Blender Git Loki

Cycles: Use adaptive sampling division in the film kernel

Matches the pass data access in the buffers.cpp.

Unfortunately, the code is not currently running because there are no
adaptive sampling passes added to the viewport render.

Will work on enabling required passes in the viewport next.

Cycles: Use more const qualifiers in the film kernel

Cycles: Fix adaptive sampling possibly being false enabled

Add an explicit check with the configuration rather than relying
on the sampling pattern.

Cycles: Streamline naming in the film kernel

No functional changes.

Indicated possible issue with the difference in stride in the render
buffer and in the display half-float buffer.

Cycles: Use more clear variable name

The offsets are calculated using stride during update, but semantically
those are offsets.

Cycles: Remove unused adaptive sampling settings from KernelIntegrator

Cycles: Unify terminology/naming

Follow Cycles convention of using get/set as prefix.
This is the proper way ;)

Cycles: Further improvements to the viewport rendering

The idea is to schedule more samples at a lower resolution, which
gives more usable visual feedback. For example, The Junk Shop scene
becomes a bit less resolution but is way less noisy (almost no black
pixels) while steel feeling interactive.

Currently schedule 4 samples during navigation, 2 samples for the
resolution which is the previous to the final one. When denoising is
enabled during navigation, schedule single sample (which makes it so
higher resolution will be rendered). This is because both denoisers
we have behave better on more noisy image but on higher resolution.

Additionally, lower the update rate in viewport. Non-trivial scenes
can not keep up 60fps without dropping resolution too much. So prefer
less often updates but have more clean updates during navigation.

Fix wrong MIS weights for mixed BSDF and BSSRDF

Fix subsurface scattering not working with bump mapping in new integrator

Cycles: store ray differentials as scale instead of vector

For bump mapping at least this makes little visual difference, and reduces memory
usage and access cost on the GPU.

A further optimizations could also make the differentials in ShaderData smaller.

Ray differentials for NDC coordinates are disabled by this, to be brought back
later.

Fix missing object light visibility flags in new integrator

Do it as part of BSDF evaluation now, with advantage that we can skip evaluating
some BSDFs this way.

Fix wrong MIS weights leading to fireflies

Must keep track of distance to where the BSDF was sampled.

Cycles: restore ray differentials in new integrator

Accept the extra memory usage for now, we'll try to compress these later.
This fixes bump mapping for more accurate comparisons.

Cycles: Fix transmission pass leaking into Combined

Was happening when there are transmission objects in the scene and the
transmission passes are not enabled. The pass offset in the KernelFilm
were left at 0, making it so `kernel_accum_light` believed the pass is
to be written, and the writing was happening into the combined pass.

Maybe we should initialize all passes offsets to PASS_UNUSED, so that
on an incorrect usage we'll have a memory error instead of silently
corrupting the combined pass?

This change fixes the startup file of myself, which is a glass monkey
on a plane.

Cycles: Fix missing updates with adaptive sampling

Was a mistake in the calculations, which was delaying the update for
too long and did not guarantee and update at the end of render.

Cycles: Partial fix for the viewport passes

Blender's addon should use same constants as the PassType.

One thing which is still not working in viewport is Emission and
Background passes when there is an object with Glass shader: the
object is still shaded. The final F12 render of the Emission and
background is correct. So not immediately clear what's the issue.

Cycles: Fix threading issues in the PathTrace render cancel

The actual flow is a bit of a mystery to me, but there were cases
when `PathTrace::cancel()` was waiting for the `PathTrace::render()`
to finish, but the `render_cancel_.is_requested` was cleared so that
the render_pipeline was not quitting.

The new code seems more logical to me, but really wouldn't mind a
second pair of eyes here.

Cycles: Fix wrong OpenGL context manipulation in some cases

When the BlenderSession is destructed from an initialization of a
temporary are the current active context is window, not the draw
manager.

Made it so context release/activation goes correct in this case.

Steps to reproduce an original issue:
- Compile in debug mode, so that asserts are active.
- Start Cycles viewport render.
- Hit shortcut to open file (Cmd-O / Ctrl-O).

Cycles: Fix samples scheduling on retina screens