In 2.81 there was a change to increase the performance of viewport
animation rendering. This change would perform the color management on the
GPU if the only 8bit was needed. This saved CPU cycles and data
transfer.
The issue is that in the image editor or when saving the image the CM
will be reapplied. Although the speed is desired, exporting the actual
colors has more priority.
In the ticket there is an analysis that shows that shows that this fix
is the correct short term step to take. It would be better that the
render result is aware of the color space of its buffers so the applying
color management could be skipped when saving to disk or drawing in the
image editor.
The issue with this change is the performance penalty it has.
Reviewed By: Brecht van Lommel
Maniphest Tasks: T79999
Differential Revision: https://developer.blender.org/D10371
Dividing the workload by number of tasks in float is imprecise and
lead in some cases to particles not being calculated at all
(example: 20000 particles, 144 tasks).
Switching this calculation to integer makes sure we don't lose count.
Differential Revision: https://developer.blender.org/D10157
The root of the issue was caused by the dependency graph using ID pointer
to map evaluated state from old depsgraph to new one upon relations update.
This was failing when IDs were re-allocated rapidly: was possible that
Object ID's evaluated state assigned to Mesh and vice versa.
Now depsgraph uses Session UUID to identify which IDs to restore evaluated
state to. The session UUID is stored in the IDNode, so that id_orig is not
dereferenced on depsgraph update since the ID might be freed.
The root of the issue is identified by Campbell, original patch was done
by Bastien, thanks! Also thanks to Oliver and Ray and everyone else for
testing!
Issue appears to be caused by AMD graphics driver later than 20.11.1 and
affects older GPUs (Polaris/FIJI cards). Wireframe drawing uses the same
OpenGL data type for storing normals what is known to be faulty.
This patch enabled storing the normals using GPU_COMP_I16. It also
solves the normals drawing in edit mode for vertex and loop normals.
This is part two of the fix for T84459.
Issue appears to be caused by AMD graphics driver later than 20.11.1 and
affects older GPUs (Polaris/FIJI cards). Drawing normals in edit mode
uses the same OpenGL data type for storing normals that is known to be
faulty.
This change fixes the face dot normals by using GPU_COMP_I16.
THe high quality normals work around is enabled for Polaris cards using
the official drivers. Since driver version 2.11.2 they fail to render
using low quality normals.
The detection of polaris cards is done by matching the opengl renderer.
The renderer strings have been extracted from various reports linked to
{T82856} but isn't complete as some reports are missing the exact
renderer as users don't always report via the help menu.
This change makes it possible for platforms to only support high quality
normal rendering. This is part of {T82856} where current AMD drivers
running on the polaris architecture does not support the low quality
setting due to a driver bug.
In a next commit the work around will be enabled.
This adds high quality normals for non meshes. These include
* Volumetric Object Wireframe
* Metaballs
* Extracted Curves
* Curves in edit mode
This is in preparation to fix a regression in recent AMD
drivers where the `GL_INT_2_10_10_10_REV` data type isn't
working in Polaris cards.
Some GPU platforms don't support having more than one underscore in
sequence in an attribute name. This change will remove the underscore
as a possible character when encoding to save names.
Root of the issue was missing management of ID pointers in the cloth
modifier data stored in ParticleSystem for hair physics, in the
'foreach_id' particle system code.
Using modifier's 'foreach_id' code in psys one unfortunately requires
some ugly conversion gymnastics, but this is still better than having
dedicated code for that case.
Note that this is actually a fairly critical issue, fix should be
backported to 2.91.1 should we do it, and to 2.83 LTS as well I think.
In `calc_curve_deform` a factor is calculated without checking if
the divisior is zero or close to zero. This patch adds the missing
checks and sets the factor to zero if the division shouldn't be
computed.
Reviewed By: mont29
Differential Revision: https://developer.blender.org/D9645
Looks like prior to the introduction of mantaflow, the former
SmokeModifierData always had a domain initialized (even if its type was
set to None). Since mataflow, the FluidModifierData type needs to be set
to MOD_FLUID_TYPE_DOMAIN (otherwise domain is invalid)
Maniphest Tasks: T82729
Differential Revision: https://developer.blender.org/D9644
This was reported in the form of the eyedropper of the 'Parent' property
creating a custom property 'parent' if self was picked.
Problem arises when certain checks for setting rna pointer properties
failed (for example: the PROP_ID_SELF_CHECK check) and then a different
code path was entered (which was only meant for IDProperties).
Problem was introduced in rBa7b3047cefcb.
To solve, now first enter the branch for rna-based pointer properties,
then perform the sanity-checks (and if these fail: dont enter the other
unrelated codepath but instead do nothing)
Maniphest Tasks: T83055
Differential Revision: https://developer.blender.org/D9652
Endian switching when loading 64bit blend files on a 64bit system was
crashing as the endian switching is only applicable when loading
on 32 bit systems.
This crash goes back to 2.7x, it looks like this never worked
all the way back to the first commit.
Regular rendering uses a custom blend mode, but render passes renders to
2 separate textures. This wasn't configured correctly inside the
fragment shaders. This patch adds a switch to configure the fragment
shader with the correct attachments.
Backport to Blender 2.83.
Reviewed By: Clément Foucault
Differential Revision: https://developer.blender.org/D9038
Calling PyEval_ReleaseLock() was crashing with Python 3.9
because it accessed the NULL pointer set by PyThreadState_Swap().
This happened when calling ViewLayer.update() for example.
While the existing logic could be fixed by swapping the thread-state
back before calling PyEval_ReleaseLock(), this depends on functions
which are tagged to be removed by v4.0.
Replace use of deprecated functions by calling PyEval_SaveThread(),
instead of inlining the logic, using _PyThreadState_UncheckedGet()
to prevent Python aborting.
The call to PyEval_ThreadsInitialized has been removed
as threads are now initialized with Python.
This could be replaced with Py_IsInitialized() however it doesn't look
like this is necessary.
This is compatible with Python 3.7 & 3.9.
The logic of `BKE_sculpt_update_object_for_edit` was not correct. such
low-level functions should typically never preform depsgraph evaluation
themselves, they should be able to rely on getting a fully evaluated
depsgraph and just get needed data from there.
Supporting that required fixing other broken code higher in the
callstack, namely:
* `ED_object_sculptmode_enter_ex` was freeing evaluated data, for no
valid reason it would seem.
* `sculpt_undosys_step_decode` was ensuring an evaluated depsgraph
**before** calling `ED_object_mode_generic_exit`, which would
invalidate a lot of evaluated data.
Note that it is fairly difficult to track down all code paths leading to
`BKE_sculpt_update_object_for_edit`, so there may be still cases where
this gets called with improperly evaluated depsgraph.
Reviewed By: sergey
Maniphest Tasks: T81854
Differential Revision: https://developer.blender.org/D9270
