Make expm1(::Complex64) type stable.
#21896
Conversation
|
LGTM. |
|
Should this be backported to 0.6? |
|
Ah, probably. I wasn't sure if 0.5 has the necessary |
|
Actually |
(cherry picked from commit a34fd51)
| wr = isfinite(er) ? erm1 - 2.0*er*(sin(0.5*zi))^2 : er*cos(zi) | ||
| Complex(wr, er*sin(zi)) | ||
| if isfinite(er) | ||
| wr = erm1 - 2 * er * (sin(convert(Tf, 0.5) * zi))^2 |
There was a problem hiding this comment.
I know I'm a bit too late, but why not using zi / 2 instead of converting 0.5? The performance difference is sensible for BigFloat for this operation (not the whole function, though).
There was a problem hiding this comment.
I'm just trying to not change the result for machine types and being conservative about this transformation. I assume this transformation is exact for Float16, Float32 and Float64 (though floating point corner cases frequently surprises me...)? If that's the case then I think we can change this to / 2.
Noticed by accident when debugging #21831 (comment) .... The issue is apparently there ever since
expm1(::Complex)is defined in 0.3....