Eltwise Master Tracking

[eyonland]

https://docs.google.com/spreadsheets/d/1MWzDkcmI_Jq4xeQAcrXnd7BpyD7o66RpvuE6dKNoBb8/edit?gid=0#gid=0

PyTorch2 Tracing Generality

• Pytorch Tracing sweeps - debugs & fixes #13521

Master issue for work needed on LLKs

• LLK Bugs & Features for Binary Generality  #9702
• Handle Inf, Nan at Kernel level for blocked ops #14077

Layout Issues

• ttnn.div_bw unary inconsistent documentation regarding input tensors layout #13854
• ttnn.ldexp_bw operation does not work with row major layout #13928

Scalar Issues

• [Bug Report] Binary op with scalar is very slow #13644
• Support for tensor-scalar variant in ttnn.maximum and ttnn.minimum #13864

Broadcasting

• Eltwise / Reduce / Broadcast Related tasks/bugs - MCW #6445
• [Bug Report] Multiply and divide has different broadcasting rules #12798
• eltwise ops fail PCC when broadcasting [GS][WH] #13464
• [Bug Report] Binary operations which require ttnn::repeat are very slow #13643
• TTNN throws an exception if second operand of binary Ops require broadcast #13566

Blackhole Issues

• misc/test_unary_ops_ttnn.py::test_unary_pow_ttnn #12577
• sweep_tests/pytests/tt_dnn/test_bitwise_or.py mismatch #12584
• sweep_tests/pytests/tt_dnn/test_bitwise_xor.py mismatch #12585
• sweep_tests/pytests/tt_dnn/test_div.py only supported on WH #12586
• sweep_tests/pytests/tt_dnn/test_unary_div.py only supported on WH #12588
• sweep_tests/pytests/tt_dnn/test_floor_div.py only supported on WH #12589
• sweep_tests/pytests/tt_dnn/test_frac.py unsupported on BH #12590
• sweep_tests/pytests/tt_dnn/test_remainder.py unsupported on BH #12591
• sweep_tests/pytests/tt_dnn/test_right_shift.py mismatching on BH #12592

Bugs: Unsupported Operations or Features Generality

• Support floating-point exponents for ttnn.pow #13779
• [Feature Request] DPS version of logical_xor op #13758
• [Bug Report] Ttnn::multiply does not broadcast in second dimension #13646
• [Bug Report] ttnn.clip interface does not follow Pytorch #14099

Bugs: Mixed Precisions

• ttnn.lerp operation does not work with mixed precaution #13625
• ttnn.logaddexp_bw operation fails with low PCC when at least one tensor is bfloat8_b #13930

Invalid Results

• [Bug Report] invalid acosh backward result #6583
• [Feature Request] Improvement Needed for Unit Tests #6633
• [Bug Report] invalid isclose result #6672 -->
• [Bug Report] Invalid recip result #6720
• [Bug Report] Invalid log result  #6721
• [Bug Report] invalid asin result #6722
• [Bug Report] invalid acos result #6723
• [Bug Report] invalid isfinite result #6724
• [Bug Report] invalid isinf result #6725
• [Bug Report] invalid isposinf result #6726
• [Bug Report] invalid isneginf result #6727
• [Bug Report] invalid signbit result #6728
• [Bug Report] invalid tan result  #6735
• [N300] TTNN Unit Test Failures: Op PCC failures #6991
• Multiple operations are handling inf/nan values differently that pytorch #7488
• Implement PReLU correctly #8544
• TTNN pow - fail when input tensor BFLOAT8_B and exponent is scalar float #8593
• TTNN pow breaks when exponent is tensor #8594
• LLK Bugs & Features for Binary Generality  #9702
• TT_LIB eltwise_addcmul - fail if one of the tensor arguments is on the host device (GS and WH) #9985
• Binary Shift operators #10034
• Binary Bitwise ops #10035
• TTNN eltwise_acosh - fail when input tensor has ROW_MAJOR layout (GS and WH) #10081
• TTNN eltwise_atanh - fail when input tensor has ROW_MAJOR layout (GS and WH) #10141
• TTNN eltwise_acosh fails with Fatal Python error: floating point exception when input tensor has a shape [3, 1, 22, 12] and a ROW_MAJOR layout #10160
• Batch normalization support #12253
• Support clamping with min/max tensor #13234
• Some pointwise unary ops return 2-D results from 1-D inputs #12671
• ttnn.mish operation kernel duration is high #12815
• Incorrect results with fmod and remainder from negative inputs #12835
• Tensor–tensor bitwise operators #12907
• Unsupported cases for aten.amin/amax conversion #12936
• Unsupported cases for aten.var.correction conversion #12937
• Unsupported cases for aten.prod.dim_int conversion #12939
• [Bug Report] ttnn.max - Unsupported dim  #13186
• Suboptimal trigonometric functions #13337
• PyTorch Tracing Sweeps #13373
• Integer support of eltwise ops #13374
• Optimize ttnn.round with a direct implementation #13385
• ttnn.bitwise_left_shift gives results different from the torch version when the bit used to represent a sign differs from the bit on position 31 - shift_bits #13415
• Pytorch Tracing sweeps - debugs & fixes #13521
• Values in (2^15, 2^23) are not rounded by rounding ops #13544
• ttnn.add example shared in TTNN API documentation fails from main and from v0.52.0 tag  #14448

Bugs: Operational Failures and Hangs

• multiple ttnn operations hang when run many times in loop on WH cards #4702
• TTNN eltwise_logical_xor sweep test execution hangs (GS & WH) #10144

Bugs: Low PCC Errors

• ? tt_lib.tensor.rpow fails with low PCC [Grayskull] [Doc] #3309
• ttl.tensor.recip operation breaks with low PCC [Wormhole] [Doc] #3561
• Low PCC for negative values in exp #3612
• * isfinite, isinf, isposinf, isneginf and isnan fail with low PCC on Wormhole cards #4409
• Some TTNN operations fail with low PCC error [Grayskull and Wormhole] #5893
• ttnn.exp, exp2, expm1 and ldexp fail with low PCC on both GS and WH in larger value ranges [Bug Report][GS][WH] #6391
• ttnn.signbit operation failed with low PCC [Grayskull and Wormhole] #7749
• ttnn.log, log2 and log10 operation fail with low PCC on TILE layout, bfloat8_b and random shapes like (198, 216)  #8634
• ttnn.tan_bw operation fails with low PCC  #13675
• ttnn.i0_bw operation fails with low PCC #13676
• ttnn.cos_bw operation fails with low PCC on Grayskull #13739
• ttnn.subalpha_bw low PCC #13774
• ttnn.hypot_bw low PCC #13776
• ttnn.bias_gelu_bw unary low PCC #13856
• ttnn.ldexp_bw operation fails with low PCC #13929
• ttnn.rdiv_bw low PCC #13937

Unsupported dtypes

• ttnn.lgamma operation does not work bfloat8_b #13626
• ttnn.acosh_bw operation does not work with bfloat8_b tensors #13674
• softshrink_bw operation does not work bfloat8_b #13927
• ttnn.logit_bw operation does not work bfloat8_b #13926

Documentation Issues

• ttnn.polar_bw binary inconsistent documentation regarding the input tensors datatype #14144
• ttnn.angle_bw binary inconsistent documentation regarding the input tensors datatype #14145
• ttnn.sub_bw binary inconsistent documentation regarding the input tensors layout #14141
• ttnn.remainder_bw binary inconsistent documentation regarding the supported architecture types #14293
• ttnn.remainder_bw binary inconsistent documentation regarding the input tensors layout #14291
• Improve documentation for ttnn glu, reglu, and geglu operations #7487
• TTNN hard*-alike ops - incomplete documentation #7711
• TTNN trigonometric hyperbolic ops - wrong types info in documentation #7792
• TTNN log/exp-alike ops - incomplete documentation #7955
• TTNN tensor creation ops - zeros, zeros_like, empty - incomplete documentation #7957
• TTNN logical binary ops - incomplete documentation #7961
• TTNN sub and mul ops - incomplete/inconsistent documentation #7996
• TTNN log-alike ops incomplete documentation #8142
• TTNN reglu incomplete documentation #8151
• TTNN multigammaln - incomplete documentation #8155
• TTNN mish - incomplete documentation  #8156
• TTNN leaky relu - incomplete documentation #8157
• tt_lib.tensor.glu, geglu, reglu, swiglu - inconsistent documentation  #8343
• TTNN backward ops - incomplete documentation #10585
• Update documentation for Unary bw ops #11086
• Incomplete documentation for ttnn.normalize_global and ttnn.normalize_hw regarding input_tensor rank #13682
• ttnn.mul_bw unary inconsistent documentation regarding input tensors layout #13683
• ttnn.embedding_bw inconsistent documentation regarding input tensors dtypes #13685
• ttnn.acosh_bw operation does not work with bfloat8_b tensors #13674
• ttnn.lgamma operation does not work bfloat8_b #13626
• ttnn.logit_bw operation does not work bfloat8_b #13926
• ttnn.add example shared in TTNN API documentation fails from main and from v0.52.0 tag  #14448
• ttnn.mul_bw binary inconsistent documentation regarding the input tensors layout #13977