User control of block size in parallel_for for HIP execution policy?

[rocm]

User control of block size in parallel_for for HIP execution policy?

[doctorcolinsmith]

256 should be optimal for the default case, but @nolmoonen will review.

[nolmoonen]

Hi @cwsmith, I agree with Colin here: a block size of 256 is generally a good choice for AMD cards. It would be helpful to know why you would want to change it.

As it stands, it is not possible to change the kernel configuration for parallel_for with a custom policy (or in any other way). rocThrust must maintain compatiblity with Thrust, which also does not offer this functionality.

Perhaps you would want to have a look at rocPRIM's device_transform which does the same thing as parallel_for, but does take in a kernel configuration as parameter.

[cwsmith]

Hello @doctorcolinsmith and @nolmoonen.

I was looking at register usage for some of our kernels and wanted to experiment with different block sizes.

Thank you for the info on the default [roc]Trust size and the link to the rocPRIM API.