Generating new blocksizes to LIBCUSMM

[Abhishek Bagusetty]

I have tried adding the triples of required block sizes but the build fails 
because parameters couldn't be found in the parameters.txt . I have also 
tried adding some default parameters to the parameters.txt file for the new 
block sizes but that didn't really work out. The libcusmm.cu which involves 
kernels for matrix transpose and other functions, fail as it couldn't find 
the new block sizes that were recently added. 

BTW, from my past experiences, the libcusmm.cu looks like a auto-generated 
file created while building libsmm package.  

On Monday, October 27, 2014 4:57:20 AM UTC-4, Samuel Andermatt wrote:
>
> Adding new Kernels in generate.py is just adding something like:
>
> triples += combinations(41, 7, 15)
>
> This will add all combinations of the block sizes specified.
> The parameters in parameters.txt are generated using autotuning. I think 
> if you do not specify them they will use some default parameters. I am not 
> sure how you could generate the parameters yourself (the calculations are 
> also somewhat expensive). However there are a lot more parameters available 
> in the development version of cp2k. So maybe you can just copy the 
> optimised parameters from there.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20141028/c1cd697b/attachment.htm>