Files
test2/source/blender/blenkernel/intern/node_tree_reference_lifetimes.cc

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1045 lines
44 KiB
C++
Raw Normal View History

Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* SPDX-FileCopyrightText: 2024 Blender Foundation
*
* SPDX-License-Identifier: GPL-2.0-or-later */
#include <iostream>
#include <fmt/format.h>
#include "BKE_node_legacy_types.hh"
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
#include "BKE_node_runtime.hh"
#include "BKE_node_tree_dot_export.hh"
#include "BKE_node_tree_reference_lifetimes.hh"
#include "BKE_node_tree_zones.hh"
#include "BLI_bit_group_vector.hh"
#include "BLI_bit_span_ops.hh"
#include "NOD_node_declaration.hh"
#include "NOD_socket.hh"
#include "BLI_resource_scope.hh"
namespace blender::bke::node_tree_reference_lifetimes {
using bits::BitInt;
using nodes::NodeDeclaration;
namespace aal = nodes::aal;
std::ostream &operator<<(std::ostream &stream, const ReferenceSetInfo &info)
{
switch (info.type) {
case ReferenceSetType::GroupOutputData:
stream << "Group Output Data: " << info.index;
break;
case ReferenceSetType::GroupInputReferenceSet:
stream << "Group Input Reference: " << info.index;
break;
case ReferenceSetType::LocalReferenceSet:
stream << "Local: " << info.socket->name;
break;
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
case ReferenceSetType::ClosureInputReferenceSet:
stream << "Closure Input Reference: " << info.socket->name;
break;
case ReferenceSetType::ClosureOutputData:
stream << "Closure Output Data: " << info.socket->name;
break;
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
stream << " (";
for (const bNodeSocket *socket : info.potential_data_origins) {
stream << socket->name << ", ";
}
stream << ")";
return stream;
}
static bool socket_may_have_reference(const bNodeTree &tree, const bNodeSocket &socket)
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
{
return tree.runtime->field_states[socket.index_in_tree()] == FieldSocketState::IsField;
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
static bool or_into_each_other_masked(MutableBoundedBitSpan a,
MutableBoundedBitSpan b,
const BoundedBitSpan mask)
{
if (bits::spans_equal_masked(a, b, mask)) {
return false;
}
bits::inplace_or_masked(a, mask, b);
bits::inplace_or_masked(b, mask, a);
return true;
}
static bool or_into_each_other(MutableBoundedBitSpan a, MutableBoundedBitSpan b)
{
if (bits::spans_equal(a, b)) {
return false;
}
a |= b;
b |= a;
return true;
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
bool can_contain_reference(const eNodeSocketDatatype socket_type)
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
{
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
return nodes::socket_type_supports_fields(socket_type) ||
ELEM(socket_type, SOCK_BUNDLE, SOCK_CLOSURE);
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
bool can_contain_referenced_data(const eNodeSocketDatatype socket_type)
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
{
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
return ELEM(socket_type, SOCK_GEOMETRY, SOCK_BUNDLE, SOCK_CLOSURE);
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
static const bNodeTreeZone *get_zone_of_node_if_full(const bNodeTreeZones *zones,
const bNode &node)
{
if (!zones) {
return nullptr;
}
const bNodeTreeZone *zone = zones->get_zone_by_node(node.identifier);
if (!zone) {
return nullptr;
}
if (!zone->input_node_id || !zone->output_node_id) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
return nullptr;
}
return zone;
}
static Array<const aal::RelationsInNode *> prepare_relations_by_node(const bNodeTree &tree,
ResourceScope &scope)
{
Array<const aal::RelationsInNode *> relations_by_node(tree.all_nodes().size());
for (const bNode *node : tree.all_nodes()) {
const aal::RelationsInNode *node_relations = nullptr;
switch (node->type_legacy) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
case GEO_NODE_SIMULATION_INPUT:
case GEO_NODE_SIMULATION_OUTPUT:
case GEO_NODE_BAKE: {
/* The relations of these nodes depend on field evaluation to avoid unnecessary
* relations, but besides that they don't need special handling. */
aal::RelationsInNode &relations = scope.construct<aal::RelationsInNode>();
{
/* Add eval relations. */
int prev_geometry_index = -1;
for (const int i : node->input_sockets().index_range()) {
const bNodeSocket &socket = node->input_socket(i);
if (socket.type == SOCK_GEOMETRY) {
prev_geometry_index = i;
continue;
}
if (prev_geometry_index == -1) {
continue;
}
if (socket_may_have_reference(tree, socket)) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
relations.eval_relations.append({i, prev_geometry_index});
}
}
}
{
/* Add available relations. */
int prev_geometry_index = -1;
for (const int i : node->output_sockets().index_range()) {
const bNodeSocket &socket = node->output_socket(i);
if (socket.type == SOCK_GEOMETRY) {
prev_geometry_index = i;
}
if (prev_geometry_index == -1) {
continue;
}
if (socket_may_have_reference(tree, socket)) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
relations.available_relations.append({i, prev_geometry_index});
}
}
}
node_relations = &relations;
break;
}
case GEO_NODE_REPEAT_INPUT:
case GEO_NODE_REPEAT_OUTPUT: {
aal::RelationsInNode &relations = scope.construct<aal::RelationsInNode>();
for (const bNodeSocket *socket : node->output_sockets()) {
if (can_contain_referenced_data(eNodeSocketDatatype(socket->type))) {
for (const bNodeSocket *other_output : node->output_sockets()) {
if (socket_may_have_reference(tree, *other_output)) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
relations.available_relations.append({other_output->index(), socket->index()});
}
}
}
}
for (const bNodeSocket *socket : node->input_sockets()) {
if (can_contain_referenced_data(eNodeSocketDatatype(socket->type))) {
for (const bNodeSocket *other_input : node->input_sockets()) {
if (socket_may_have_reference(tree, *other_input)) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
relations.eval_relations.append({other_input->index(), socket->index()});
}
}
}
}
/* Only create propagate and reference relations for the input node. The output node
* needs special handling because attributes created inside of the zone are not directly
* referenced on the outside. */
if (node->type_legacy == GEO_NODE_REPEAT_INPUT) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int input_items_start = 1;
const int output_items_start = 1;
const int items_num = node->output_sockets().size() - 1 - output_items_start;
for (const int i : IndexRange(items_num)) {
const int input_index = input_items_start + i;
const int output_index = output_items_start + i;
const bNodeSocket &input_socket = node->input_socket(input_index);
if (can_contain_referenced_data(eNodeSocketDatatype(input_socket.type))) {
relations.propagate_relations.append({input_index, output_index});
}
else if (socket_may_have_reference(tree, input_socket)) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
relations.reference_relations.append({input_index, output_index});
}
}
}
node_relations = &relations;
break;
}
case NODE_REROUTE: {
static const aal::RelationsInNode reroute_relations = []() {
aal::RelationsInNode relations;
relations.propagate_relations.append({0, 0});
relations.reference_relations.append({0, 0});
return relations;
}();
node_relations = &reroute_relations;
break;
}
case NODE_GROUP:
case NODE_CUSTOM_GROUP: {
if (const bNodeTree *group = reinterpret_cast<const bNodeTree *>(node->id)) {
if (group->runtime->reference_lifetimes_info) {
node_relations = &group->runtime->reference_lifetimes_info->tree_relations;
}
}
break;
}
default: {
if (const NodeDeclaration *node_decl = node->declaration()) {
node_relations = node_decl->anonymous_attribute_relations();
}
break;
}
}
relations_by_node[node->index()] = node_relations;
}
return relations_by_node;
}
static Vector<ReferenceSetInfo> find_reference_sets(
const bNodeTree &tree,
const Span<const aal::RelationsInNode *> &relations_by_node,
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
Vector<int> &r_group_output_reference_sets,
MultiValueMap<const bNodeTreeZone *, int> &r_output_set_sources_by_closure_zone)
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
{
Vector<ReferenceSetInfo> reference_sets;
const Span<const bNodeTreeInterfaceSocket *> interface_inputs = tree.interface_inputs();
const Span<const bNodeTreeInterfaceSocket *> interface_outputs = tree.interface_outputs();
/* Handle references coming from field inputs. */
for (const int input_i : interface_inputs.index_range()) {
const bNodeTreeInterfaceSocket &interface_input = *interface_inputs[input_i];
const bNodeSocketType *stype = interface_input.socket_typeinfo();
const eNodeSocketDatatype socket_type = stype ? eNodeSocketDatatype(stype->type) : SOCK_CUSTOM;
if (can_contain_reference(socket_type)) {
reference_sets.append({ReferenceSetType::GroupInputReferenceSet, input_i});
}
}
/* Handle references required by output geometries. */
for (const int output_i : interface_outputs.index_range()) {
const bNodeTreeInterfaceSocket &interface_output = *interface_outputs[output_i];
const bNodeSocketType *stype = interface_output.socket_typeinfo();
const eNodeSocketDatatype socket_type = stype ? eNodeSocketDatatype(stype->type) : SOCK_CUSTOM;
if (can_contain_referenced_data(socket_type)) {
r_group_output_reference_sets.append(
reference_sets.append_and_get_index({ReferenceSetType::GroupOutputData, output_i}));
}
}
/* All references referenced by the sources found so far can exist on all geometry inputs. */
for (const int input_i : interface_inputs.index_range()) {
const bNodeTreeInterfaceSocket &interface_input = *interface_inputs[input_i];
const bNodeSocketType *stype = interface_input.socket_typeinfo();
const eNodeSocketDatatype socket_type = stype ? eNodeSocketDatatype(stype->type) : SOCK_CUSTOM;
if (can_contain_referenced_data(socket_type)) {
for (const bNode *node : tree.group_input_nodes()) {
const bNodeSocket &socket = node->output_socket(input_i);
for (ReferenceSetInfo &source : reference_sets) {
source.potential_data_origins.append(&socket);
}
}
}
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
/* Handle references created by nodes in the current tree. */
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
for (const bNode *node : tree.all_nodes()) {
if (node->is_muted()) {
continue;
}
if (const aal::RelationsInNode *relations = relations_by_node[node->index()]) {
for (const aal::AvailableRelation &relation : relations->available_relations) {
const bNodeSocket &data_socket = node->output_socket(relation.geometry_output);
const bNodeSocket &reference_socket = node->output_socket(relation.field_output);
if (!reference_socket.is_available() || !reference_socket.is_available()) {
continue;
}
if (!reference_socket.is_directly_linked() || !data_socket.is_directly_linked()) {
continue;
}
reference_sets.append({ReferenceSetType::LocalReferenceSet, &reference_socket});
reference_sets.last().potential_data_origins.append(&data_socket);
}
}
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
const bNodeTreeZones *zones = tree.zones();
if (!zones) {
return reference_sets;
}
for (const bNodeTreeZone *zone : zones->zones) {
const bNode &input_node = *zone->input_node();
const bNode &output_node = *zone->output_node();
if (output_node.type_legacy != GEO_NODE_CLOSURE_OUTPUT) {
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
continue;
}
const auto &storage = *static_cast<const NodeGeometryClosureOutput *>(output_node.storage);
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
const int old_reference_sets_count = reference_sets.size();
/* Handle references coming from field inputs in the closure. */
for (const int input_i : IndexRange(storage.input_items.items_num)) {
const bNodeSocket &socket = input_node.output_socket(input_i);
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
if (can_contain_reference(eNodeSocketDatatype(socket.type))) {
reference_sets.append({ReferenceSetType::ClosureInputReferenceSet, &socket});
}
}
/* Handle references required by output geometries in the closure. */
for (const int output_i : IndexRange(storage.output_items.items_num)) {
const bNodeSocket &socket = output_node.input_socket(output_i);
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
if (can_contain_referenced_data(eNodeSocketDatatype(socket.type))) {
r_output_set_sources_by_closure_zone.add(
zone,
reference_sets.append_and_get_index({ReferenceSetType::ClosureOutputData, &socket}));
}
}
/* All references referenced passed into this zone may exist on the geometry inputs. */
MutableSpan<ReferenceSetInfo> new_reference_sets = reference_sets.as_mutable_span().drop_front(
old_reference_sets_count);
for (const int input_i : IndexRange(storage.input_items.items_num)) {
const bNodeSocket &socket = input_node.output_socket(input_i);
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
if (can_contain_referenced_data(eNodeSocketDatatype(socket.type))) {
for (ReferenceSetInfo &source : new_reference_sets) {
source.potential_data_origins.append(&socket);
}
}
}
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
return reference_sets;
}
static void set_initial_data_and_reference_bits(const bNodeTree &tree,
const Span<ReferenceSetInfo> reference_sets,
BitGroupVector<> &r_potential_data_by_socket,
BitGroupVector<> &r_potential_reference_by_socket)
{
for (const int reference_set_i : reference_sets.index_range()) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
for (const bNodeSocket *socket : reference_set.potential_data_origins) {
r_potential_data_by_socket[socket->index_in_tree()][reference_set_i].set();
}
switch (reference_set.type) {
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
case ReferenceSetType::LocalReferenceSet:
case ReferenceSetType::ClosureInputReferenceSet: {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
r_potential_reference_by_socket[reference_set.socket->index_in_tree()][reference_set_i]
.set();
break;
}
case ReferenceSetType::GroupInputReferenceSet: {
for (const bNode *node : tree.group_input_nodes()) {
const bNodeSocket &socket = node->output_socket(reference_set.index);
r_potential_reference_by_socket[socket.index_in_tree()][reference_set_i].set();
}
break;
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
case ReferenceSetType::GroupOutputData:
case ReferenceSetType::ClosureOutputData: {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* Nothing to do. */
break;
}
}
}
}
static BitVector<> get_references_coming_from_outside_zone(
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
const bNodeTreeZone &zone, const Span<const BitGroupVector<> *> sources)
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
{
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
BitVector<> found(sources.first()->group_size(), false);
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* Gather references that are passed into the zone from the outside, either through the input
* node or border links. */
for (const bNodeSocket *socket : zone.input_node()->input_sockets()) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int src = socket->index_in_tree();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
for (const BitGroupVector<> *source : sources) {
found |= (*source)[src];
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
for (const bNodeLink *link : zone.border_links) {
const int src = link->fromsock->index_in_tree();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
for (const BitGroupVector<> *source : sources) {
found |= (*source)[src];
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
return found;
}
/**
* \return True when propagation needs to be done again.
*/
static bool pass_left_to_right(const bNodeTree &tree,
const Span<const aal::RelationsInNode *> &relations_by_node,
BitGroupVector<> &r_potential_data_by_socket,
BitGroupVector<> &r_potential_reference_by_socket)
{
bool needs_extra_pass = false;
const bNodeTreeZones *zones = tree.zones();
for (const bNode *node : tree.toposort_left_to_right()) {
for (const bNodeSocket *socket : node->input_sockets()) {
if (!socket->is_available()) {
continue;
}
const int dst_index = socket->index_in_tree();
for (const bNodeLink *link : socket->directly_linked_links()) {
if (!link->is_used()) {
continue;
}
const int src_index = link->fromsock->index_in_tree();
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
r_potential_reference_by_socket[dst_index] |= r_potential_reference_by_socket[src_index];
}
}
if (node->is_muted()) {
for (const bNodeLink &link : node->internal_links()) {
const bNodeSocket &input_socket = *link.fromsock;
const bNodeSocket &output_socket = *link.tosock;
if (!input_socket.is_available() || !output_socket.is_available()) {
continue;
}
const int src_index = input_socket.index_in_tree();
const int dst_index = output_socket.index_in_tree();
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
r_potential_reference_by_socket[dst_index] |= r_potential_reference_by_socket[src_index];
}
continue;
}
if (const aal::RelationsInNode *relations = relations_by_node[node->index()]) {
/* Propagate references. */
for (const aal::ReferenceRelation &relation : relations->reference_relations) {
const bNodeSocket &from_socket = node->input_socket(relation.from_field_input);
const bNodeSocket &to_socket = node->output_socket(relation.to_field_output);
if (!from_socket.is_available() || !to_socket.is_available()) {
continue;
}
const int src_index = from_socket.index_in_tree();
const int dst_index = to_socket.index_in_tree();
r_potential_reference_by_socket[dst_index] |= r_potential_reference_by_socket[src_index];
}
/* Propagate data. */
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
for (const aal::PropagateRelation &relation : relations->propagate_relations) {
const bNodeSocket &from_socket = node->input_socket(relation.from_geometry_input);
const bNodeSocket &to_socket = node->output_socket(relation.to_geometry_output);
if (!from_socket.is_available() || !to_socket.is_available()) {
continue;
}
const int src_index = from_socket.index_in_tree();
const int dst_index = to_socket.index_in_tree();
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
}
}
switch (node->type_legacy) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* This zone needs additional special handling because attributes from the input geometry
* are propagated to the output node. */
case GEO_NODE_FOREACH_GEOMETRY_ELEMENT_OUTPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode *input_node = zone->input_node();
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const bNode *output_node = node;
const auto *storage = static_cast<const NodeGeometryForeachGeometryElementOutput *>(
node->storage);
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int src_index = input_node->input_socket(0).index_in_tree();
for (const bNodeSocket *output_socket : output_node->output_sockets()) {
if (output_socket->type == SOCK_GEOMETRY) {
const int dst_index = output_socket->index_in_tree();
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
}
}
/* Propagate references from the inside to the outside. Like in the repeat zone, new
* references created in the zone stay local inside the zone and are not propagated to the
* outside. Instead, the foreach-element output node creates new references. */
const BitVector<> outside_references = get_references_coming_from_outside_zone(
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
*zone, {&r_potential_data_by_socket, &r_potential_reference_by_socket});
for (const int item_i : IndexRange(storage->generation_items.items_num)) {
const int src_index =
node->input_socket(storage->main_items.items_num + item_i).index_in_tree();
const int dst_index =
node->output_socket(1 + storage->main_items.items_num + item_i).index_in_tree();
bits::inplace_or_masked(r_potential_data_by_socket[dst_index],
outside_references,
r_potential_data_by_socket[src_index]);
bits::inplace_or_masked(r_potential_reference_by_socket[dst_index],
outside_references,
r_potential_reference_by_socket[src_index]);
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
break;
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
case GEO_NODE_CLOSURE_INPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode &input_node = *zone->input_node();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
/* Data referenced by border links may also be passed into the closure as input. */
const BitVector<> outside_references = get_references_coming_from_outside_zone(
*zone, {&r_potential_data_by_socket, &r_potential_reference_by_socket});
for (const int i : node->output_sockets().index_range()) {
const int dst_index = input_node.output_socket(i).index_in_tree();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
r_potential_data_by_socket[dst_index] |= outside_references;
r_potential_reference_by_socket[dst_index] |= outside_references;
}
break;
}
case GEO_NODE_CLOSURE_OUTPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode &output_node = *zone->output_node();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
/* References passed through border links are referenced by the closure. */
const BitVector<> passed_in_references = get_references_coming_from_outside_zone(
*zone, {&r_potential_reference_by_socket});
const int dst_index = output_node.output_socket(0).index_in_tree();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
for (const int i : node->input_sockets().index_range()) {
const int src_index = output_node.input_socket(i).index_in_tree();
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
r_potential_reference_by_socket[dst_index] |= r_potential_reference_by_socket[src_index];
r_potential_reference_by_socket[dst_index] |= passed_in_references;
}
break;
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
case GEO_NODE_REPEAT_OUTPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode &input_node = *zone->input_node();
const bNode &output_node = *zone->output_node();
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int items_num = output_node.output_sockets().size() - 1;
/* Handle data propagation in the case when the iteration count is zero. */
for (const int i : IndexRange(items_num)) {
const int src_index = input_node.input_socket(i + 1).index_in_tree();
const int dst_index = output_node.output_socket(i).index_in_tree();
r_potential_data_by_socket[dst_index] |= r_potential_data_by_socket[src_index];
r_potential_reference_by_socket[dst_index] |= r_potential_reference_by_socket[src_index];
}
const BitVector<> outside_references = get_references_coming_from_outside_zone(
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
*zone, {&r_potential_data_by_socket, &r_potential_reference_by_socket});
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* Propagate within output node. */
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
for (const int i : IndexRange(items_num)) {
const int src_index = output_node.input_socket(i).index_in_tree();
const int dst_index = output_node.output_socket(i).index_in_tree();
bits::inplace_or_masked(r_potential_data_by_socket[dst_index],
outside_references,
r_potential_data_by_socket[src_index]);
bits::inplace_or_masked(r_potential_reference_by_socket[dst_index],
outside_references,
r_potential_reference_by_socket[src_index]);
}
/* Ensure that references and data on the input and output node are equal. Since this may
* propagate information backwards, an additional pass can be necessary. */
for (const int i : IndexRange(items_num)) {
const bNodeSocket &body_input_socket = input_node.output_socket(i + 1);
const bNodeSocket &body_output_socket = output_node.input_socket(i);
const int in_index = body_output_socket.index_in_tree();
const int out_index = body_input_socket.index_in_tree();
needs_extra_pass |= or_into_each_other_masked(r_potential_data_by_socket[in_index],
r_potential_data_by_socket[out_index],
outside_references);
needs_extra_pass |= or_into_each_other_masked(r_potential_reference_by_socket[in_index],
r_potential_reference_by_socket[out_index],
outside_references);
}
break;
}
}
}
return needs_extra_pass;
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
static void prepare_required_data_for_group_outputs(
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const bNodeTree &tree,
const Span<ReferenceSetInfo> reference_sets,
const Span<int> group_output_set_sources,
const BitGroupVector<> &potential_data_by_socket,
const BitGroupVector<> &potential_reference_by_socket,
BitGroupVector<> &r_required_data_by_socket)
{
if (const bNode *group_output_node = tree.group_output_node()) {
const Span<const bNodeSocket *> sockets = group_output_node->input_sockets().drop_back(1);
for (const int reference_set_i : group_output_set_sources) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
BLI_assert(reference_set.type == ReferenceSetType::GroupOutputData);
const int index = sockets[reference_set.index]->index_in_tree();
r_required_data_by_socket[index][reference_set_i].set();
}
BitVector<> potential_output_references(reference_sets.size(), false);
for (const bNodeSocket *socket : sockets) {
potential_output_references |= potential_reference_by_socket[socket->index_in_tree()];
}
for (const bNodeSocket *socket : sockets) {
if (!can_contain_referenced_data(eNodeSocketDatatype(socket->type))) {
continue;
}
const int index = socket->index_in_tree();
r_required_data_by_socket[index] |= potential_output_references;
/* Make sure that only available data is also required. This is enforced in the end anyway,
* but may reduce some unnecessary work. */
r_required_data_by_socket[index] &= potential_data_by_socket[index];
}
}
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
static void prepare_required_data_for_closure_outputs(
const bNodeTree &tree,
const Span<ReferenceSetInfo> reference_sets,
MultiValueMap<const bNodeTreeZone *, int> &output_set_sources_by_closure_zone,
const BitGroupVector<> &potential_data_by_socket,
const BitGroupVector<> &potential_reference_by_socket,
BitGroupVector<> &r_required_data_by_socket)
{
const bNodeTreeZones *zones = tree.zones();
if (!zones) {
return;
}
for (const bNodeTreeZone *zone : zones->zones) {
if (!zone->input_node_id || !zone->output_node_id) {
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
continue;
}
const bNode &output_node = *zone->output_node();
if (output_node.type_legacy != GEO_NODE_CLOSURE_OUTPUT) {
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
continue;
}
const Span<int> closure_output_set_sources = output_set_sources_by_closure_zone.lookup(zone);
for (const int reference_set_i : closure_output_set_sources) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
BLI_assert(reference_set.type == ReferenceSetType::ClosureOutputData);
r_required_data_by_socket[reference_set.socket->index_in_tree()][reference_set_i].set();
}
BitVector<> potential_output_references(reference_sets.size(), false);
const Span<const bNodeSocket *> sockets = output_node.input_sockets().drop_back(1);
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
for (const bNodeSocket *socket : sockets) {
potential_output_references |= potential_reference_by_socket[socket->index_in_tree()];
}
for (const bNodeSocket *socket : sockets) {
if (!can_contain_referenced_data(eNodeSocketDatatype(socket->type))) {
continue;
}
const int index = socket->index_in_tree();
r_required_data_by_socket[index] |= potential_output_references;
/* Make sure that only available data is also required. This is enforced in the end anyway,
* but may reduce some unnecessary work. */
r_required_data_by_socket[index] &= potential_data_by_socket[index];
}
}
}
static void prepare_required_data_for_outputs(
const bNodeTree &tree,
const Span<ReferenceSetInfo> reference_sets,
const Span<int> group_output_set_sources,
MultiValueMap<const bNodeTreeZone *, int> &output_set_sources_by_closure_zone,
const BitGroupVector<> &potential_data_by_socket,
const BitGroupVector<> &potential_reference_by_socket,
BitGroupVector<> &r_required_data_by_socket)
{
prepare_required_data_for_group_outputs(tree,
reference_sets,
group_output_set_sources,
potential_data_by_socket,
potential_reference_by_socket,
r_required_data_by_socket);
prepare_required_data_for_closure_outputs(tree,
reference_sets,
output_set_sources_by_closure_zone,
potential_data_by_socket,
potential_reference_by_socket,
r_required_data_by_socket);
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
static bool pass_right_to_left(const bNodeTree &tree,
const Span<const aal::RelationsInNode *> &relations_by_node,
const BitGroupVector<> &potential_reference_by_socket,
BitGroupVector<> &r_required_data_by_socket)
{
bool needs_extra_pass = false;
const bNodeTreeZones *zones = tree.zones();
for (const bNode *node : tree.toposort_right_to_left()) {
for (const bNodeSocket *socket : node->output_sockets()) {
if (!socket->is_available()) {
continue;
}
const int dst_index = socket->index_in_tree();
for (const bNodeLink *link : socket->directly_linked_links()) {
if (!link->is_used()) {
continue;
}
const int src_index = link->tosock->index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
}
if (node->is_muted()) {
for (const bNodeLink &link : node->internal_links()) {
const bNodeSocket &input_socket = *link.fromsock;
const bNodeSocket &output_socket = *link.tosock;
if (!input_socket.is_available() || !output_socket.is_available()) {
continue;
}
const int dst_index = input_socket.index_in_tree();
const int src_index = output_socket.index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
continue;
}
if (const aal::RelationsInNode *relations = relations_by_node[node->index()]) {
for (const aal::PropagateRelation &relation : relations->propagate_relations) {
const bNodeSocket &output_socket = node->output_socket(relation.to_geometry_output);
const bNodeSocket &input_socket = node->input_socket(relation.from_geometry_input);
if (!input_socket.is_available() || !output_socket.is_available()) {
continue;
}
const int dst_index = input_socket.index_in_tree();
const int src_index = output_socket.index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
for (const aal::EvalRelation &relation : relations->eval_relations) {
const bNodeSocket &data_socket = node->input_socket(relation.geometry_input);
const bNodeSocket &reference_socket = node->input_socket(relation.field_input);
if (!data_socket.is_available() || !reference_socket.is_available()) {
continue;
}
r_required_data_by_socket[data_socket.index_in_tree()] |=
potential_reference_by_socket[reference_socket.index_in_tree()];
}
}
switch (node->type_legacy) {
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
/* Propagate from the geometry outputs to the geometry input. */
case GEO_NODE_FOREACH_GEOMETRY_ELEMENT_INPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode *input_node = node;
const bNode *output_node = zone->output_node();
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int dst_index = input_node->input_socket(0).index_in_tree();
for (const bNodeSocket *output_socket : output_node->output_sockets()) {
if (output_socket->type == SOCK_GEOMETRY) {
const int src_index = output_socket->index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
}
break;
}
case GEO_NODE_FOREACH_GEOMETRY_ELEMENT_OUTPUT: {
const bNode *output_node = node;
const auto *storage = static_cast<NodeGeometryForeachGeometryElementOutput *>(
output_node->storage);
for (const int item_i : IndexRange(storage->generation_items.items_num)) {
const int src_index =
node->output_socket(1 + storage->main_items.items_num + item_i).index_in_tree();
const int dst_index =
node->input_socket(storage->main_items.items_num + item_i).index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
break;
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
case GEO_NODE_REPEAT_OUTPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
/* Propagate within output node. */
const int items_num = node->output_sockets().size() - 1;
for (const int i : IndexRange(items_num)) {
const int src_index = node->output_socket(i).index_in_tree();
const int dst_index = node->input_socket(i).index_in_tree();
r_required_data_by_socket[dst_index] |= r_required_data_by_socket[src_index];
}
break;
}
case GEO_NODE_REPEAT_INPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
const bNode *input_node = node;
const bNode *output_node = zone->output_node();
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const int items_num = output_node->output_sockets().size() - 1;
for (const int i : IndexRange(items_num)) {
const bNodeSocket &body_input_socket = input_node->output_socket(i + 1);
const bNodeSocket &body_output_socket = output_node->input_socket(i);
const int in_index = body_input_socket.index_in_tree();
const int out_index = body_output_socket.index_in_tree();
needs_extra_pass |= or_into_each_other(r_required_data_by_socket[in_index],
r_required_data_by_socket[out_index]);
}
break;
}
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
case GEO_NODE_EVALUATE_CLOSURE: {
/* Data referenced by the closure is required on all the other inputs. */
const bNodeSocket &closure_socket = node->input_socket(0);
const BoundedBitSpan required_references =
potential_reference_by_socket[closure_socket.index_in_tree()];
for (const bNodeSocket *input_socket : node->input_sockets().drop_front(1)) {
const int dst_index = input_socket->index_in_tree();
r_required_data_by_socket[dst_index] |= required_references;
}
break;
}
case GEO_NODE_CLOSURE_OUTPUT: {
const bNodeTreeZone *zone = get_zone_of_node_if_full(zones, *node);
if (!zone) {
break;
}
/* Data that's required on the closure is also required on all inputs of the closure. */
const bNodeSocket &output_socket = node->output_socket(0);
const BoundedBitSpan required_data =
r_required_data_by_socket[output_socket.index_in_tree()];
for (const bNodeSocket *input_socket : node->input_sockets()) {
r_required_data_by_socket[input_socket->index_in_tree()] |= required_data;
}
break;
}
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
}
}
return needs_extra_pass;
}
class bNodeTreeBitGroupVectorOptions : public bNodeTreeToDotOptions {
private:
Vector<BitGroupVector<>> bit_groups_;
public:
bNodeTreeBitGroupVectorOptions(const Span<BitGroupVector<>> &bit_groups)
: bit_groups_(bit_groups)
{
}
std::string socket_name(const bNodeSocket &socket) const override
{
Vector<std::string> extra_data;
for (const BitGroupVector<> &bit_group : bit_groups_) {
const BoundedBitSpan bits = bit_group[socket.index_in_tree()];
Vector<int> indices;
bits::foreach_1_index(bits, [&](const int i) { indices.append(i); });
extra_data.append(fmt::format("({})", fmt::join(indices, ",")));
}
return fmt::format("{} {}", socket.name, fmt::join(extra_data, " "));
}
};
static aal::RelationsInNode get_tree_relations(
const bNodeTree &tree,
const Span<ReferenceSetInfo> reference_sets,
const BitGroupVector<> &potential_data_by_socket,
const BitGroupVector<> &potential_reference_by_socket,
const BitGroupVector<> &required_data_by_socket)
{
aal::RelationsInNode tree_relations;
const bNode *group_output_node = tree.group_output_node();
for (const int input_i : tree.interface_inputs().index_range()) {
const bNodeTreeInterfaceSocket &interface_input = *tree.interface_inputs()[input_i];
const eNodeSocketDatatype socket_type = eNodeSocketDatatype(
interface_input.socket_typeinfo()->type);
if (can_contain_referenced_data(socket_type)) {
BitVector<> required_data(required_data_by_socket.group_size(), false);
for (const bNode *input_node : tree.group_input_nodes()) {
required_data |=
required_data_by_socket[input_node->output_socket(input_i).index_in_tree()];
}
bits::foreach_1_index(required_data, [&](const int reference_set_i) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
switch (reference_set.type) {
case ReferenceSetType::GroupOutputData: {
tree_relations.propagate_relations.append_non_duplicates(
{input_i, reference_set.index});
break;
}
case ReferenceSetType::GroupInputReferenceSet: {
tree_relations.eval_relations.append_non_duplicates({reference_set.index, input_i});
break;
}
default:
break;
}
});
}
}
if (group_output_node) {
for (const int output_i : tree.interface_outputs().index_range()) {
const bNodeSocket &socket = group_output_node->input_socket(output_i);
const eNodeSocketDatatype socket_type = eNodeSocketDatatype(socket.type);
if (can_contain_reference(socket_type)) {
const BoundedBitSpan potential_references =
potential_reference_by_socket[socket.index_in_tree()];
bits::foreach_1_index(potential_references, [&](const int reference_set_i) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
switch (reference_set.type) {
case ReferenceSetType::GroupInputReferenceSet: {
tree_relations.reference_relations.append_non_duplicates(
{reference_set.index, output_i});
break;
}
default: {
break;
}
}
});
}
if (can_contain_referenced_data(socket_type)) {
const BoundedBitSpan potential_data = potential_data_by_socket[socket.index_in_tree()];
bits::foreach_1_index(potential_data, [&](const int reference_set_i) {
const ReferenceSetInfo &reference_set = reference_sets[reference_set_i];
switch (reference_set.type) {
case ReferenceSetType::LocalReferenceSet: {
for (const bNodeSocket *other_socket :
group_output_node->input_sockets().drop_back(1))
{
if (!can_contain_reference(eNodeSocketDatatype(other_socket->type))) {
continue;
}
const BoundedBitSpan potential_references =
potential_reference_by_socket[other_socket->index_in_tree()];
if (potential_references[reference_set_i].test()) {
tree_relations.available_relations.append_non_duplicates(
{other_socket->index(), output_i});
}
}
break;
}
default: {
break;
}
}
});
}
}
}
return tree_relations;
}
static std::unique_ptr<ReferenceLifetimesInfo> make_reference_lifetimes_info(const bNodeTree &tree)
{
tree.ensure_topology_cache();
tree.ensure_interface_cache();
if (tree.has_available_link_cycle()) {
return {};
}
const bNodeTreeZones *zones = tree.zones();
if (zones == nullptr) {
return {};
}
std::unique_ptr<ReferenceLifetimesInfo> reference_lifetimes_info =
std::make_unique<ReferenceLifetimesInfo>();
ResourceScope scope;
Array<const aal::RelationsInNode *> relations_by_node = prepare_relations_by_node(tree, scope);
Vector<int> group_output_set_sources;
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
MultiValueMap<const bNodeTreeZone *, int> output_set_sources_by_closure_zone;
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
reference_lifetimes_info->reference_sets = find_reference_sets(
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
tree, relations_by_node, group_output_set_sources, output_set_sources_by_closure_zone);
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
const Span<ReferenceSetInfo> reference_sets = reference_lifetimes_info->reference_sets;
const int sockets_num = tree.all_sockets().size();
const int reference_sets_num = reference_sets.size();
BitGroupVector<> potential_data_by_socket(sockets_num, reference_sets_num, false);
BitGroupVector<> potential_reference_by_socket(sockets_num, reference_sets_num, false);
set_initial_data_and_reference_bits(
tree, reference_sets, potential_data_by_socket, potential_reference_by_socket);
/* Propagate data and reference from left to right. This may need to be done multiple times
2024-10-08 09:54:04 +11:00
* because there may be some back-links. */
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
while (pass_left_to_right(
tree, relations_by_node, potential_data_by_socket, potential_reference_by_socket))
{
}
BitGroupVector<> required_data_by_socket(sockets_num, reference_sets_num, false);
prepare_required_data_for_outputs(tree,
reference_sets,
group_output_set_sources,
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
output_set_sources_by_closure_zone,
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
potential_data_by_socket,
potential_reference_by_socket,
required_data_by_socket);
while (pass_right_to_left(
tree, relations_by_node, potential_reference_by_socket, required_data_by_socket))
{
}
/* Make sure that all required data is also potentially available. */
required_data_by_socket.all_bits() &= potential_data_by_socket.all_bits();
/* Only useful when debugging the reference lifetimes analysis. */
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
#if 0
std::cout << "\n\n"
<< node_tree_to_dot(tree,
bNodeTreeBitGroupVectorOptions(
Geometry Nodes: add Closures and Bundles behind experimental feature flag This implements bundles and closures which are described in more detail in this blog post: https://code.blender.org/2024/11/geometry-nodes-workshop-october-2024/ tl;dr: * Bundles are containers that allow storing multiple socket values in a single value. Each value in the bundle is identified by a name. Bundles can be nested. * Closures are functions that are created with the Closure Zone and can be evaluated with the Evaluate Closure node. To use the patch, the `Bundle and Closure Nodes` experimental feature has to be enabled. This is necessary, because these features are not fully done yet and still need iterations to improve the workflow before they can be officially released. These iterations are easier to do in `main` than in a separate branch though. That's because this patch is quite large and somewhat prone to merge conflicts. Also other work we want to do, depends on this. This adds the following new nodes: * Combine Bundle: can pack multiple values into one. * Separate Bundle: extracts values from a bundle. * Closure Zone: outputs a closure zone for use in the `Evaluate Closure` node. * Evaluate Closure: evaluates the passed in closure. Things that will be added soon after this lands: * Fields in bundles and closures. The way this is done changes with #134811, so I rather implement this once both are in `main`. * UI features for keeping sockets in sync (right now there are warnings only). One bigger issue is the limited support for lazyness. For example, all inputs of a Combine Bundle node will be evaluated, even if they are not all needed. The same is true for all captured values of a closure. This is a deeper limitation that needs to be resolved at some point. This will likely be done after an initial version of this patch is done. Pull Request: https://projects.blender.org/blender/blender/pulls/128340
2025-04-03 15:44:06 +02:00
{potential_reference_by_socket, required_data_by_socket}))
Refactor: Geometry Nodes: improve lifetime analysis for anonymous attributes This refactors the lifetime analysis of anonymous attributes in geometry nodes. The refactor has a couple of goals: * Use a better and simpler abstraction that can be used when building the lazy-function graph. We currently have a bunch of duplicate code to handle "field source" and "caller propagation" attributes. This is now unified so that one only has to worry about one kind of "reference sets". * Make the abstraction compatible with handling bundles and closures in case we want to support them in the future. Both types can contain geometries and fields so they need to be taken into account when determining lifetimes. * Make more parts independent of the concept of "anonymous attributes". In theory, there could be more kinds of referenced data whose lifetimes need to be managed. I don't have any concrete plans for adding any though. At its core, deterministic anonymous attributes still work the same they have been since they became deterministic [0]. Even the generated lazy-function graph is still pretty much or even exactly the same as before. The patch renames `AnonymousAttributeSet` to the more general `GeometryNodesReferenceSet` which is more. This also makes more places independent of the concept of anonymous attributes. Functionally, this still the same though. It's only used in the internals of geometry nodes nowadays. Most code just gets an `AttributeFilter` that is based on it. [0]: https://archive.blender.org/developer/D16858 Pull Request: https://projects.blender.org/blender/blender/pulls/128667
2024-10-07 12:59:39 +02:00
<< "\n\n";
#endif
reference_lifetimes_info->tree_relations = get_tree_relations(tree,
reference_sets,
potential_data_by_socket,
potential_reference_by_socket,
required_data_by_socket);
reference_lifetimes_info->required_data_by_socket = std::move(required_data_by_socket);
return reference_lifetimes_info;
}
bool analyse_reference_lifetimes(bNodeTree &tree)
{
std::unique_ptr<ReferenceLifetimesInfo> reference_lifetimes_info = make_reference_lifetimes_info(
tree);
std::unique_ptr<ReferenceLifetimesInfo> &stored_reference_lifetimes_info =
tree.runtime->reference_lifetimes_info;
if (!reference_lifetimes_info) {
if (!tree.runtime->reference_lifetimes_info) {
return false;
}
stored_reference_lifetimes_info.reset();
return true;
}
const bool interface_changed = stored_reference_lifetimes_info &&
stored_reference_lifetimes_info->tree_relations !=
reference_lifetimes_info->tree_relations;
stored_reference_lifetimes_info = std::move(reference_lifetimes_info);
return interface_changed;
}
} // namespace blender::bke::node_tree_reference_lifetimes