Mesh: Calculate edges with VectorSet instead of Map #120224

Merged

Hans Goudey merged 11 commits from mod_moder/blender:calc_edges_vector_set into main

2024-04-11 04:33:36 +02:00

Iliya Katushenock commented

2024-04-03 18:52:35 +02:00

Member

Due to legacy reasons (MEdge), edge calculation was being done with idea that edges cannot be temporarily copied.
But today, edges are just int2, so using of edge * instead of edge actually made things worst.
And since OrderedEdge itself is the same thing as int2, there it does not make sense to use Map for edges.
So, now edges are in a hash set. To be able to take index of edges, vector set is used.

The only functional change now is that original edges will be reordered as well. This should be okay just like
an unintentional but stable indices change, but could be avoided by using weak version of OrderedEdge.

For 2'000 x 2'000 x 2'000 cube edges calculation, change is around 3703.47 -> 2911.18 ms.

In order to reduce memory usage, a template parameter is added to VectorSet slots, so they can
use a 32 instead of 64 bit index type. Without that, the performance change is not consistent and
might not be better on a computer with more memory bandwidth.

Co-authored-by: @HooglyBoogly

Due to legacy reasons (`MEdge`), edge calculation was being done with idea that edges cannot be temporarily copied. But today, edges are just `int2`, so using of `edge *` instead of `edge` actually made things worst. And since `OrderedEdge` itself is the same thing as `int2`, there it does not make sense to use `Map` for edges. So, now edges are in a hash set. To be able to take index of edges, vector set is used. The only functional change now is that original edges will be reordered as well. This should be okay just like an unintentional but stable indices change, but could be avoided by using weak version of `OrderedEdge`. For 2'000 x 2'000 x 2'000 cube edges calculation, change is around `3703.47` -> `2911.18` ms. In order to reduce memory usage, a template parameter is added to `VectorSet` slots, so they can use a 32 instead of 64 bit index type. Without that, the performance change is not consistent and might not be better on a computer with more memory bandwidth. Co-authored-by: @HooglyBoogly

corner_edges_order_test.blend

901 KiB

calc_edges_becnh.blend

913 KiB

alembic_invalid_corners_for_edge_calc.rar

1.1 MiB

Iliya Katushenock added the

label 2024-04-03 18:52:35 +02:00

Iliya Katushenock added 1 commit 2024-04-03 18:52:45 +02:00

56f8f42727 init

Iliya Katushenock added this to the Nodes & Physics project 2024-04-03 18:52:53 +02:00

Iliya Katushenock requested review from Hans Goudey 2024-04-03 18:57:43 +02:00

Hans Goudey requested changes 2024-04-03 20:24:46 +02:00

Dismissed

Hans Goudey left a comment

Member

I like the code in your PR more, but I can't reproduce the speedup with your calc_edges_becnh.blend file (thanks for including that here BTW!)

BEFORE: Timer 'mesh_calc_edges': (Average: 969.48 ms, Min: 953.64 ms, Last: 972.75 ms)
AFTER:  Timer 'mesh_calc_edges': (Average: 1029.37 ms, Min: 999.79 ms, Last: 1068.50 ms)

Maybe it's worth trying in intermediate version without the change from Map to VectorSet?
Or the problem might be that VectorSet uses int64_t for its indices. Maybe it needs a template
specialization to use int instead for this use case.

I like the code in your PR more, but I can't reproduce the speedup with your `calc_edges_becnh.blend` file (thanks for including that here BTW!) ``` BEFORE: Timer 'mesh_calc_edges': (Average: 969.48 ms, Min: 953.64 ms, Last: 972.75 ms) AFTER: Timer 'mesh_calc_edges': (Average: 1029.37 ms, Min: 999.79 ms, Last: 1068.50 ms) ``` Maybe it's worth trying in intermediate version without the change from `Map` to `VectorSet`? Or the problem might be that `VectorSet` uses `int64_t` for its indices. Maybe it needs a template specialization to use `int` instead for this use case.

source/blender/blenkernel/intern/mesh_calc_edges.cc Outdated

						
				@ -120,1 +94,3 @@

				    }

				    const Span<OrderedEdge> task_edges = edge_map.as_span();

				    MutableSpan<int2> result_edges = new_edges.slice(edge_offsets[task_index]);

				    std::transform(task_edges.begin(),

Hans Goudey commented

2024-04-03 20:24:13 +02:00

Member

I'd try changing this to result_edges.copy_from(task_edges.cast<int2>())

I'd try changing this to `result_edges.copy_from(task_edges.cast<int2>())`

Iliya Katushenock commented

2024-04-03 21:38:38 +02:00

Author

Member

Hope OrderedEdge will be refactored to be based on BaseVec<int, 2> to make it more safe.
I also noticed some change by this way: 3517.18 -> 3260.55 ms (befor / after of copy_from).
And next, i noticed that test file is 3x faster for you. This leed in less real-world case, but i wonder if proof of performance still can be visible on 3k cuboid resolution (or more).
I check 2.7k (limit that can be holded by my RAM) both main and latest state of pr: 7.1 -> 6.7s.

It seems i was wrong to point out the performance of this refactoring at the begin, since it mainly legacy reducing, but i still think it would be better if it at least seemed better\

Hope `OrderedEdge` will be refactored to be based on `BaseVec<int, 2>` to make it *more safe*. I also noticed some change by this way: `3517.18` -> `3260.55` ms (befor / after of `copy_from`). And next, i noticed that test file is 3x faster for you. This leed in less real-world case, but i wonder if proof of performance still can be visible on 3k cuboid resolution (or more). I check 2.7k (limit that can be holded by my RAM) both main and latest state of pr: `7.1` -> `6.7`s. It seems i was wrong to point out the performance of this refactoring at the begin, since it mainly legacy reducing, but i still think it would be better if it at least seemed better\

mod_moder marked this conversation as resolved

Iliya Katushenock commented

2024-04-03 20:35:52 +02:00

Author

Member

Thanks for benchmark check, i think speed improvement is related with memory bandwidth (less number of pointers, smaller size of value in map/set container, more trivial loops for read/write).
Might this is small enough change to be not visible (just like in #114733) for you on your pc (so not sure how to prove speedup for you then).

Thanks for benchmark check, i think speed improvement is related with memory bandwidth (less number of pointers, smaller size of value in map/set container, more trivial loops for read/write). Might this is small enough change to be not visible (just like in #114733) for you on your pc (so not sure how to prove speedup for you then).

Iliya Katushenock added 3 commits 2024-04-03 21:30:44 +02:00

5c064f3de8 Merge branch 'main' into calc_edges_vector_set

2e9da27a11 cast to int2

5e8def697a fix finalize

Iliya Katushenock requested review from Hans Goudey 2024-04-04 05:45:13 +02:00

Hans Goudey commented

2024-04-04 20:07:22 +02:00

Member

Using int for SimpleVectorSetSlot works here and improves the performance for me:
main: Timer 'mesh_calc_edges': (Average: 976.20 ms, Min: 955.63 ms, Last: 1002.69 ms)
PR with change: Timer 'mesh_calc_edges': (Average: 902.40 ms, Min: 885.50 ms, Last: 904.97 ms)

I'll share the diff in chat.

Using `int` for `SimpleVectorSetSlot` works here and improves the performance for me: `main: Timer 'mesh_calc_edges': (Average: 976.20 ms, Min: 955.63 ms, Last: 1002.69 ms)` `PR with change: Timer 'mesh_calc_edges': (Average: 902.40 ms, Min: 885.50 ms, Last: 904.97 ms)` I'll share the diff in chat.

Iliya Katushenock added 2 commits 2024-04-04 21:45:42 +02:00

c3baca4833 Merge branch 'main' into calc_edges_vector_set

bf25f704b1 Hans' change

Iliya Katushenock commented

2024-04-04 21:46:55 +02:00

Author

Member

I can see 3711.94 -> 2911.18 ms changes with apply your changes @HooglyBoogly thanks for spending time for this!

I can see `3711.94` -> `2911.18` ms changes with apply your changes @HooglyBoogly thanks for spending time for this!

Iliya Katushenock added 1 commit 2024-04-04 22:26:26 +02:00

7ddab02290 cleanup

Iliya Katushenock changed title from ~~Refactor: Mesh calculate edges: Use VectorSet instead of Map~~ to Refactoring: Mesh calculate edges: Use VectorSet instead of Map

2024-04-04 22:27:16 +02:00

Hans Goudey changed title from ~~Refactoring: Mesh calculate edges: Use VectorSet instead of Map~~ to Mesh: Calculate edges with VectorSet instead of Map

2024-04-04 22:31:37 +02:00

Hans Goudey requested review from Jacques Lucke 2024-04-04 23:04:11 +02:00

Iliya Katushenock referenced this pull request

2024-04-06 00:04:16 +02:00

Fix: Reading of empty span #120330

Hans Goudey referenced this issue from a commit

2024-04-10 02:36:37 +02:00

Fix: Crash calculating edges of empty mesh

Hans Goudey added 3 commits 2024-04-11 04:28:57 +02:00

19f9015cb1 Merge branch 'main' into calc_edges_vector_set

f961779ba6 Make diff slightly smaller

7dc06e11d0 Update tests submodule hash

Hans Goudey added 1 commit 2024-04-11 04:31:02 +02:00

7ba0dc4e43 Update tests hash again

Hans Goudey referenced this issue from a commit

2024-04-11 04:31:10 +02:00

Mesh: Update tests for possible edge index changes

Hans Goudey approved these changes 2024-04-11 04:31:18 +02:00

Hans Goudey merged commit 6bafe65d28 into main

2024-04-11 04:33:36 +02:00

Hans Goudey referenced this issue from a commit

2024-04-11 04:33:37 +02:00

Mesh: Calculate edges with VectorSet instead of Map

Iliya Katushenock deleted branch calc_edges_vector_set

2024-04-11 09:48:28 +02:00

Iliya Katushenock referenced this pull request

2024-04-12 12:25:01 +02:00

Sculpt: Improve Expand performance #120125

Sign in to join this conversation.

No reviewers

No Label

Animation & Rigging

Asset Browser Project Overview

Automated Testing

Blender Asset Bundle

Dependency Graph

Development Management

EEVEE & Viewport

Images & Movies

Motion Tracking

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Render Pipeline

Sculpt, Paint & Texture

Video Sequencer

Virtual Reality

Blender 2.8 Project

Milestone 1: Basic, Local Asset Browser

Good First Issue

Animation & Rigging

Development Management

EEVEE & Viewport

Nodes & Physics

Pipeline, Assets & IO

Platforms, Builds & Tests

Render & Cycles

Sculpt, Paint & Texture

Needs Info from Developers

Needs Information from User

No Milestone

No project

No Assignees

2 Participants

Notifications

Due Date

The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: blender/blender#120224

No description provided.