Modifiers can use the same bake directory now. If multiple modifiers
using conflicting paths are baked at the same time there is now an
error popup.
Modifiers can also bake to directories that already contain data
(including their own). To give users a heads-up there is now a
confirmation popup in that case. It only comes up once per bake operator
invoke to strike a balance between silently overwriting data and not
being too obnoxious when users want to rebake the same simulation many
times.
Pull Request: https://projects.blender.org/blender/blender/pulls/108288
Fixed an issue where meshes without loose elements or hidden triangle
faces were not taking into account the relevant masks.
From the user's perspective, this might not have seemed like a
significant problem. However, hidden faces in edit mode were
occasionally still snappable.
Duplicating will otherwise keep the same bake directory, which causes
a path conflict and requires manual fixing by users. Clearing the path
on copy makes it so the default path is generated on baking.
This does not change the path when doing internal copies, like
- Depsgraph "evaluated" object vs "original"
- Appending or linking objects
- Making a linked object local
Pull Request: https://projects.blender.org/blender/blender/pulls/109014
Better to disable than crashing, as we are not expecting a quick fix. The cause
is likely similar to issues with the light tree, which was already disabled.
Ref #104013
Keep the PTX compute capability at the same level as the latest
architecture-optimized CUDA kernel.
Should help performance of the future cards when running older
Blender, and maybe will allow to perform JIT optimization faster.
Pull Request: https://projects.blender.org/blender/blender/pulls/109007
The counters for polygons count and face corners count were inverted,
leading to various out of bounds accesses due to wrong data size
allocation.
Not sure where the bug comes from, whether it's original, or a typo from
some refactor.
There should be no functional changes for the typical usecase,
but it allows to have more tricky setups like pointing to a
BAT script to override some configuration.
The issue is that BAT scripts do not support new lines in the
command line arguments. That's where single-line python expression
helps.
For example, it is possible to point benchmark script to a blender.bat
which contains
blender.exe --python-expr "import bpy; bpy.context.preferences.addons['cycles'].preferences.use_oneapirt = False" %*
to have side-by-side numbers of oneAPI with and without HW RT.
Without this change the %* is which did not work: the BAT
script did not "see" part of the command line past the new line.
Pull Request: https://projects.blender.org/blender/blender/pulls/109006
The Alembic procedural was refrenced here even if USD was loaded.
Now check the CacheFile type and let the user know Render Procedurals
are only supported for Alembic atm.
Pull Request: https://projects.blender.org/blender/blender/pulls/108701
Since we scale the GPU matrix a certain way (in order to draw a round
circle even for different UI x/y scaling), we have to also compensate
with GPU matrix translation (instead of drawing the raw
frame/value as coords untranslated).
There might be ways to solve this differently (e.g. manipulating the passed
coords instead of the GPU matrix or having a version of imm_drawcircball
that can do this in a way that compensates UI view scaling), but this seems
to be the most straightforward.
Thx @mano-wii hinting at the way to do this with GPU matrix alone.
Should probably go into 3.3 LTS as well.
Pull Request: https://projects.blender.org/blender/blender/pulls/108979
As Alembic stores hierarchies as paths separated by a forward slash, such
character cannot be used in a name. This resulted in an uncaught thrown
exception. To fix this, replace '/' with '_' like for other illegal
characters.
Somehow the implementation for the main function to load point clouds
data was missing although everything else to support point clouds was
there. Compilers were more than happy to convert the IPointsSchema to
another schema type for the compilation to succeed, and the crash
occurs because the points schema does not contain the same data as the
compiler's chosen schema (in this case an ICurvesSchema).
Another crash was found due to the radius array not being properly
initialized and left with a size of 0, when Cycles expects a full array.
The Alembic data streaming can optionally interpolate between vertex of
two adjacent frames in order to smooth out the transition between
frames.
However, the decision to interpolate is only based on the vertex count.
This is not too robust as topology/connectivity can still differ even if
the number of vertices is the same (for example physics simulations and
videogrammetry can be set to output the same vertex count, but optimize
the triangle placement). This lead to vertices of unrelated polygons
being interpolated across frames.
To fix this, we now also check if the connectivity across frames is the
same, instead of just checking the topology counters. Although the bug
is revealed by the vertex interpolation routine, a similar fix is applied
to the check on topology change used to decide if the modifier has to be
evaluated for orco evaluation.
Pull Request: #105867
This fixes#108487 by applying upstream OIIO commit 8da473e254
This commit fixes the deps builder to include the patch, the actual
bug will not be fixed until the platform maintainers update the OIIO
library in SVN.
Happens with systems which do not provide GOLD linker: the linking state
would failing with some missing symbols and print about missing libsycl.so.6.
Seems that BFD linker expects to resolve all symbols, even the indirectly
used ones. This is somewhat counter-intuitive and is not how LLD, GOLD,
or MOLD worls.
The current state of the CMakeLists.txt does request the cycles_bvh to be
linked against SYCL_LIBRARIES. However, the SYCL was only requested to be found
if WITH_CYCLES_DEVICE_ONEAP is true.
Arguably the SYCL_LIBRARIES should only be linked-in into cycles_bvh if
EMBREE_STATIC_LIB, but that does not solve the issue with BFD.
This change makes it so the SYCL is requested to be found if the oneAPI
device is enabled, or if the Embree is detected to require/use SYCL
support.
Pull Request: https://projects.blender.org/blender/blender/pulls/108965
The scene contains some interesting names, which requires to be
written as utf-8. And on Windows file descriptor is not guaranteed
to be using utf-8. Or, will error out if the invalid utf-8 sequence
is written.
This change makes it so running benchmarks on windows it fully successful.
Pull Request: https://projects.blender.org/blender/blender/pulls/108982
Caused by cad897de16
The crash happens because the `SnapData_EditMesh` is removed from the
map but still accessed afterwards.
However, after cad897de16, invalidating an object's unique snap date is
no longer required.
Sometimes the modifier has never been evaluated by an active depsgraph
before it is used for rendering. In this case, the baked data was never
loaded. Now also allow checking for baked data in a non-active depsgraph.
The locking that is in-place already should probably already be enough
to make this thread-safe but maybe that could be made more explicit.
With this the simulation cache pointer is copied over to the evaluated modifier.
This allows the original modifier to be removed without breaking the evaluated
modifier, which results in better decoupling. This can avoid issues when a
non-active depsgraph is evaluated in the background while the user is manipulating
the scene.
Also, it is now assumed that the simulation cache is always allocated even if
there is no simulation (similar to run-time data). This simplifies the code.
Pull Request: https://projects.blender.org/blender/blender/pulls/108976
Windows does not really have an idea of shebangs, and it needs to
go via a file extension to see that the script is to be executed
by Python.
This change simplifies execution from `python3 benchmark ...`
to `benchmark.py ...`.
Pull Request: https://projects.blender.org/blender/blender/pulls/108971
The face normal was used in local space whereas it is needed as a global
direction (since other vectors here are in worldspace as well --
knife_input_ray_segment() returns worldspace).
Could go into 3.3 LTS.
Pull Request: https://projects.blender.org/blender/blender/pulls/108938
The problem here was that when direct light contibutions to baking were
disabled, the kernel just skipped all direct lighting evaluation.
However, at secondary bounces, "direct light" would actually end up
being indirect (since there's an extra bounce along the way), but
we're still skipping it.
Therefore, only apply direct lighting skipping at the first bounce.
e5ec04d73c changed the way crease vakyes are stored in 4.0.
Add versioning for reading the new files that replaces the new generic
attributes with the old non-generic custom data layers. The code is
paranoid with lots fo checks I expect will typically not be necessary.
Similar to f75af7cbf5
Interpolation from edge attributes is unsupported, and the data
of the new point attribute was uninitialized. As a fix, just avoid
interpolating edge attributes in the first place.
PyUnicode_AsUTF8AndSize is used when renaming a custom python property,
this method stores the size of the string without including the null
terminator in the size.
Renaming a custom python property now includes the null terminator when
copying the new string name.
Pull Request: https://projects.blender.org/blender/blender/pulls/107983
This fixes an issue where the light tree sampling algorithm would
discard light samples from groups of distance lights with an angle
greater than 0 when it shouldn't.
Pull Request: https://projects.blender.org/blender/blender/pulls/108832
The split edges code had a complex method of merging duplicate edges,
going backwards to avoid shifting elements in a vector. Sometimes it
could result in incorrect corner edge indices though, if it moved an
index that matched one of the local variables (I think! I've bee
trying to understand this all day and still struggling). Instead,
replace it with a `VectorSet` that handles the deduplication by
itself, and avoid creating the new edges until the end.
I think this code could still be simpler if we tried to reduce the
amount of things happening at the same time, making more code
deal with the input or final state rather than an in-between one.
But to avoid making the change too complicated I stopped here.
Pull Request: https://projects.blender.org/blender/blender/pulls/108826
So far, each closure in Cycles was either diffuse OR glossy OR
transmissive, and its color and contributions were assigned
to the corresponding direct/indirect/color passes.
However, since Glass is a single closure now, that is no longer enough,
since glass has both a glossy and a transmissive component.
Therefore, this commit adds support for splitting contributions from
the Glass closure between the two types.
For 4.0, we might want to also use this for Principled Hair since it
also technically has both types, but that would be a change from
the existing result so it's not part of 3.6 yet.
Before 4369627e71, the Mirror modifier merge did not interpolate
customdata.
Previously, the customdata was simply copied from the original vertex
during the merge, resulting in the retention of flags such as
`MVERT_SKIN_ROOT`.
This fix could just repeat that behavior for mirror. But these flags
can be useful in other cases, so this commit re-adds them to the
resulting merged vertex.
Although this can generate geometries with multiple roots on the same
island, there are different scenarios in which this can happen.
Therefore, multiple roots on the same island is not necessarily
considered a bug.
idiff sometimes locks up while shutting down when the CPU is
oversubscribed. While blender does not rely on the idiff tool
the tests that run on the CI environment do, which causes tests
to occasionally fail due to a timeout.
The root cause is a bit complex but can be found on the oiio tracker
at https://github.com/OpenImageIO/oiio/issues/3851
This change fixes idiff by :
1- Shutting down the thread pool before the main function exits
2- Have the shutdown wait for the pool threads to actually join, to
prevent the OS from forcefully terminating them while they could
potentially still be holding a lock.
Socket shapes are used in geometry nodes to indicate field types.
In other tree types (e.g. shaders) the socket shape is not touched at
all. Since nodes can be copied from geometry node trees, we need to
reset the socket shape when copying to a shader, compositor, or texture
tree.
Pull Request: https://projects.blender.org/blender/blender/pulls/108412
