Metal: Removing fences as they seem to not be needed.

[ivarflakstad]

Updating/waiting for fences was initially added to help tackle GPU race condition bugs. However I think that the bug was fixed and that these fences are no longer needed.

@Narsil what do you think? Seems to work perfectly on the m1 pro (faster too)

[Narsil]

No it's keeps failing on m3. You can repro with my gist https://gist.github.com/Narsil/13bfeb66959a0b7e2750a23be9ecc957.

Also how did you test and what did you measure. On m3 the fence is quite negligible (as long as the scheduler is kept happy by leaving a few critical ones)

Also you cannot remove fences inside a single command encoder. This is specifically written as non ordered by default.

[ivarflakstad]

No it's keeps failing on m3. You can repro with my gist https://gist.github.com/Narsil/13bfeb66959a0b7e2750a23be9ecc957.

I get the following - I assume you do not?

First matmul is OK [2.0, 3.0, 6.0, 11.0, 46.0, 55.0, 66.0, 79.0]
This should be filled [6.0, 11.0, 22.0, 39.0, 514.0, 615.0, 738.0, 883.0]

Also how did you test and what did you measure. On m3 the fence is quite negligible (as long as the scheduler is kept happy by leaving a few critical ones)

Ran some models and the difference was significant. However I just did a reboot and the difference is gone 🤷
Guess I'll close

[ivarflakstad]

Reopening as it seems the fence situation has changed for M3.
Is that correct, @Narsil ?😊