Changes in cufftdx vs 0.3.1 patch

Accelerated Computing GPU-Accelerated Libraries
1

I’m looking at the changes in the 0.3.1 patch and wanted to confirm I see the important bits (I hope I didn’t miss a changelog in the docs somewhere.)

  1. I’m sad to see sm86 is still not supported, is this correct?

  2. It looks like most of the changes are in include/database, to definitions and kernel ptx and the addition of several look up tables.

    • are these mainly performance related?
    • how significant are they?

  3. I see in the examples a call to set the MaxDynamicSharedMemorySize is added. (Glad to see this as i’ve been doing the same for quite a while, see below). Even though in your case the amount of memory is known at compile time (FFT::shared_memory_size) the call cudaFuncSetAttribute() is a runtime function right? Do you know if the compiler makes any different decisions?

Thanks!

Ben

2

  1. The sole intent of 0.3.1 was to fix a compiler issue. You can try to manually remove the HW checks and see if the kernel works.

  2. I don’t recall many performance improvements, in any.

  3. Yes, cudaFuncSetAttribute(). Nothing is different with compiler

3

Thanks for the info Matt, I hacked out the HW checks and was able to get things to run on sm86. I won’t comment on performance since it isn’t actually supported yet, but it is good to know I can run on most of our GPUs now : )

1 Like