Regression: Box Hide occasionally causes artifacts #121107

New Issue

Sean Kim · 2024-04-26T07:47:21+02:00

Sean Kim commented

2024-04-26 07:47:21 +02:00

System Information
Operating system: Linux-6.8.6-200.fc39.x86_64-x86_64-with-glibc2.38 64 Bits, X11 UI
Graphics card: NVIDIA GeForce RTX 3060 Ti/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 550.67

Blender Version
Broken: version: 4.2.0 Alpha, branch: main, commit date: 2024-04-25 19:53, hash: 095d37c186b6
Worked: version: 3.6.11, branch: blender-v3.6-release, commit date: 2024-04-15 07:27, hash: 0d97ea8e3dc2

Also tested and verified broken behavior on 4.1.1

Short description of error
Rapid Box Hide actions on a dense mesh can sometimes cause weird artifacts

3.6.11	4.2

Exact steps for others to reproduce the error

Open attached blend file
Switch to Box Hide
Repeatedly perform Box Hide action on mesh

With a developer build, I get an assert crash on a MutableSpan operator[] call.

BLI_assert failed: source/blender/blenlib/BLI_span.hh:578, operator[](), at 'index < size_'

But the rest of the stacktrace provides nothing readable. I noticed that running blender with only a single thread seems to prevent the issue from occurring.

**System Information** Operating system: Linux-6.8.6-200.fc39.x86_64-x86_64-with-glibc2.38 64 Bits, X11 UI Graphics card: NVIDIA GeForce RTX 3060 Ti/PCIe/SSE2 NVIDIA Corporation 4.6.0 NVIDIA 550.67 **Blender Version** Broken: version: 4.2.0 Alpha, branch: main, commit date: 2024-04-25 19:53, hash: `095d37c186b6` Worked: version: 3.6.11, branch: blender-v3.6-release, commit date: 2024-04-15 07:27, hash: `0d97ea8e3dc2` Also tested and verified broken behavior on 4.1.1 **Short description of error** Rapid *Box Hide* actions on a dense mesh can sometimes cause weird artifacts | 3.6.11 | 4.2 | | --- | --- | | ![3.6.11_hide_glitch.gif](/attachments/a356d962-a893-4194-84bf-9056eff3ae69) | ![4.2_hide_glitch.gif](/attachments/218a3a91-cc75-44f8-9ca8-6d82cd1cc426) | **Exact steps for others to reproduce the error** 1. Open attached blend file 2. Switch to *Box Hide* 3. Repeatedly perform Box Hide action on mesh --- With a developer build, I get an assert crash on a `MutableSpan` `operator[]` call. ``` BLI_assert failed: source/blender/blenlib/BLI_span.hh:578, operator[](), at 'index < size_' ``` But the rest of the stacktrace provides nothing readable. I noticed that running blender with only a single thread seems to prevent the issue from occurring.

2million_verts.blend

124 MiB

3.6.11_hide_glitch.gif

3.9 MiB

4.2_hide_glitch.gif

9.8 MiB

Sean Kim added the

labels 2024-04-26 07:47:22 +02:00

Sean Kim commented

2024-04-26 08:44:28 +02:00

Running with a debugger gets this particular block as triggering the assert:

  vert_hide_update(*object, nodes, [&](const Span<int> verts, MutableSpan<bool> hide) {
    for (const int i : verts.index_range()) {
      if (gesture::is_affected(gesture_data, positions[verts[i]], normals[verts[i]])) {
        hide[i] = value;
      }
    }
  });

on hide[i] = value

Running with a debugger gets this particular block as triggering the assert: ```Cpp vert_hide_update(*object, nodes, [&](const Span<int> verts, MutableSpan<bool> hide) { for (const int i : verts.index_range()) { if (gesture::is_affected(gesture_data, positions[verts[i]], normals[verts[i]])) { hide[i] = value; } } }); ``` on `hide[i] = value`

Sean Kim commented

2024-04-26 09:28:59 +02:00

Adding this assert into paint_hide.cc in vert_hide_update shows the problem I think.

  threading::EnumerableThreadSpecific<Vector<bool>> all_new_hide;
  threading::parallel_for(nodes.index_range(), 1, [&](const IndexRange range) {
    Vector<bool> &new_hide = all_new_hide.local();
    for (PBVHNode *node : nodes.slice(range)) {
      const Span<int> verts = bke::pbvh::node_unique_verts(*node);
      new_hide.reinitialize(verts.size());
      array_utils::gather(hide_vert.span.as_span(), verts, new_hide.as_mutable_span());
      BLI_assert(verts.size() == new_hide.size());
      calc_hide(verts, new_hide);

I suspect that the usage of the the thread local vector in this way is not thread-safe for cases where there are fewer threads than PBVH nodes.

Adding this assert into `paint_hide.cc` in `vert_hide_update` shows the problem I think. ```Cpp threading::EnumerableThreadSpecific<Vector<bool>> all_new_hide; threading::parallel_for(nodes.index_range(), 1, [&](const IndexRange range) { Vector<bool> &new_hide = all_new_hide.local(); for (PBVHNode *node : nodes.slice(range)) { const Span<int> verts = bke::pbvh::node_unique_verts(*node); new_hide.reinitialize(verts.size()); array_utils::gather(hide_vert.span.as_span(), verts, new_hide.as_mutable_span()); BLI_assert(verts.size() == new_hide.size()); calc_hide(verts, new_hide); ``` I suspect that the usage of the the thread local vector in this way is not thread-safe for cases where there are fewer threads than PBVH nodes.

Sean Kim commented

2024-04-26 10:41:32 +02:00

Adding this

    for (PBVHNode *node : nodes.slice(range)) {
      std::stringstream msg;
      msg << std::this_thread::get_id() << "\t" << node << "\t" << &new_hide << std::endl;
      std::cout << msg.str();

Shows that multiple different nodes are using the same thread id and ~~therefore~~ the same Vector<bool> reference.

A dump of the data is in the attached sorted_dump.tsv

Adding this ```Cpp for (PBVHNode *node : nodes.slice(range)) { std::stringstream msg; msg << std::this_thread::get_id() << "\t" << node << "\t" << &new_hide << std::endl; std::cout << msg.str(); ``` Shows that multiple different nodes are using the same thread id and ~~therefore~~ the same `Vector<bool>` reference. A dump of the data is in the attached `sorted_dump.tsv`

sorted_dump.tsv

2.3 KiB

Pratik Borhade commented

2024-04-26 11:42:43 +02:00

Unable to confirm the assert (neither artifacts) in debug build: 71e7775bcbf4f53ffeade9b34f6e6a8c6946e0e0

Unable to confirm the assert (neither artifacts) in debug build: `71e7775bcbf4f53ffeade9b34f6e6a8c6946e0e0`

Iliya Katushenock added the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Regression: Box Hide occasionally causes artifacts #121107