dynamic update of texture in a kernel is it worth it

_DK · May 9, 2009, 6:29pm

Hi,

I had a simple question. I have a kernel which fills an array in global memory. I’d like the kernel executed after that to read the data generated by the previous kernel and produce something else. Is it worth binding this global array to texture each frame? Is it even valid and do I need to unbind the texture each time I update global memory?

It could look like this:

unbindTexture(resultTexture);

launch_firstKernel(); // fills mem d_result

bindTexture(d_result, resultTexture);

launch_secondKernel();// uses tex1dfetch(resultTexture)

I thought this could be beneficial since the texture cache could be enabled, but is it worth the effort? I don’t know what’s really going on behind the scene when texture binding is performed.

Thanks!

eyalhir74 · May 9, 2009, 7:34pm

unbindTexture(resultTexture);

launch_firstKernel(); // fills mem d_result

bindTexture(d_result, resultTexture);

launch_secondKernel();// uses tex1dfetch(resultTexture)
I thought this could be beneficial since the texture cache could be enabled, but is it worth the effort? I don’t know what’s really going on behind the scene when texture binding is performed.

Thanks!

I havent tried this but this seems valid and possible. As far as I saw binding/unbinding shouldn’t take time/too much time.

I guess the best answer is to test it on your code :)

Its all a matter of how much it will take oppose to the time your kernels run.

would be interesting to see what you got though :)

eyal

MisterAnderson42 · May 9, 2009, 8:55pm

It is a lot of trouble for little reason. You can write to device memory that is bound to a texture, leave it bound and read it in the next kernel.

Tovi · May 10, 2009, 12:20am

Hi,

I had a simple question. I have a kernel which fills an array in global memory. I’d like the kernel executed after that to read the data generated by the previous kernel and produce something else. Is it worth binding this global array to texture each frame? Is it even valid and do I need to unbind the texture each time I update global memory?

It could look like this:
unbindTexture(resultTexture);

launch_firstKernel(); // fills mem d_result

bindTexture(d_result, resultTexture);

launch_secondKernel();// uses tex1dfetch(resultTexture)
I thought this could be beneficial since the texture cache could be enabled, but is it worth the effort? I don’t know what’s really going on behind the scene when texture binding is performed.

Thanks!

I asked myself the same thing a few days ago

http://forums.nvidia.com/index.php?showtopic=96101

_DK · May 11, 2009, 9:10am

Thanks Tovi, looks like it’s worth giving textures a try.