[LLVMGPU] Remove some op's polynomial approximation

[lialan]

Except:

• erf, which is under investigation.
• powf, need to also update many tests. Doing it in another PR.

[lialan]

@bjacob @kuhar This is just a follow up to remove some unaffected ops for poly approx on LLVMGPU.

[kuhar]

Looks good. Have you confirmed the numerics?

[lialan]

Looks good. Have you confirmed the numerics?

Not seeing CI complain about numerics except something is off with pow. I will wait until CI is clear.

[bjacob]

The problem is that at the moment, this is not only dropping polynomial approximation for erf, this is dropping promotion from f16 to f32 for all math functions. That is because that promotion is done by mlir::populateMathPolynomialApproximationPatterns upstream, and not exposed separately from it, and the only place where we call that function is in the polynomial approximation path for erf. I believe that to be the cause of accuracy issues currently observed when trying to drop the polynomial approximation for erf.

[kuhar]

The problem is that at the moment, this is not only dropping polynomial approximation for erf, this is dropping promotion from f16 to f32 for all math functions. That is because that promotion is done by mlir::populateMathPolynomialApproximationPatterns upstream, and not exposed separately from it, and the only place where we call that function is in the polynomial approximation path for erf. I believe that to be the cause of accuracy issues currently observed when trying to drop the polynomial approximation for erf.

This PR doesn't drop polynomial approx for erf.

[bjacob]

This PR doesn't drop polynomial approx for erf.

Oh, I see.

[bjacob]

But I mean, the whole premise of singling out erf is depending on that ongoing investigation. It would make more sense to resolve that first.

Do we have a short term need for this PR? (Is it producing an improvement in a metric we're tracking) ?

[lialan]

It would make more sense to resolve that first.

This is separate so we can wait until you resolve erf and then merge this.

[lialan]

Using @__ocml_pow_f64 (with fptosi) return 31 for calculating 2 ** 5.0 rather than 32. other values are okay.

[bjacob]

Interesting!!

[krzysz00]

Side note: there is a standalone pass and pattern set for promoting math operations to wider types