Using block sharded config with height < 32 leads to low PCC on multiple ops

[nemanjagrujic]

Using block sharded config with height < 32 leads to low PCC on isfinite and similar ops. Provided unit test gives multiple configurations where we observe problem. Some of them are:

input_shape: [32, 4, 8, 768]
ttnn.bfloat16,
ttnn.TILE_LAYOUT,
ttnn.ShardStrategy.BLOCK,
ttnn.ShardOrientation.COL_MAJOR,
hw_as_shard_shape: False

input_shape: [32, 4, 8, 768]
ttnn.bfloat16,
ttnn.TILE_LAYOUT,
ttnn.ShardStrategy.BLOCK,
ttnn.ShardOrientation.ROW_MAJOR,
hw_as_shard_shape: False

input_shape: [32, 4, 8, 768]
ttnn.bfloat8_b,
ttnn.TILE_LAYOUT,
ttnn.ShardStrategy.BLOCK,
ttnn.ShardOrientation.ROW_MAJOR,
hw_as_shard_shape: False

Used grid size is: (8, 8).

Ops found affected:

ttnn.exp
ttnn.cos
ttnn.sin
ttnn.abs
ttnn.isfinite
ttnn.isnan
ttnn.isinf
ttnn.isneginf
ttnn.isposinf

Also problem is observed when no op is called (just copy tensor back and forth).

Example shapes that lead to problem can be:

[256, 2, 5, 1536]
[1, 256, 2, 2304]
[32, 4, 8, 768]

• Steps to reproduce Checkout ngrujic/unit-tests-1branch (soon to be merged into main). You can run unit test which can showcase this problem with this command:

pytest  tests/ttnn/python_api_testing/non_working_unit_tests/wormhole/test_eltwise_block_sharded.py

[umadevimcw]

We need more information on this issue like grid size , row_major or col_major etcc..

[nemanjagrujic]

@umadevimcw I updated ticket with required information. You can find those in unit test as well.