This was very simple to reproduce, just turn on Freestyle and press render.
Now to the truth of things. Most (if not all) of
~BlenderStrokeRenderer() can be removed. I believe this was done back
when freestyle was using G.main, and since we gave freestyle its own
main we can just leave the cleanup for later.
I will leave this for freestyle maintainers to think over though.
Note: There is a chance this was the issue reported on T57890. I will
wait for the reporter to confirm this as fixed though.
This commit adds another optional check (when `--debug-io` is set) on
write .blend process, to check and ensure all shape keys have their
'from' pointer properly set to their respective user ID.
This is intended to be used as debuging tool mostly (to try to detect
when/why some of those pointers can become NULL).
For now, it also systematically perform same checks/fixes when loading a
.blend file, to fix all broken ones laying around. Later we might move
that usage to a do_version instead, but for now think it's safer to
always perfom it (and it's rather cheap process anyway).
Does not make sense to keep that with BLO_writefile.h, this can also be
used by read code, and some other parts of Blender (like ed_undo.c
currently)...
- Compute samples positions on CPU.
- Use 3x3 Box blur instead of 2x2.
- Implement bokeh parameters.
With this commit, dof performance is almost negligeable.
The quality is a bit lower than before but can be improve. Also now big
Circle of confusion are supported (up to 200px).
Cost is ~1.25ms on AMD Vega with a 2560p viewport than full HD and
pretty shallow depth of field.
Coc downsampling and dilation is not used anymore for now (commented).
The algorithm used is borrowed from :
http://tuxedolabs.blogspot.com/2018/05/bokeh-depth-of-field-in-single-pass.html
This makes it possible to have a decent blur for foreground over defocused
background in one pass only.
The algorithm is using a gather approach that is much faster
than the scatter approach used in Eevee. This makes it possible to have
custom bokeh shapes (not implemented yet) which would be impossible with
a separable gaussian technique.
The blur is done in 2 steps. The first one define the shape of the bokeh
and the second that fill the undersampling.
A downsample max-CoC tile texture speed up the gathering process.
Was generating INVALID_FRAMEBUFFER here instead of failled texture alloc.
Add safety asserts in gpu_texture.c and clamp minimum size to 1 inside
GPU_offscreen_create.
Artists requested to show the stroke while drawing a new stroke using a material with fill color only, because it's very difficult to see the stroke.
Now the stroke shows always but using the fill color, not the stroke color because maybe is not set.
The issue was caused by the hair step checking whether
particle system needs to have path cache. This was done
in a way which was traversing an entire scene and was
checking every object for particle instance modifier.
Ideally, path cache should be an own operation in the
dependency graph. Or at least, this flag should be set
by dependency graph builder, similar to curve's path.
Since the code was broken already (it was only checking
first particle instance modifier), it is easier to
remove the buggy code, solve the crash and move on for
now.
If this causes an issue, simply set particle system to
be rendered as path.
Fixes crash with playback of Spring scenes.
This reverts commit 1d908bffdd.
Enough uses of repeat last expect skip-save properties to be set,
transform being the most obvious example T60777#605681.
I wanted to avoid operators having account for two kinds of 'skip-save'
but this may be unavoidable.
Originally I wanted to avoid adding draw manager specific ifdef's all
over generic shaders however this isn't needed in so many places.
Also there are shaders that are only used by the draw manager so
duplicating them only to have the original unused doesn't make sense.
