Cycles Denoising: Fix reading of the new properties

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

Cycles Denoising: Add option to store feature passes in the render result

Useful for debugging and possibly in the future in combination with a standalone animation denoising mode

Cycles Denoising: Move denoising properties to the Cycles addon and improve the UI

Cycles Denoising: Fix crash in CPUDevice

The code initialized allocated memory by assigning the KernelGlobals to it. However, that calls the assignment operator, which frees previous elements which were never allocated.

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

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

Cycles Denoising: Fix shadowing pass calculation

Cycles Denoising: Cleanup offset calulation

Cycles Denoising: Only store lower-triangular parts of XtWX matrix

This saves 55 floats per pixel, which is quite significant with GPU tile sizes.

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

Cycles Denoising: Merge all the changes from the version that's being reviewed

No functional changes.

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

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

Cycles Denoising: Disable local (aka shared) memory on OpenCL

On GPU architectures, storing the design row in local memory improves performance due to lower global memory bandwidth requirements.
However, if the GPU doesn't have enough local memory available, occupancy suffers which makes it even slower than the global memory version.

On CUDA, the amount of available local memory (shared memory in CUDA terminology) can be controlled, but that's not possible on OpenCL. So, to avoid a huge performance hit when the local memory isn't enough, it's disabled on OpenCL.

Cycles Denoising: Fix compilation on AMD CPU OpenCL

Cycles Denoising: Remove variable for split kernel variance

The devices know whether they're using the split kernel or not, there's no need to store the information.

Cycles Denoising: Remove unused filter_compat headers

Cycles Denoising: Remove obsolete filter kernel and argument

Cycles Denoising: Fix filter kernel build system