Python CUDA parallize multiple SVD's of small matrices

Question

I've seen a similar post on stackoverflow which tackles the problem in C++: Parallel implementation for multiple SVDs using CUDA I want to do exactly the same in python, is that possible? I have multiple matrices (approximately 8000 with size 15x3) and each of them I want to decompose using the SVD. This takes years on a CPU. Is it possible to do that in python? My computer has an NVIDIA GPU installed. I already had a look at several libraries such as numba, pycuda, scikit-cuda, cupy but didnt found a way to implement my plan with that libraries. I would be very glad for some help.

as it turns out, cusolver has a batched implementation. I missed it the first time I looked thru the doc. — Robert Crovella, Oct 30 '20 at 20:58

score 2 · Answer 1 · edited Apr 16 '22 at 08:03

2

cuPy gives access to cuSolver, including a batched SVD:

https://docs.cupy.dev/en/stable/reference/generated/cupy.linalg.svd.html

edited Apr 16 '22 at 08:03

talonmies

70,661
34
192
269

answered Oct 30 '20 at 13:36

Stripedbass

194
1
8

Yepp I have seen that one, but this only make sense for big matrices. My problem is that I want that each thread is doing such an SVD and if I am not wrong cuPy does not allow me to do so in python. (Control each thread via its ID) – horsti Oct 30 '20 at 13:41
I haven’t tried it in cuPy, but in numpy the input array can be > 2 dimensions ( 8000 x 15 x 3 in this case) and it broadcasts. In theory, cuPy does similar? https://numpy.org/doc/stable/reference/routines.linalg.html#routines-linalg-broadcasting – Stripedbass Oct 30 '20 at 23:16

Python CUDA parallize multiple SVD's of small matrices

1 Answers1