Compute Accelerator Forum - SYCL, HLS4ML

Wednesday 13 Jan 2021, 16:30 → 18:00 Europe/Zurich

Virtual (Zoom)

 

To receive annuoncements and information about this forum please subscribe to compute-accelerator-forum-announce@cern.ch

 

Live notes

Video Recording

ZOOM Info

ZOOM URL

Heterogeneous Modern C++ using Khronos SYCL and oneAPI

• What is the plan for non-C++ language support ? OpenCL handled this reasonably well by virtue of being a C API which you could bind to any language of your choosing, but as SyCL is forking away from OpenCL to avoid dying along with it (see e.g. CUDA backends, oneAPI Zero…), it raises the question of what unified backend other popular programming languages like Java, C#, Python, Julia, Rust, Go… will be able to use in order to program GPUs, FPGAs, and so on. Binding to a C API is easy, binding to C++ that is preprocessed by a special SyCL compiler sounds… unpleasant. [Hadrien Grasland]

  • For a poor-man’s “pySYCL”, something like cppyy may help [Vince]

  • Is OpenCL really dying ? Open source driver implementations are emerging. POCL, intel, nvidia and amd runtimes, as well as  pyopencl, work really well these days.  [Riccardo]

• There is a oneAPI MatterMost channel for those who wish to discuss sycl. [Laurence]

  • Can you put the link? [Graeme]

  • https://mattermost.web.cern.ch/it-dep/channels/oneapi [Laurence]

    • This redirects to a https://mattermost.web.cern.ch/error?type=team_not_found error for me, maybe the team is not publicly accessible? [Hadrien]

• Compiler backends: all llvm-based? What does "openMP" mean - source generator? [Axel]

16:30 → 16:35 News

Speakers: Benjamin Morgan (University of Warwick (GB)), Graeme A Stewart (CERN), Dr Maria Girone (CERN), Michael Bussmann (Helmholtz-Zentrum Dresden - Rossendorf), Stefan Roiser (CERN)

Compute Accelerator Forum Introduction, 2021-01.pdf

16:35 → 17:05 Heterogeneous Modern C++ using Khronos SYCL ad one API

Khronos SYCL has exploded in recent years as the go-to frameworkfor heterogeneous compute, graphics, and AI building on top of modern ISO C++, and extending it for use in all forms of accelerators. Its recent adoption in Intel’s one API as the heterogeneous programming model for all forms of processors, and inclusion in US Argonne National Lab in their upcoming AURORA Exascale supercomputer has certainly cemented SYCL as the programming frameworke for Modern C++ HPC workloads. Its strength is that enables modern C++, even template programming techniques such as that used in Tensorflow. More then that it is open, collaborative, cross-vendor and non-proprietary with a rich ecosystem and multiple implementations following Khronos’ rich portfolio of compute and graphics languages. This talk will introduce you to SYCL, its current status, while looking ahead to future upcoming SYCL 2020 features.

Speaker: Michael Wong (Codeplay Software Ltd., ISOCPP, Khronos)

Cern 2021 Parallel Programming model landscape4.pdf

17:05 → 17:35 Fast Deep Learning Inference for Particle Physics and Beyond

Speaker: Vladimir Loncar (CERN)

hls4ml_CAF_13.01.2021.pdf