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
This is part of D16858 but is also useful for other purposes.
The changes to the node declaration in this commit allow us to figure
out which fields might be evaluated on which geometries statically (without
executing the node tree). This allows for deterministic anonymous attribute
handling, which will be committed separately. Furthermore, this is necessary
for usability features that help the user to avoid creating links that don't
make sense (e.g. because a field can't be evaluated on a certain geometry).
This also allows us to better separate fields which depend or don't depend
on anonymous attributes.
The main idea is that each node defines some relations between its sockets.
There are four relations:
* Propagate relation: Indicates that attributes on a geometry input can be
propagated to a geometry output.
* Reference relation: Indicates that an output field references an inputs field.
So if the input field depends on an anonymous attribute, the output field
does as well.
* Eval relation: Indicates that an input field is evaluated on an input geometry.
* Available relation: Indicates that an output field has anonymous attributes
that are available on an output geometry.
These relations can also be computed for node groups automatically, but that
is not part of this commit.
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.
Geometry nodes used to log all socket values during evaluation.
This allowed the user to hover over any socket (that was evaluated)
to see its last value. The problem is that in large (nested) node trees,
the number of sockets becomes huge, causing a lot of performance
and memory overhead (in extreme cases, more than 70% of the
total execution time).
This patch changes it so, that only socket values are logged that the
user is likely to investigate. The simple heuristic is that socket values
of the currently visible node tree are logged.
The downside is that when the user changes the visible node tree, it
won't have any logged values until it is reevaluated. I updated the
tooltip message for that case to be a bit more precise.
If user feedback suggests that this new behavior is too annoying, we
can always add a UI option to log all socket values again. That shouldn't
be done without an actual need though because it takes up UI space.
Differential Revision: https://developer.blender.org/D16884
Use socket indices to keep track of logged values instead of their
identifiers. This decreases memory pressure when there are many
sockets. In cases with many cheap nodes, this can give a relatively
large improvement to overall performance. We observed a 15% increase
in a case with many math nodes and a larger increase in an experimental
softbody node setup. The log is invalidated when we add/remove/move
sockets anyway.
This is the best way I found to make building socket declarations without
the builder helper class work. Besides a vague hope for non-leaky
abstractions, I don't think there's any reason for these fields not to be
accessible directly.
Ref D16850
Separate the "insert nodes into group" operation into more distinct
phases. This helps to clarify what is actually happening, to avoid
redundant updates to group nodes every time a new socket is discovered,
and to make use of the topology cache to avoid the "accidentally
quadratic" alrogithms that we have slowly been removing from node
editing.
The change is motivated by the desire to use dynamic node declarations
for group nodes and group input/output nodes, where it is helpful to
avoid updating the declaration and sockets multiple times.
Previously, the code tried to keep node groups working even if some of
their input/output sockets had undefined type. This caused some
complexity with no benefit because not all places outside of this file
would handle the case correctly. Now node groups with undefined
interface sockets are disabled and have to be fixed manually before
they work again.
Undefined interface sockets are mostly caused by invalid Python
API usage and incomplete forward compatibility (e.g. when newer
versions introduce new socket types that the older version does
not know).
This adds an explicit post processing step to node declarations.
The purpose of this is to keep the actual node declaration functions
concise by avoiding to specify redundant information. Also it improves
the separation of the creation of the declaration from using it.
This patch adds an integer identifier to nodes that doesn't change when
the node name changes. This identifier can be used by different systems
to reference a node. This may be important to store caches and simulation
states per node, because otherwise those would always be invalidated
when a node name changes.
Additionally, this kind of identifier could make some things more efficient,
because with it an integer is enough to identify a node and one does not
have to store the node name.
I observed a 10% improvement in evaluation time in a file with an extreme
number of simple math nodes, due to reduced logging overhead-- from
0.226s to 0.205s.
Differential Revision: https://developer.blender.org/D15775
The main goal here is to move towards more self contained node
definitions. Previously, one would have to change `blenkernel` to
add a new node which is not necessary anymore. There is no need
for all these register functions to "leak out" of the nodes module.
Differential Revision: https://developer.blender.org/D16612
Instead of generating a dependency sorted node list whenever evaluating
texture or EEVEE/viewport shader nodes, use the existing sorted array
from the topology cache. This may be more efficient because the
algorithm isn't quadratic. It's also the second-to-last place to
use `node.runtime->level`, which can be removed soon.
Differential Revision: https://developer.blender.org/D16565
* This patch just moves runtime data to the runtime struct to cleanup
the dna struct. Arguably, some of this data should not even be there
because it's very use case specific. This can be cleaned up separately.
* `miniwidth` was removed completely, because it was not used anywhere.
The corresponding rna property `width_hidden` is kept to avoid
script breakage, but does not do anything (e.g. node wrangler sets it).
* Since rna is in C, some helper functions where added to access the
C++ runtime data from rna.
* This size of `bNode` decreases from 432 to 368 bytes.
* Support bidirectional type lookups. E.g. finding the base type of a
field was supported, but not the other way around. This also removes
the todo in `get_vector_type`. To achieve this, types have to be
registered up-front.
* Separate `CPPType` from other "type traits". For example, previously
`ValueOrFieldCPPType` adds additional behavior on top of `CPPType`.
Previously, it was a subclass, now it just contains a reference to the
`CPPType` it corresponds to. This follows the composition-over-inheritance
idea. This makes it easier to have self-contained "type traits" without
having to put everything into `CPPType`.
Differential Revision: https://developer.blender.org/D16479
Use the node topology cache and avoid modifying the node tree
in a non-threadsafe way to improve the predictability of using
the helper function. Replaces the implementation from
e0d4047136.
This makes instance handling more consistent with all the other geometry
component types. For example, `MeshComponent` contains a `Mesh *` and
now `InstancesComponent` has a `Instances *`.
Differential Revision: https://developer.blender.org/D16137
