2016-12-06 HPC Strategy Meeting Minutes

Attendance

Committee Members

Name	Present
Kanta Vekaria (OCTO, Linaro)
Martin Stadtler (Director of LEG, Linaro)
Darren Cepulis (ARM)
Eric Van Hensbergen (ARM)
Kangkang Shen (HiSilicon)
Larry Wikelius (Cavium)
Gary Yurcak (Qualcomm)
Elsie Wahlig (Qualcomm)
Grant Likely (HPE)
Jon Masters (RedHat)
Koichi Hirai (Fujitsu)
Takeharu Kato Fujitsu)

Guests

Name	Present
David Rusling (CTO, Linaro)
Mark Orvek (EVP, Linaro Engineering)
Andrea Gallo (VP of Segment Groups, Linaro)
Anoop Saxena (Project Manager, Segments)
Francoise Ozog (Director of LNG, Linaro)
Renato Golin
(ARM) Ashok Bhat, Richard Sandiford, , Alan Hayward, Geraint North, Guilherme Marshall
Masami Hiramatsu (Linaro)
Frank Wang (Linaro)

LLVM has a lot of downstream users and has uses of vector types that we do not know - as a result we need community input to determine what is acceptable/not acceptable for IR
Unsure if toolchain team appropriate to optimise BLAS, etc, FFTW need expertise in this area to do this. Team are not experts in HPC
Without the SVE spec, e.g. the SVE additions to ISA/ABI/ Linaro is not in a position to push patches to the maintainers or as maintainers.
SVE support for QEMU in opensrc can only happen with the publically information available.
Fast models can be used, but only for bootstrap or continuous integration, we really need a QEMU implementation to validate.
No support for SVE in the debugger, patches for gdb are up for review on the gdb mailing list, nothing from the llvm side. None of the opensource debuggers currently to allow clustered or mpi support This is available in commercial debuggers.
If needed a new team will need to be formulated
Toolchain - Cannot staff that effort unless priorities are reconsidered by the TSC
What is the role for Linaro to bootstrap the SVE solution. We need to find the balance between what Linaro does and what is done by the members own internal teams.

What should Linaro be doing, where Linaro is acting as the glue for all member efforts.

Geraint(ARM) - Accurate analysis. Linaros contribution on QEMU yes! as that is something that ARM will not do and there is no rush.. Agree that arch specs need to be published and public ASAP, before upstream actually happens. Plans are underway to open the specs up.

Lot’s of work to be done in the libraries just with Neon, which we can get started sooner without waiting for the SVE spec and simulation model.

Renato: Do you have any visibility on the gcc changes 7 vs 8? Is there any backporting required?

Geraint: SVE support in GCC - It’s a nice to have not a must have. No urgency on pushing SVE support in GCC 7. GCC 7 is not an absolute requirement, 8 would be perfectly fine. We need "a" compiler with SVE support, not necessarily the *system* compiler
Richard: Even if we push the changes to GCC 7, they wouldn't be complete anyway

Geraint: FLANG and PGI - ARM has access and are tracking it making sure it builds on ARM and works with SVE. Will come "for free" when PGI open sources their LLVM bridge
Renato: we need to decide if we want to staff and grow the toolchain team

Kanta: the team can grow via member engineers as well and the HPC SIG may be in the position to decide on this. Which engineer profiles would we need?
Martin: There is "at least one member" that will provide a member engineer to TCWG

Larry: Neon lots of room to improve Neon. Need to have a balance on what is available today for the users

Renato Depends on the implementation with have SVE or not SVE
Geraint: SVE is not a replacement for NEON. They will have both. Work on the assumption that both will be present and that there is significant performance boost with Neon.
Eric: SVE and Neon has proven to increase performance

ARM slides (Darren Cepulis)
Good growth of the attendance for the meeting (85 attendees, over approx 70 from last year).