpos = float4(
-      xystart.x,
-      xystart.x + g_StepXY.x,
-      xystart.x + g_StepXY.x * 2.0,
-      xystart.x + g_StepXY.x * 3.0);
+  float2 xystart;
+  xystart.x = In.TextureUV.x;
+  xystart.y = (-1.0 - f.y) * g_StepXY_P1.y + In.TextureUV.y;
 
-  OUT.RGBColor.rgb =
-    getLine(xystart.y                   , xpos, linetaps) * columntaps.r +
-    getLine(xystart.y + g_StepXY.y      , xpos, linetaps) * columntaps.g +
-    getLine(xystart.y + g_StepXY.y * 2.0, xpos, linetaps) * columntaps.b +
-    getLine(xystart.y + g_StepXY.y * 3.0, xpos, linetaps) * columntaps.a;
+  float4 ypos = float4(
+      xystart.y,
+      xystart.y + g_StepXY_P1.y,
+      xystart.y + g_StepXY_P1.y * 2.0,
+      xystart.y + g_StepXY_P1.y * 3.0);
 
+  OUT.RGBColor.rgb = getRow(xystart.x, ypos, columntaps);
   OUT.RGBColor.a = 1.0;
-  return OUT;
-}
 
-PS_OUTPUT Nothing(VS_OUTPUT In)
-{
-  PS_OUTPUT OUT;
-
-  OUT.RGBColor.rgb = tex2D(IntermediateSampler, In.TextureUV).rgb;
-  OUT.RGBColor.a = 1.0;
   return OUT;
 }
 
@@ -174,7 +175,7 @@ technique SCALER_T
   }
   pass P1
   {
-    PixelShader  = compile ps_3_0 Nothing();
+    PixelShader  = compile ps_3_0 CONVOLUTION4x4Vert();
     ZEnable = False;
     FillMode = Solid;
     FogEnable = False;
diff --git a/system/shaders/convolutionsep-6x6_d3d.fx b/system/shaders/convolutionsep-6x6_d3d.fx
index 52cc641664..07073e860e 100644
--- a/system/shaders/convolutionsep-6x6_d3d.fx
+++ b/system/shaders/convolutionsep-6x6_d3d.fx
@@ -22,7 +22,8 @@
 texture g_Texture;
 texture g_KernelTexture;
 texture g_IntermediateTexture;
-float2  g_StepXY;
+float2  g_StepXY_P0;
+float2  g_StepXY_P1;
 
 sampler RGBSampler =
   sampler_state {
@@ -82,98 +83,96 @@ half3 weight(float pos)
 #endif
 }
 
-half3 pixel(float xpos, float ypos)
+half3 pixel(sampler samp, float xpos, float ypos)
 {
-  return tex2D(RGBSampler, float2(xpos, ypos)).rgb;
+  return tex2D(samp, float2(xpos, ypos)).rgb;
 }
 
+// Code for first pass - horizontal
+
 half3 getLine(float ypos, float3 xpos1, float3 xpos2, 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;
+    pixel(RGBSampler, xpos1.r, ypos) * linetaps1.r +
+    pixel(RGBSampler, xpos1.g, ypos) * linetaps2.r +
+    pixel(RGBSampler, xpos1.b, ypos) * linetaps1.g +
+    pixel(RGBSampler, xpos2.r, ypos) * linetaps2.g +
+    pixel(RGBSampler, xpos2.g, ypos) * linetaps1.b +
+    pixel(RGBSampler, xpos2.b, ypos) * linetaps2.b;
 }
 
 PS_OUTPUT CONVOLUTION6x6Horiz(VS_OUTPUT In)
 {
   PS_OUTPUT OUT;
 
-  float2 pos = In.TextureUV + g_StepXY * 0.5;
-  float2 f = frac(pos / g_StepXY);
+  float2 pos = In.TextureUV + g_StepXY_P0 * 0.5;
+  float2 f = frac(pos / g_StepXY_P0);
 
-  half3 linetaps1   = weight((1.0 - f.x) / 2.0);
-  half3 linetaps2   = weight((1.0 - f.x) / 2.0 + 0.5);
-  half3 columntaps1 = weight((1.0 - f.y) / 2.0);
-  half3 columntaps2 = weight((1.0 - f.y) / 2.0 + 0.5);
+  half3 linetaps1 = weight((1.0 - f.x) / 2.0);
+  half3 linetaps2 = weight((1.0 - f.x) / 2.0 + 0.5);
 
   // kernel generation code made sure taps add up to 1, no need to adjust here.
 
-  float2 xystart = (-2.0 - f) * g_StepXY + In.TextureUV;
+  float2 xystart;
+  xystart.x = (-2.0 - f.x) * g_StepXY_P0.x + In.TextureUV.x;
+  xystart.y = In.TextureUV.y;
+
   float3 xpos1 = float3(
       xystart.x,
-      xystart.x + g_StepXY.x,
-      xystart.x + g_StepXY.x * 2.0);
+      xystart.x + g_StepXY_P0.x,
+      xystart.x + g_StepXY_P0.x * 2.0);
   float3 xpos2 = half3(
-      xystart.x + g_StepXY.x * 3.0,
-      xystart.x + g_StepXY.x * 4.0,
-      xystart.x + g_StepXY.x * 5.0);
-
-  OUT.RGBColor.rgb = getLine(xystart.y                   , xpos1, xpos2, linetaps1, linetaps2) * columntaps1.r +
-					 getLine(xystart.y + g_StepXY.y      , xpos1, xpos2, linetaps1, linetaps2) * columntaps2.r +
-					 getLine(xystart.y + g_StepXY.y * 2.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.g +
-					 getLine(xystart.y + g_StepXY.y * 3.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.g +
-					 getLine(xystart.y + g_StepXY.y * 4.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.b +
-					 getLine(xystart.y + g_StepXY.y * 5.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.b;
+      xystart.x + g_StepXY_P0.x * 3.0,
+      xystart.x + g_StepXY_P0.x * 4.0,
+      xystart.x + g_StepXY_P0.x * 5.0);
 
+  OUT.RGBColor.rgb = getLine(xystart.y, xpos1, xpos2, linetaps1, linetaps2);
   OUT.RGBColor.a = 1.0;
+
   return OUT;
 }
 
+// Code for second pass - vertical
+
+half3 getRow(float xpos, float3 ypos1, float3 ypos2, half3 columntaps1, half3 columntaps2)
+{
+  return
+    pixel(IntermediateSampler, xpos, ypos1.r) * columntaps1.r +
+    pixel(IntermediateSampler, xpos, ypos1.g) * columntaps2.r +
+    pixel(IntermediateSampler, xpos, ypos1.b) * columntaps1.g +
+    pixel(IntermediateSampler, xpos, ypos2.r) * columntaps2.g +
+    pixel(IntermediateSampler, xpos, ypos2.g) * columntaps1.b +
+    pixel(IntermediateSampler, xpos, ypos2.b) * columntaps2.b;
+}
+
 PS_OUTPUT CONVOLUTION6x6Vert(VS_OUTPUT In)
 {
   PS_OUTPUT OUT;
 
-  float2 pos = In.TextureUV + g_StepXY * 0.5;
-  float2 f = frac(pos / g_StepXY);
+  float2 pos = In.TextureUV + g_StepXY_P1 * 0.5;
+  float2 f = frac(pos / g_StepXY_P1);
 
-  half3 linetaps1   = weight((1.0 - f.x) / 2.0);
-  half3 linetaps2   = weight((1.0 - f.x) / 2.0 + 0.5);
   half3 columntaps1 = weight((1.0 - f.y) / 2.0);
   half3 columntaps2 = weight((1.0 - f.y) / 2.0 + 0.5);
 
   // kernel generation code made sure taps add up to 1, no need to adjust here.
 
-  float2 xystart = (-2.0 - f) * g_StepXY + In.TextureUV;
-  float3 xpos1 = float3(
-      xystart.x,
-      xystart.x + g_StepXY.x,
-      xystart.x + g_StepXY.x * 2.0);
-  float3 xpos2 = half3(
-      xystart.x + g_StepXY.x * 3.0,
-      xystart.x + g_StepXY.x * 4.0,
-      xystart.x + g_StepXY.x * 5.0);
+  float2 xystart;
+  xystart.x = In.TextureUV.x;
+  xystart.y = (-2.0 - f.y) * g_StepXY_P1.y + In.TextureUV.y;
 
-  OUT.RGBColor.rgb = getLine(xystart.y                   , xpos1, xpos2, linetaps1, linetaps2) * columntaps1.r +
-					 getLine(xystart.y + g_StepXY.y      , xpos1, xpos2, linetaps1, linetaps2) * columntaps2.r +
-					 getLine(xystart.y + g_StepXY.y * 2.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.g +
-					 getLine(xystart.y + g_StepXY.y * 3.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.g +
-					 getLine(xystart.y + g_StepXY.y * 4.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps1.b +
-					 getLine(xystart.y + g_StepXY.y * 5.0, xpos1, xpos2, linetaps1, linetaps2) * columntaps2.b;
+  float3 ypos1 = float3(
+      xystart.y,
+      xystart.y + g_StepXY_P1.y,
+      xystart.y + g_StepXY_P1.y * 2.0);
+  float3 ypos2 = half3(
+      xystart.y + g_StepXY_P1.y * 3.0,
+      xystart.y + g_StepXY_P1.y * 4.0,
+      xystart.y + g_StepXY_P1.y * 5.0);
 
+  OUT.RGBColor.rgb = getRow(xystart.x, ypos1, ypos2, columntaps1, columntaps2);
   OUT.RGBColor.a = 1.0;
-  return OUT;
-}
 
-PS_OUTPUT Nothing(VS_OUTPUT In)
-{
-  PS_OUTPUT OUT;
-
-  OUT.RGBColor.rgb = tex2D(IntermediateSampler, In.TextureUV).rgb;
-  OUT.RGBColor.a = 1.0;
   return OUT;
 }
 
@@ -188,7 +187,7 @@ technique SCALER_T
   }
   pass P1
   {
-    PixelShader  = compile ps_3_0 Nothing();
+    PixelShader  = compile ps_3_0 CONVOLUTION6x6Vert();
     ZEnable = False;
     FillMode = Solid;
     FogEnable = False;