Samuli
Samuli
The handling of the `w` component in the SIMD code is a bit inconsistent, on SSE it should be ignored for all operations so denormal/NAN/INF should only cause performance regression....
Initial dataset testing indicates that even though errors are rare, some badly behaving files will spam an astounding amount of errors (one file so far has issued a whopping 29088...
Limited the maximum errors to 10 per type. Also added element names (like nodes or meshes that prompt the warning) to the warnings.
The dataset looked fine at a glance, didn't have the time to do a proper labeling this time. The only open issue is the `use_blender_pbr_material` mentioned above with the caveats...
Disabled it for now in that case
> Could you amend in this documentation change too? Sure, thanks! My bad for missing that one, should be fixed now.
Ran through all the dataset files containing skeletons: ``` Loaded 2264 files in 598min 50s. Processed 5.41GB at 0.15MB/s. Ignored 0 invalid files. ``` After this fix there were no...
> We need to debate adding a new option, but that can review that in a separate pr. That means you can stack the two prs together into that pr....
@lyuma provided some test cases with multiple bind poses per bone. Added explicit support for this in the form of warning the user if a bone has multiple conflicting bind...
Thanks! I'll squash the commits after seeing that the CI passes etc. Also ran through all the skeletons in my dataset (+ the entirety of the public dataset) again with...