-
Notifications
You must be signed in to change notification settings - Fork 738
2019 04 15 meeting notes
- Meeting schedule: 29 Apr/13 May.
- [SYCL] Driver/ToolChain Extension (https://github.com/intel/llvm/issues/53).
- As Hal suggested, I created a wiki page with the list of SYCL patches and their status: https://github.com/intel/llvm/wiki/SYCL-patches-upstream-status. I need to look through them carefully and update “comment” column in the table if there are any pre-requisites for upstreaming.
- Feedback on https://reviews.llvm.org/D60455.
Alexey Bader, James Brodman, Andrew Kaylor, Michael Kinsner, Oleg Maslov, Mariya Podchishchaeva, Andrew Gozillon, Kevin Petit, Ronan Keryell, Hal Finkel
Most people on the call are travelling on Apr 29 from Khronos F2F meeting. Cancelled.
Alexey Bader is travelling on May 13 also. Alexey is to move the meeting to May 14.
Long term plans for SYCL support in clang driver (https://github.com/intel/llvm/issues/53)
Andrew Gozillon: What is the right way to integrate 3rd party back-ends (e.g. Xilix FPGA compiler, Intel FPGA compiler, Intel GPU compiler etc.)? Should we add support for tools used in tool-chain like SPIR-V translator, Offload bundlers, wrappers, 3rd party back-ends?
Andrew Gozillon: Is SYCL tool chain assume SPIR-V?
Ronan: We should use getToolChain() method and overload it for our needs.
In abstract we should follow what OpenMP infrastructure is doing.
Andrew is okay to adjust SYCL implementation to be more flexible for 3rd party back-ends and test with Xilix compiler. If it works, he will send PR.
Alexey Bader created wiki with the list of SYCL patches: https://github.com/intel/llvm/wiki/SYCL-patches-upstream-status.
Please, review and send your feedback if any.
https://reviews.llvm.org/D60455
Concerns with this patch are:
- Anastasia: why not reuse existing OpenCL keywords instead of new SYCL attributes?
- AlexeyB: Any concerns with re-using current OpenCL keywords?
- Ronan: it is better to have SYCL specific attributes than re-use OpenCL keywords/attributes
- Aaron: to add logic using new attributes. WIP.
No major issues with existing patch.
AlexeyB: Are CUDA/OpenMP require explicit marks in the code to identify device code or compiler can implicitly compile
Hal: In CUDA functions are marked explicitly, but there are cases when compiler transitively propagates use of __device__
/__host__
attributes. For OpenMP implicit declare target
AlexeyB: AR: investigate if we can re-use code written for CUDA/OpenMP for SYCL.