MAINT: align random double_fill to float_fill, simplify distributions.c #14611
Conversation
|
ping @bashtage |
|
Seems right to me, but @bashtage might know if there was something motivating that implementation. |
|
It is an optimization. The 0 parameter functions are unique in the sense that it is possible to fill blocks without function call overhead since these can be inlined. Once upon a time, I did some tests and it seemed worth it for large samples. I don't have any saved results to back this up now. |
| @@ -920,7 +920,7 @@ cdef class Generator: | |||
| """ | |||
| key = np.dtype(dtype).name | |||
| if key == 'float64': | |||
| return double_fill(&random_gauss_zig_fill, &self._bitgen, size, self.lock, out) | |||
| return double_fill(&random_gauss_zig, &self._bitgen, size, self.lock, out) | |||
There was a problem hiding this comment.
I think you can go further if the fill-methods are being removed and you can remove double_fill and just use
cont(&random_gauss_zig, &self._bitgen, size, self.lock, 0,
0.0, '', CONS_NONE, None,0.0, '', CONS_NONE, None, 0.0, '', CONS_NONE, None
out)
OK, makes sense, since the function calls cannot be inlined. Let's put this off until after gh-14608 and we can see if eliminating |
Part of the C-API overhaul in gh-14608 and gh-14604, and should be merged first to simplify the distribution function renaming.
double_fillhad an interface that required afillvariant of the sampling function, whilefloat_fillandfloat_fill_from_doublewere different. Align the interfaces and remove thefill variants indistributions.canddistributions.h`.