Ecp-aurora-sdk-discuss

Download

[email protected]

May 2024

2 participants
2 discussions

Re: [Ecp-aurora-sdk-discuss] New oneAPI release and eng-compiler versions for sunspot.
by John Mellor-Crummey 08 May '24

08 May '24

I hate to complain all the time about unexpected changes in the Intel software stack, but I’m tired of reverse engineering what Intel’s software does differently with each release so that I can build tools. The new software stack both changed how the generated code uses level0 command lists AND it changed important characteristics of GPU binaries. I just figured out that the reason HPCToolkit’s binary analyzer is having trouble with GPU binaries generated by the new release is that in prior releases, all text segments for GPU kernels in a binary each had a unique address range that doesn’t overlap with any other text segments. In the new release, all text segments for GPU kernels start with address 0. I hate that we don’t know what the invariants are for anything. In the binaries generated by prior releases of the software stack, kernels were at non-overlapping addresses and we wrote our software accordingly. Should we have relied on that? Apparently not. Did we have any idea that this might change? Absolutely not. Was the change documented in release notes? Not as far as I can tell. What can be done? Intel should document invariants and what may change. -- John Mellor-Crummey Professor Dept of Computer Science Rice University email: johnmc(a)rice.edu phone: 713-348-5179 > On Mar 29, 2024, at 11:27 AM, Chan-nui, Christopher via Ecp-aurora-sdk-discuss <ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov> wrote: > > This Message Is From an External Sender > This message came from outside your organization. > With the Sunspot update, a new version of the Aurora SDK is now available on Sunspot. > > The module files for this release, which are loaded by default: > > oneapi/eng-compiler/2023.12.15.002 > > This is based on the 2024.0 release of the oneAPI SDK. > > Release notes for the publicly available portions are available at: > https://urldefense.us/v3/__https://www.intel.com/content/www/us/en/develope… <https://urldefense.us/v3/__https://www.intel.com/content/www/us/en/develope…> > https://urldefense.us/v3/__https://www.intel.com/content/www/us/en/develope… <https://urldefense.us/v3/__https://www.intel.com/content/www/us/en/develope…> > > The engineering build contains bug fixes and performance improvements. > > Colleen will provide caveats, workarounds and other useful information about these releases shortly. > > > -- > Christopher Chan-Nui > christopher.chan-nui(a)intel.com > > -- > Ecp-aurora-sdk-discuss mailing list > Ecp-aurora-sdk-discuss(a)lists.jlse.anl.gov > https://urldefense.us/v3/__https://lists.jlse.anl.gov/mailman/listinfo/ecp-…

1 0

Frameworks 2024.1 'Preview' Release with updated Pytorch feature release available on Sunspot
by Ramer, Rebecca 03 May '24

03 May '24

There is now an updated 2024.1 DL Frameworks release with support for pytorch and tensorflow on the Sunspot deployment. This is a 'preview' release which has an update release of Intel pytorch DL Framework, along with the same Intel tensorflow DL Framework provided in the DL Frameworks deployment announced in April. This deployment uses the 'preview' version oneAPI SDK release, Intel's 2024.1.0 oneapi basekit. This release is available by loading the frameworks modulefile: module use /soft/preview-modulefiles/24.086.0 module load frameworks/2024.04.15.002 Loading this module activates the aurora_nre_models_frameworks-2024.1_preview_u1 conda environment. Along with the pytorch and tensorflow packages, all necessary python packages needed to run the 11 'Aurora NRE' models have been installed into this conda environment. An experimental version of Intel's xpu backend for triton is also provided. More information about this preview release can be found here: https://urldefense.us/v3/__https://apps.cels.anl.gov/confluence/display/int… Please let me know of any issues encountered. Rebecca Ramer rebecca.ramer(a)intel.com<mailto:[email protected]>

1 0