[MCW] Data Parallel Implementation of ShallowUnet in n300

[saichandax]

Currently, on main we have Shallow Unet in old conv.

End goal is to have a data parallel implementation of Shallow Unet in n300 with new conv API.

Executive summary:

• Implemented model to n150 with new ConvAPI.
• Implemented model to n300 with new ConvAPI via data-parallel.
• Working on optimising the model using 2cq and trace.
• Resolution = 4,1056,160, for n150 batch_size=2, for n300 batch_size = 4.

sub-tasks

• [x] #11811
• [x] To run n150 version on n300 as is
• [x] #11812
• [x] #11813

[saichandax]

ToDo (Aug-27)

• Double check model running on both devices (add perf sheets) - confirmed that model runs on 2 devices
• Device perf to be 750 FPS (we get ~11 for All ops on device)
• Host perf using trace with 2cqs (on both n150 and n300) [on hold]
• Validate Host side perf. Check with Austin. [on hold]
• generalise and switch between n150 and n300 [on hold]

[saichandax]

Currently, this task is on hold. Updates are as of 28th August.

Currently Evan is working on Shallow Unet in #11912