This is an extension of 131404a1bd. The inferencing time is now down
to 4-5ms from 24ms originally.
Further improvements are likely possible. Especially, caching the results
can be very beneficial, but cache invalidation may not be entirely trivial.
This was found to be a bottleneck in the file from #144756 which has 235
group inputs and 174 group input nodes, which leads to 40,890 sockets.
Most of those are never used, so it's wasteful to add them to the map
of input values.
The inferencing time improved from 24ms to 16ms. Can likely still be
improved more, but that's a good improvement for such a small change
already.
Previously, we always just used `int` when dealing with menu values on the C++
side. It's currently the only type where we have the same base type (`int`) for
two different socket types (integer and menu sockets). This has some downsides:
* It requires special cases in some places.
* There is no function from static base type to socket type (which is useful for
some utilities like `SocketValueVariant`).
* It implicitly allows operations on menu values that shouldn't be allowed such
as arithmetic operations and conversions to and from other types.
This patch adds a new `MenuValue` type that is used for menu sockets in Geometry
Nodes and the (CPU) Compositor, clarifying the distinction between integer and
menu values.
Pull Request: https://projects.blender.org/blender/blender/pulls/144476
This patch replaces the Composite node with the Group Output node as the
primary compositor output. The old node was removed and versioned. This
was done for consistency with Geometry Nodes and in preparation for more
generic use of the compositor in VSE modifiers, layered compositing, NPR
multi-stage compositing, and more.
The Group Output node relies on the node tree interface, so we now have
a default interface of a single input and a single output. For now, only
the first input is considered while the rest are ignored, just like the
Geometry Nodes design. Furthermore, the input is required to be of type
color. Warnings and errors are issues if any of those are not met, also
similar to Geometry Nodes.
This introduces a new limitation: Composite outputs can no longer exist
in node groups, since they obviously then act as their respective group
outputs.
A refactor for the compositor scheduler is needed to simplify the logic
after this change, but this will be done in a separate patch.
Pull Request: https://projects.blender.org/blender/blender/pulls/142232
So far, only node group were able to have menu input sockets. Built-in nodes did
not support them. Currently, all menus of built-in nodes are stored on the node
instead of on the sockets. This limits their flexibility because it's not
possible to expose these inputs.
This patch adds initial support for having menu inputs in built-in nodes. For
testing purposes, it also changes a couple built-in nodes to use an input socket
instead of a node property: Points to Volume, Transform Geometry, Triangulate,
Volume to Mesh and Match String.
### Compatibility
Forward and backward compatibility is maintained where possible (it's not
possible when the menu input is linked in 5.0). The overall compatibility
approach is the same as what was done for the compositor with two differences:
there are no wrapper RNA properties (not necessary for 5.0, those were removed
for the compositor already too), no need to version animation (animation on the
menu properties was already disabled).
This also makes menu sockets not animatable in general which is kind of brittle
(e.g. doesn't properly update when the menu definition changes). To animate a
menu it's better to animate an integer and to drive an index switch with it.
### Which nodes to update?
Many existing menu properties can become sockets, but it's currently not the
intention to convert all of them. In some cases, converting them might restrict
future improvements too much. This mainly affects Math nodes.
Other existing nodes should be updated but are a bit more tricky to update for
different reasons:
* We don't support dynamic output visibility yet. This is something I'll need to
look into at some point.
* They are shared with shader/compositor nodes, which may be more limited in
what can become a socket.
* There may be performance implications unless extra special cases are
implemented, especially for multi-function nodes.
* Some nodes use socket renaming instead of dynamic socket visibility which
isn't something we support more generally yet.
### Implementation
The core implementation is fairly straight forward. The heavy lifting is done by
the existing socket visibility inferencing. There is a new simple API that
allows individual nodes to implement custom input-usage-rules based on other
inputs in a decentralized way.
In most cases, the nodes to update just have a single menu, so there is a new
node-declaration utility that links a socket to a specific value of the menu
input. This internally handles the usage inferencing as well as making the
socket available when using link-drag-search.
In the modified nodes, I also had to explicitly set the "main input" now which
is used when inserting the node in a link. The automatic behavior doesn't work
currently when the first input is a menu. This is something we'll have to solve
more generally at some point but is out of scope for this patch.
Pull Request: https://projects.blender.org/blender/blender/pulls/140705
This removes the need to deal with void pointers and explicit casts in many
places. Also it simplifies possibly extending what kind of values can be stored
for each socket in the future.
Pull Request: https://projects.blender.org/blender/blender/pulls/140844
Function nodes can be evaluated when determining which inputs are used. However,
for general function nodes the output value is considered to be unknown if any
of the inputs is considered to be unknown. For some built-in nodes we can do
better than that.
This patch improves the behavior for the following cases, previously the output
was always unknown:
* Float Math:
* `0.0 * x = 0.0`
* `x * 0.0 = 0.0`
* Integer Math:
* `0 * x = 0`
* `x * 0 = 0`
* Boolean Math:
* `false AND x = false`
* `x AND false = false`
* `true OR x = true`
* `x OR true = true`
* `false NAND x = true`
* `x NAND false = true`
* `true NOR x = false`
* `x NOR true = false`
* `false IMPLY x = true`
* `x IMPLY true = true`
* `false NIMPLY x = false`
* `x NIMPLY true = false`
This helps the inferencing code to be smarter on some cases.
This partially solves #139319. Although in the provided file even more complex
reasoning is necessary to fully support it. For integers it would need to keep
track of the current possible values e.g. `{0, 1}` so that when the integer is
compared to 2, it can be determined that this will always be false. This kind of
inferencing is a bit out of reach for the time being, it's not impossible
though.
Pull Request: https://projects.blender.org/blender/blender/pulls/139331
This is used as base class for the compute contexts for group and evaluate
closure nodes. Furthermore, in the future this can be used for the compute
context that is passed into field evaluation.
Pull Request: https://projects.blender.org/blender/blender/pulls/139377
Previously, when a socket was detected to be unused, it was just grayed out.
This patch adds support for automatically hiding unused sockets based on this
convention: Menu inputs control visibility while other inputs only control
whether something is grayed out.
More specifically, an input is visible if any of these conditions is met:
* It affects the output currently.
* It never affects the output. In this case its usage does not depend on any
menu input.
* It is used if all non-menu inputs are considered to be unknown.
In the future, we could support customizing which inputs are allowed to control
visibility. For now it's good to use the convention that Blender generally
follows itself.
As before, panels are grayed out if they only contain grayed out sockets and
panels are hidden when they don't contain any visible sockets.
Hiding inputs works in group nodes, the Geometry Nodes modifier and node
operators. In theory it will work for all node tree types, but since only
Geometry Nodes supports the Menu Switch node currently, this patch currently
only makes a difference there.
The implementation reuses the existing `SocketUsageInferencer` with a different
sets of inputs. So no new core-inferencing logic was needed.
Design task: #132706.
Pull Request: https://projects.blender.org/blender/blender/pulls/138186
This reduces verbosity when using `LinearAllocator` or `ResourceScope` to
allocate values for a `CPPType`. Now, this is simplified and one also does not
have to manually add a destructor call anymore.
Pull Request: https://projects.blender.org/blender/blender/pulls/137685
This makes accessing these properties more convenient. Since we only ever have
const references to `CPPType`, there isn't really a benefit to using methods to
avoid mutation.
Pull Request: https://projects.blender.org/blender/blender/pulls/137482
For various purposes we traverse the computation done by a node tree (e.g. for
gizmos and socket usage infeferencing). For that we generally have to keep track
of the compute context we're in at any given time. During the traversal, it's
common to enter and exist the same compute contexts multiple times. Currently,
we'd always build a new compute context when that happens. That happens even
though the old one is generally still around, because other data may reference
it. This patch implements a `ComputeContextHash` type that avoids rebuilding the
same compute contexts over and over again.
I'm considering to also replace the usage of `ComputeContextBuilder` with this
cache somehow, but will see how that works out.
The reason I'm working on this now is that I have to traverse the node tree a
bit again to find where closures might be evaluated. I wanted to be able to
cache the compute contexts for a while already.
Pull Request: https://projects.blender.org/blender/blender/pulls/137360
Under some circumstances the `value_task` run by node tree updates
needs access to a node group topology cache. Unlike the `usage_task`
it was not ensuring the cache for the node group, so a stale cache
could be used (triggers assert).
Pull Request: https://projects.blender.org/blender/blender/pulls/136878
Resolves#136183
To avoid quadratic worst case runtime when gathering values from
the modifier properties, build a temporary VectorSet of the modifier's
IDProperties. In the file from #136183, this change improves playback
performance by 1.4x for me, from 50 to 70ms.
Ideally IDProperty groups would have constant time lookup on their
own, but that's a much larger change, and this smaller change for just
Geometry Nodes is not so invasive.
Pull Request: https://projects.blender.org/blender/blender/pulls/136463
If a panel has a toggle, then it's a best practice that everything inside it is
grayed out of the toggle is off. It's not something we can enforce for node
groups, but should enforce it for built-in nodes. Right now, we don't have
built-in nodes with panels, but that may change e.g. with #135990.
This patch makes it so that the socket usage inferencing takes panel toggles
into account when determining if an input socket in a built-in node in a panel
is used.
Pull Request: https://projects.blender.org/blender/blender/pulls/135993
The main reason that didn't work before because undefined custom node groups
have the undefined `bNodeType` (and thus are ignored in various places) but are
actually still groups that can be evaluated. The fix is just to handle custom
node groups a bit more explicitly.
In the future, we may want to have a separate "undefined custom group"
`bNodetype`, but that might a be a bit bigger project.
Pull Request: https://projects.blender.org/blender/blender/pulls/135974
The issue was that the code did not handle input sockets that are fields
implicitly correctly. It just retrieved a single from them, instead of
treating them as having an unknown value.
This builds on top of 80441190c6 to gray output
unused inputs of the Mix and Mix Shader node. These nodes have unused inputs if
their mix factor is 0 or 1 (+ some special cases for non-uniform vector mixing
and blend modes).
Pull Request: https://projects.blender.org/blender/blender/pulls/133369
This patch automatically grays out input values which can't affect the output
currently. It works with inputs of group nodes, geometry nodes modifiers and
node tools.
To achieve this, it analyses the node tree and partially evaluates it to figure
out which group inputs are currently not linked to an output or are disabled by e.g.
some switch node.
Original proposal: https://devtalk.blender.org/t/dynamic-socket-visibility/31874
Related info in blog post:
https://code.blender.org/2023/11/geometry-nodes-workshop-november-2023/#dynamic-socket-visibility
Follow up task for designing a UI that allows hiding sockets: #132706
Limitations:
* The inferencing does not update correctly when a socket starts being
animated/driven. I haven't found a good way to invalidate the cache in a good
way reliably yet. It's only a very short term problem though. It fixes itself
after the next modification of the node tree and is only noticeable when
animating some specific sockets such as the switch node condition.
* Whether a socket is grayed out is not exposed in the Python API yet. That will
be done separately.
* Only a partial evaluation is done to determine if an input affects an output.
There should be no cases where a socket is found to be unused when it can actually
affect the output. However, there can be cases where a socket is inferenced to be used
even if it is not due to some complex condition. Depending on the exact circumstances,
this can either be improved or the condition in the node tree should be simplified.
Pull Request: https://projects.blender.org/blender/blender/pulls/132219
