[BLAS] Add BLAS ARM performance libraries backend.#629
[BLAS] Add BLAS ARM performance libraries backend.#629Rbiessy merged 7 commits intouxlfoundation:developfrom
Conversation
Signed-off-by: Augustin Degomme <augustin.degomme@sipearl.com> Co-authored-by: Nicolas Bouton <nicolas.bouton@sipearl.com> Co-authored-by: Romain Dolbeau <romain.dolbeau@sipearl.com>
andrewtbarker
left a comment
andrewtbarker
left a comment
There was a problem hiding this comment.
Thank you for this contribution! I have a few initial comments/questions.
andrewtbarker
left a comment
andrewtbarker
left a comment
There was a problem hiding this comment.
Looks good! Just to double-check: you have tested both the Netlib backend and the Arm Performance Libraries backend on ARM hardware?
Indeed, see log attached for the netlib one (on an ARM neoverse v2 platform) |
|
It looks like some tables are off:
Please update the tables. Thanks! |
good catch, it shoud be fixed now. Thanks! |
|
By the way, I only spotted the tables by chance. The PR can be merged without my approval. |
4 participants
Description
This adds the support for aarch64 CPUs using ARM performance libraries backend, with BLAS domain for now (LAPACK to come later).
Support for most functions is native, some batch ones are implemented directly.
It also enables NETLIB backend on aarch64 CPUs.
This has been tested on Neoverse N1, V1, V2 CPUs, with dpcpp compiler and pocl backend.
100% tests passed, 0 tests failed out of 1960
128 tests are skipped, due to unimplemented omatadd/copy, batch or int8/bfloat16 unsupported features.
AdaptiveCpp also has been tested to compile and run.
log.txt