2021 03 29 sync

Agenda

• Opens
• Feedback on https://reviews.llvm.org/D99190.

Meeting notes

• Opens
  • Victor: The documentation is uploaded for review in Markdown document. Should we use ReST format instead? LLVM project enabled Markdown support rendering, but clang project. AR Alexey: convert Markdown to ReST.
• Feedback on https://reviews.llvm.org/D99190.
  • Victor suggests splitting the document into multiple parts.
  • AR Alexey: leave device front-end part in https://reviews.llvm.org/D99190 and move other sections to separate reviews. E.g. driver part should be reviewed with people familiar with Driver implementation.
  • Victor suggests replacing integration header with OpenMP-like approach i.e. host compiler emits type traits required to.
  • OpenMP-like approach pros: no issues with unnecessary code parsing, no forward-declarable requirement for kernel name, no order dependency between host and device compiler invocations(?)
  • NOTE: adding OpenMP-like approach doesn't mean we should drop intermediate header immediately. Two approaches can co-exist (as long as needed).
  • AR Alexey: create an issue to prototype this approach for DPC++ compiler.
  • Lowering to SPIR-like kernels should be helpful for other back-ends. Dead kernel argument elimination can applied if lambda captures are split into basic types and improve performance by minimizing host<=>device data transfer.