Why is sampling from large textures so slow?

What kind of sampling pattern does your shader use? Does each pixel sample a different part of the texture? Are the samples randomly distributed? Is there a 1-to-1 mapping between pixels and texture-samples?

What format is your texture? Is it compressed (e.g. DXT)?


It's possibly to do with the texture cache. When you fetch a texel, a section of that texture (nearby the fetched texel) is pulled into the texture cache. If you subsequently fetch the same texel, or nearby texels, it's much faster as they're already in the cache.

When using smaller textures, you're much more likely to get a cache-hit, because a larger percentage of the texture will be resident in the cache at any one time.

I use the following shader:

// Input image
Texture2D inputImage : register(t0);
SamplerState inputImageSampler : register(s0);

float4 main(float2 nonused : TEXCOORD0, float2 uv : TEXCOORD1) : COLOR
{
return inputImage.Sample(inputImageSampler, uv);
}

This is 1-to-1 mapping between pixels and texture-samples (uv are calculated to make it). Format is BGRA 32bit, not compressed.

Can it occur due to large amount of GPU memory.. I mean maybe GPU has to move the texture in memory in case of large textures before sampling?

It's based on two things. One is the cache as mentioned previously and the other is the dimensions. Non pow2 dimensions are stored in a linear fashion that is also not at all cache friendly, especially when bilnear is filtering them.

There's a good chance that if you round them up and make them bigger then you'd see a small win, but the best thing to do is use smaller pow2 textures