[Blender 2.8] Evee Baking Indirect Lighitng causes scene to crash #63969

New Issue

Karthik Hebbar · 2019-04-28T20:38:46+02:00

Karthik Hebbar commented

2019-04-28 20:38:46 +02:00

Baking Indirect Light with dimension 8x8x8 and few Area lights causes the application to freeze up with progress at 0% . I waited for an hour or so. But no progress. The RAM memory usage was at 90% and CPU of around 50%.

Workarounds I tried.
Initially I had around 12 spotlights and 4 area lights. Since the problem started to occur. I deleted the spot lights and tried again. But invain .

I have attached the blend file also for reference

Scene.blend

Do check

Thanks

Baking Indirect Light with dimension 8x8x8 and few Area lights causes the application to freeze up with progress at 0% . I waited for an hour or so. But no progress. The RAM memory usage was at 90% and CPU of around 50%. Workarounds I tried. Initially I had around 12 spotlights and 4 area lights. Since the problem started to occur. I deleted the spot lights and tried again. But invain . I have attached the blend file also for reference [Scene.blend](https://drive.google.com/open?id=1eC5yWhzx9q4iOSJEeZ2D5z1P2luLzzVI) Do check Thanks

Karthik Hebbar commented

2019-04-28 20:38:47 +02:00

Added subscriber: @deltamish

Marco commented

2019-04-28 21:46:19 +02:00

Added subscriber: @nacioss

Marco commented

2019-04-28 21:46:19 +02:00

I confirm, even at 4x4x4 it is suuuper slow (1% after 1 minute), but you also have a lot of extra unnecessary geometry in the scene.

I confirm, even at 4x4x4 it is suuuper slow (1% after 1 minute), but you also have a lot of extra unnecessary geometry in the scene. ![Screenshot.jpg](https://archive.blender.org/developer/F6989492/Screenshot.jpg)

Philipp Oeser commented

2019-04-29 15:46:11 +02:00

Added subscriber: @lichtwerk

Philipp Oeser commented

2019-04-29 15:46:11 +02:00

Can confirm it is pretty slow [I have a comparable interior scene around 4M tris that bakes a lot quicker and also eats a lot less RAM -- ]

Needs some more checking which components are actually slowing it down that much...

Can confirm it is pretty slow [I have a comparable interior scene around 4M tris that bakes a lot quicker and also eats a lot less RAM -- ] Needs some more checking which components are actually slowing it down that much...

Stefano Bonicatti commented

2019-04-30 01:19:13 +02:00

Added subscriber: @Smjert

Stefano Bonicatti commented

2019-04-30 01:19:13 +02:00

I've debugged it a bit for fun (because I don't know the code, it's the first time I see it), but the part that is taking most of the time is the rendering of the Irradiance grid.
The part of the code is in evee_lightcache.c:1198, there's a loop there and the highest loop count to do is the grid_sample_len which is 1200 (which seems calculated from the current LightProbe grid_resolution_x, y, z).
I tried to understand why 1200 and where that resolution is taken/calculated but I couldn't understand how it was related to the UI values.

If the grid_sample_len number seems fine, what I've also seen is that the lightbake_do_sample() call inside that loop takes obviously all the time, and inside, passing through a callback, it calls render_cubemap (in evee_lightprobes), which takes from 10 to 15s each call.
In the render_cubemap it loops all the 6 faces, calling lightbake_render_scene_face, and there you could say that the majority of time is spent by EEVEE_draw_shadows().

EDIT:
Forgot to say, I did the test with a Release build with Debug Info
OS: ArchLinux
CPU: Intel i7 6700K
GPU: NVIDIA GTX 1080

I don't know how much is related but loading the scene in blender shows these on the terminal:
GPU failed to find function node_bsdf_hair
GPUShader: compile error:
0(12226) : error C0000: syntax error, unexpected ',', expecting "::" at token ","

I've debugged it a bit for fun (because I don't know the code, it's the first time I see it), but the part that is taking most of the time is the rendering of the Irradiance grid. The part of the code is in evee_lightcache.c:1198, there's a loop there and the highest loop count to do is the grid_sample_len which is 1200 (which seems calculated from the current LightProbe grid_resolution_x, y, z). I tried to understand why 1200 and where that resolution is taken/calculated but I couldn't understand how it was related to the UI values. If the grid_sample_len number seems fine, what I've also seen is that the lightbake_do_sample() call inside that loop takes obviously all the time, and inside, passing through a callback, it calls render_cubemap (in evee_lightprobes), which takes from 10 to 15s each call. In the render_cubemap it loops all the 6 faces, calling lightbake_render_scene_face, and there you could say that the majority of time is spent by EEVEE_draw_shadows(). EDIT: Forgot to say, I did the test with a Release build with Debug Info OS: ArchLinux CPU: Intel i7 6700K GPU: NVIDIA GTX 1080 I don't know how much is related but loading the scene in blender shows these on the terminal: GPU failed to find function node_bsdf_hair GPUShader: compile error: 0(12226) : error C0000: syntax error, unexpected ',', expecting "::" at token ","

Stefano Bonicatti commented

2019-05-08 16:27:09 +02:00

I've continued to have a look at it (mind, I have no hope of having a fix, I don't have enough knowledge) and I've noticed some things:

Since EEVEE_draw_shadows is the one taking the majority of time (like 80% of the time) and specifically the update of the cascaded shadows, I've noticed that after ~30min a cascade was never deemed as not visible. Maybe it's just a case, but probably it has to do with the fact that the visibility test is done with a plane at local Z = 0 (as the comment on the code states...)?. Sorry if I'm saying a stupid thing but, basically if I understand it right and that the plane for instance is representing the floor, there's no limit of distance on the visibility?
By putting some profiling code around lightbake_do_sample basically I noticed that on my hardware it takes ~10-15 seconds to do one grid sample, and each EEVEE_draw_shadows call is ~1.5s (and there are 6 of them per sample). I've tried to force cascaded shadows to be always invisible, so basically to avoid them to be rendered, and each sample now is down to ~3s, which though at the current resolution (which I eventually found to be taken from the Irradiance Volume object added in the scene) it's still 1200 samples x 3 light bounces x 2 lights aka a lot. Also half of the time for each sample is spent in eevee_lightbake_cache_create
This is more a (noob) question than a discovery, but I was wondering why the entire shadow cubemap is rendered every face of the irradiance cubemap. Namely, EEVEE_draw_shadows is called 6 times and as far as I understand, 6 shadow map cubes are created. Shouldn't those be created once per grid sample? (Also here https://docs.blender.org/manual/en/dev/render/eevee/lightprobes/irradiance_volumes.html#visibility it says that each sample, creates a VSM, which isn't true?).

I've continued to have a look at it (mind, I have no hope of having a fix, I don't have enough knowledge) and I've noticed some things: 1) Since `EEVEE_draw_shadows` is the one taking the majority of time (like 80% of the time) and specifically the update of the cascaded shadows, I've noticed that after ~30min a cascade was never deemed as not visible. Maybe it's just a case, but probably it has to do with the fact that the visibility test is done with a plane at local Z = 0 (as the comment on the code states...)?. Sorry if I'm saying a stupid thing but, basically if I understand it right and that the plane for instance is representing the floor, there's no limit of distance on the visibility? 2) By putting some profiling code around `lightbake_do_sample` basically I noticed that on my hardware it takes ~10-15 seconds to do one grid sample, and each `EEVEE_draw_shadows` call is ~1.5s (and there are 6 of them per sample). I've tried to force cascaded shadows to be always invisible, so basically to avoid them to be rendered, and each sample now is down to ~3s, which though at the current resolution (which I eventually found to be taken from the Irradiance Volume object added in the scene) it's still 1200 samples x 3 light bounces x 2 lights aka a lot. Also half of the time for each sample is spent in `eevee_lightbake_cache_create` 3) This is more a (noob) question than a discovery, but I was wondering why the entire shadow cubemap is rendered every face of the irradiance cubemap. Namely, EEVEE_draw_shadows is called 6 times and as far as I understand, 6 shadow map cubes are created. Shouldn't those be created once per grid sample? (Also here https://docs.blender.org/manual/en/dev/render/eevee/lightprobes/irradiance_volumes.html#visibility it says that each sample, creates a VSM, which isn't true?).

Jeroen Bakker commented

2019-07-09 15:00:14 +02:00

Added subscriber: @Jeroen-Bakker

Jeroen Bakker commented

2019-07-09 15:00:14 +02:00

In #63969#668220, @Smjert wrote:
EDIT:
I don't know how much is related but loading the scene in blender shows these on the terminal:
GPU failed to find function node_bsdf_hair

This is as the shader tree uses Pricipled Hair Shader that is only supported by Cycles.

GPUShader: compile error:
0(12226) : error C0000: syntax error, unexpected ',', expecting "::" at token ","

This is due to an error in the shadergraph of Designedwall1 material, where BSDF socket are linked directly to color socket. This is not supported.

> In #63969#668220, @Smjert wrote: > EDIT: > I don't know how much is related but loading the scene in blender shows these on the terminal: > GPU failed to find function node_bsdf_hair This is as the shader tree uses Pricipled Hair Shader that is only supported by Cycles. > GPUShader: compile error: > 0(12226) : error C0000: syntax error, unexpected ',', expecting "::" at token "," This is due to an error in the shadergraph of `Designedwall1` material, where BSDF socket are linked directly to color socket. This is not supported.

Richard Antalik commented

2020-01-06 12:13:40 +01:00

Added subscriber: @iss

Richard Antalik commented

2020-01-06 12:13:40 +01:00

Changed status from 'Needs Triage' to: 'Confirmed'

Richard Antalik commented

2020-01-06 12:13:40 +01:00

Baking still freezes (unable to stop it) in 2.82 56ef761381

Baking still freezes (unable to stop it) in 2.82 56ef761381ec

Dalai Felinto commented

2020-05-12 16:30:29 +02:00

Added subscriber: @dfelinto

Dalai Felinto commented

2020-05-12 16:30:29 +02:00

Changed status from 'Confirmed' to: 'Archived'

Dalai Felinto closed this issue

2020-05-12 16:30:29 +02:00

Dalai Felinto self-assigned this 2020-05-12 16:30:29 +02:00

Dalai Felinto commented

2020-05-12 16:30:29 +02:00

Thanks for the report. Unfortunately the sample file is too complex (>600MB) which is too time consuming for us to track down, we require the bug reporter to narrow down the problem.

Normally .blend files can be simplified by removing most objects and disabling settings, until the problem reveals itself more clearly.

If you can reproduce this issue with a simple file feel free to report it as a new issue (you can mention this report).

Thanks for the report. Unfortunately the sample file is too complex (>600MB) which is too time consuming for us to track down, we require the bug reporter to narrow down the problem. Normally .blend files can be simplified by removing most objects and disabling settings, until the problem reveals itself more clearly. If you can reproduce this issue with a simple file feel free to report it as a new issue (you can mention this report).

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

[Blender 2.8] Evee Baking Indirect Lighitng causes scene to crash #63969