texture with CUBLAS

Accelerated Computing CUDA CUDA Programming and Performance
1

Greetings!

I’m using linear array on libraries like CUBLAS and CUDPP. I read some technical report that using texture memory is faster than global memory. My question is:

Does it work if I bind linear array (1D normally) on texture memory, and use them with CUBLAS and CUDPP? Or these libraries are only valid on global memory?

Thanks!

Deyuan

2

Texture memory is only faster than global memory when it is being accessed in a somewhat random fashion where you can get the benefits of the cache.

I believe CUBLAS and CUDPP will bind and use textures internally where appropriate. You can check out their source for yourself to confirm.

3

Compared to coalesced access to global mem, linear array binded to texture is slower.