Blender Git Loki

Cleanup: clang-format

Fix PMJ pattern automatically enabling adaptive sampling

It should only be the other way around.

Merge branch 'master' into cycles-x

Temporarily disabling per-object fast GI AO distance as it needs to
be implemented a bit different in the Cycles X.

Fix denoising of shadow catcher pass in Cycles X

Something what slipped through misread of the documentation:
normal pass implies requirement of the albedo pass.

Cycles X: Log OIDN error

Ideally would communicate this to the interface, but those errors
are mainly useful for development and should happen in the release
builds.

Fix OIDN guiding passes filtered multiple times in Cycles X

Was happening for shadow catcher passes, caused by missing filtered
flag assignment.

Fix missing space in debug logging

Merge branch 'master' into cycles-x

Cycles X: Restore OptiX SDK 7.3 requirement

Got lost with recent master branch merge. Hopefully this will fix a
missing OptiX compute device issue.

Merge branch 'master' into cycles-x

Fix Cycles X true displacement with no faces

Do an explicit check for the number of triangles in the input mesh.

This allows to skip some computation, and avoids a tricky decision
about whether shading of 0 points is a success or a failure.

Test file: F10247756

Differential Revision: https://developer.blender.org/D12088

Fix missing pre-filtering execute in the OIDN Cycles X

The `execute()` call got missed in some cleanup of the initial hacky
implementation and indicatives of it did not get proper attention.

The pre-filter takes about 0.5sec per pass on i9-11900k Full-HD image
which indeed makes it not suitable for the preview render.

The file used to verify this change: F10247300

Result:
|No filter|With filter|
|{F10247303}|{F10247304}|

Cycles X: Support OIDN with guiding pass prefiltering

Straightforward implementation based on example from the OIDN 1.4
release page. Always configure the beauty pass denoising with the
`cleanAux true`, so that when guiding passes are noise-free we
don't filter them and don't introduce possible artifacts. When the
pre-filtering is enabled, then on top of cleanAux option we will
filter the guiding passes.

Has a separate option for viewport and final rendering, as pre-filter
is not coming for free and for viewport we want to be as interactive
as possible by default. From quick tests it seems that pre-filter is
rather quick, so maybe it can be enabled for viewport as well.

The final render has the option enabled by default to help with cases
when the result is tricky and noisy. If the passes are known to be
noise free is better to disable filtering manually. Can re-iterate
over the default value after gaining some more statistics.

Test file: F10242785
Ground truth: {F10242777}

|Samples|NLM|OIDN Old|OIDN new no filter|OIDN new with filter|
|64|{F10245231}|{F10245242}|{F10245244}|{F10245243}|
|256|{F10245232}|{F10245248}|{F10245233}|{F10245249}|

Extra result with the `pvt_flat.blend` scene:
|Samples|NLM|OIDN Old|OIDN new no filter|OIDN new with filter|
|256|{F10245234}|{F10245237}|{F10245235}|{F10245236}|
|512|{F10245238}|{F10245241}|{F10245239}|{F10245240}|

NOTE: NLM is done in the master branch, where lighting is handled a bit
different. Hence the difference in the light condition.

Differential Revision: https://developer.blender.org/D12043

Merge branch 'master' into cycles-x

Cycles X: restore CPU sampling profiler

Stats now reflect the new kernels. Also change implementation to unset active
shader and object, previously there were some hard to verify assumptions about
for which kernels we can assume they are correctly set.

Differential Revision: https://developer.blender.org/D12042

Fixes for pass handling in Cycles X

- Register Shadow Catcher pass as 3-component RGB color.
Matches the way how it is stored, solving uninitialized alpha channel.

- Don't write alpha channel in denoiser when the denoising input is a
3-component color pass.

- Add safety asserts in the film conversion kernels.

Differential Revision: https://developer.blender.org/D12078

Cycles X: fix/improve handling of mismatches between render pass components

Rules are as follows now:
* Single channel pass: can be read into buffer with 1, 3 or 4 channels
* Motion and cryptomatte passes: must be read into buffer with 4 channels
* RGB, RGBA and vector passes: can be read into buffer with 3 or 4 channels

Differential Revision: https://developer.blender.org/D12063

Merge branch 'master' into cycles-x

Merge branch 'master' into cycles-x

Merge branch 'master' into cycles-x