Generating high-performance OpenCL kernels

[ghost]

Dear Lift-developers,
I would like to gain experience with Lift. I was able to install Lift by following the instructions at http://lift-project.readthedocs.io/en/latest/. All unit tests were succesfull. I have now written a couple of high-level expressions but don't know how to generate high-performance OpenCL kernels from these. I couldn't find any documentation on this and would greatly appreciate your support.

[ghost]

I noticed you extended the documentation and now provide information on how to generate and execute kernels. Thank you! I tried the steps and have the following questions:

1. The provided steps seem to be specific for matrix multiplication and a few other BLAS routines. How can Lift be used for arbitrary programs (“high-level expressions” in Lift terminology) that are not BLAS?
2. I ran into an issue with the time required for executing the generated kernels. Executing the last command
   for i in `seq 1 250`; do find . -mindepth 1 -type d -exec sh -c '(cd {} && timeout 5m harness_mm -k 1024 -n 1024 -m 1024 --transpose-A -d $DEVICE -p $PLATFORM)' ';'; done
   took more than 48h to execute and exceeds the time available to me on the cluster I am using. Is there a way to break this step down into smaller tasks, that can be executed successively?

Many thanks in advance.

[tremmelg]

For 2, the command just finds all the subfolders containing OpenCL kernels and executes them. You can either divide the subfolders into several folders and run them one by one or just run the command again, as it will skip kernels it has already run.