Blender Git Loki

Cycles: use CUDA toolkit 11.3 for buildbot in cycles-x branch

Cycles: add initial specialization of svm_eval_nodes

To leave out shaders nodes from rendering kernels where they are known to
be unused. Reduces GPU render time between 3-12% depending on the scene in
our tests.

Differential Revision: https://developer.blender.org/D11131

Cleanup: add mechanism for devices to not have a megakernel

Cleanup: move number of concurrent GPU paths to device queue

So different devices can set it to different values, though it's still
hardcoded as of now.

Cycles: add GPU queue kernel time stats

* Print summary of kernel execution times for GPU queues with --verbose 3
* This is for combinations of kernels rather than individual ones, since
measuring individual ones has a performance impact.
* Refactor code to be independent of CUDA queue

Differential Revision: https://developer.blender.org/D11112

Fix wrong order of graphics interop destruction

The GPU display needs to be destroyed after path trace works, otherwise
the CUDA interop resource destruction will error with an invalid graphics
context.

Additionally, added guards to ensure the CUDA graphics interop object is
never copied.

Fix wrong number of integrator kernels

The value was off-by-one, causing buffer overflow on accessing the queue
counters.

Cycles: Extend information provided in the full report

Include actual devices used for path tracing and denoising.

useful for in the following cases:

- Verify that the benchmark was made on the device it was intended to
be done.

- Easily see which exactly device was used for OptiX denoiser.

- Share and compare different reports, by making it explicit in the
report itself what conditions the benchmark was run in.

The report is printed at the end of rendering when Blender is run with
the `--debug-cycles` command line argument (verbosity level 2).

Differential Revision: https://developer.blender.org/D11105

Restore submodule URLs for pushing branch to git.blender.org.

Cycles: force disable OpenCL building since it has errors

Avoid costly OptiX module creation when only doing denoising

Cycles: Show message about loading denoising kernels

It is possible that denoiser will create device on-demand, and the
creation might take a while due to possible runtime kernel compilation
or optimization of ptx for OptiX.

This change makes it so a nice and meaningful status is reported to the
user instead of status being stuck at o samples.

Fix wrong render with Cycles persistent data

Need to always zero render buffers, otherwise it will have previous
result stored in it, making integration wrong (image will be over
exposed) and will make it so that adaptive sampling will pretend that
it adds samples, causing an early convergence.

Probably the render buffer is to be freed at the end of the render
anyway, but this would need a deeper look. For now just mimicking
what happens in master.

Fix compilation error with strict flags

It could be seen as too much of a boiler plate code to explicitly
mention all kernels, but when one does not constantly work with the
related code it is a bit hidden to see where modifications are to
be made when new kernel is added. Having compiler to nag about places
where modification might be needed is personally found to be very
helpful.

Cycles: Refactor of OptiX queue

- Use early output instead of indented if-else block.
- Use nullptr instead of NULL.

Should be no functional changes, just making code a bit more modern
and easier to follow.

Cycles: Use best suitable BVH layout for OptiX

Restore back the decision between fast-to-trace and fast-to-update
BVH types. Based on the BVH type which is available via the BVH
parameters.

The BVHType is now moved from render/ to bvh/ so that the device
can easily access the enum values.

Fix OptiX denoiser being broken after addition of CUDA streams

Enable megakernel path in OptiX

Fix launch params not being allocated and updated

Remove OptiX code that is no longed in use