1922223 - Use gemmology threading support to speed up matrix multiplication for translations

Status: RESOLVED WONTFIX

(Firefox :: Translations, enhancement)

Comment 1

[[:sergesanspaille]]

Attachment: Bug 1922223 - Minimal parallel engine for gemmology r=marco!

Comment 2

[Phabricator Automation]

Comment on attachment 9428483 [details]
Bug 1922223 - Minimal parallel engine for gemmology r=marco!

Revision D224296 was moved to bug 1713066. Setting attachment 9428483 [details] to obsolete.

Comment 3

[[:sergesanspaille]]

Attachment: Bug 1922223 - Minimal parallel engine for gemmology r=marco!

With static scheduling but without threadpool

Comment 4

[[:sergesanspaille]]

Attachment: Bug 1922223 - Minimal parallel engine for gemmology r=padenot!

The parallel engine uses a thread pool that's lazily started and where
threads stop on their own if not receiving extra tasks. This suits
current behavior of bergamot which submits many small matrix multiplies.

To avoid contention on the shared index, thread process tasks per chunk.

Comment 5

[BugBot [:suhaib / :marco]]

The following patches are waiting for review from a reviewer who resigned from the review:

ID	Title	Author	Reviewer Status
D234831	Bug 1922223 - Minimal parallel engine for gemmology r=marco!	sergesanspaille	padenot: Resigned from review
D234880	Bug 1922223 - Minimal parallel engine for gemmology r=padenot!	sergesanspaille	padenot: Resigned from review

:sergesanspaille, could you please find another reviewer?

For more information, please visit BugBot documentation.

Comment 6

[[:sergesanspaille]]

After extensive testing and benchmarking with Paul Adenot, we concluded that the granularity of the task didn"t make it possible to efficiently accelerate translation at the matrix multiplication level, closing as WONTFIX