Blender Git Loki

Cycles Denoising: Implement OpenCL denoising kernels

Unfortunately, as always when making kernel code OpenCL compatible, that mainly means adding a huge amount of annoying pointer qualifiers...

Cycles Denoising: Rework memory offset function for upcoming OpenCL integration

Cycles Denoising: Remove useless include

Cycles Denoising: Cleanup safe_isfinite

Cycles Denoising: Print file and line when opencl_assert fails

Cycles Denoising: Add separate set_tiles device callback

For CPU and CUDA, it was possible to determine the pointers to the tile buffers on the host and just fill the TilesInfo there.
However, for OpenCL the actual pointer is only known inside the kernel, so a separate kernel for setting them is needed.

Cycles Denoising: Use kernel_config.h for CUDA as in the regular kernel

Cycles Denoising: Bring back UI option after recent merge

Cycles Denoising: Fix variance pass generation in the split kernel

Cycles Denoising: Remove unused functions left over from the denoise-after-rendering removal

Cycles Denoising: Remove obsolete Keep Denoising Data option

Cycles Denoising: Remove tile overscan when using GPU rendering

Denoising a pixel requires access to the other pixels surrounding it. On the CPU, this is solved by waiting for the neighboring tiles to be rendered before the central tile is denoised.
On the GPU, it was handled by rendering larger tiles internally and discarding the overscan area after denoising. That saved a bit of memory, but wasted computation (with 256x256 tiles and a half-window of 8, 13% of rendered pixels were never actually seen).

Also, supporting overscan tiles made the code more complex. So, this commit removes the overscan code and uses the CPU approach on GPUs as well.

Cycles Denoising: Implement support in the split kernel

Merge updated shadowcatcher into temp-cycles-denoising

Update the shadowcatcher patch and make it look like a merge so that Git is happy

Cycles Denoising: Deduplicate split/mega kernel code in the CUDADevice

Cycles Denoising: Deduplicate split/mega kernel code in the CPUDevice

This also means that denoising is now also executed when using the split kernel, but the result is black since the kernel doesn't write the data yet.

Cycles Denoising: Fix NaN results in a few scenes

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

Conflicts:
intern/cycles/device/CMakeLists.txt
intern/cycles/device/device_cpu.cpp
intern/cycles/device/device_cuda.cpp
intern/cycles/device/device_task.h
intern/cycles/device/opencl/opencl_split.cpp
intern/cycles/kernel/CMakeLists.txt
intern/cycles/kernel/kernel_path_branched.h
intern/cycles/kernel/kernel_path_surface.h
intern/cycles/kernel/kernel_shader.h
intern/cycles/kernel/kernels/cpu/kernel_cpu_impl.h
intern/cycles/kernel/kernels/cuda/kernel.cu
intern/cycles/kernel/split/kernel_buffer_update.h
intern/cycles/kernel/svm/svm_closure.h
intern/cycles/util/util_atomic.h

Cycles Denoising: Store feature pass into in PathRadiance