How does texture memory work on the C1060

Accelerated Computing CUDA CUDA Programming and Performance
1

I have some computational science experience and know that there are a number of strategies as to how CPU cache memory works.

But how does texture memory work as a cache, in particular on the C1060?

For example, with CPUs if a particular array element is used then that array element and its neighbours are also brought to cache, and cache members can also be kicked out of cache under different strategies. With texture memory is it just the individual element that is stored in texture, or as suggested above, are its neighbours also brought to texture? And if so what is the exact strategy?

2

Nvidia does not document these things. You can find quite some info in the Demystifying GPU Microarchitecture through Microbenchmarking paper though.

Basically the 2d coordinates are transformed by an unknown function to 1d addresses that preserve 2d locality as much as possible, and then a standard cache architecture with cachelines is used. So not all neighbors are fetched and discarded at the same time but only those who happen to share the same cacheline.

3

That paper looks very interesting. Thanks.