- the initial code will fail on systems where libxsmm is not installed
because the cpp test assumes the macro LIBXSMM_VERSION4 exists. The
macro is present in global namespace only when libxsmm.h is included
Calling it with multiple filenames was broken, revised and modernized
the script in the process. The cloud-based precommit was running it on
all files even though it was never intended to work on shell scripts.
Fixed that as well.
- Add support for AMD gpus in the toolchain
- Works with both Nvidia and AMD gpus
- generic kernels with templated types
- Needs ROCM 4.2 and CUDA 11 (for compiling on nvidia hardware)
- task list creation is hardware agnostic
