Autotuning tile size in OpenACC

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

Hello!
I would like to know if there is any functionality in OpenACC that allows for autotuning of the parameters in the tile clause.
I’ve searched the documentation but couldn’t find anything. My goal is to empirically determine the best tile size for a kernel.
The intuitive solution is to build a script to test a set of candidate sizes, collect the time for each, and select the smallest.
Before going ahead with implementation, I’d like to make sure that there is nothing similar already implemented in OpenACC or any other tool.

Thanks for any help!

2

Hi Gabriel,

I’ve not heard of anything like this. The tile clause isn’t used too often and I’d be surprised if anything beyond using tile(32,4) would be significantly faster, but it’s worth the experiment. It would be highly dependent on the code.

-Mat

1 Like