[llvm-dev] RFC: Proposing an LLVM subproject for parallelism runtime and support libraries

[Hal Finkel via llvm-dev]

----- Original Message -----
> From: "Arpith C Jacob" <acjacob at us.ibm.com>
> To: llvm-dev at lists.llvm.org
> Cc: jhen at google.com, "Hal J. Finkel" <hfinkel at anl.gov>
> Sent: Thursday, March 10, 2016 10:38:46 AM
> Subject: Re: [llvm-dev] RFC: Proposing an LLVM subproject for parallelism runtime and support libraries
> 
> Hi Jason,
> 
> I'm trying to better understand your StreamExecutor proposal and how
> it relates to other parallel programming models and runtimes such as
> RAJA [1], KOKKOS [2], or some hypothetical SPARK C++ API.
> 
> Please correct me if I'm misunderstanding your proposal, but I think
> the essence of what you want from the compiler is type safety for
> accelerator kernel launches, i.e., you would like the frontend to
> parse, check, and codegen for the construct:
> add_mystery_value<<<1, 1>>>(kernel_input_argument, *result.ptr());
> 
> Is that a correct understanding?
> 

Without answering your question, I'll point out that, as I understand it, StreamExecutor completely replaces the CUDA userspace library runtime components and talks directly to the drivers. Jason, please correct me if I'm wrong.

 -Hal

> Thanks,
> Arpith
> 
> [1]
> http://computation.llnl.gov/projects/raja-managing-application-portability-next-generation-platforms
> [2] https://github.com/kokkos/kokkos
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory