Freezes when jumping in front of a keyframe. #38870

New Issue

Rick Globe · 2014-02-27T14:25:22+01:00

Rick Globe commented

2014-02-27 14:25:22 +01:00

System Information
Win7 Enterprise, Radeon X850

Blender Version
Broken: 2.69

Short description of error
Freezing when jumping before a frame.

Exact steps for others to reproduce the error

see file for details: vis.blend

I have a cube that's Z scale is controlled by an F-curve that was baked from a sound (mp3 to be precise, using the default settings, RHCP:Body of Water, if that means anything :) )
I also added the song to a speaker that I parented to the camera (for easier editing).
When I jump after a certain frame and want to go back before it, Blender freezes.

**System Information** Win7 Enterprise, Radeon X850 **Blender Version** Broken: 2.69 **Short description of error** Freezing when jumping before a frame. **Exact steps for others to reproduce the error** see file for details: [vis.blend](https://archive.blender.org/developer/F79125/vis.blend) I have a cube that's Z scale is controlled by an F-curve that was baked from a sound (mp3 to be precise, using the default settings, RHCP:Body of Water, if that means anything :) ) I also added the song to a speaker that I parented to the camera (for easier editing). When I jump after a certain frame and want to go back before it, Blender freezes.

Rick Globe commented

2014-02-27 14:25:22 +01:00

Changed status to: 'Open'

Rick Globe commented

2014-02-27 14:25:22 +01:00

Added subscriber: @PeterPronai

Bastien Montagne commented

2014-02-27 17:12:08 +01:00

Added subscriber: @mont29

Bastien Montagne commented

2014-02-27 17:12:08 +01:00

Confirmed in current master as well…

Sergey Sharybin was assigned by Bastien Montagne

2014-02-27 19:29:50 +01:00

Bastien Montagne commented

2014-02-27 19:29:50 +01:00

Added subscriber: @Sergey

Bastien Montagne commented

2014-02-27 19:29:50 +01:00

Note: to reproduce, you need a rather large gap, e.g. in given file, first click on frame 240, then on frame 10, locks 100% of time here. Also note the file has drivers…

Looks like we have a race condition here, in scene_update_objects, BLI_task_pool_work_and_wait() (line 1494) seems to wait indefinitely on BLI_condition_wait(&pool->num_cond, &pool->num_mutex); (line 383 in BLI's task.c). At least, that’s what says gdb.

@Sergey, think this is your area?

Note: to reproduce, you need a rather large gap, e.g. in given file, first click on frame 240, then on frame 10, locks 100% of time here. Also note the file has drivers… Looks like we have a race condition here, in scene_update_objects, BLI_task_pool_work_and_wait() (line 1494) seems to wait indefinitely on BLI_condition_wait(&pool->num_cond, &pool->num_mutex); (line 383 in BLI's task.c). At least, that’s what says gdb. @Sergey, think this is your area?

Sergey Sharybin commented

2014-02-28 08:35:31 +01:00

Developer note: that's not related on threaded update. Root of the issue is caused by count driver producing really huge values in some cases which makes array modifier work quite some time.

Sergey Sharybin was unassigned by Bastien Montagne

2014-02-28 09:17:13 +01:00

Bastien Montagne self-assigned this 2014-02-28 09:17:13 +01:00

Bastien Montagne was unassigned by Sergey Sharybin

2014-02-28 10:56:33 +01:00

Sergey Sharybin self-assigned this 2014-02-28 10:56:33 +01:00

Sergey Sharybin commented

2014-02-28 10:56:33 +01:00

Added subscriber: @brecht

Sergey Sharybin commented

2014-02-28 10:56:33 +01:00

Developer note:
Ok, the issue is coming from BKE_constraint_mat_convertspace() when converting world space to local one. This will get object world matrix and invert it. Now, when object does have zero scale across some axis it makes matrix uninvertible and gives verido results.

Now, the question is why do we need to get object matrix and normalize it? Wouldn't it be more mathematically robust to get rotation matrix and flip components which corresponds to negative scale components?

@brecht, afair you had quite same issues in Cycles, what was your solution there?

Developer note: Ok, the issue is coming from `BKE_constraint_mat_convertspace()` when converting world space to local one. This will get object world matrix and invert it. Now, when object does have zero scale across some axis it makes matrix uninvertible and gives verido results. Now, the question is why do we need to get object matrix and normalize it? Wouldn't it be more mathematically robust to get rotation matrix and flip components which corresponds to negative scale components? @brecht, afair you had quite same issues in Cycles, what was your solution there?

Brecht Van Lommel commented

2014-03-01 19:14:51 +01:00

What we do in Cycles is add a small value to the diagonal in case computing the matrix inverse fails:

  if(!transform_matrix4_gj_inverse(R, M)) {
      /* matrix is degenerate (e.g. 0 scale on some axis), ideally we should

     * never be in this situation, but try to invert it anyway with tweak */

      M[0][0] += 1e-8f;
      M[1][1] += 1e-8f;
      M[2][2] += 1e-8f;

      if(!transform_matrix4_gj_inverse(R, M))
          return transform_identity();
  }

There's also a pseudoinverse_m4_m4 function in blenlib that you could use as a fallback instead of adding a value to the diagonal. That will actually keep the axis zero in the inverted matrix so I think it's more accurate but also slower..

What we do in Cycles is add a small value to the diagonal in case computing the matrix inverse fails: ``` if(!transform_matrix4_gj_inverse(R, M)) { /* matrix is degenerate (e.g. 0 scale on some axis), ideally we should ``` * never be in this situation, but try to invert it anyway with tweak */ ``` M[0][0] += 1e-8f; M[1][1] += 1e-8f; M[2][2] += 1e-8f; ``` ``` if(!transform_matrix4_gj_inverse(R, M)) return transform_identity(); } ``` There's also a `pseudoinverse_m4_m4` function in blenlib that you could use as a fallback instead of adding a value to the diagonal. That will actually keep the axis zero in the inverted matrix so I think it's more accurate but also slower..

Sergey Sharybin commented

2014-03-25 10:33:57 +01:00

Seems this http://www.pasteall.org/50452/diff solves the issue, but not sure whether it's really good idea to use pseudo-inverse performance-wise.