Blender Git Loki

Cycles Denoising: Remove obsolete kernel pass writing option

It was used for cross-buffer denoising, but that was removed a while ago.

Merge remote-tracking branch 'origin/master' into temp-cycles-denoising

Cycles Denoising: Remove unused flag

Cycles Denoising: Only write shadowing info if denoising or shadowcatcher is enabled

Cycles Denoising: Avoid useless calculation in branched path tracing if denoising is disabled

Cycles Denoising: Move the denoising feature evaluation before the subsurface bounce

Previously, SSS bounces would skip the denoising feature evaluation, basically acting like a specular surface.

Cycles Denoising: Move closure classification to a separate function

Cycles Denoising: Move offset pointer handling to a separate class

This also fixes the memory leak on OpenCL.

Cycles Denoising: Move back into the kernel/ folder

Cycles Denoising: CPUDevice tweaks from the review

Cycles Denoising: Add back the feature strength option

Note: The control is now different, reusing old values might give surprising results

Cycles Denoising: Fix Multi-GPU viewport rendering

Cycles Denoising: Fix Multi-GPU OpenCL denoising

Cycles Denoising: Fix AMD OpenCL support

Credit for this patch goes to nirved.

Merge remote-tracking branch 'origin/master' into temp-cycles-denoising

Merge remote-tracking branch 'origin/master' into temp-cycles-denoising

Cycles Denoising: Remove a few unused utility functions

Cycles Denoising: Refactor design row construction

This change improves the shared memory access pattern, reduces the local memory requirementy by 11 floats per thread and saves some memory copying.

Cycles Denoising: Allocate more shared memory for the CUDA reconstruction kernel

The small shared memory size limited execution to a single block, which caused slowdowns in CUDA denoising due to bad occupancy.

Cycles Denoising: Remove cross- and gradient-denoising option, replace PCA threshold with a checkbox

Cross denoising is significantly slower, needs more memory and didn't really produce better results in my tests.
Gradient denoising sometimes helped, but tends to produce artifacts and was broken for a few weeks already anyways.

The extremely confusing "Filter strength" (negative values used to map to an absolute threshold of 10^(2*-strength), positive ones to a relative threshold of 10^(2*strength), 0 to relative 1e-3) was replaced by a checkbox that selects between an absolute threshold of 1 and a relative threshold of 1e-3.
Eventually, I'd like to completely remove the option, but it's not clear yet which one is the better approach.