Refactoring: Geometry Node: Avoid copy last buffer in result for Blur Attribute node #106860

Merged
Jacques Lucke merged 7 commits from mod_moder/blender:blur_avoid_extra_copy into main 2023-04-17 08:09:15 +02:00
1 changed files with 65 additions and 74 deletions

View File

@ -246,16 +246,19 @@ static Array<Vector<int>> create_mesh_map(const Mesh &mesh,
}
template<typename T>
static void blur_on_mesh_exec(const Span<float> neighbor_weights,
const Span<Vector<int>> neighbors_map,
const int iterations,
MutableSpan<T> main_buffer,
MutableSpan<T> tmp_buffer)
static Span<T> blur_on_mesh_exec(const Span<float> neighbor_weights,
const Span<Vector<int>> neighbors_map,
const int iterations,
const MutableSpan<T> buffer_a,
const MutableSpan<T> buffer_b)
{
MutableSpan<T> src = main_buffer;
MutableSpan<T> dst = tmp_buffer;
/* Source is set to buffer_b even though it is actually in buffer_a because the loop below starts
mod_moder marked this conversation as resolved Outdated

Add comment Source is set to buffer_b even though it is actually in buffer_a because the loop below starts with swapping both.

Add comment `Source is set to buffer_b even though it is actually in buffer_a because the loop below starts with swapping both.`
* with swapping both. */
MutableSpan<T> src = buffer_b;
MutableSpan<T> dst = buffer_a;
for ([[maybe_unused]] const int64_t iteration : IndexRange(iterations)) {
std::swap(src, dst);
attribute_math::DefaultMixer<T> mixer{dst, IndexMask(0)};
threading::parallel_for(dst.index_range(), 1024, [&](const IndexRange range) {
for (const int64_t index : range) {
@ -268,56 +271,48 @@ static void blur_on_mesh_exec(const Span<float> neighbor_weights,
}
mixer.finalize(range);
});
std::swap(src, dst);
}
/* The last computed values are in #src now. If the main buffer is #dst, the values have to be
* copied once more. */
if (dst.data() == main_buffer.data()) {
threading::parallel_for(dst.index_range(), 1024, [&](const IndexRange range) {
initialized_copy_n(
src.data() + range.start(), range.size(), main_buffer.data() + range.start());
});
}
return dst;
}
static void blur_on_mesh(const Mesh &mesh,
const eAttrDomain domain,
const int iterations,
const Span<float> neighbor_weights,
GMutableSpan main_buffer,
GMutableSpan tmp_buffer)
static GSpan blur_on_mesh(const Mesh &mesh,
const eAttrDomain domain,
const int iterations,
const Span<float> neighbor_weights,
const GMutableSpan buffer_a,
const GMutableSpan buffer_b)
{
Array<Vector<int>> neighbors_map = create_mesh_map(mesh, domain, neighbor_weights.index_range());
if (neighbors_map.is_empty()) {
return;
return buffer_a;
}
attribute_math::convert_to_static_type(main_buffer.type(), [&](auto dummy) {
GSpan result_buffer;
attribute_math::convert_to_static_type(buffer_a.type(), [&](auto dummy) {
using T = decltype(dummy);
if constexpr (!std::is_same_v<T, bool>) {
blur_on_mesh_exec<T>(neighbor_weights,
neighbors_map,
iterations,
main_buffer.typed<T>(),
tmp_buffer.typed<T>());
result_buffer = blur_on_mesh_exec<T>(
neighbor_weights, neighbors_map, iterations, buffer_a.typed<T>(), buffer_b.typed<T>());
}
});
return result_buffer;
}
template<typename T>
static void blur_on_curve_exec(const bke::CurvesGeometry &curves,
const Span<float> neighbor_weights,
const int iterations,
MutableSpan<T> main_buffer,
MutableSpan<T> tmp_buffer)
static Span<T> blur_on_curve_exec(const bke::CurvesGeometry &curves,
const Span<float> neighbor_weights,
const int iterations,
const MutableSpan<T> buffer_a,
const MutableSpan<T> buffer_b)
{
MutableSpan<T> src = main_buffer;
MutableSpan<T> dst = tmp_buffer;
MutableSpan<T> src = buffer_b;
MutableSpan<T> dst = buffer_a;
const OffsetIndices points_by_curve = curves.points_by_curve();
const VArray<bool> cyclic = curves.cyclic();
for ([[maybe_unused]] const int iteration : IndexRange(iterations)) {
std::swap(src, dst);
attribute_math::DefaultMixer<T> mixer{dst, IndexMask(0)};
threading::parallel_for(curves.curves_range(), 256, [&](const IndexRange range) {
for (const int curve_i : range) {
@ -339,53 +334,43 @@ static void blur_on_curve_exec(const bke::CurvesGeometry &curves,
const float first_neighbor_weight = neighbor_weights[first_i];
const int last_i = points.last();
const float last_neighbor_weight = neighbor_weights[last_i];
/* First point. */
mixer.set(first_i, src[first_i], 1.0f);
mixer.mix_in(first_i, src[first_i + 1], first_neighbor_weight);
/* Last point. */
mixer.set(last_i, src[last_i], 1.0f);
mixer.mix_in(last_i, src[last_i - 1], last_neighbor_weight);
if (cyclic[curve_i]) {
/* First point. */
mixer.set(first_i, src[first_i], 1.0f);
mixer.mix_in(first_i, src[first_i + 1], first_neighbor_weight);
mixer.mix_in(first_i, src[last_i], first_neighbor_weight);
/* Last point. */
mixer.set(last_i, src[last_i], 1.0f);
mixer.mix_in(last_i, src[last_i - 1], last_neighbor_weight);
mixer.mix_in(last_i, src[first_i], last_neighbor_weight);
}
else {
/* First point. */
mixer.set(first_i, src[first_i], 1.0f);
mixer.mix_in(first_i, src[first_i + 1], first_neighbor_weight);
/* Last point. */
mixer.set(last_i, src[last_i], 1.0f);
mixer.mix_in(last_i, src[last_i - 1], last_neighbor_weight);
}
}
mixer.finalize(points_by_curve[range]);
});
std::swap(src, dst);
}
/* The last computed values are in #src now. If the main buffer is #dst, the values have to be
* copied once more. */
if (dst.data() == main_buffer.data()) {
threading::parallel_for(dst.index_range(), 1024, [&](const IndexRange range) {
initialized_copy_n(
src.data() + range.start(), range.size(), main_buffer.data() + range.start());
});
}
return dst;
}
static void blur_on_curves(const bke::CurvesGeometry &curves,
const int iterations,
const Span<float> neighbor_weights,
GMutableSpan main_buffer,
GMutableSpan tmp_buffer)
static GSpan blur_on_curves(const bke::CurvesGeometry &curves,
const int iterations,
const Span<float> neighbor_weights,
const GMutableSpan buffer_a,
const GMutableSpan buffer_b)
{
attribute_math::convert_to_static_type(main_buffer.type(), [&](auto dummy) {
GSpan result_buffer;
attribute_math::convert_to_static_type(buffer_a.type(), [&](auto dummy) {
using T = decltype(dummy);
if constexpr (!std::is_same_v<T, bool>) {
blur_on_curve_exec<T>(
curves, neighbor_weights, iterations, main_buffer.typed<T>(), tmp_buffer.typed<T>());
result_buffer = blur_on_curve_exec<T>(
curves, neighbor_weights, iterations, buffer_a.typed<T>(), buffer_b.typed<T>());
}
});
return result_buffer;
}
class BlurAttributeFieldInput final : public bke::GeometryFieldInput {
@ -408,39 +393,41 @@ class BlurAttributeFieldInput final : public bke::GeometryFieldInput {
{
const int64_t domain_size = context.attributes()->domain_size(context.domain());
GArray<> main_buffer(*type_, domain_size);
GArray<> buffer_a(*type_, domain_size);
FieldEvaluator evaluator(context, domain_size);
evaluator.add_with_destination(value_field_, main_buffer.as_mutable_span());
evaluator.add_with_destination(value_field_, buffer_a.as_mutable_span());
evaluator.add(weight_field_);
evaluator.evaluate();
/* Blurring does not make sense with a less than 2 elements. */
if (domain_size <= 1) {
return GVArray::ForGArray(std::move(main_buffer));
return GVArray::ForGArray(std::move(buffer_a));
}
if (iterations_ <= 0) {
return GVArray::ForGArray(std::move(main_buffer));
return GVArray::ForGArray(std::move(buffer_a));
}
VArraySpan<float> neighbor_weights = evaluator.get_evaluated<float>(1);
GArray<> tmp_buffer(*type_, domain_size);
GArray<> buffer_b(*type_, domain_size);
GSpan result_buffer;
switch (context.type()) {
case GEO_COMPONENT_TYPE_MESH:
if (ELEM(context.domain(), ATTR_DOMAIN_POINT, ATTR_DOMAIN_EDGE, ATTR_DOMAIN_FACE)) {
if (const Mesh *mesh = context.mesh()) {
blur_on_mesh(
*mesh, context.domain(), iterations_, neighbor_weights, main_buffer, tmp_buffer);
result_buffer = blur_on_mesh(
*mesh, context.domain(), iterations_, neighbor_weights, buffer_a, buffer_b);
}
}
break;
case GEO_COMPONENT_TYPE_CURVE:
if (context.domain() == ATTR_DOMAIN_POINT) {
if (const bke::CurvesGeometry *curves = context.curves()) {
blur_on_curves(*curves, iterations_, neighbor_weights, main_buffer, tmp_buffer);
result_buffer = blur_on_curves(
*curves, iterations_, neighbor_weights, buffer_a, buffer_b);
}
}
break;
@ -448,7 +435,11 @@ class BlurAttributeFieldInput final : public bke::GeometryFieldInput {
break;
}
return GVArray::ForGArray(std::move(main_buffer));
BLI_assert(ELEM(result_buffer.data(), buffer_a.data(), buffer_b.data()));
if (result_buffer.data() == buffer_a.data()) {
return GVArray::ForGArray(std::move(buffer_a));
}
return GVArray::ForGArray(std::move(buffer_b));
}
void for_each_field_input_recursive(FunctionRef<void(const FieldInput &)> fn) const override