RFC: Enabling oneDNN operations for x86 CPUs

[penpornk]

This RFC will be open for comment until Tuesday, November 9th, 2021.

Status	Proposed
RFC #	400
Author(s)	Penporn Koanantakool (penporn@google.com)
Sponsor	Tatiana Shpeisman (shpeisman@google.com)
Updated	2021-10-25

Objective

Default TensorFlow Linux packages can efficiently utilize CPU hardware features. Make it easier to support new features.

[penpornk]

We have tentatively scheduled a public design review meeting for this RFC.
Time: Thursday, November 11th, 2021 from 3-4pm PT
Meeting link: Google Meet

Meeting participants are expected to read the RFC ahead of time as the meeting will focus on the remaining issues/questions.

If there are no outstanding issues on this RFC by the time the feedback phase closes on Tuesday November 9th, 2021, we may cancel the design review meeting. (We will update here.)

[penpornk]

The design review meeting has been canceled since there are no outstanding issues that require discussion.

[penpornk]

A quick follow-up that oneDNN operations were enabled by default on recent x86 CPUs on Linux starting from TF 2.9.
Next, we plan to also enable them by default on:

• aarch64 CPUs on Linux (starting with Neoverse V1)
• x86 CPUs on Windows

If anyone has any concerns, please post them here by July 12, 2023.

Thank you!