I think the argument nworkers = -1 to scipy.fft.fft2 and scipy.fft.ifft2 is in the wrong places in the notebook. Le lun. 11 mars 2024, à 21 h 25, via NumPy-Discussion < numpy-discussion@python.org> a écrit :
Good afternoon, Ralf.
We have done some of the measurements you recommended, for your convenience we have created a separate folder with notebooks where we measured memory usage and performance of our interpretation against Scipy. Separately you can run the tests on your hardware and separately measure memory. I've left the link below.
https://github.com/2D-FFT-Project/2d-fft/tree/main/notebooks
We measured efficiency for 4 versions - with multithreading and data type conversion. According to the results of the tests, our algorithm has the greatest lead in the case with multithreading and without data type conversion - 75%, the worst performance without multithreading and with data type conversion - 14%. In terms of memory usage we beat NumPy and Scipy by 2 times in all cases, I think this is a solid achievement at this point.
I can generalise that our mathematical approach still has a serious advantage, nevertheless we lose always to Scipy in inverse operation case, we haven't figured out the reasons yet, we are discussing it at the moment, but we will fix it.
It is important to note that at this stage our algorithm shows the above perfomance on matrices of size powers of two. This is a specificity of the mathematical butterfly formula. We are investigating ways to remove this limitation, we already assessed the effect of element imputation and column dropping, the result is not accurate enough. Otherwise, we can suggest putting our version to work only in cases of the mentioned matrices, it'll still be an upgrade for NumPy.
At this point I can say that we are willing to work and improve the existing version within our skills, knowledge and available resources. We still live with the idea of adding our interpretation or idea to the existing NumPy package, as in theoretical perspective within the memory usage and efficiency, it can give a serious advantage on other projects built on NumPy.
Thank you for your time, we will continue our work and look forward to your review. _______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: george.trojan@gmail.com