Pierre Paleo
Pierre Paleo
Thanks! From what I see in `requirements.txt` and `pyproject.toml`, the minimum required numpy version is `1.12`, not `1.8`. To simplify version comparison a little bit, you can use `from pkg_resources...
Test fails for `numpy == 1.19.5`. Seems to be because [not all normalization were available for this version](https://numpy.org/doc/stable/release/1.20.0-notes.html#norm-backward-forward-keyword-options-for-numpy-fft-functions). Just tested on my machine: - `test_norms` fails with `numpy==1.19.5` - `test_norms`...
Tested with pyfftw 0.12 and 0.13.
Hi @weichengkai The purpose of a `SiftPlan` is to be created once for a given image shape, and then to be used multiple times over different images. It is similar...
I have nothing against it. Let's hope POCL is mature enough to be used in the CI without causing more headaches.
Thanks, I'll look into it
Last commit seems to have fixed the problem. Tried on - Titan V, Debian 11, cuda 11.2, Python 3.9.2 - Tesla V100, Ubuntu 20.04, cuda 10.1, Python 3.8.10 Interestingly, the...
Tests are now passing on gitlab CI. Good! Although the ones from this PR are skipped because none of the card has enough memory. Before merging I'd like to rework...
Should be better. The tests now require "only" 4.5 GB device memory (skipped otherwise) and should be faster.
Thanks! test passed on all 3 CI GPUs. This PR should be good now.