[ttl] Add eq, ne, gt and lt ops

[vtsilytskyiTT]

[ttl] Add FP32 eq, ne, gt and lt ops

How?

By using corresponding ttnn ops.
Test method changed. ULP and PCC are not applicable to bool output. 1 ≠ 0.
So a new boolean comparison function introduced.
Only FP32 is supported for now. Bf16 should be added in tt-metal first.

Dependency

tenstorrent/tt-metal#41304

How to Test?

pytest test/python/test_elementwise_ops.py test/python/test_elementwise_shapes.py test/python/test_nd_shapes.py
pytest test/me2e/test_compare_ops.py test/me2e/ops/test_binary.py

Checklist:

• Self-reviewed (style, logic)
• Added tests (or justified none needed)
• PR is small and focused (one task)

[vtsilytskyiTT]

With this approach I hit a catastrophic cancellation . float32-1x1_bfloat16_buf2_interleaved-lt test fails because elements at index [25][9] are very close, (-0.6108) - (-0.6106) < 0, so "lt" must produce "true" (which it does in golden function), but ttnn produces fasle due to cancellation.

[zoecarver]

So excited to use these! Thank you, Vlad!

With this approach I hit a catastrophic cancellation . float32-1x1_bfloat16_buf2_interleaved-lt test fails because elements at index [25][9] are very close, (-0.6108) - (-0.6106) < 0, so "lt" must produce "true" (which it does in golden function), but ttnn produces fasle due to cancellation.

Maybe we can reduce the PCC for these ops? This seems fine within bf16 tolerance

[zoecarver]

LGTM!

[vtsilytskyiTT]

Thanks, @zoecarver ! I decided to go with the proper comparison. There is a tt-metal PR tenstorrent/tt-metal#41304 and I'm afraid that I'll have to commit to tt_llk. So this PR is in draft.

[vtsilytskyiTT]

PR tenstorrent/tt-metal#41304 merged, now tt-metal has a great support for comparison OPs.