Blender Git Loki

Cycles: make megakernel execution work again

After addition of more kernels the logic didn't work well anymore, for example
shade_light was holding up all paths going to the intersect_closest state.

Rather than trying to get all paths to the intersect_closest state, make the
megakernel support paths in different states.

This seems to help viewport rendering by about 10% here. I also tweaked the
threshold to what gave me best performance, but hard to say in general what this
value should be.

I noticed get_num_active_paths() used for adding more work tiles is not correct
since it's double counting some paths. Not fixing it as part of this patch, only
adding a comment about it.

Cycles: Tweaks to the resolution divider

- Base initial resolution divider on a pixel size.

Using hardcoded low resolution provides rather annoying blobby
viewport on first enter of rendered viewport. Set the divider in
this case to 8 of the native pixel sizes. A bit of heuristic, which
seems to give more appealing pixel size, and kind of matches what
is possible to setup as fastest preview with the pixel size setting.

- Fix miscalculation of how long viewport number of samples will take.
Need to properly take number of samples during update into account.
Seems to make Junk Shop more interactive on home PC. Although, it
was already feeling much better than master.

- Always render 4 samples for highest resolution divider.
Before it was clamped to the number of samples. There is an
explanation in the code but the idea is that this is not slower and
does bring extra information during navigation.

Fix fireflies in barbershop due to self intersection

With the new handling of lights as regular intersections, float precision
could lead to self-intersection, causing the light to be counted many times
over and leading to fireflies.

This could be generalized to work for all primitives instead of just lights,
but that's more involved and could be done later. For now this is a simple
isolated change.

Cleanup: remove line of code that has been replaced

Cycles: Suppress adaptive sampling logs when adaptive sampling is not used

This way nobody gets confused why there are some calculations about adaptive
sampling when adaptive sampling is not used.

Cleanup: remove some unnecessary includes from kernel API

Only needed for implementation.

Cycles: Fix compilation error after merge

There is no more kernel_globals.h, those are in the kernel/device now.
The state of the file did not include this header, so assuming it is
safe to remove: it works on my macOS, but the PC is currently busy
compiling master's CUDA kernels, so can't easily verify GPU.

Merge branch 'master' into cycles-x

Cycles: Correction to previous commit

Committing directly because those are rather trivial fixes for something
what got broken rather badly.

- Fix compilation error in debug mode.
- Fix non-combined passes in viewport.
- Fix adaptive sampling for final rendering.

Cycles: Move use adaptive sampling flag to integrator

Makes it closer to the rest of the adaptive sampling settings and
allows access from Scene.

The current downside is that rendering does need an extra integrator
synchronize call which will be removed by a followup development.

There should be no functional changes.

Cycles: Add idea of automatically created passes

Those are the opposite of the ones which are configured by users in the
interface.

Currently unused, preparing for taking over of automatically creating
adaptive sampling passes.

Cycles: Use early output and avoid over-commenting

Cycles: Shorten function name

It is inside of viewport parameters, no need to repeat this in the
function name.

Cycles: Remove passes from BufferParams

Now it is the Scene which is the source of truth for passes.

The RenderBuffers will store pre-calculated offsets and strides
which are set once when the passes are known. There are few things
which are good:

- Data is no longer duplicated, is clear where passes are to be
accessed from.
- Code which sues BufferParams can access offsets and strides
directly, without re-calculating anything.
- DenoiserBuferParams are removed, since it does not bring any
new functionality.

Some downsides:

- Construction of new RenderBuffers is to be done with a bit more
care now.

- Passes needs to be passed to `get_pass_rect`.
A bit annoying, and almost like the RenderBuffers is the wrong place
for this function.

- `set_pass_rect` is disabled for now. Is unclear whether it will be
needed in the future, and what would it final placement be.

Cycles: Remove baking tile reading code

The baking needs to be brought back in a better shape.

Removing the code which is to be re-written to simplify some upcoming
development.

Cycles: Automatically create adaptive sampling and viewport passes

Fixes non-working adaptive sampling in the viewport: the missing passes
was the only reason for that.

Now the synchronization code is supposed to store passes in Scene and
never provide them via BufferParams.

The passes are assigned to the BufferParams when dopin session reset.
This way the rendering code has access to proper stride information.
In the future we can remove passes from BufferParams and pre-compute
strides and offsets before passing the BufferParams down the road.

Had to explicitly remove the passes: can not rely on the fact that the
synchronization code overrides passes in scene as this is not the case
for viewport rendering.

Cycles: Simplify delayed reset API

Avoids passing arguments around.

Currently no functional changes, just simplifies some things in the
future in the terms of naming.

Cycles: Use range based loop

Cycles: Clarity comment in integrator synchronization

Fix multi-device (for denoising) crash with background shader evaluation

And prepare code a bit support multi-devices in the future, by filling the
buffers through a callback per device.