#15221: Post completion messages to dispatch_s

[jbaumanTT]

Ticket

#15221

Problem description

We're seeing a latency of around 500ns from when the dispatch message is sent to when it's received, in the case where a many are sent at the same time.

What's changed

We never wait on the acks from these completion messages, so make them posted to avoid contention from a lot of replies being sent at once. In the case where every worker is sending them at the same time, this can halve the latency from 500ns to 250ns (on wormhole).

Checklist

• Post commit CI passes
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• (For models and ops writers) Full new models tests passes
• New/Existing tests provide coverage for changes

[tt-asaigal]

Unrelated to your change but a potential bug with the non-posted atomic incs previously: Looking at the additional args we're now setting.. I realize that we never correctly configured the return addr of the non-posted atomic inc.

The pre-incremented value would have been written over the NOC at worker address 0. Seems like we were lucky that this address wasn't being actively used?

This is not an issue with your change, since response traffic for the atomic inc won't be generated anymore.

Edit: Ah this issue is only with DYNAMIC_NOC. I guess this feature was never enabled.

[ubcheema]

I am assuming this is just speeding up the noc itself.
I don't see any atomic barrier, so fw/kernel does not seem to be held up due to an inflight response.