Handle Inf, Nan at Kernel level for blocked ops

[VirdhatchaniKN]

Issue :

Need to handle ops at the kernel level to Handle Nan, inf cases

Blocks the following issues :

• isfinite, isinf, isposinf, isneginf and isnan fail with low PCC on Wormhole cards #4409
• [Bug Report] invalid isinf result #6725
• [Bug Report] invalid isneginf result #6727
• [Bug Report] invalid isposinf result #6726
• [Bug Report] invalid isfinite result #6724
• [Bug Report] invalid isclose result #6672
• Incorrect ttnn.acos and ttnn.asin result #14356
• Multiple operations are handling inf/nan values differently that pytorch #7488
• Integer support of eltwise ops #13374

Issues to check:

• ttnn.sqrt - should return nan for negative values.

Issue Description :

• For ops developed on composite file structure, the nan. inf issue was solved using torch.nan_to_num : PR Ref
• Need to handle ops at the kernel level to Handle Nan, inf cases. Reference to discussion related to this :
  • Identify the unique small number that will represent a NaN across data types #8945,
  • Tech Report on Special values
• While running the test mentioned in the below issus, it was found that while passing
  • inf or -inf as inputs passes into the kernel as inf or -inf ( But still the result produces is not right )

• But the same for nan, it gets stored in dst_reg as -inf. Need inputs on why nan is getting stored as -inf in the kernel

[prajaramanTT]

@VirdhatchaniKN is this still an open issue ? If not, can you please close this ? Thanks.