Next Meeting : April 22nd
Opens:
- Johannes requested developers to participate in reviews to make code robust.
- stream pool
- Shilei has posted support for fixing the correctness of device synchronization with streams here https://reviews.llvm.org/D77412
- Async support in target library
- To support async, the current tgt API for plugin was changed to add an additional argument. https://reviews.llvm.org/D77005
We had agreed previous any exposed API should not change but new APIs will be added and the old API can internally call the new API
- AMD plugin
- Nice if AMD plugin can be committed to llvm.org. This will enable testing the rewrite of device rtl into common/device specific partition
- Greg said they are working on upstreaming the code. This code is currently available for all to access in their github
- NEC SX-Aurora plugin
- New plugin for NEC SX-Aurora is being developed. Can be found here https://reviews.llvm.org/D76843
- The activity of partitioning device rtl was on hold as Ron was spending quality time as proud dad. Congratulations.
- Cuda requirement has been removed
- Shared data is supported using declare target and device code using target pragmas.
Development Activity:-
- Unshackled threads
- Shilei presented the design of a prototype to add unshackled threads in openmp runtime to support async target tasks. The presentation is attached
There are still issue with how to synchronize with other task and with task wait.
Alex mentioned we already have support for detach task for use with target task.
We need to create examples and check if they are handled in both the schemes. (unshackled or existing implementation)
- Fat Static Library
- George has uploaded the support in offload bundler here https://reviews.llvm.org/D74262
- He will be uploading the driver changes.
- math support in target regions
- Support has been added.
- omp loop, omp metadirective, omp interop
- Johannes has prototypes implemented omp loop and metadirectives
- Was asking if anyone want to participate in omp interop.
- Declare variants
- Johannes has implemented support for begin/end declare variants.
Roll Call :
unshackled_task.pdf (430 KB)