4a35c4d38a
Summary: The Offloading library wraps around the underlying plugins. The problem is that we currently initialize all plugins we find, even if they are not needed for the program. This is very expensive for trivial uses, as fully heterogenous usage is quite rare. In practice this means that you will always pay a 200 ms penalty for having CUDA installed. This patch changes the behavior to provide accessors into the plugins and devices that allows them to be initialized lazily. We use a once_flag, this should properly take a fast-path check while still blocking on concurrent use. Making full use of this will require a way to filter platforms more specifically. I'm thinking of what this would look like as an API. I'm thinking that we either have an extra iterate function that takes a callback on the platform, or we just provide a helper to find all the devices that can run a given image. Maybe both? Fixes: https://github.com/llvm/llvm-project/issues/159636
Liboffload
This directory contains the implementation of the work-in-progress new API for Offload. It builds on top of the existing plugin implementations but provides a single level of abstraction suitable for implementation of many offloading language runtimes, rather than just OpenMP.
Testing liboffload
The main test suite for liboffload can be run with the check-offload-unit
target, which runs the offload.unittests executable. The test suite will
automatically run on every available device, but can be restricted to a single
platform (CUDA, AMDGPU) with a command line argument:
$ ./offload.unittests --platform=CUDA
Tracing of Offload API calls can be enabled by setting the OFFLOAD_TRACE
environment variable. This works with any program that uses liboffload.
$ OFFLOAD_TRACE=1 ./offload.unittests
---> olInit()-> OL_SUCCESS
# etc
The host plugin is not currently supported.
Modifying liboffload
The main header (OffloadAPI.h) and some implementation details are
autogenerated with tablegen. See the API definition README
for implementation details.