b: block number,
t: thread number per block,
c: the size of a continuous data that each block reads from memory.
experiment on a big (2G) loop of texture reading:
32b, 128t, c varies 4k-256k: timings are the same.
conclusion: texture has no real cache.
why??
Yes, texture is cached in 0.8.
If all your tests take the same time, then I wouldn’t conclude there is no cache. Since you read an increasing number of values from texture, I would expect time to increase with or without a texture cache.
If your times are all equal, there is probably another bottleneck masking the texture read cost.
Mark
Thank you very much! I was retailing other’s work, so I was not fully clear about the experiment. His block num=32, i guess this is the bottleneck that masked the texture read cost. I have suggested he using 64~256 blocks.
Thanks a lot.