Thu, Jun 21
The problem here is:
- pdInitEffectors() is called when building depsgraph relations for each particle system, looping over all objects in the scene.
- In this file, it creates no effector relations because there are no effectors in the scene.
- When evaluating the particle system pdInitEffectors() is called again to get all the evaluated effectors.
- Since there are no relations, looping over all objects in the scene can access data that has not been copied yet, causing a crash.
Appears to be fixed by rBb350edc20750: Depsgraph: Add missing relation from proxy group.
This appears to be the same issue that @Sergey Sharybin (sergey) has been investigating, where the transform matrix of the empty instancing the proxy group is used, but there is no relation in the depsgraph so sometimes it copies uninitialized data.
These crashes were fixed somewhere between monday and now, did not try to find which commit.
Wed, Jun 20
Tue, Jun 19
Mon, Jun 18
To fix this bug, we should add exception handling somewhere so we can report an error instead of crashing.
Sun, Jun 17
It's no big deal, just an unnecessary merge commit that doesn't break anything. Use rebase next time to avoid it.
Sat, Jun 16
Looks generally fine.
If we want to follow other Blender then uchar is the more common one, though uint8_t perhaps makes it more clear that we are not dealing with raw bytes and not strings? Either is fine with me.
Fri, Jun 15
Tested performance on CPU/CUDA GTX 1080, found no significant slowdown.
- Keep Color input and add Color output, for convenience and backwards compat.
- Default to Only Local, it's faster and sufficient for most cases.
- Rename Radius to Distance.
- Add OSL support. Quite ugly but we have no better solution in the OSL API.
- Placeholder GPU shader for master, will need more work when merging into 2.8 for Eevee.
- Don't use SVM AO code for global and baking AO. These support transparency which the new node does not, I prefer to keep that feature available at least for baking.
- Remove VNDF code since sampling it is quite noisy and the discontinuous sampling is quite visible with low number of samples. I also don't see really bad artifacts when just sampling the hemisphere around the smooth normal.
It would also be useful if Blender could automatically "auto advance" to the next slot upon rendering.
I can't envision a situation where you'd intentionally want to overwrite an existing slot, but one often wants to compare sequential images and forgets to change slots before, unintentionally overwriting the previous one.
Thu, Jun 14
This needed some build fixes, but other than that it looks good to me as the first step.
Build fixes for standalone and tests.
Assigning to @Clément Foucault (fclem) for advice.
Thanks for the suggestion. We don't track feature requests in this tracker, http://rightclickselect.com/ is the right place for that.
Updated with some UI tweaks.
- Add Render Slots panel with list widget and +/-/x icons.
- Only create render slots for render result images.
- Always initialize slot names.
Looks great, just minor comments about naming and properties.
Wed, Jun 13
Tue, Jun 12
There was a further problem where performance of manipulators degrades over time. This should be fixed as well now.
Mon, Jun 11
Committed fix for the Link/Append behavior.
With the two committed fixes it's already noticeably faster, but I did not measure it exactly.