Texture access performance

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

In your tests, which texture access pattern is the fastest ?

with :

  • linear memory
  • 1D Arrays
  • 2D Arrays

I thought that access with 2D arrays would be the faster but in my case (none closely spaced access), the fastest is with linear memory.

And for you ? and in what situation (closely spaced access or not) ?

Thanks

2

My application has closely spaced accesses. Texture fetches from linear memory are faster than 2D arrays, but only by 2% or so. A few more % of total performance is obtained since using textured linear memory removes the need for device-device memory copies between iterations.

I did not try 1D arrays.