Curves: Port endpoint selection node to new data-block

The node should be faster than in 3.1, for a few reasons:
- It doesn't need to calculate and allocate the curve offsets.
- It doesn't need to de-reference a pointer for each curve.
- The inputs are accessed from the virual arrays fewer times.

On top of that, I added two other performance improvements:
- The node is multi-threaded when there are many curves.
- There are generated special cases for single value and span inputs.

**Performance**
With a set position node affecting 1 million splines with a selection
based on this node, on an Intel i5 8250U (times are approximate):
| Before | After | Speedup |
| 760 ms | 60 ms | 13x     |

Differential Revision: https://developer.blender.org/D14233
This commit is contained in:
Hans Goudey 2022-03-03 11:55:13 -05:00
parent 908ac7dce2
commit 471f27d66b
Notes: blender-bot 2023-02-14 00:57:33 +01:00
Referenced by issue #95443, Refactor curve nodes to use new data structure
1 changed files with 19 additions and 31 deletions

View File

@ -1,6 +1,6 @@
/* SPDX-License-Identifier: GPL-2.0-or-later */
#include "BKE_spline.hh"
#include "BKE_curves.hh"
#include "UI_interface.h"
#include "UI_resources.h"
@ -27,16 +27,6 @@ static void node_declare(NodeDeclarationBuilder &b)
N_("The selection from the start and end of the splines based on the input sizes"));
}
static void select_by_spline(const int start, const int end, MutableSpan<bool> r_selection)
{
const int size = r_selection.size();
const int start_use = std::min(start, size);
const int end_use = std::min(end, size);
r_selection.slice(0, start_use).fill(true);
r_selection.slice(size - end_use, end_use).fill(true);
}
class EndpointFieldInput final : public GeometryFieldInput {
Field<int> start_size_;
Field<int> end_size_;
@ -63,37 +53,35 @@ class EndpointFieldInput final : public GeometryFieldInput {
return nullptr;
}
const std::unique_ptr<CurveEval> curve = curves_to_curve_eval(*curve_component.get_for_read());
Array<int> control_point_offsets = curve->control_point_offsets();
if (curve == nullptr || control_point_offsets.last() == 0) {
const Curves &curves_id = *curve_component.get_for_read();
const bke::CurvesGeometry &curves = bke::CurvesGeometry::wrap(curves_id.geometry);
if (curves.points_size() == 0) {
return nullptr;
}
GeometryComponentFieldContext size_context{curve_component, ATTR_DOMAIN_CURVE};
fn::FieldEvaluator evaluator{size_context, curve->splines().size()};
fn::FieldEvaluator evaluator{size_context, curves.curves_size()};
evaluator.add(start_size_);
evaluator.add(end_size_);
evaluator.evaluate();
const VArray<int> &start_size = evaluator.get_evaluated<int>(0);
const VArray<int> &end_size = evaluator.get_evaluated<int>(1);
const int point_size = control_point_offsets.last();
Array<bool> selection(point_size, false);
int current_point = 0;
Array<bool> selection(curves.points_size(), false);
MutableSpan<bool> selection_span = selection.as_mutable_span();
for (int i : IndexRange(curve->splines().size())) {
const SplinePtr &spline = curve->splines()[i];
if (start_size[i] <= 0 && end_size[i] <= 0) {
selection_span.slice(current_point, spline->size()).fill(false);
}
else {
int start_use = std::max(start_size[i], 0);
int end_use = std::max(end_size[i], 0);
select_by_spline(start_use, end_use, selection_span.slice(current_point, spline->size()));
}
current_point += spline->size();
}
devirtualize_varray2(start_size, end_size, [&](const auto &start_size, const auto &end_size) {
threading::parallel_for(curves.curves_range(), 1024, [&](IndexRange curves_range) {
for (const int i : curves_range) {
const IndexRange range = curves.range_for_curve(i);
const int start = std::max(start_size[i], 0);
const int end = std::max(end_size[i], 0);
selection_span.slice(range.take_front(start)).fill(true);
selection_span.slice(range.take_back(end)).fill(true);
}
});
});
return VArray<bool>::ForContainer(std::move(selection));
};