Hi,

I have a pixel shader for a terrain tile which samples a texture atlas for various ground cover.

I have around 8 different ground cover types which I pass to the PS via a Texture Blend Map using Texture2D NormalizedByte4 with one channel per texture weight.

Most pixels will blend perhaps 3 of the 8 possible textures, exceptionally 4.

What is the optimal way to sample the textures ?

I followed the standard "linear" "add them all up" approach, shown below;

textureColor += tex2D(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(0,0))) * textureWeights1.a;
textureColor += tex2D(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(1,0))) * textureWeights1.r ;
textureColor += tex2D(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(0,1))) * textureWeights1.g ;
textureColor += tex2D(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(1,1))) * textureWeights1.b ;
textureColor += tex2D(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(0,0))) * textureWeights2.a;
textureColor += tex2D(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(1,0))) * textureWeights2.r;
textureColor += tex2D(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(0,1))) * textureWeights2.g;
textureColor += tex2D(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(1,1))) * textureWeights2.b;

However I could have bypassed the many lookups by checking the texture weight prior to sampling. Like this

if(textureWeights1.r > 0) textureColor += tex2D(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(1,0))) * textureWeights1.r ;

Which is the recommended approach for performance in a pixel shader ? I understand that the PS can bypass texture lookups sometimes, but I dont know whether doing this manually by checking the texture weight myself is a better solution, especially because I expect only 50% maximum of the texture lookups to actually achieve anything.

Phillip

Thanks phil_t.

Unfortunately I'm using SM2 so cannot use tex2DLod. Fortunately it seems I can use text2DGrad to get the same outcome, in combination with ddx() and ddy() helpfully discussed here http://www.gamedev.net/topic/470659-anyone-tell-me-the-difference-of-tex2dgrad--and-tex2d/

Even more unfortunately although text2DGrad is supported in SM2 (and therefore Reach Profile) ddx() and ddy() are not supported until SM2.1.

It seems I need to abandon Reach profile in order to get optimal performance in branched texture lookups. Ah well, its only a hobby game, so moving to HiDef wont be an issue.

Phillip

Hmm,

I followed my nose and have the following now; which works nicely enough.

if(textureWeights1.a > 0) textureColor += tex2Dgrad(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(0,0)), texDdx, texDdy) * textureWeights1.a;
if(textureWeights1.r > 0) textureColor += tex2Dgrad(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(1,0)), texDdx, texDdy) * textureWeights1.r;
if(textureWeights1.g > 0) textureColor += tex2Dgrad(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(0,1)), texDdx, texDdy) * textureWeights1.g;
if(textureWeights1.b > 0) textureColor += tex2Dgrad(TextureSamplerAtlas1,getAtlasTextureCoords(textureCoords, float2(1,1)), texDdx, texDdy) * textureWeights1.b;

if(textureWeights2.a > 0) textureColor += tex2Dgrad(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(0,0)), texDdx, texDdy) * textureWeights2.a;
if(textureWeights2.r > 0) textureColor += tex2Dgrad(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(1,0)), texDdx, texDdy) * textureWeights2.r;
if(textureWeights2.g > 0) textureColor += tex2Dgrad(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(0,1)), texDdx, texDdy) * textureWeights2.g;
if(textureWeights2.b > 0) textureColor += tex2Dgrad(TextureSamplerAtlas2,getAtlasTextureCoords(textureCoords, float2(1,1)), texDdx, texDdy) * textureWeights2.b;

Phillip