[Embench] Notes from Embench call 19 October 2020

Jeremy Bennett jeremy.bennett at embecosm.com
Tue Oct 20 15:47:29 CEST 2020


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi all,

Notes from yesterday's meeting.

0. Welcome new participants
===========================

Peter Alexander, Southampton University
Aaryman Bhattacharya, Southampton University
Robert Chyla, IAR
Peter Gielda, Antmicro
Karol Gugala, Antmicro
Matthew Johns, Southampton University
Bryon Theobald, Southampton University

1. Actions from previous meetings
=================================

ACTION: Jeremy Bennett, Stephan Wallentowitz to complete interview blog
post on website.
- - in progress

ACTION: Jeremy Bennett to circulate new generic slide deck
- - complete

ACTION: David Patterson to prepare draft archive paper
- - abstract prepared, more discussion, see agenda item 5

ACTION: Jeremy to add tutorial example to reproduce baseline results
in documentation.
- - captured as GitHub issue 68
- - in progress

ACTION (ongoing): Roger Shepherd to provide Embench results for ST dev
boards, Apple Macs and Raspberry Pi.
- - in progress

ACTION: Jeremy Bennett/Paolo Savini to attempt to get results for
Atmel AVR using Arduino.
- - in progress

ACTION: Jeremy Bennett to explore alternatives to Google Docs for
presentations
- - Open Document Format version prepared
- - hieroglyph version in progress

ACTION (ongoing): Ray Simar to set class project to come up with new
benchmarks based on the TinyML book.
- - Ray to report progress in agenda item 4

ACTION: Jeremy Bennett to update documentation to advise that Python
minimum version 3.6 is required.
- - in progress

2. GitHub pull requests
=======================

- From embench-iot
- ----------------

None

- From embench-iot-results
- ------------------------

None

3. New GitHub issues
====================

- From embench-iot
- ----------------

https://github.com/embench/embench-iot/issues/95
- - Does Wikisort need to use double precision constants
- - responses from Roger Shepherd and Jeremy Bennett
- - thanks to user axiomlegend for raising this
- - thanks to all who provided responses

https://github.com/embench/embench-iot/issues/96
- - Explanation / Origin of benchmarks
  (beebs.eu link broken, due to a failure to renew domain)
- - website moved to https://beebs.mageec.org/
- - marked as bug, assigned to Jeremy Bennett to fix links
- - thanks to Nagendra Gulur for raising this

https://github.com/embench/embench-iot/issues/97
- - Issue Running Embench
- - advice provided by Roger Shepherd, Jeremy Bennett and Paolo Savini
- - marked as an enhancement for documentation assigned to Paolo Savini
- - thanks to George Onwubuya for raising this
- - thanks to all who provided responses

https://github.com/embench/embench-iot/issues/98
- - Help needed for Cortex-m55f
- - initial advice from Jeremy Bennett
- - asssigned to Paolo Savini for further advice
- - thanks to user sheenagoel for raising this

- From embench-iot-results
- ------------------------

None

Thanks to all those who raised issues and provided responses.

4. Update on benchmarks, including Embench 0.6
==============================================

Ray Sinar's class on benchmarking
- ---------------------------------

Ray Sinar is running classes at Rice University on Embench 0.6. Ray
was unable to attend the call, his report was presented on his behalf
by Jeremy Bennett as follows:

DSP Embench:
- - A team of 15 students, grads, and undergrads, has been assembled.
- - Initial focus is the the DSP big three algorithms of FIR, IIR, and
FFT.  We have an initial IIR up and running.  And we will expand the list.
- - Initial work is using floating-point.
- - Interested in running the benchmarks on hardware platforms of
interest to the Embench team.
  - What hardware would we want to recommend to my students?
- - Would like to have Jeremy come speak to the class.
  - We are very interested in his work on how to reproduce the
baseline - Several students are very interested in continuing their
work into the spring.
  - I will be offering a special research project class on
benchmarking, for credit, to carry this work forward.

TinyML Embench:
- - Have about 30 students, grads and undergrads, working on multiple
projects from Pete Warden’s TinyML book.
- - Pete Warden spoke to the students about current and future TinyML work.
- - Will end up with at least three examples from Pete’s book.
- - Reached out to Pete to find out his thoughts on the MLPerf inference
benchmarks.  My thought is that these are still too big for a
microcontroller benchmark, but I will defer to Pete.
- - As above, several students are interested in continuing this work in
the spring and I will fold them into my benchmarking research class.

There was a discussion on the question of which hardware platform to use:
- - for a baseline platform possibly Arm Cortex-M7
  - good to stick to M class, since still embedded
  - Arm A class is a different architecture
- - for comparison
  - could compare to Raspberry Pi
  - what about RISC-V platforms
    - RIOS, Alibaba, SiFive, MicroChip all developing suitable ones
    - PolarFire (MicroChip) is available today
    - SparkFun may have suitable plaforms

Noted that RISC-V has an academic community. Contact Mark Himelstein
(CTO, RISC-V International) for more details.

Southampton University RISC-V AI ISA extension
- ----------------------------------------------

Peter Alexander, Aaryman Bhattacharya, Matthew Johns and Bryon
Theobald, masters students at Southampton University joined the call.
They are developing extensions to RISC-V to support AI/ML inference
and training and plan to use TinyML as their benchmark. Ray's report
expressed a welcome to collaborate

Antmicro CPU testing with Embench
- ---------------------------------

Karol Gugula of Antmicro presented their work on automated CPU testing
using Embench, as described in the blog post at:

https://antmicro.com/blog/2020/09/automatic-cpu-testing-with-embench/

Questions:

- - Q. All in simulation. How long does it take to run.
  - A. SERV (bit serial minimalist RISC-V) is slowest and takes 1.5h
on a typical PC.

- - Q. Can you model interrupts
  - A. Yes

- - Q. Looks better than the IoT results repository, at least for
softcores. Will then contribute to Embench.
  - A. Probably, but the code needs more work before it is ready for this.

- - Q. Can you get data for Arm models?
  - A. You can integrate with any cores - not a technical limit. Might
be licensing limit.

- - Q. What next?
  - A. exploration of different SoC components (types of flash,
multiple clocks etc).

- - Q. Can we use Amazon FPGAs in the cloud?
  - A. Sounds like a good idea.

ACTION: Mark Himelstein to arrange recorded presentation to wider
RISC-V community with Karol Gugula.

Karol's presentation is attached to these minutes.

A reminder that we need volunteers to lead particular aspects of
current and future efforts.

5. Publicity
============

Recent talks
- ------------

3rd Workshop on RISC-V Activities held by Edacentrum, Germany. The
talk will be looking at Embench within the context of the RISC-V
Software Ecosystem

VSDOpen. The talk considered Embench as an example of software that
bridges into the hardware world as part of a talk on the history of
open source software and hardware.

Blog posts
- ----------

As noted in the previous item, Antmicro have published a blog post on
automated CPU testing with Embench

arXiv paper
- -----------

Following on from the suggestion of a reference arXiv paper on
Embench, David Patterson has put together an abstract:

https://docs.google.com/document/d/1Lb-g0NP6dRL-UbS_t8w_RfI7thFeqGMZVCSi9JhUq_0

ACTION: All to provide comments and input. We have data for soft cores
in cycle accurate modeling and FPGA, but we need data on real ASIC
cores (other than the baseline Cortex-M4 data).

Interview blog post
- -------------------

Stefan Wallentowitz reported by email that the interview style
blog post about Embench had been delayed due to the start of
University term, but was in progress.

Standard presentation
- ---------------------

Jeremy Bennett presented an updated version of the generic Embench
presentation. This includes up to date data comparing RV32I, RV32IC,
RV32IM and RV32IMC and a fuller historical record of LLVM performances
(all prepared by Paolo Savini). It still lacks good data on ARC and
AVR. The presentation is attached.

New version, spell out ToT

6. Next meeting
===============

We currently meet regularly on the third Monday of each month 8-9am
California time. Our future meetings will be. Note that the Decmember
date has been changed to avoid the week before Christmas.
- - Mon 16 Nov 2020
- - Mon 14 Dec 2020 (changed)

The following dates are proposed for 2021

- - Mon 11 Jan 2021 (week 2 to keep clear of Martin Luther King Jr day)
- - Mon 15 Feb 2021
- - Mon 15 Mar 2021
- - Mon 19 Apr 2021
- - Mon 17 May 2021
- - Mon 21 Jun 2021
- - Mon 19 Jul 2021
- - Mon 16 Aug 2021
- - Mon 21 Sep 2021
- - Mon 18 Oct 2021
- - Mon 15 Nov 2021
- - Mon 13 Dec 2021 (week 2, to keep clear of Christmas)

I've created a calendar entry for our regular attendees, to provide
automated reminders. Please ask if you wish to be added to this.

7. AOB
======

Robert Chyla is looking at Embench for trace benchmarking.

Ofer talked about work on benchmarking interrupts/context switching.
This is the other aspect of Embench, initiated by Cesare Garlati

Question. Do we have a spec for what we are trying to achieve?

Noted we have repository with Cesare Garlati's original work on GitHub

Best wishes,


Jeremy
Vice Chair

- -- 
Cell: +44 (7970) 676050
SkypeID: jeremybennett
Twitter: @jeremypbennett
Email: jeremy.bennett at embecosm.com
Web: www.embecosm.com
PGP key: 1024D/BEF58172FB4754E1 2009-03-20
-----BEGIN PGP SIGNATURE-----

iF0EARECAB0WIQRASGDWqmhRZUfAaPW+9YFy+0dU4QUCX47qbwAKCRC+9YFy+0dU
4fpGAJ0WSr66t63MnN46Ho/YYbz+zXS4ygCfawyZKuIkm8wHdQ84QJeR+MNexq0=
=Ks+q
-----END PGP SIGNATURE-----
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Embench tester.pdf
Type: application/pdf
Size: 1124102 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment.pdf>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: embench-generic-oct-2020.odp
Type: application/vnd.oasis.opendocument.presentation
Size: 1246972 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment.odp>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: embench-generic-oct-2020.pdf
Type: application/pdf
Size: 452920 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment-0001.pdf>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Embench tester.pdf.sig
Type: application/octet-stream
Size: 95 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: embench-generic-oct-2020.odp.sig
Type: application/octet-stream
Size: 95 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment-0001.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: embench-generic-oct-2020.pdf.sig
Type: application/octet-stream
Size: 95 bytes
Desc: not available
URL: <http://lists.librecores.org/pipermail/embench/attachments/20201020/25679504/attachment-0002.obj>


More information about the Embench mailing list