Array of texture references

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi,

In my application, I need a variable number of texture references (number ‘n’ specified by user at runtime). Like this:

texture<float, 3, cudaReadModeElementType> *texrefs;

texrefs = (texture<float, 3, cudaReadModeElementType> *) malloc(n*sizeof(texture<float, 3, cudaReadModeElementType>));

A dynamic array for that purpose is good. Is it possible to do it and how do I access the texrefs inside a kernel?

Every member of the texref will be bound to its own cudaArray.

I’m fine even if it is somehow possible to unbind/bind to different cudaArrays inside a kernel to a single texref, but I don’t think that’s possible.

Any suggestion will be highly appreciated.

-Oj

#2

http://forums.nvidia.com/index.php?showtopic=80654

Says that it is not possible. Is it still the case with cuda 2.1? Do I have any option then?

-Oj

#3

Yes, this is a hardware limitation (fixed in DirectX 11, btw).

Your options are:

  1. use a switch statement like in the other thread
  2. use a 3D texture
#4

Thanks for reply. In fact I needed an array of 3D tex references.

-Oj

#5

Sorry, missed that. Unfortunately we don’t have 4D textures. Yet!

#6

The number of texture units is dependent on the hardware. Is there any limitation on number of 3D textures that can be simultaneously used in CUDA (did I miss that in the CUDA manuals anywhere??)? or is it again the same as the hardware limit?

I should then query CUDA to return me that limit? Is there an API for that?

#7

As a side note, I just checked on my card (8600 M GT that supports 4 tex units) that I could use 5 texture references bound to 5 cudaArrays at the same time in a kernel!

Did I understand something wrong here??

#8

I think I have a clue of what to do.
I can concatenate my 3D textures for form a larger (rather longer) 3D texture and kinda emulate a 4D texture!
(http://forums.nvidia.com/index.php?showtopic=44424)

Thanks Simon for the info.

-Oj

#9

Yes, packing multiple textures into a larger texture atlas is a good solution if this works for you.

BTW, I don’t see it documented anywhere, but I believe CUDA supports up to 16 texture references.