Next Meeting : Jan 8th (Jan 1st is New Year’s Day and on Jan 15th we have the F2F)
Opens:
- None
Development Activity:
-
Francesco submitted a patch on IR vectorization of functions (https://reviews.llvm.org/D675720), but it had to be reverted. Will resubmit a revised version after the Christmas break.
-
Ongoing work on “declare variant” – enable use of overload-centric declare variants (https://reviews.llvm.org/D71241).
-
OpenMPOpt transformation pass (https://reviews.llvm.org/D69930) is under implementation. We need more people to help with OpenMP-specific optimizations.
Implementation of “declare variant”:
- Progress of https://reviews.llvm.org/D71241 is stuck due to opposing opinions on the implementation. More reviewers are needed to step in and help resolve the disagreement.
OpenMP-IR-Builder status:
-
The IR-Builder (https://reviews.llvm.org/D69785, https://reviews.llvm.org/D69922) and OpenMPConstants (https://reviews.llvm.org/D69853) patches have already landed.
-
Patches on lowering of “taskwait” (https://reviews.llvm.org/D69828) and “taskyield” (https://reviews.llvm.org/D70799) using the IR-Builder and codegen for the “parallel” construct (https://reviews.llvm.org/D70109) have been accepted but not committed yet.
-
Work on the “target” directive is currently in progress.
-
Johannes has started a spreadsheet with pending OpenMP-IR-Builder tasks to help coordinate who works on what. People should feel free to claim any tasks they want. Link: https://docs.google.com/spreadsheets/d/1FvHPuSkGbl4mQZRAwCIndvQx9dQboffiD-xD0oqxgU0/edit?usp=sharing
LLVM OpenMP community review etiquette:
- There is a request for more community members to participate in the discussions around reviews of various patches because there is a feeling that currently only a handful of people get actively involved.
HIP/CUDA/OpenMP target compilation:
- AMD proposed the idea of using a thread sanitizer with device RTLs by emulating runs of the RTLs on the host (because thread sanitizers cannot run on the device). Proposal was welcomed.
DeviceRTL redesign to support sharing code:
-
Jon has moved more code out of the nvptx RTL into the common library.
-
Implemented more functionality for the amdgcn target.
-
There is consensus on deprecating support for nvcc when building the nvptx device RTL. nvcc was the only robust solution when the library was initially written, but now clang can handle it just fine. This helps resolve a problem with compiler command-line options; clang, gcc and nvcc use the same options with different meanings. nvcc will not be dropped completely; instead a warning message will be emitted that compiling the nvptx RTL with nvcc has been deprecated and no support is provided.
Roll Call :