Files
test/source/blender/blenkernel/intern/mesh_sample.cc

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

497 lines
19 KiB
C++
Raw Normal View History

/* SPDX-FileCopyrightText: 2023 Blender Authors
*
* SPDX-License-Identifier: GPL-2.0-or-later */
#include "BKE_attribute_math.hh"
#include "BKE_bvhutils.h"
#include "BKE_mesh.hh"
#include "BKE_mesh_runtime.hh"
#include "BKE_mesh_sample.hh"
#include "DNA_mesh_types.h"
#include "DNA_meshdata_types.h"
#include "BLI_math_geom.h"
#include "BLI_rand.hh"
#include "BLI_task.hh"
namespace blender::bke::mesh_surface_sample {
template<typename T>
BLI_NOINLINE static void sample_point_attribute(const Span<int> corner_verts,
const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const VArray<T> &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const MutableSpan<T> dst)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
const MLoopTri &tri = looptris[looptri_indices[i]];
dst[i] = attribute_math::mix3(bary_coords[i],
src[corner_verts[tri.tri[0]]],
src[corner_verts[tri.tri[1]]],
src[corner_verts[tri.tri[2]]]);
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
void sample_point_normals(const Span<int> corner_verts,
const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const Span<float3> src,
const IndexMask mask,
const MutableSpan<float3> dst)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
const MLoopTri &tri = looptris[looptri_indices[i]];
const float3 value = attribute_math::mix3(bary_coords[i],
src[corner_verts[tri.tri[0]]],
src[corner_verts[tri.tri[1]]],
src[corner_verts[tri.tri[2]]]);
dst[i] = math::normalize(value);
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
void sample_point_attribute(const Span<int> corner_verts,
const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const GVArray &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const GMutableSpan dst)
{
BLI_assert(src.type() == dst.type());
const CPPType &type = src.type();
attribute_math::convert_to_static_type(type, [&](auto dummy) {
using T = decltype(dummy);
sample_point_attribute<T>(corner_verts,
looptris,
looptri_indices,
bary_coords,
src.typed<T>(),
mask,
dst.typed<T>());
});
}
template<typename T, bool check_indices = false>
BLI_NOINLINE static void sample_corner_attribute(const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const VArray<T> &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const MutableSpan<T> dst)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
if constexpr (check_indices) {
if (looptri_indices[i] == -1) {
dst[i] = {};
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
return;
}
}
const MLoopTri &tri = looptris[looptri_indices[i]];
dst[i] = sample_corner_attribute_with_bary_coords(bary_coords[i], tri, src);
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
void sample_corner_normals(const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const Span<float3> src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const MutableSpan<float3> dst)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
const MLoopTri &tri = looptris[looptri_indices[i]];
const float3 value = sample_corner_attribute_with_bary_coords(bary_coords[i], tri, src);
dst[i] = math::normalize(value);
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
void sample_corner_attribute(const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> bary_coords,
const GVArray &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const GMutableSpan dst)
{
BLI_assert(src.type() == dst.type());
const CPPType &type = src.type();
attribute_math::convert_to_static_type(type, [&](auto dummy) {
using T = decltype(dummy);
sample_corner_attribute<T>(
looptris, looptri_indices, bary_coords, src.typed<T>(), mask, dst.typed<T>());
});
}
template<typename T>
void sample_face_attribute(const Span<int> looptri_faces,
const Span<int> looptri_indices,
const VArray<T> &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const MutableSpan<T> dst)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
const int looptri_index = looptri_indices[i];
const int face_index = looptri_faces[looptri_index];
dst[i] = src[face_index];
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
void sample_face_attribute(const Span<int> looptri_faces,
const Span<int> looptri_indices,
const GVArray &src,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
const GMutableSpan dst)
{
BLI_assert(src.type() == dst.type());
const CPPType &type = src.type();
attribute_math::convert_to_static_type(type, [&](auto dummy) {
using T = decltype(dummy);
sample_face_attribute<T>(looptri_faces, looptri_indices, src.typed<T>(), mask, dst.typed<T>());
});
}
template<bool check_indices = false>
static void sample_barycentric_weights(const Span<float3> vert_positions,
const Span<int> corner_verts,
const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> sample_positions,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
MutableSpan<float3> bary_coords)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
if constexpr (check_indices) {
if (looptri_indices[i] == -1) {
bary_coords[i] = {};
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
return;
}
}
const MLoopTri &tri = looptris[looptri_indices[i]];
bary_coords[i] = compute_bary_coord_in_triangle(
vert_positions, corner_verts, tri, sample_positions[i]);
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
template<bool check_indices = false>
static void sample_nearest_weights(const Span<float3> vert_positions,
const Span<int> corner_verts,
const Span<MLoopTri> looptris,
const Span<int> looptri_indices,
const Span<float3> sample_positions,
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
const IndexMask &mask,
MutableSpan<float3> bary_coords)
{
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
mask.foreach_index([&](const int i) {
if constexpr (check_indices) {
if (looptri_indices[i] == -1) {
bary_coords[i] = {};
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
return;
}
}
const MLoopTri &tri = looptris[looptri_indices[i]];
bary_coords[i] = MIN3_PAIR(
math::distance_squared(sample_positions[i], vert_positions[corner_verts[tri.tri[0]]]),
math::distance_squared(sample_positions[i], vert_positions[corner_verts[tri.tri[1]]]),
math::distance_squared(sample_positions[i], vert_positions[corner_verts[tri.tri[2]]]),
float3(1, 0, 0),
float3(0, 1, 0),
float3(0, 0, 1));
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
});
}
int sample_surface_points_spherical(RandomNumberGenerator &rng,
const Mesh &mesh,
const Span<int> looptri_indices_to_sample,
const float3 &sample_pos,
const float sample_radius,
const float approximate_density,
Vector<float3> &r_bary_coords,
Vector<int> &r_looptri_indices,
Vector<float3> &r_positions)
{
Mesh: Move positions to a generic attribute **Changes** As described in T93602, this patch removes all use of the `MVert` struct, replacing it with a generic named attribute with the name `"position"`, consistent with other geometry types. Variable names have been changed from `verts` to `positions`, to align with the attribute name and the more generic design (positions are not vertices, they are just an attribute stored on the point domain). This change is made possible by previous commits that moved all other data out of `MVert` to runtime data or other generic attributes. What remains is mostly a simple type change. Though, the type still shows up 859 times, so the patch is quite large. One compromise is that now `CD_MASK_BAREMESH` now contains `CD_PROP_FLOAT3`. With the general move towards generic attributes over custom data types, we are removing use of these type masks anyway. **Benefits** The most obvious benefit is reduced memory usage and the benefits that brings in memory-bound situations. `float3` is only 3 bytes, in comparison to `MVert` which was 4. When there are millions of vertices this starts to matter more. The other benefits come from using a more generic type. Instead of writing algorithms specifically for `MVert`, code can just use arrays of vectors. This will allow eliminating many temporary arrays or wrappers used to extract positions. Many possible improvements aren't implemented in this patch, though I did switch simplify or remove the process of creating temporary position arrays in a few places. The design clarity that "positions are just another attribute" brings allows removing explicit copying of vertices in some procedural operations-- they are just processed like most other attributes. **Performance** This touches so many areas that it's hard to benchmark exhaustively, but I observed some areas as examples. * The mesh line node with 4 million count was 1.5x (8ms to 12ms) faster. * The Spring splash screen went from ~4.3 to ~4.5 fps. * The subdivision surface modifier/node was slightly faster RNA access through Python may be slightly slower, since now we need a name lookup instead of just a custom data type lookup for each index. **Future Improvements** * Remove uses of "vert_coords" functions: * `BKE_mesh_vert_coords_alloc` * `BKE_mesh_vert_coords_get` * `BKE_mesh_vert_coords_apply{_with_mat4}` * Remove more hidden copying of positions * General simplification now possible in many areas * Convert more code to C++ to use `float3` instead of `float[3]` * Currently `reinterpret_cast` is used for those C-API functions Differential Revision: https://developer.blender.org/D15982
2023-01-10 00:10:43 -05:00
const Span<float3> positions = mesh.vert_positions();
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
const Span<int> corner_verts = mesh.corner_verts();
const Span<MLoopTri> looptris = mesh.looptris();
const float sample_radius_sq = pow2f(sample_radius);
const float sample_plane_area = M_PI * sample_radius_sq;
/* Used for switching between two triangle sampling strategies. */
const float area_threshold = sample_plane_area;
const int old_num = r_bary_coords.size();
for (const int looptri_index : looptri_indices_to_sample) {
const MLoopTri &looptri = looptris[looptri_index];
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
const float3 &v0 = positions[corner_verts[looptri.tri[0]]];
const float3 &v1 = positions[corner_verts[looptri.tri[1]]];
const float3 &v2 = positions[corner_verts[looptri.tri[2]]];
const float looptri_area = area_tri_v3(v0, v1, v2);
if (looptri_area < area_threshold) {
/* The triangle is small compared to the sample radius. Sample by generating random
* barycentric coordinates. */
const int amount = rng.round_probabilistic(approximate_density * looptri_area);
for ([[maybe_unused]] const int i : IndexRange(amount)) {
const float3 bary_coord = rng.get_barycentric_coordinates();
const float3 point_pos = attribute_math::mix3(bary_coord, v0, v1, v2);
const float dist_to_sample_sq = math::distance_squared(point_pos, sample_pos);
if (dist_to_sample_sq > sample_radius_sq) {
continue;
}
r_bary_coords.append(bary_coord);
r_looptri_indices.append(looptri_index);
r_positions.append(point_pos);
}
}
else {
/* The triangle is large compared to the sample radius. Sample by generating random points
* on the triangle plane within the sample radius. */
float3 normal;
normal_tri_v3(normal, v0, v1, v2);
float3 sample_pos_proj = sample_pos;
project_v3_plane(sample_pos_proj, normal, v0);
const float proj_distance_sq = math::distance_squared(sample_pos_proj, sample_pos);
const float sample_radius_factor_sq = 1.0f -
std::min(1.0f, proj_distance_sq / sample_radius_sq);
const float radius_proj_sq = sample_radius_sq * sample_radius_factor_sq;
const float radius_proj = std::sqrt(radius_proj_sq);
const float circle_area = M_PI * radius_proj_sq;
const int amount = rng.round_probabilistic(approximate_density * circle_area);
const float3 axis_1 = math::normalize(v1 - v0) * radius_proj;
const float3 axis_2 = math::normalize(math::cross(axis_1, math::cross(axis_1, v2 - v0))) *
radius_proj;
for ([[maybe_unused]] const int i : IndexRange(amount)) {
const float r = std::sqrt(rng.get_float());
const float angle = rng.get_float() * 2.0f * M_PI;
const float x = r * std::cos(angle);
const float y = r * std::sin(angle);
const float3 point_pos = sample_pos_proj + axis_1 * x + axis_2 * y;
if (!isect_point_tri_prism_v3(point_pos, v0, v1, v2)) {
/* Sampled point is not in the triangle. */
continue;
}
float3 bary_coord;
interp_weights_tri_v3(bary_coord, v0, v1, v2, point_pos);
r_bary_coords.append(bary_coord);
r_looptri_indices.append(looptri_index);
r_positions.append(point_pos);
}
}
}
return r_bary_coords.size() - old_num;
}
int sample_surface_points_projected(
RandomNumberGenerator &rng,
const Mesh &mesh,
BVHTreeFromMesh &mesh_bvhtree,
const float2 &sample_pos_re,
const float sample_radius_re,
const FunctionRef<void(const float2 &pos_re, float3 &r_start, float3 &r_end)>
region_position_to_ray,
const bool front_face_only,
const int tries_num,
const int max_points,
Vector<float3> &r_bary_coords,
Vector<int> &r_looptri_indices,
Vector<float3> &r_positions)
{
Mesh: Move positions to a generic attribute **Changes** As described in T93602, this patch removes all use of the `MVert` struct, replacing it with a generic named attribute with the name `"position"`, consistent with other geometry types. Variable names have been changed from `verts` to `positions`, to align with the attribute name and the more generic design (positions are not vertices, they are just an attribute stored on the point domain). This change is made possible by previous commits that moved all other data out of `MVert` to runtime data or other generic attributes. What remains is mostly a simple type change. Though, the type still shows up 859 times, so the patch is quite large. One compromise is that now `CD_MASK_BAREMESH` now contains `CD_PROP_FLOAT3`. With the general move towards generic attributes over custom data types, we are removing use of these type masks anyway. **Benefits** The most obvious benefit is reduced memory usage and the benefits that brings in memory-bound situations. `float3` is only 3 bytes, in comparison to `MVert` which was 4. When there are millions of vertices this starts to matter more. The other benefits come from using a more generic type. Instead of writing algorithms specifically for `MVert`, code can just use arrays of vectors. This will allow eliminating many temporary arrays or wrappers used to extract positions. Many possible improvements aren't implemented in this patch, though I did switch simplify or remove the process of creating temporary position arrays in a few places. The design clarity that "positions are just another attribute" brings allows removing explicit copying of vertices in some procedural operations-- they are just processed like most other attributes. **Performance** This touches so many areas that it's hard to benchmark exhaustively, but I observed some areas as examples. * The mesh line node with 4 million count was 1.5x (8ms to 12ms) faster. * The Spring splash screen went from ~4.3 to ~4.5 fps. * The subdivision surface modifier/node was slightly faster RNA access through Python may be slightly slower, since now we need a name lookup instead of just a custom data type lookup for each index. **Future Improvements** * Remove uses of "vert_coords" functions: * `BKE_mesh_vert_coords_alloc` * `BKE_mesh_vert_coords_get` * `BKE_mesh_vert_coords_apply{_with_mat4}` * Remove more hidden copying of positions * General simplification now possible in many areas * Convert more code to C++ to use `float3` instead of `float[3]` * Currently `reinterpret_cast` is used for those C-API functions Differential Revision: https://developer.blender.org/D15982
2023-01-10 00:10:43 -05:00
const Span<float3> positions = mesh.vert_positions();
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
const Span<int> corner_verts = mesh.corner_verts();
const Span<MLoopTri> looptris = mesh.looptris();
int point_count = 0;
for ([[maybe_unused]] const int _ : IndexRange(tries_num)) {
if (point_count == max_points) {
break;
}
const float r = sample_radius_re * std::sqrt(rng.get_float());
const float angle = rng.get_float() * 2.0f * M_PI;
float3 ray_start, ray_end;
const float2 pos_re = sample_pos_re + r * float2(std::cos(angle), std::sin(angle));
region_position_to_ray(pos_re, ray_start, ray_end);
const float3 ray_direction = math::normalize(ray_end - ray_start);
BVHTreeRayHit ray_hit;
ray_hit.dist = FLT_MAX;
ray_hit.index = -1;
BLI_bvhtree_ray_cast(mesh_bvhtree.tree,
ray_start,
ray_direction,
0.0f,
&ray_hit,
mesh_bvhtree.raycast_callback,
&mesh_bvhtree);
if (ray_hit.index == -1) {
continue;
}
if (front_face_only) {
const float3 normal = ray_hit.no;
if (math::dot(ray_direction, normal) >= 0.0f) {
continue;
}
}
const int looptri_index = ray_hit.index;
const float3 pos = ray_hit.co;
Mesh: Remove redundant custom data pointers For copy-on-write, we want to share attribute arrays between meshes where possible. Mutable pointers like `Mesh.mvert` make that difficult by making ownership vague. They also make code more complex by adding redundancy. The simplest solution is just removing them and retrieving layers from `CustomData` as needed. Similar changes have already been applied to curves and point clouds (e9f82d3dc7ee, 410a6efb747f). Removing use of the pointers generally makes code more obvious and more reusable. Mesh data is now accessed with a C++ API (`Mesh::edges()` or `Mesh::edges_for_write()`), and a C API (`BKE_mesh_edges(mesh)`). The CoW changes this commit makes possible are described in T95845 and T95842, and started in D14139 and D14140. The change also simplifies the ongoing mesh struct-of-array refactors from T95965. **RNA/Python Access Performance** Theoretically, accessing mesh elements with the RNA API may become slower, since the layer needs to be found on every random access. However, overhead is already high enough that this doesn't make a noticible differenc, and performance is actually improved in some cases. Random access can be up to 10% faster, but other situations might be a bit slower. Generally using `foreach_get/set` are the best way to improve performance. See the differential revision for more discussion about Python performance. Cycles has been updated to use raw pointers and the internal Blender mesh types, mostly because there is no sense in having this overhead when it's already compiled with Blender. In my tests this roughly halves the Cycles mesh creation time (0.19s to 0.10s for a 1 million face grid). Differential Revision: https://developer.blender.org/D15488
2022-09-05 11:56:34 -05:00
const float3 bary_coords = compute_bary_coord_in_triangle(
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
positions, corner_verts, looptris[looptri_index], pos);
r_positions.append(pos);
r_bary_coords.append(bary_coords);
r_looptri_indices.append(looptri_index);
point_count++;
}
return point_count;
}
Mesh: Move positions to a generic attribute **Changes** As described in T93602, this patch removes all use of the `MVert` struct, replacing it with a generic named attribute with the name `"position"`, consistent with other geometry types. Variable names have been changed from `verts` to `positions`, to align with the attribute name and the more generic design (positions are not vertices, they are just an attribute stored on the point domain). This change is made possible by previous commits that moved all other data out of `MVert` to runtime data or other generic attributes. What remains is mostly a simple type change. Though, the type still shows up 859 times, so the patch is quite large. One compromise is that now `CD_MASK_BAREMESH` now contains `CD_PROP_FLOAT3`. With the general move towards generic attributes over custom data types, we are removing use of these type masks anyway. **Benefits** The most obvious benefit is reduced memory usage and the benefits that brings in memory-bound situations. `float3` is only 3 bytes, in comparison to `MVert` which was 4. When there are millions of vertices this starts to matter more. The other benefits come from using a more generic type. Instead of writing algorithms specifically for `MVert`, code can just use arrays of vectors. This will allow eliminating many temporary arrays or wrappers used to extract positions. Many possible improvements aren't implemented in this patch, though I did switch simplify or remove the process of creating temporary position arrays in a few places. The design clarity that "positions are just another attribute" brings allows removing explicit copying of vertices in some procedural operations-- they are just processed like most other attributes. **Performance** This touches so many areas that it's hard to benchmark exhaustively, but I observed some areas as examples. * The mesh line node with 4 million count was 1.5x (8ms to 12ms) faster. * The Spring splash screen went from ~4.3 to ~4.5 fps. * The subdivision surface modifier/node was slightly faster RNA access through Python may be slightly slower, since now we need a name lookup instead of just a custom data type lookup for each index. **Future Improvements** * Remove uses of "vert_coords" functions: * `BKE_mesh_vert_coords_alloc` * `BKE_mesh_vert_coords_get` * `BKE_mesh_vert_coords_apply{_with_mat4}` * Remove more hidden copying of positions * General simplification now possible in many areas * Convert more code to C++ to use `float3` instead of `float[3]` * Currently `reinterpret_cast` is used for those C-API functions Differential Revision: https://developer.blender.org/D15982
2023-01-10 00:10:43 -05:00
float3 compute_bary_coord_in_triangle(const Span<float3> vert_positions,
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
const Span<int> corner_verts,
const MLoopTri &looptri,
const float3 &position)
{
Mesh: Replace MLoop struct with generic attributes Implements #102359. Split the `MLoop` struct into two separate integer arrays called `corner_verts` and `corner_edges`, referring to the vertex each corner is attached to and the next edge around the face at each corner. These arrays can be sliced to give access to the edges or vertices in a face. Then they are often referred to as "poly_verts" or "poly_edges". The main benefits are halving the necessary memory bandwidth when only one array is used and simplifications from using regular integer indices instead of a special-purpose struct. The commit also starts a renaming from "loop" to "corner" in mesh code. Like the other mesh struct of array refactors, forward compatibility is kept by writing files with the older format. This will be done until 4.0 to ease the transition process. Looking at a small portion of the patch should give a good impression for the rest of the changes. I tried to make the changes as small as possible so it's easy to tell the correctness from the diff. Though I found Blender developers have been very inventive over the last decade when finding different ways to loop over the corners in a face. For performance, nearly every piece of code that deals with `Mesh` is slightly impacted. Any algorithm that is memory bottle-necked should see an improvement. For example, here is a comparison of interpolating a vertex float attribute to face corners (Ryzen 3700x): **Before** (Average: 3.7 ms, Min: 3.4 ms) ``` threading::parallel_for(loops.index_range(), 4096, [&](IndexRange range) { for (const int64_t i : range) { dst[i] = src[loops[i].v]; } }); ``` **After** (Average: 2.9 ms, Min: 2.6 ms) ``` array_utils::gather(src, corner_verts, dst); ``` That's an improvement of 28% to the average timings, and it's also a simplification, since an index-based routine can be used instead. For more examples using the new arrays, see the design task. Pull Request: https://projects.blender.org/blender/blender/pulls/104424
2023-03-20 15:55:13 +01:00
const float3 &v0 = vert_positions[corner_verts[looptri.tri[0]]];
const float3 &v1 = vert_positions[corner_verts[looptri.tri[1]]];
const float3 &v2 = vert_positions[corner_verts[looptri.tri[2]]];
float3 bary_coords;
interp_weights_tri_v3(bary_coords, v0, v1, v2, position);
return bary_coords;
}
BaryWeightFromPositionFn::BaryWeightFromPositionFn(GeometrySet geometry)
: source_(std::move(geometry))
{
source_.ensure_owns_direct_data();
static const mf::Signature signature = []() {
mf::Signature signature;
mf::SignatureBuilder builder{"Bary Weight from Position", signature};
builder.single_input<float3>("Position");
builder.single_input<int>("Triangle Index");
builder.single_output<float3>("Barycentric Weight");
return signature;
}();
this->set_signature(&signature);
const Mesh &mesh = *source_.get_mesh();
vert_positions_ = mesh.vert_positions();
corner_verts_ = mesh.corner_verts();
looptris_ = mesh.looptris();
}
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
void BaryWeightFromPositionFn::call(const IndexMask &mask,
mf::Params params,
mf::Context /*context*/) const
{
const VArraySpan<float3> sample_positions = params.readonly_single_input<float3>(0, "Position");
const VArraySpan<int> triangle_indices = params.readonly_single_input<int>(1, "Triangle Index");
MutableSpan<float3> bary_weights = params.uninitialized_single_output<float3>(
2, "Barycentric Weight");
sample_barycentric_weights<true>(vert_positions_,
corner_verts_,
looptris_,
triangle_indices,
sample_positions,
mask,
bary_weights);
}
CornerBaryWeightFromPositionFn::CornerBaryWeightFromPositionFn(GeometrySet geometry)
: source_(std::move(geometry))
{
source_.ensure_owns_direct_data();
static const mf::Signature signature = []() {
mf::Signature signature;
mf::SignatureBuilder builder{"Nearest Weight from Position", signature};
builder.single_input<float3>("Position");
builder.single_input<int>("Triangle Index");
builder.single_output<float3>("Barycentric Weight");
return signature;
}();
this->set_signature(&signature);
const Mesh &mesh = *source_.get_mesh();
vert_positions_ = mesh.vert_positions();
corner_verts_ = mesh.corner_verts();
looptris_ = mesh.looptris();
}
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
void CornerBaryWeightFromPositionFn::call(const IndexMask &mask,
mf::Params params,
mf::Context /*context*/) const
{
const VArraySpan<float3> sample_positions = params.readonly_single_input<float3>(0, "Position");
const VArraySpan<int> triangle_indices = params.readonly_single_input<int>(1, "Triangle Index");
MutableSpan<float3> bary_weights = params.uninitialized_single_output<float3>(
2, "Barycentric Weight");
sample_nearest_weights<true>(vert_positions_,
corner_verts_,
looptris_,
triangle_indices,
sample_positions,
mask,
bary_weights);
}
BaryWeightSampleFn::BaryWeightSampleFn(GeometrySet geometry, fn::GField src_field)
: source_(std::move(geometry))
{
source_.ensure_owns_direct_data();
this->evaluate_source(std::move(src_field));
mf::SignatureBuilder builder{"Sample Barycentric Triangles", signature_};
builder.single_input<int>("Triangle Index");
builder.single_input<float3>("Barycentric Weight");
builder.single_output("Value", source_data_->type());
this->set_signature(&signature_);
}
BLI: refactor IndexMask for better performance and memory usage Goals of this refactor: * Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an `int64_t` for each index which is more than necessary in pretty much all practical cases currently. Using `int32_t` might still become limiting in the future in case we use this to index e.g. byte buffers larger than a few gigabytes. We also don't want to template `IndexMask`, because that would cause a split in the "ecosystem", or everything would have to be implemented twice or templated. * Allow for more multi-threading. The old `IndexMask` contains a single array. This is generally good but has the problem that it is hard to fill from multiple-threads when the final size is not known from the beginning. This is commonly the case when e.g. converting an array of bool to an index mask. Currently, this kind of code only runs on a single thread. * Allow for efficient set operations like join, intersect and difference. It should be possible to multi-thread those operations. * It should be possible to iterate over an `IndexMask` very efficiently. The most important part of that is to avoid all memory access when iterating over continuous ranges. For some core nodes (e.g. math nodes), we generate optimized code for the cases of irregular index masks and simple index ranges. To achieve these goals, a few compromises had to made: * Slicing of the mask (at specific indices) and random element access is `O(log #indices)` now, but with a low constant factor. It should be possible to split a mask into n approximately equally sized parts in `O(n)` though, making the time per split `O(1)`. * Using range-based for loops does not work well when iterating over a nested data structure like the new `IndexMask`. Therefor, `foreach_*` functions with callbacks have to be used. To avoid extra code complexity at the call site, the `foreach_*` methods support multi-threading out of the box. The new data structure splits an `IndexMask` into an arbitrary number of ordered `IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The indices within a segment are stored as `int16_t`. Each segment has an additional `int64_t` offset which allows storing arbitrary `int64_t` indices. This approach has the main benefits that segments can be processed/constructed individually on multiple threads without a serial bottleneck. Also it reduces the memory requirements significantly. For more details see comments in `BLI_index_mask.hh`. I did a few tests to verify that the data structure generally improves performance and does not cause regressions: * Our field evaluation benchmarks take about as much as before. This is to be expected because we already made sure that e.g. add node evaluation is vectorized. The important thing here is to check that changes to the way we iterate over the indices still allows for auto-vectorization. * Memory usage by a mask is about 1/4 of what it was before in the average case. That's mainly caused by the switch from `int64_t` to `int16_t` for indices. In the worst case, the memory requirements can be larger when there are many indices that are very far away. However, when they are far away from each other, that indicates that there aren't many indices in total. In common cases, memory usage can be way lower than 1/4 of before, because sub-ranges use static memory. * For some more specific numbers I benchmarked `IndexMask::from_bools` in `index_mask_from_selection` on 10.000.000 elements at various probabilities for `true` at every index: ``` Probability Old New 0 4.6 ms 0.8 ms 0.001 5.1 ms 1.3 ms 0.2 8.4 ms 1.8 ms 0.5 15.3 ms 3.0 ms 0.8 20.1 ms 3.0 ms 0.999 25.1 ms 1.7 ms 1 13.5 ms 1.1 ms ``` Pull Request: https://projects.blender.org/blender/blender/pulls/104629
2023-05-24 18:11:41 +02:00
void BaryWeightSampleFn::call(const IndexMask &mask,
mf::Params params,
mf::Context /*context*/) const
{
const VArraySpan<int> triangle_indices = params.readonly_single_input<int>(0, "Triangle Index");
const VArraySpan<float3> bary_weights = params.readonly_single_input<float3>(
1, "Barycentric Weight");
GMutableSpan dst = params.uninitialized_single_output(2, "Value");
attribute_math::convert_to_static_type(dst.type(), [&](auto dummy) {
using T = decltype(dummy);
sample_corner_attribute<T, true>(
looptris_, triangle_indices, bary_weights, source_data_->typed<T>(), mask, dst.typed<T>());
});
}
void BaryWeightSampleFn::evaluate_source(fn::GField src_field)
{
const Mesh &mesh = *source_.get_mesh();
looptris_ = mesh.looptris();
/* Use the most complex domain for now, ensuring no information is lost. In the future, it should
* be possible to use the most complex domain required by the field inputs, to simplify sampling
* and avoid domain conversions. */
domain_ = ATTR_DOMAIN_CORNER;
source_context_.emplace(bke::MeshFieldContext(mesh, domain_));
const int domain_size = mesh.attributes().domain_size(domain_);
source_evaluator_ = std::make_unique<fn::FieldEvaluator>(*source_context_, domain_size);
source_evaluator_->add(std::move(src_field));
source_evaluator_->evaluate();
source_data_ = &source_evaluator_->get_evaluated(0);
}
} // namespace blender::bke::mesh_surface_sample