0001219: Ourlier Rejection with Bidir+Random/LD hangs the renderer and eats up all RAM

(0003472)
Lord Crc (administrator)
2012-07-26 08:13

I can't seem to reproduce it.

The attached scene uses glossy_lossy which has been deprecated, perhaps it's material related. Could you update the scene please?

(0003473)
Lord Crc (administrator)
2012-07-26 08:14

Or, how long should I expect to wait for the bug to occur?

(0003475)
J the Ninja (developer)
2012-07-26 08:32

Ever time it's happened to me, it's been on the first pass of the sampler, usually at a particular location on the image, but not ever time you run the render. (it is not glossy_lossy related, i've had it happen with several other materials)

(0003514)
Lord Crc (administrator)
2012-08-09 21:42

I'm pretty certain the issue is due to Inf or NaN colors being added to the BSH.

(0003577)
jeanphi (administrator)
2012-08-27 15:51

I've added some checks to the BSH code and fixed a few issues in the outlier code that I had found.
When I test it on my box, it can eat quite a lot of memory with very complex scenes and high resolutions, but that's because the BSH has to remember a lot more data than for a simple scene.

(0003581)
SATtva (developer)
2012-08-29 05:17

At least i can't reproduce the problem neither with the attached testscene, nor with a newly created one. Moreover, the dragon testscene provided in 0001266 doesn't exhibit the spontaneous RAM leakage anymore.

Lord Crc has mentioned the root of the issue may lie deeper in material or integrators code, so i won't mark the task as resolved, leaving it to your discretion.

(0003583)
Lord Crc (administrator)
2012-08-29 09:11

Initially I had the outlier rejection code inside the main for loop, where you've put it back to.

There was a reason why I moved it out before that, but unfortunately I can't recall right now, perhaps I was clever enough to comment it in my commit message. Perhaps it was just something trivial as better cache utilization.

(0003624)
jeanphi (administrator)
2012-09-02 12:33

Maybe cache coherence, but that shouldn't make a huge difference anyway given the amount of data processed, and putting it inside the for loop allows to take advantage of the existing abnormal values rejection.
Do you think the issue can be marked resolved? I'll leave it to you.

(0003643)
Lord Crc (administrator)
2012-09-06 01:02

Ok I've looked over the old commits and I haven't documented anything. As I mentioned I do recall thinking about cache coherence, so I'm guessing that was all. I'm resolving this then.

Issue History
Date Modified	Username	Field	Change
2012-04-13 22:50	SATtva	New Issue
2012-04-13 22:50	SATtva	File Added: bug.blend
2012-05-12 13:24	Lord Crc	Product Version	1.0 => 1.0RC1
2012-05-12 13:24	Lord Crc	Assigned To	=> Lord Crc
2012-05-12 13:24	Lord Crc	Status	new => assigned
2012-07-26 08:13	Lord Crc	Note Added: 0003472
2012-07-26 08:13	Lord Crc	Status	assigned => feedback
2012-07-26 08:14	Lord Crc	Note Added: 0003473
2012-07-26 08:32	J the Ninja	Note Added: 0003475
2012-08-09 21:42	Lord Crc	Note Added: 0003514
2012-08-09 21:42	Lord Crc	Status	feedback => confirmed
2012-08-25 23:35	SATtva	Relationship added	parent of 0001266
2012-08-27 15:51	jeanphi	Note Added: 0003577
2012-08-27 15:51	jeanphi	Status	confirmed => feedback
2012-08-29 03:35	jeanphi	Relationship deleted	parent of 0001266
2012-08-29 03:35	jeanphi	Relationship added	child of 0001266
2012-08-29 05:17	SATtva	Note Added: 0003581
2012-08-29 05:17	SATtva	Status	feedback => assigned
2012-08-29 09:11	Lord Crc	Note Added: 0003583
2012-09-02 12:33	jeanphi	Note Added: 0003624
2012-09-02 12:33	jeanphi	Status	assigned => feedback
2012-09-06 01:02	Lord Crc	Note Added: 0003643
2012-09-06 01:03	Lord Crc	Mercurial Changeset #	000000000000 => f600a54980d
2012-09-06 01:03	Lord Crc	Status	feedback => resolved
2012-09-06 01:03	Lord Crc	Fixed in Version	=> 1.0
2012-09-06 01:03	Lord Crc	Resolution	open => fixed
2012-09-11 06:10	jeanphi	Status	resolved => closed

Notes
(0003472) Lord Crc (administrator) 2012-07-26 08:13	I can't seem to reproduce it. The attached scene uses glossy_lossy which has been deprecated, perhaps it's material related. Could you update the scene please?

(0003473) Lord Crc (administrator) 2012-07-26 08:14	Or, how long should I expect to wait for the bug to occur?

(0003475) J the Ninja (developer) 2012-07-26 08:32	Ever time it's happened to me, it's been on the first pass of the sampler, usually at a particular location on the image, but not ever time you run the render. (it is not glossy_lossy related, i've had it happen with several other materials)

(0003514) Lord Crc (administrator) 2012-08-09 21:42	I'm pretty certain the issue is due to Inf or NaN colors being added to the BSH.

(0003577) jeanphi (administrator) 2012-08-27 15:51	I've added some checks to the BSH code and fixed a few issues in the outlier code that I had found. When I test it on my box, it can eat quite a lot of memory with very complex scenes and high resolutions, but that's because the BSH has to remember a lot more data than for a simple scene.

(0003581) SATtva (developer) 2012-08-29 05:17	At least i can't reproduce the problem neither with the attached testscene, nor with a newly created one. Moreover, the dragon testscene provided in 0001266 doesn't exhibit the spontaneous RAM leakage anymore. Lord Crc has mentioned the root of the issue may lie deeper in material or integrators code, so i won't mark the task as resolved, leaving it to your discretion.

(0003583) Lord Crc (administrator) 2012-08-29 09:11	Initially I had the outlier rejection code inside the main for loop, where you've put it back to. There was a reason why I moved it out before that, but unfortunately I can't recall right now, perhaps I was clever enough to comment it in my commit message. Perhaps it was just something trivial as better cache utilization.

(0003624) jeanphi (administrator) 2012-09-02 12:33	Maybe cache coherence, but that shouldn't make a huge difference anyway given the amount of data processed, and putting it inside the for loop allows to take advantage of the existing abnormal values rejection. Do you think the issue can be marked resolved? I'll leave it to you.

(0003643) Lord Crc (administrator) 2012-09-06 01:02	Ok I've looked over the old commits and I haven't documented anything. As I mentioned I do recall thinking about cache coherence, so I'm guessing that was all. I'm resolving this then.