Hi Scott,

may I ask you which kind of vector / matrix operation in extended precision (np.longdouble) is supported in 'pint' ? It can't be backed by the underlying blas library as extended precision functions are not supported there.


Am Di., 6. Dez. 2022 um 15:24 Uhr schrieb Charles R Harris <charlesr.harris@gmail.com>:

On Sun, Dec 4, 2022 at 9:36 AM <abhisrkckl@gmail.com> wrote:

I am interested in contributing to this. I am a contributor in the PINT pulsar timing package Scott mentioned. But I am new to contributing to numpy, and I am not sure where or how to begin. Are there beginner-level resources that I can look at?


Abhimanyu Susobhanan
Postdoctoral fellow
Centre for Gravitation, Cosmology, and Astrophysics
University of Wisconsin Milwaukee

I think the first thing to do is have a discussion in the astropy community. Dependence on extended precision is fragile at best, it isn't available on all platforms and probably doesn't offer enough precision to be future proof. Then gather some performance data.
  1. What computations are needed?
  2. Can the computational algorithms be improved?
  3. Would parallel computation help?
  4. How much faster is extended precision?
  5. Where are the performance bottlenecks in the double-double time type?
  6. Is there hope of improving the double-double performance.
My own hope would be that a combination parallel computation and lower level code for double-double would be a viable solution.



NumPy-Discussion mailing list -- numpy-discussion@python.org
To unsubscribe send an email to numpy-discussion-leave@python.org
Member address: cmkleffner@gmail.com