[petsc-dev] Fwd: ArmPL - Batched interfaces

Barry Smith bsmith at petsc.dev
Wed Oct 7 21:50:25 CDT 2020


> Begin forwarded message:
> From: Christopher Armstrong <Christopher.Armstrong at arm.com>
> Subject: ArmPL - Batched interfaces
> Date: October 6, 2020 at 10:34:05 AM CDT
> To: "bsmith at mcs.anl.gov" <bsmith at mcs.anl.gov>
> Cc: Nicholas Dingle <Nicholas.Dingle at arm.com>
> Message-Id: <DBBPR08MB4901BA0F0BF65CABA6C788FBFF0D0 at DBBPR08MB4901.eurprd08.prod.outlook.com>
> Hi Barry,
> I work on Arm Performance Libraries, which is the Arm vendor tuned maths library for BLAS, LAPACK, FFTs. Si Hammond at Sandia suggested I get in touch with you about the work that we've been doing with them on batched linear algebra. We're looking for feedback on some new interfaces we've developed with their help, which are designed to give high performance for large batches of small linear algebra problems.
> Si said that either you might be interested, or you might be able to put us in touch with other developers at Argonne who are interested, so feel free to forward on my mail.
> Our new functions are analogous to MKL's compact BLAS and LAPACK functions, but with a few tweaks such as allowing arbitrary strides in both matrix dimensions and tying the input format to a particular vector length.
> We have double precision, real implementations of matrix multiplication and rank-revealing QR factorization available in our 20.3 release of the library (undocumented for now!). We also have in the pipeline LU factorization with threshold pivoting, Cholesky, TRSM and TRMM, again all double real at this stage. Our benchmarking so far shows that we can achieve some decent performance improvements over repeated calls to BLAS.
> If you or someone else at Argonne/PETSc are interested in reviewing the work we've been doing and giving some constructive feedback before we publish our API, we'd be happy to hear from you. We can provide prototype documentation and help if you want to experiment with some examples. Just let me know if you're interested!
> Best Regards,
> Chris Armstrong
> Tech Lead, Arm Performance Libraries
> IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20201007/2079fa02/attachment.html>

More information about the petsc-dev mailing list