#14257: few more optimization for yolo

[shwetankTT]

Ticket

#14257:

Problem description

YoloV4 Optimization.
Improvement:--> Getting 200fps from 170fps.

What's changed

Adding conv optimization flag for few layers. Removed few sharding which are not needed.

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)
• New/Existing tests provide coverage for changes

[shwetankTT]

Nightly Regression:https://github.com/tenstorrent/tt-metal/actions/runs/12103838737

[pavlejosipovic]

Is there any perf impact of this change?
Does trace size needs to be adjusted?

[shwetankTT]

Is there any perf impact of this change? Does trace size needs to be adjusted?

Yes. The current perf is 200fps from 170. I ran the nightly CI no need for trace size adjastment.

[pavlejosipovic]

Is there any perf impact of this change? Does trace size needs to be adjusted?

Yes. The current perf is 200fps from 170. I ran the nightly CI no need for trace size adjastment.

We don't have a perf CI in place for yolo ATM?

[shwetankTT]

Is there any perf impact of this change? Does trace size needs to be adjusted?

Yes. The current perf is 200fps from 170. I ran the nightly CI no need for trace size adjastment.

We don't have a perf CI in place for yolo ATM?

Yeah we don't hv perf CI for yolo. I am getting the numbers from IRD machines(1st sheet with final name.).
https://docs.google.com/spreadsheets/d/1MRNswbQJeF-v9vzA_ajSRlBRHFTY8dAdDlnN6FiMRHQ/edit?gid=174224023#gid=174224023

[pavlejosipovic]

Is there any perf impact of this change? Does trace size needs to be adjusted?

Yes. The current perf is 200fps from 170. I ran the nightly CI no need for trace size adjastment.

We don't have a perf CI in place for yolo ATM?

Yeah we don't hv perf CI for yolo. I am getting the numbers from IRD machines(1st sheet with final name.). https://docs.google.com/spreadsheets/d/1MRNswbQJeF-v9vzA_ajSRlBRHFTY8dAdDlnN6FiMRHQ/edit?gid=174224023#gid=174224023

Ok than, document the perf diff in PR description/commit message.
Btw I would prioritise adding perf CI before doing any perf optimisations.

[shwetankTT]

To calculate the FPS we take the sum of device kernel duration(ns) Column divided by 10^9 (sec --> nano sec conversion).

[tt-rkim]

Please run:

• Nightly fast dispatch

[shwetankTT]

Please run:

• Nightly fast dispatch

https://github.com/tenstorrent/tt-metal/actions/runs/12103838737

[mywoodstock]

@shwetankTT Lets get this on the perf CI so that this does not regress.

[shwetankTT]

@mywoodstock Yeah. Added the perf on CI. Validating it.