Texture cache hit rate statistics

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hello everyone,

I’m trying to optimize spatial locality in my textures to improve the cache hit rate. For this reason, I’d like to have some kind of quantitative statistics to verify or discard my ideas. Does anybody know a way to profile cache efficiency? I didn’t find any standard solution in the doc or the forum.

Kind regards,
Marco

#2

some random thoughts:

Measure this indirectly by recording the kernel execution times.

or by comparing the contents of the clock register before and after individual texture accesses

#3

I created a visualization of the memory access pattern that a warp makes over its lifetime based on the known memory access patterns of the kernel.