Repository navigation
Worse performance using datashader? #296
Description
I wrote some benchmarks available here #295 (they can simply run as tests) and I have noticed that the datashader performance is worse than the matplotlib based one.
I think this maybe be due to the size of the canvas used by datashader since in the MERFISH example here #243 the performance was (as expected) better.
Therefore using a smaller default canvas size may fixed the issue. @Sonja-Stockhaus could you please have a look into this?
Activity
LucaMarconato commentedon Jul 13, 2024
Here are the results of a (single) run of the tests (the timing are consistent across multiple manual runs).
LucaMarconato commentedon Jul 13, 2024
With the fix that I proposed to the performance bug here #297 the performance gap is much bigger
timtreis commentedon Jul 14, 2024
@Sonja-Stockhaus my "didn't-look-at-the-code" theory is that datashader generates too large of an image which then bypasses the rasterisation-downsampling logic. Wdyt?
Sonja-Stockhaus commentedon Jul 15, 2024
Yep, datashader generates an image that is exactly the size of the extent (large extent = large image = long runtime). I'll think of sth so that we can use a smaller canvas size and then maybe rasterize or so to bring it back to the original scale.
Do we want a heuristic again to decide on the "smaller canvas size"?
I also noticed that for datashader, e.g. the radius of the points is relative to the axes which is not the case for matplotlib. So for a large extent you need extremely large point sizes to even make them visible at all with datashader. That should be consistent with matplotlib.
LucaMarconato commentedon Jul 16, 2024
Thanks for the explanation. I would reuse the logic of
_rasterize_if_necessary()or_multiscale_to_spatial_image()to take thedpiof the figure and thefig_sizeinto consideration, since the extent could be extremely large, but in the end we are limited by the pixels available on screen/paper for plotting.LucaMarconato commentedon Jul 16, 2024
Btw, off-topic comment, when plotting Visium HD data as points/circles I noticed a Moire pattern due to the presence of a small rotation in the raw data. With datashader rasterization the Moire pattern disappears, which is great! So using datashader could have also this nice use case beyond improved performance.