[Bug 2040027] Re: [SPR][ICX] OpenBLAS: Issues while testing sbgemm for Sapphire Rapids (SPR)

[Dave Jones]

This has been sponsored, so I'm unsubscribing ubuntu-sponsors.

One thing I would advise is that the sponsored patches do now include
"Fix syscall error on non-x86 platform", yet the section on regression
potential suggests little possibility of any change on other platforms,
and the test case is specifically limited to the sapphire rapids
platform. That is unlikely to sit well with the SRU team, and I would
recommend that the test case is expanded to cover all potentially
affected platforms. Looking at the patch it would appear to remove a
section of code from consideration on non-x86 platforms, but that's
still capable of affecting such platforms in an adverse way.

-- 
You received this bug notification because you are a member of Ubuntu
Sponsors, which is subscribed to the bug report.
https://bugs.launchpad.net/bugs/2040027

Title:
  [SPR][ICX] OpenBLAS: Issues while testing sbgemm for Sapphire Rapids
  (SPR)

Status in intel:
  Fix Released
Status in openblas package in Ubuntu:
  Fix Released
Status in openblas source package in Jammy:
  In Progress
Status in openblas source package in Mantic:
  Won't Fix

Bug description:
  [Impact]

  The sbgemm algorithm of openblas in Mantic and Jammy encounters 2 issues on Intel sapphire rapids platform:
  - the wrong detection of sapphire rapids as copperlake makes the sbgemm algorithm output wrong calculation.
  - the sbgemm test program encounters an illegal execution error.

  In addition to these errors, the OpenBLAS build with DYNAMIC_ARCH (see https://fossies.org/linux/OpenBLAS/README.md)
  does not work for saphhire rapids.

  [Test Plan]

  On sapphire rapids platform:

  - To check the runtime issue, run the dbgemm test program
  - To check the DYNAMIC_ARCH build support, build OpenBLAS with DYNAMIC_ARCH enabled and
    check that the sapphire rapids is supported

  [Where problems could occur]

  For the fix, most of the change is related to Intel SPR (Sapphire rapids) so i do not expect any regression on other Intel (or others) platforms.
  For sapphire rapids, the change impacts the sbgemm algorithm, so that is where i
  can see potential problems to occur

  [Other Info]

  This patchset has been integrated upstreamed for the version 0.3.24 of OpenBLAS.
  The original upstream PR can be found here : https://github.com/OpenMathLib/OpenBLAS/pull/4048/files
  The patchset contains 7 commits and modifies 11 files. Most of the changes are adding code (platform support, build support),
  the rest is modifying small/precise portion of the code (change GCC version macro, remove a constant in formula)
  The patchset applies without conflict for Mantic and it applies with a very light conflict resolution for Jammy.

  ==========
  [Description]
    Fix issues with sbgemm in Sapphire Rapids
    https://github.com/OpenMathLib/OpenBLAS/pull/4048

  [Hardware Information]
    Architecture:
      Intel / AMD (x86_64)
    Platform(s):
      Sapphire Rapids
    Date HW is expected at Canonical:

    Component(s):
      Performance and Scalability

  [Software Information]
    Target Version:
      22.04.4
    Target Kernel:
      N/A
    Commit IDs:
      TBD
    External Links:
      https://github.com/OpenMathLib/OpenBLAS

  [Business Justification]

  [Testing guidance]
  Some apps that use OpenBlas are rbenchmark, numpy, and cp2k.  These can we found via the Phoronix Test Suite.  Verify that the patches don't result in any regression.

  [External ID]
    OSVE-5181

To manage notifications about this bug go to:
https://bugs.launchpad.net/intel/+bug/2040027/+subscriptions