@Michael-Jones Hey. It this PR still relevant?
Yes, I think that Metal renders still have an upper texture limit of 16384. IIRC, I ran into performance & spill limit issues when trying enable…
@Michael-Jones Would you have time to tackle this in the coming days, or shall we take over?
I will push changes this afternoon, with the two other changes split out as per your breakdown.…
Thanks a lot @Iain-1. You can email me at michael underscore p underscore jones at apple dot com (after substituting the the right symbols in every other word!)
Hi @Iain-1 , thanks for the update. If you do have any logs from a bad run it might help, depending on the nature of the crash. In particular, if you are seeing any GPU faults or command buffer…
I'd like it to land please. Either here or as a separate PR if you prefer. I see no good reason to bake in the seed as a kernel constant.
It doesn't - my apologies. It fixes a genuine crash I was seeing when doing top-level BVH builds of a motion blurred scene, caused by an indexing mismatch with motion_transform_index in a later…
I'm going to re-check this. I added some hacks into the geometry syncing layer to induce automatic edits (I tried using bpy for this but hit a wall). I encountered some crashes here due to trying…
I'm proposing that we ignore this for now. It's an existing member definition which has been relocated, and changing it would bloat the diff.
More generally, the Metal backend has evolved its…
My soak test involved playing an animation in live viewport mode and I noticed that seed was changing and causing compilation churn whenever I had kernel specialisation enabled.
Testing the patch seems indeed to fix the issue.
However, I am not really happy with the way how the PR is organized and presented:
- It seems to mix some functional changes and…
@Alaska, I'm not sure. From the description It sounds like GPU denoise is not helping with stability, but other factors may be at play. It's curious that it seems to progressing at least part way…
Yes, I would like to get this fixed in 4.2. I suspect that there's some hazardous updating of BVHs.