Blender gives weird ouput when baking (4096*4096) resolution on GPU #41222

New Issue

Dalai Felinto · 2014-07-28T12:37:48+02:00

RadiationG0D commented

2014-07-28 12:37:48 +02:00

(Migrated from localhost:3001)

System Information
Desktop PC:
Win 7 Home Premium
CPU: AMD FX-4170 (Quad, 4.20 GHz)
GPU: (Asus) GTX 560 Ti DirectCUII Top

Laptop:
Win 7 Professional
CPU: Intel Core i7 3520m (2,9GHz)
GPU: Nvidia Quadro K2000m

Blender Version
Broken: 2.71 9337574

Short description of error
Blender gives weird ouput when baking uv-map on GPU at a resolution of 4096*4096. CPU works normal but takes a lot more time than GPU would've.
Higher resolutions are not tested as of now.

Output can be seen here

Exact steps for others to reproduce the error
Bake the cube in the blend file with GPU on a resolution of 4096*4096

brokenuvcube.blend

**System Information** Desktop PC: Win 7 Home Premium CPU: AMD FX-4170 (Quad, 4.20 GHz) GPU: (Asus) GTX 560 Ti DirectCUII Top Laptop: Win 7 Professional CPU: Intel Core i7 3520m (2,9GHz) GPU: Nvidia Quadro K2000m **Blender Version** Broken: 2.71 9337574 **Short description of error** Blender gives weird ouput when baking uv-map on GPU at a resolution of 4096*4096. CPU works normal but takes a lot more time than GPU would've. Higher resolutions are **not** tested as of now. Output can be seen [here](http://i.stack.imgur.com/nqIOJ.jpg) **Exact steps for others to reproduce the error** Bake the cube in the blend file with GPU on a resolution of 4096*4096 [brokenuvcube.blend](https://archive.blender.org/developer/F99994/brokenuvcube.blend)

RadiationG0D commented

2014-07-28 12:37:49 +02:00

(Migrated from localhost:3001)

Changed status to: 'Open'

RadiationG0D commented

2014-07-28 12:37:49 +02:00

(Migrated from localhost:3001)

Added subscriber: @RadiationG0D

RadiationG0D commented

2014-07-28 12:50:37 +02:00

(Migrated from localhost:3001)

The bake was attempted with cycles

Dalai Felinto was assigned by Sergey Sharybin

2014-07-28 14:00:47 +02:00

Sergey Sharybin commented

2014-07-28 14:00:47 +02:00

Added subscriber: @Sergey

Sergey Sharybin commented

2014-07-28 14:00:47 +02:00

For some reason i can't reproduce the issue here on linux, but form the code side it's really bad idea to multiply image width by height and store it in int, size_t is to be used in this cases.

I've started moving to size_t there http://www.pasteall.org/53120/diff but the issue is -- we can't pass it via API. So afraid the only way is to pass width,height to the baker and multiply them in the baker.

For some reason i can't reproduce the issue here on linux, but form the code side it's really bad idea to multiply image width by height and store it in int, size_t is to be used in this cases. I've started moving to size_t there http://www.pasteall.org/53120/diff but the issue is -- we can't pass it via API. So afraid the only way is to pass width,height to the baker and multiply them in the baker.

Dalai Felinto commented

2014-07-28 21:14:37 +02:00

Added subscriber: @ThomasDinges

Dalai Felinto commented

2014-07-28 21:14:37 +02:00

@Sergey at the moment we can bake an object with multiple materials/target images at once, which means there is no real width and height. We can get some fictional values, but often we will end up with width being 1 or 2. So the problem will likely still be there (assuming that's the issue). (that would work for the case reported here though).

See D688 - in particular calculate_buffer_dimension() in object_bake_api.c
(@RadiationG0D can you see if this patch fix the issue for you)?

Another option is to not pass num_pixels, and tweak rna_BakePixel_next_get() to return NULL, so in Cycles we populate the baking data dynamically until we get a NULL from bp->next (in populate_bake_data).

Or ... to find a way to pass size_t to python. Isn't 'k' enough for that?
https://docs.python.org/3/c-api/arg.html

@ThomasDinges can you reproduce this issue in your CUDA-friendly machine?

@Sergey at the moment we can bake an object with multiple materials/target images at once, which means there is no real width and height. We can get some fictional values, but often we will end up with width being 1 or 2. So the problem will likely still be there (assuming that's the issue). (that would work for the case reported here though). See [D688](https://archive.blender.org/developer/D688) - in particular calculate_buffer_dimension() in object_bake_api.c (@RadiationG0D can you see if this patch fix the issue for you)? Another option is to not pass num_pixels, and tweak rna_BakePixel_next_get() to return NULL, so in Cycles we populate the baking data dynamically until we get a NULL from bp->next (in populate_bake_data). Or ... to find a way to pass size_t to python. Isn't 'k' enough for that? https://docs.python.org/3/c-api/arg.html @ThomasDinges can you reproduce this issue in your CUDA-friendly machine?

RadiationG0D commented

2014-07-28 21:20:20 +02:00

(Migrated from localhost:3001)

Added subscriber: @dfelinto

RadiationG0D commented

2014-07-28 21:20:20 +02:00

(Migrated from localhost:3001)

@dfelinto - I'm afraid I don't know how to test the patch..? I was directed here from blender.stackexchange.com and though intrested in programming and such, I'm not sure what you would like me to do. I suppose it would not be enough to download the latest blender from blender.org?

blender-admin commented

2014-07-28 21:43:32 +02:00

This issue was referenced by 1607bcebc65fd5296f9f695e76580a1ae7ade9a2

Dalai Felinto commented

2014-07-28 21:45:28 +02:00

Changed status from 'Open' to: 'Resolved'

Dalai Felinto closed this issue

2014-07-28 21:45:28 +02:00

Dalai Felinto commented

2014-07-28 21:45:28 +02:00

Closed by commit 1607bcebc6.

Dalai Felinto commented

2014-07-28 22:14:18 +02:00

Changed status from 'Resolved' to: 'Open'

Dalai Felinto reopened this issue

2014-07-28 22:14:18 +02:00

Dalai Felinto commented

2014-07-28 22:14:18 +02:00

closed by 'mistake' while pushing the test to build bot. re-opening now.

@RadiationG0D can you try downloading a build from here [ link ]? This has the patch applied to it.

(note, the build may be overridden any soon, so please download it as soon as you can).

closed by 'mistake' while pushing the test to build bot. re-opening now. @RadiationG0D can you try downloading a build from here [ [link ](https://builder.blender.org/download/testbuild-blender-2.71-1607bce-win64.zip) ]? This has the patch applied to it. (note, the build may be overridden any soon, so please download it as soon as you can).

Dalai Felinto commented

2014-07-28 22:18:24 +02:00

For comparison you can test the above build against a more recent one, such as [ link ].

For comparison you can test the above build against a more recent one, such as [ [link ](https://builder.blender.org/download/blender-2.71-34937f6-win64.zip) ].

RadiationG0D commented

2014-07-28 22:31:35 +02:00

(Migrated from localhost:3001)

The ouptut is better, but far from good. Also, the progressbar seemed to show progress now which it has not before :)

Same .blend:
Output:

The ouptut is better, but far from good. Also, the progressbar seemed to show progress now which it has not before :) Same .blend: Output: ![4096.png](https://archive.blender.org/developer/F100062/4096.png)

Dalai Felinto commented

2014-07-28 22:36:17 +02:00

Also, the progressbar seemed to show progress now which it has not before :)

This is unrelated, it's one of the new features for the upcoming Blender 2.72.

How about the other pass types? I think you are hitting some memory limit in your system, though I don't think it's relate to the patch. Did you try the other build too? (the more recent one from master).

> Also, the progressbar seemed to show progress now which it has not before :) This is unrelated, it's one of the new features for the upcoming Blender 2.72. How about the other pass types? I think you are hitting some memory limit in your system, though I don't think it's relate to the patch. Did you try the other build too? (the more recent one from master).

RadiationG0D commented

2014-07-28 22:39:38 +02:00

(Migrated from localhost:3001)

Other pass types, like bake types?
I've set up cache at 4096MB at maximum though my gpu has only 1GB I think. (System has 8GB RAM otherwise)
I will try the other build now...

Other pass types, like bake types? I've set up cache at 4096MB at maximum though my gpu has only 1GB I think. (System has 8GB RAM otherwise) I will try the other build now...

ronan ducluzeau commented

2014-07-28 22:40:10 +02:00

Added subscriber: @zeauro

ronan ducluzeau commented

2014-07-28 22:40:10 +02:00

No problem to bake 4096 on ubuntu 14.04 with CUDA 6 Nvidia driver 340.24 with Titan Black

RadiationG0D commented

2014-07-28 22:46:10 +02:00

(Migrated from localhost:3001)

The second link gives similar output.

Wolfgang Faehnle commented

2014-07-28 23:10:56 +02:00

Added subscriber: @mib2berlin

Wolfgang Faehnle commented

2014-07-28 23:10:56 +02:00

Hi, bake work with my GTX 760 4 GB but not with my GTX 560Ti 448 1.28 GB.
Result with GTX 560Ti 448:

http://www.pasteall.org/pic/show.php?id=74660

Opensuse 13.1/64
Intel i5 3770K
GTX 760 4 GB (Display)
GTX 560Ti 1.28 GB 448 Cores
Driver 340.24

Cheers, mib

Hi, bake work with my GTX 760 4 GB but not with my GTX 560Ti 448 1.28 GB. Result with GTX 560Ti 448: http://www.pasteall.org/pic/show.php?id=74660 Opensuse 13.1/64 Intel i5 3770K GTX 760 4 GB (Display) GTX 560Ti 1.28 GB 448 Cores Driver 340.24 Cheers, mib

Sergey Sharybin commented

2014-07-29 10:12:23 +02:00

I can easily bake 4K x 4K texture on my gt520m with 1gig ram, but that's on linux, will test win7 later a bit. It looks like some particular setup issue or something up to the particular kernel (can't redo the issue forcing sm_20 for usage tho).

So some questions just to be sure:

Can you consistently reproduce the issue o both desktop and laptop?
Did you try installing latest video drivers and CUDA toolkit?
Am i right lower resolution rendering works just fine to you?

Will also check gtx560ti this evening from home, maybe that'd give some clues as well.

I can easily bake 4K x 4K texture on my gt520m with 1gig ram, but that's on linux, will test win7 later a bit. It looks like some particular setup issue or something up to the particular kernel (can't redo the issue forcing sm_20 for usage tho). So some questions just to be sure: - Can you consistently reproduce the issue o both desktop and laptop? - Did you try installing latest video drivers and CUDA toolkit? - Am i right lower resolution rendering works just fine to you? Will also check gtx560ti this evening from home, maybe that'd give some clues as well.

RadiationG0D commented

2014-07-29 13:59:50 +02:00

(Migrated from localhost:3001)

I can consistently reproduce the problem on the GTX but the quadro seems to be working fine today.

I'm currently running driver v. 337.88 on my GTX 560 Ti (latest). The laptop drivers are probably not the latest but I'm not sure since the update mechanism for Quadro is a bit diffrent.

I have not tried the CUDA toolkit, but I've never heard of it before though..

Lower resolution solves the problem, yes.

I can consistently reproduce the problem on the GTX but the quadro seems to be working fine today. I'm currently running driver v. 337.88 on my GTX 560 Ti (latest). The laptop drivers are *probably* not the latest but I'm not sure since the update mechanism for Quadro is a bit diffrent. I have not tried the CUDA toolkit, but I've never heard of it before though.. Lower resolution solves the problem, yes.

Wolfgang Faehnle commented

2014-07-29 15:17:57 +02:00

Hi, try again with own file and can reproduce error on GTX 560Ti 448.
Just for info, the GTX 560Ti 448 is a cut back GTX 570 and a sm_20 card.
Regular GTX 560Ti is a sm_21 card.

Cheers, mib

Hi, try again with own file and can reproduce error on GTX 560Ti 448. Just for info, the GTX 560Ti 448 is a cut back GTX 570 and a sm_20 card. Regular GTX 560Ti is a sm_21 card. Cheers, mib

Dalai Felinto commented

2014-07-29 16:57:57 +02:00

@Sergey the problem is that we copy a big chunk of memory and pass that to CUDA. See D690 (it bakes a chunk of 3k * 3k at a time). This works fine in mi2berlin's GPUs.

That said I think the fix should be in device_cuda.cpp::task_add(). It would make more sense to me if for baking CUDA could split the task according to the memory size. So, to do the equivalent of D690 but internally, in the device code.

[about size_t I agree that we should change that, but let's keep that discussion in the patch - D688 - since I don't think it's entirely related to this issue]

@Sergey the problem is that we copy a big chunk of memory and pass that to CUDA. See [D690](https://archive.blender.org/developer/D690) (it bakes a chunk of 3k * 3k at a time). This works fine in mi2berlin's GPUs. That said I think the fix should be in device_cuda.cpp::task_add(). It would make more sense to me if for baking CUDA could split the task according to the memory size. So, to do the equivalent of [D690](https://archive.blender.org/developer/D690) but internally, in the device code. [about size_t I agree that we should change that, but let's keep that discussion in the patch - [D688](https://archive.blender.org/developer/D688) - since I don't think it's entirely related to this issue]

blender-admin commented

2014-08-05 18:50:50 +02:00

This issue was referenced by a48b372b04

This issue was referenced by a48b372b04421b00644a0660bfdf42229b5ffceb

Dalai Felinto commented

2014-08-05 18:51:18 +02:00

Changed status from 'Open' to: 'Resolved'

Dalai Felinto closed this issue

2014-08-05 18:51:18 +02:00

Dalai Felinto commented

2014-08-05 18:51:18 +02:00

Closed by commit a48b372b04.

Closed by commit a48b372b04.

Wolfgang Faehnle commented

2014-08-06 08:06:19 +02:00

Hi, make another tests with buildbot with commit a48b372b04 from dfelinto and got same weird output with GTX 560Ti 448 1.28 GB and 4096x4096 pixel.
3072x3072 is working.
Testing my own build with change device_cuda.cpp to 512 * 512 or 256 * 256 give same error.

Testfile brokenuvcube.blend and own file.

Opensuse 13.1/64
Intel i5 3770K
GTX 760 4 GB (Display)
GTX 560Ti 1.28 GB 448 Cores
Driver 340.24

Cheers, mib

Hi, make another tests with buildbot with commit a48b372b04 from dfelinto and got same weird output with GTX 560Ti 448 1.28 GB and 4096x4096 pixel. 3072x3072 is working. Testing my own build with change device_cuda.cpp to 512 * 512 or 256 * 256 give same error. Testfile brokenuvcube.blend and own file. Opensuse 13.1/64 Intel i5 3770K GTX 760 4 GB (Display) GTX 560Ti 1.28 GB 448 Cores Driver 340.24 ![4097.jpeg](https://archive.blender.org/developer/F101583/4097.jpeg) Cheers, mib

Sergey Sharybin commented

2014-08-06 09:48:31 +02:00

Changed status from 'Resolved' to: 'Open'

Sergey Sharybin reopened this issue

2014-08-06 09:48:31 +02:00

Sergey Sharybin commented

2014-08-06 09:48:31 +02:00

Worked for me yesterday, will check with buildbot.