Goals:
* Better high level control over where devirtualization occurs. There is always
a trade-off between performance and compile-time/binary-size.
* Simplify using array devirtualization.
* Better performance for cases where devirtualization wasn't used before.
Many geometry nodes accept fields as inputs. Internally, that means that the
execution functions have to accept so called "virtual arrays" as inputs. Those
can be e.g. actual arrays, just single values, or lazily computed arrays.
Due to these different possible virtual arrays implementations, access to
individual elements is slower than it would be if everything was just a normal
array (access does through a virtual function call). For more complex execution
functions, this overhead does not matter, but for small functions (like a simple
addition) it very much does. The virtual function call also prevents the compiler
from doing some optimizations (e.g. loop unrolling and inserting simd instructions).
The solution is to "devirtualize" the virtual arrays for small functions where the
overhead is measurable. Essentially, the function is generated many times with
different array types as input. Then there is a run-time dispatch that calls the
best implementation. We have been doing devirtualization in e.g. math nodes
for a long time already. This patch just generalizes the concept and makes it
easier to control. It also makes it easier to investigate the different trade-offs
when it comes to devirtualization.
Nodes that we've optimized using devirtualization before didn't get a speedup.
However, a couple of nodes are using devirtualization now, that didn't before.
Those got a 2-4x speedup in common cases.
* Map Range
* Random Value
* Switch
* Combine XYZ
Differential Revision: https://developer.blender.org/D14628
Remembering the number of curves of every type makes it fast to know
whether processing specific to a single curve type has to be done.
This information was accessed in quite a few places, so this should be
an overall reduction in overhead for the new curves type.
The cache is computed eagerly, in other words every time after changing
the curve types. In order to reduce verbosity I added helper functions
for some common ways to set the types.
Differential Revision: https://developer.blender.org/D14732
This is mostly a cleanup to avoid hardcoding the eager calculation of
normals it isn't necessary, by reducing calls to `BKE_mesh_calc_normals`
and by removing calls to `BKE_mesh_normals_tag_dirty` when the mesh
is newly created and already has dirty normals anyway. This reduces
boilerplate code and makes the "dirty by default" state more clear.
Any regressions from this commit should be easy to fix, though the
lazy calculation is solid enough that none are expected.
If the `wpoly` vector was small, the `wpoly_new` pointer could point
to part of its inline buffer on the stack, which becomes invalid out of
that scope. Instead, store `wpoly_new` as a span, and assign it properly
from the moved vector.
If all islands had a size of zero, a division by zero would occur in
`GEO_uv_parametrizer_pack`, causing the UV coordinates to be set to
NaN. An alternative approach would be to skip packing islands with a
zero size, but If UV coordinates are for example outside the 0-1 range,
it's better if they get moved into that range.
Differential Revision: https://developer.blender.org/D14522
Add "for_write" on function names that retrieve mutable data arrays.
Though this makes function names longer, it's likely worth it because
it allows more easily using the const functions in a non-const context,
and reduces cases of mistakenly retrieving with edit access.
In the long term, this situation might change more if we implement
attributes storage that is accessible directly on `CurvesGeometry`
without duplicating the attribute API on geometry components,
which is currently the rough plan.
Differential Revision: https://developer.blender.org/D14562
Rename "size" variables and functions to use "num" instead,
based on T85728 (though this doesn't apply to simple C++
containers, it applies here). Rename "range" to "points" in
some functions, so be more specific.
Differential Revision: https://developer.blender.org/D14431
The result handle attributes for non-bezier types are zeroed.
By mistake though, the entire array was zeroed, not just the
area corresponding to that curves source.
Realizing and copying attributes of meshes, curves, and points are very
similar processes, but currently the logic is duplicated three times in
the realize instances code. This commit combines the implementation
for copying generic attributes and creating the result id attribute.
The functions for threaded copying and filling should ideally be in
some file elsewhere, since they're not just useful here. But it's not
clear where they would go yet.
Differential Revision: https://developer.blender.org/D14294
The realize instances code used "assign", but the attribute buffers on
the result aren't necessarily initialized. This doesn't make a difference
for trivial types like `int`, but it would with more complex types.
This commit replaces the temporary conversion to `CurveEval` with
use of the new curves data-block. The end result is that the
process looks more like the other components-- somewhere in between
meshes and point clouds in terms of complexity.
The final result is that the logic between meshes and curves is
very similar. There are a few different strategies to reduce
duplication here, so I'll investigate that separately.
There is some special behavior for the radius and handle position
attributes. I used the attribute API to store spans of these
attributes temporarily. Using access methods on `CurvesGeometry`
would be reasonable to, storing spans separately feels a bit more
predictable for now though.
There should be significant performance improvements in some cases,
I haven't tested that specifically though.
Differential Revision: https://developer.blender.org/D14247
The main improvement is a code simplification, because attributes don't
have to be transferred separately for each curve, and all attributes can
be handled generically. Performance improves significantly when the
output contains many curves. Basic testing with a 2 million curve output
shows an approximate 10x performance improvement.
This commit changes `CurveComponent` to store the new curve
type by adding conversions to and from `CurveEval` in most nodes.
This will temporarily make performance of curves in geometry nodes
much worse, but as functionality is implemented for the new type
and it is used in more places, performance will become better than
before.
We still use `CurveEval` for drawing curves, because the new `Curves`
data-block has no evaluated points yet. So the `Curve` ID is still
generated for rendering in the same way as before. It's also still
needed for drawing curve object edit mode overlays.
The old curve component isn't removed yet, because it is still used
to implement the conversions to and from `CurveEval`.
A few more attributes are added to make this possible:
- `nurbs_weight`: The weight for each control point on NURBS curves.
- `nurbs_order`: The order of the NURBS curve
- `knots_mode`: Necessary for conversion, not defined yet.
- `handle_type_{left/right}`: An 8 bit integer attribute.
Differential Revision: https://developer.blender.org/D14145
This patch reverses the dependency between `BLI_math_vec_types.hh` and
`BLI_math_vector.hh`. Now the higher level `blender::math` functions
depend on the header that defines the types they work with, rather than
the other way around.
The initial goal was to allow defining an `enable_if` in the types header
and using it in the math header. But I also think this operations to types
dependency is more natural anyway.
This required changing the includes some files used from the type
header to the math implementation header. I took that change a bit
further removing the C vector math header from the C++ header;
I think that helps to make the transition between the two systems
clearer.
Differential Revision: https://developer.blender.org/D14112
Use a shorter/simpler license convention, stops the header taking so
much space.
Follow the SPDX license specification: https://spdx.org/licenses
- C/C++/objc/objc++
- Python
- Shell Scripts
- CMake, GNUmakefile
While most of the source tree has been included
- `./extern/` was left out.
- `./intern/cycles` & `./intern/atomic` are also excluded because they
use different header conventions.
doc/license/SPDX-license-identifiers.txt has been added to list SPDX all
used identifiers.
See P2788 for the script that automated these edits.
Reviewed By: brecht, mont29, sergey
Ref D14069
This commit moves the weld modifier code to the geometry module
so that it can be used in the "Merge by Distance" geometry node
from ec1b0c2014. The "All" mode is exposed in the node
for now, though we could expose the "Connected" mode in the future.
The modifier itself is responsible for creating the selections from
the vertex group. The "All" mode takes an `IndexMask` for the
selection, and the "Connected" mode takes a boolean array,
since it actually iterates over all edges.
Some disabled code for a BVH mode has not been copied over,
it's still accessible through the patches and git history anyway,
and it made the port slightly simpler.
Differential Revision: https://developer.blender.org/D13907
This implements a merge by distance operation for point clouds.
Besides the geometry input, there are two others-- a selection
input to limit the operation to certain points, and the merge
distance. While it would be a reasonable feature, the distance
does not support a field currently, since that would make
the algorithm significantly more complex.
All attributes are merged to the merged points, with the values
mixed together. This same generic method is used for all attributes,
including `position`. The `id` attribute uses the value from the
first merged index for each point.
For the implementation, most of the effort goes into creating a
merge map to speed up attribute mixing. Some parts are inherently
single-threaded, like finding the final indices accounting for the
merged points. By far most of the time is spend balancing the
KD tree.
Mesh support will be added in the next commit.
Differential Revision: https://developer.blender.org/D13649
It's better to calculate the size of a spline before creating it, and this
should simplify refactoring to a data structure that stores all point
attribute contiguously (see T94193). The mesh to curve conversion is
simplified slightly now, it creates the curve output after gathering all
of the result vertex indices. This should be more efficient too, since
it only grows an index vector for each spline, not a whole spline.
This patch refactors the instance-realization code and adds new functionality.
* Named and anonymous attributes are propagated from instances to the
realized geometry. If the same attribute exists on the geometry and on an
instance, the attribute on the geometry has precedence.
* The id attribute has special handling to avoid creating the same id on many
output points. This is necessary to make e.g. the Random Value node work
as expected afterwards.
Realizing instance attributes has an effect on existing files, especially due to the
id attribute. To avoid breaking existing files, the Realize Instances node now has
a legacy option that is enabled for all already existing Realize Instances nodes.
Removing this legacy behavior does affect some existing files (although not many).
We can decide whether it's worth to remove the old behavior as a separate step.
This refactor also improves performance when realizing instances. That is mainly
due to multi-threading. See D13446 to get the file used for benchmarking. The
curve code is not as optimized as it could be yet. That's mainly because the storage
for these attributes might change soonish and it wasn't worth optimizing for the
current storage format right now.
```
1,000,000 x mesh vertex: 530 ms -> 130 ms
1,000,000 x simple cube: 1290 ms -> 190 ms
1,000,000 x point: 1000 ms -> 150 ms
1,000,000 x curve spiral: 1740 ms -> 330 ms
1,000,000 x curve line: 1110 ms -> 210 ms
10,000 x subdivided cylinder: 170 ms -> 40 ms
10 x subdivided spiral: 180 ms -> 180 ms
```
Differential Revision: https://developer.blender.org/D13446
Goals of this refactor:
* Simplify creating virtual arrays.
* Simplify passing virtual arrays around.
* Simplify converting between typed and generic virtual arrays.
* Reduce memory allocations.
As a quick reminder, a virtual arrays is a data structure that behaves like an
array (i.e. it can be accessed using an index). However, it may not actually
be stored as array internally. The two most important implementations
of virtual arrays are those that correspond to an actual plain array and those
that have the same value for every index. However, many more
implementations exist for various reasons (interfacing with legacy attributes,
unified iterator over all points in multiple splines, ...).
With this refactor the core types (`VArray`, `GVArray`, `VMutableArray` and
`GVMutableArray`) can be used like "normal values". They typically live
on the stack. Before, they were usually inside a `std::unique_ptr`. This makes
passing them around much easier. Creation of new virtual arrays is also
much simpler now due to some constructors. Memory allocations are
reduced by making use of small object optimization inside the core types.
Previously, `VArray` was a class with virtual methods that had to be overridden
to change the behavior of a the virtual array. Now,`VArray` has a fixed size
and has no virtual methods. Instead it contains a `VArrayImpl` that is
similar to the old `VArray`. `VArrayImpl` should rarely ever be used directly,
unless a new virtual array implementation is added.
To support the small object optimization for many `VArrayImpl` classes,
a new `blender::Any` type is added. It is similar to `std::any` with two
additional features. It has an adjustable inline buffer size and alignment.
The inline buffer size of `std::any` can't be relied on and is usually too
small for our use case here. Furthermore, `blender::Any` can store
additional user-defined type information without increasing the
stack size.
Differential Revision: https://developer.blender.org/D12986
This commit adds a fields version of the mesh to curve node, with a
field for the input selection. In order to reduce code duplication,
it adds the mesh to curve conversion to the new geometry module
and calls that implementation from both places.
More details on the geometry module can be found here: T86869
Differential Revision: https://developer.blender.org/D12579
