**Changes**
As described in T93602, this patch removes all use of the `MVert`
struct, replacing it with a generic named attribute with the name
`"position"`, consistent with other geometry types.
Variable names have been changed from `verts` to `positions`, to align
with the attribute name and the more generic design (positions are not
vertices, they are just an attribute stored on the point domain).
This change is made possible by previous commits that moved all other
data out of `MVert` to runtime data or other generic attributes. What
remains is mostly a simple type change. Though, the type still shows up
859 times, so the patch is quite large.
One compromise is that now `CD_MASK_BAREMESH` now contains
`CD_PROP_FLOAT3`. With the general move towards generic attributes
over custom data types, we are removing use of these type masks anyway.
**Benefits**
The most obvious benefit is reduced memory usage and the benefits
that brings in memory-bound situations. `float3` is only 3 bytes, in
comparison to `MVert` which was 4. When there are millions of vertices
this starts to matter more.
The other benefits come from using a more generic type. Instead of
writing algorithms specifically for `MVert`, code can just use arrays
of vectors. This will allow eliminating many temporary arrays or
wrappers used to extract positions.
Many possible improvements aren't implemented in this patch, though
I did switch simplify or remove the process of creating temporary
position arrays in a few places.
The design clarity that "positions are just another attribute" brings
allows removing explicit copying of vertices in some procedural
operations-- they are just processed like most other attributes.
**Performance**
This touches so many areas that it's hard to benchmark exhaustively,
but I observed some areas as examples.
* The mesh line node with 4 million count was 1.5x (8ms to 12ms) faster.
* The Spring splash screen went from ~4.3 to ~4.5 fps.
* The subdivision surface modifier/node was slightly faster
RNA access through Python may be slightly slower, since now we need
a name lookup instead of just a custom data type lookup for each index.
**Future Improvements**
* Remove uses of "vert_coords" functions:
* `BKE_mesh_vert_coords_alloc`
* `BKE_mesh_vert_coords_get`
* `BKE_mesh_vert_coords_apply{_with_mat4}`
* Remove more hidden copying of positions
* General simplification now possible in many areas
* Convert more code to C++ to use `float3` instead of `float[3]`
* Currently `reinterpret_cast` is used for those C-API functions
Differential Revision: https://developer.blender.org/D15982
Since internal links are only runtime data, we have the flexibility to
allocating every link individually. Instead we can store links directly
in the node runtime vector. This allows avoiding many small allocations
when copying and changing node trees.
In the future we could use a smaller type like a pair of sockets
instead of `bNodeLink` to save memory.
Differential Revision: https://developer.blender.org/D16960
The bug is caused by rBb66b3f547c43e841a7d5da0ecb2c911628339f56.
From what I can see, that fix was intended to enable manual lens shift for
panorama cameras, but it appears that it also unintentionally applies
interocular shift.
This fix disables the multiview shift for panorama cameras, that way manual lens
shift still works but we get the 2.x behavior for stereoscopic renders back.
Differential Revision: https://developer.blender.org/D16950
This moves all multi-function related code in the `functions` module
into a new `multi_function` namespace. This is similar to how there
is a `lazy_function` namespace.
The main benefit of this is that many types names that were prefixed
with `MF` (for "multi function") can be simplified.
There is also a common shorthand for the `multi_function` namespace: `mf`.
This is also similar to lazy-functions where the shortened namespace
is called `lf`.
* New `build_mf` namespace for the multi-function builders.
* The type name of the created multi-functions is now "private",
i.e. the caller has to use `auto`. This has the benefit that the
implementation can change more freely without affecting
the caller.
* `CustomMF` does not use `std::function` internally anymore.
This reduces some overhead during code generation and at
run-time.
* `CustomMF` now supports single-mutable parameters.
This refactors how devirtualization is done in general and how
multi-functions use it.
* The old `Devirtualizer` class has been removed in favor of a simpler
solution. It is also more general in the sense that it is not coupled
with `IndexMask` and `VArray`. Instead there is a function that has
inputs which control how different types are devirtualized. The
new implementation is currently less general with regard to the number
of parameters it supports. This can be changed in the future, but
does not seem necessary now and would make the code less obvious.
* Devirtualizers for different types are now defined in their respective
headers.
* The multi-function builder works with the `GVArray` stored in `MFParams`
directly now, instead of first converting it to a `VArray<T>`. This reduces
some constant overhead, which makes the multi-function slightly
faster. This is only noticable when very few elements are processed though.
No functional changes or performance regressions are expected.
Profiling shows that this computation consumes a noticeable amount
of time, so it is worth moving it to an earlier part of the code
that is executed less frequently and is multithreaded.
Differential Revision: https://developer.blender.org/D16933
Both cloth object collision and self collision use a BVH tree
representing the current cloth shape. The only difference between
them is the epsilon threshold value.
If these values are the same, it is possible to use the same tree
for both uses, thus easily reducing the overhead in the case when
both collision modes are used by the same cloth object.
Differential Revision: https://developer.blender.org/D16914
Previously you had to use a workaround with the Object info node to get
evaluated data from the new curves object to an original editable mesh.
This commit makes it so that a curves object can be converted directly
to a mesh object if it has evaluated curves or an evaluated mesh.
Differential Revision: https://developer.blender.org/D16930
Socket locations are set while drawing the node tree in the editor.
They can always be recalculated this way based on the node position and
other factors. Storing them in the socket is misleading. Plus, ideally
sockets would be quite small to store, this helps us move in that
direction.
Now the socket locations are stored as runtime data of the node editor,
making use of the new node topology cache's `index_in_tree` function
to make a SoA layout possible.
Differential Revision: https://developer.blender.org/D15874
Improve animation playback performance in EEVEE for materials using Mix
nodes. Socket availability was being set and reset on every evaluation
of Mix nodes, during animation playback, this was causing the graph to
be marked dirty, and the whole graph being re-evaluated on every frame,
causing performance issues during playback.
Additionally, do a bit of cleanup, traversing the node sockets with
the next link to improve clarity and reduce errors. Also refactoring
`nodeSetSocketAvailability` to early out and increase clarity on no-op.
Differential Revision: https://developer.blender.org/D16929
The reported backtrace in T102766 strongly points at some
concurrency issues within exisitng liboverride diffing code that
restores forbidden changes to reference linked data values.
This commit instead add tags to mark liboverrides/properties that need
to be restored, and do so in a separate new step of diffing, from the
main thread only.
- Add parentheses to suppress an assertion on some compilers.
- It turns out cloth self-collision math is not precisely symmetric,
so sort the triangle index pair to restore behavior exactly identical
to the version before the new traversal.
Follow up to rB0796210c8df32
Previously cloth self-collision and other code that wants self-overlap
data was using an ordinary BVH overlap utility function. This works,
but is suboptimal because it fails to take advantage of the fact that
it is comparing two identical trees.
The standard traversal for self-collision essentially devolves into
enumerating all elements of the tree, and then matching them to the
tree starting at the root. However, the tree itself naturally partitions
the space, so overlapping elements are likely to be mostly placed nearby
in the tree structure as well.
Instead, self-overlap can be computed by recursively computing ordinary
overlap between siblings within each subtree. This only considers each
pair of leaves once, and provides significantly better locality.
It also avoids outputting every overlap pair twice in different order.
Differential Revision: https://developer.blender.org/D16915
Previously, the lifetimes of anonymous attributes were determined by
reference counts which were non-deterministic when multiple threads
are used. Now the lifetimes of anonymous attributes are handled
more explicitly and deterministically. This is a prerequisite for any kind
of caching, because caching the output of nodes that do things
non-deterministically and have "invisible inputs" (reference counts)
doesn't really work.
For more details for how deterministic lifetimes are achieved, see D16858.
No functional changes are expected. Small performance changes are expected
as well (within few percent, anything larger regressions should be reported as
bugs).
Differential Revision: https://developer.blender.org/D16858
It is a part of old-standing TODO, and code which accesses
the value was commented out for many years.
Remove the field to help with an upcoming const-correctness
improvements.
Reading or writing a vertex group is expensive enough that it's worth
parallelizing. On a Ryzen 3700x, in a grid of 250k vertices with
30 randomly assigned vertex groups (each to 10-50% of vertices),
I observed a 4x improvement for writing to a group and a 3x
improvement when reading their data. This significantly speeds
up nodes that create a new mesh from a mesh that had vertex groups.
Since 78f28b55d3, allocating on multiple threads is much
faster, making it a nice improvement to parallelize vertex group
operations. This patch adds multi-threading when removing a
vertex group from the "Remove Named Attribute" node.
On a Ryzen 3700x:
Before: `(Average: 15.6 ms, Min: 15.0 ms)`
After: `(Average: 8.1 ms, Min: 7.6 ms)`
Differential Revision: https://developer.blender.org/D16916
If we change the radius of a point or spot lamp, we also change the area lamp size.
As shown in T102853, this is bad for animating the lamp type.
The solution is to make the property point to another member of the DNA
struct `Light`.
Differential Revision: https://developer.blender.org/D16669
Use the same `".selection"` attribute for both curve and point domains,
instead of a different name for each. The attribute can now have
either boolean or float type. Some tools create boolean selections.
Other tools create float selections. Some tools "upgrade" the attribute
from boolean to float.
Edit mode tools that create selections from scratch can create boolean
selections, but edit mode should generally be able to handle both
selection types. Sculpt mode should be able to read boolean selections,
but can also and write float values between zero and one.
Theoretically we could just always use floats to store selections,
but the type-agnosticism doesn't cost too much complexity given the
existing APIs for dealing with it, and being able to use booleans is
clearer in edit mode, and may allow future optimizations like more
efficient ways to store boolean attributes.
The attribute API is usually used directly for accessing the selection
attribute. We rely on implicit type conversion and domain interpolation
to simplify the rest of the code.
Differential Revision: https://developer.blender.org/D16057
Partly a cleanup, but also iterating over spans can be faster than
linked lists. Also rewrite the multi-input socket link counting
to avoid the need for a temporary map. Overall, on my setup the changes
save about 5% (3ms) when drawing a large node tree (the mouse house file).
Expose `BKE_pose_apply_action_blend` and a simplified pose backup system
to RNA. This will make it possible to easily create some interactive
tools in Python for pose blending.
When creating a backup via this API, it is stored on the
`Object::runtime` struct. Any backup that was there before is freed
first. This way the Python code doesn't need access to the actual
`PoseBackup *`, simplifying memory management.
The limitation of having only a single backup shouldn't be too
problematic, as it is meant for things like interactive manipulation of
the current pose. Typical use looks like:
- Interactive operator starts, and creates a backup of the current pose.
- While the operator is running:
- The pose backup is restored, so that the next steps always use the
same reference pose.
- Depending on user input, determine a blend factor.
- Blend some pose from the pose library into the current pose.
- On confirmation, leave the pose as-is.
- On cancellation, restore the backup.
- Free the backup.
`BKE_pose_apply_action_blend` is exposed to RNA to make the above
possible.
An alternative approach would be to rely on the operator redo system.
However, since for poses this would use the global undo, it can get
prohibitively slow. This change is to make it easier to prototype
things; further into the future the undo system for poses should be
improved, but that's an entire project on its own.
Reviewed By: sergey
Differential Revision: https://developer.blender.org/D16900
Move `pose_backup.cc` from `editors/armature` to `blenkernel`. This is
in preparation of an upcoming change where the pose backup is going to be
owned by the `Object`. This will need to be automatically cleared when the
object is freed, which means that `blenkernel` needs the corresponding
logic.
Technically only the freeing code could be moved, but I felt it made more
sense to keep the related code together.
No functional changes.
- Move from blenkernel to the node editor, the only place it was used
- Use two vectors instead of ListBase
- Remove define for validating the clipboard, which shouldn't be skipped
- Comment formatting, other small cleanups to whitespace
Differential Revision: https://developer.blender.org/D16880
When these declarations are built without the help of the special
builder class, it's much more convenient to set them directly rather
than with a constructor, etc. In most other situations the declarations
should be const anyway, so theoretically this doesn't affect safety too
much. Most construction of declarations should still use the builder.
Whenever a node group is entered during evaluation, a new compute
context is entered which has a corresponding hash. When node groups
are entered and exited a lot, this can have some overhead. In my test
file with ~100.000 node group invocations, this patch improves performance
by about 7%.
The speedup is achieved in two ways:
* Avoid computing the same hash twice by caching it.
* Invoke the hashing algorithm (md5 currently) only once instead of twice.
