Shallow unet 512x512 resolution fails for batch>1

[keerthana-r-mcw]

Shallow unet with input resolution 512x512 is working for the dimension (1, 4, 512, 512), but fails with circular buffer issue (L1) for the dimension (2, 4, 512, 512)

To Reproduce Steps to reproduce the behavior:

1. Checkout the branch saichand/shallow_unet_512x512_and_1024x1024.
2. In /tt-metal/tests/ttnn/integration_tests/unet/test_ttnn_shallow_unet.py, change the batch = 2 in the function test_unet_512x512 (currently in line number 636) .
3. Run the following command, pytest tests/ttnn/integration_tests/unet/test_ttnn_shallow_unet.py::test_unet_512x512
4. The folowing error will occur, E RuntimeError: TT_THROW @ tt_metal/impl/program/program.cpp:496: tt::exception E info: E Statically allocated circular buffers in program 9 clash with L1 buffers on core range [(x=0,y=0) - (x=7,y=7)]. L1 buffer allocated at 595712 and static circular buffer region ends at 631680

[nsmithtt]

This is because this it too large to fit in L1.

• You have shape above (2, 4, 512, 512) NCHW
• We store this in L1 as NHWC, (1, 1, 2512512, 4) -> (1, 1, 524288, 4)
• We shard this across 64 cores so each shard is: [524288 / 64, pad(4)] -> [8192, 16], 4 gets padded to 16 so HW can work with it
• [8192, 16] is 8192162 bytes as float16b 262144, but this doesn't include the output buffer which is probably the same size and temporary buffers etc.

Anyway, the current dims of 1056160 only just fit w/ batch 2, 512512 is 1.5x bigger and 1024*1024 is 6x bigger, so unfortunately I think this is expected.

[dvartaniansTT]

@nsmithtt thanks for the prompt response on this. Do we have a plan to support higher resolutions? I heard partial sharding is meant to enable us high res support.