Ianna Osborne
Ianna Osborne
Hi @kmohrman . Thanks for bringing this up! The discrepancy could be a floating point precision. `np.exp(-eta)` applies a nonlinear transformation. Even tiny differences in eta—say, at the 10⁻⁷ level—can...
> Thanks [@ianna](https://github.com/ianna)! When I try to import `cupy` and run `cupy.exp(-eta_gpu)`, I get `TypeError: Unsupported type `, so I think I may be misunderstanding something about how to use...
> Ok, thank you! That runs, but produces the same discrepancies as reported above. Could you please try `eta_gpu = ak.to_cupy(eta_gpu).astype(cp.float64)` Thanks!
@kmohrman - Thanks for sharing those detailed comparisons - they are incredibly helpful. What you are seeing is exactly what we’d expect when comparing floating point computations: even the biggest...
> So for validation studies between CPU and GPU, I suppose I should force both to `float64`. I will need ak arrays at the end of the day, so I...
@T90REAL - thanks for reporting this! Indeed we can still do better in this range.
@T90REAL - Thanks for reporting this! I can see that you have put a lot of work in testing `awkward` and reporting the issues. Thank you! You are welcome to...
@chrisburr - Could you, please, have a look? Thanks!
Thanks, @ariostas ! I agree. Let's do this for now, and reevaluate later. Thanks!
@ikrommyd - thanks for submitting the issue!