Emit tosa::erf during lowering gelu op

[vinitdeodhar]

The PR adds support to use tosa::erf during lowering of aten.gelu op
Currently, aten.gelu uses its own custom implementation of erf which was designed before tosa support for tosa::erf : llvm/llvm-project@1fef1f9

aten.erf is already mapped to tosa::ErfOp https://github.com/vinitdeodhar/torch-mlir/blob/c785435a049a694a1814a7304ed0c34abe8b2580/lib/Conversion/TorchToTosa/TorchToTosa.cpp#L9176

The two implementations (Existing aten.gelu erf and tosa::erf) produce numerically different results.

tosa::erf is more precise approximation of the erf function. tosa::erf approximation is based on the following stackoverflow post:
https://stackoverflow.com/questions/35966695/vectorizable-implementation-of-complementary-error-function-erfcf
The stackoverflow post is in turn based on:
M. M. Shepherd and J. G. Laframboise, "Chebyshev Approximation of
(1+2x)exp(x^2)erfc x in 0 <= x < INF", Mathematics of Computation, Vol. 36,
No. 153, January 1981, pp. 249-253.
Maximum error: 2.65 ulps

[sahas3]

LGTM, good cleanup to just use tosa.erf. Can you please add an e2e test to lock down the numeric that is failing with the current implementation but passes with the change?

[sjarus]

I echo @sahas3's request - LGTM but would like to see an e2e test for the fx_importer_tosa target.

[vinitdeodhar]

I echo @sahas3's request - LGTM but would like to see an e2e test for the fx_importer_tosa target.

Thanks for the feedback. Added e2e test which fails without and pass with the cleanup changes

[vinitdeodhar]

Since I dont have access, can someone help land this change