Slow GPU bake and blown out baking cycles render #46207

New Issue

Delo Lat · 2015-09-22T12:00:12+02:00

Delo Lat commented

2015-09-22 12:00:12 +02:00

System Information
Card name: AMD Radeon (TM) R9 Fury Series
Card name: AMD FirePro W8100 (FireGL V)
Operating System: Windows Embedded 8.1 Industry Pro 64-bit (6.3, Build 9600) (9600.winblue_ltsb.150715-0840)
Processor: Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz (12 CPUs), ~3.3GHz
Memory: 32768MB RAM

Blender Version
Broken: 2.7.6rc1 and 2.7.5.4

Short description of error
When baking in GPU compute am receiving blown out Ambient Occlusion and Normal maps. Works fine in CPU.
when baked with CPU: http://snag.gy/BlvA7.jpg
when baked with GPU: http://snag.gy/SsJI3.jpg

Exact steps for others to reproduce the error
Switch mode to cycles render. Add Suzanne smart uv unwrap her. Add image texture in node editor, and create and use a texture. Switch device to GPU compute in render properties. Switch Bake type to Ambient Occlusion or Normals.

**System Information** Card name: AMD Radeon (TM) R9 Fury Series Card name: AMD FirePro W8100 (FireGL V) Operating System: Windows Embedded 8.1 Industry Pro 64-bit (6.3, Build 9600) (9600.winblue_ltsb.150715-0840) Processor: Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz (12 CPUs), ~3.3GHz Memory: 32768MB RAM **Blender Version** Broken: 2.7.6rc1 and 2.7.5.4 **Short description of error** When baking in GPU compute am receiving blown out Ambient Occlusion and Normal maps. Works fine in CPU. when baked with CPU: http://snag.gy/BlvA7.jpg when baked with GPU: http://snag.gy/SsJI3.jpg **Exact steps for others to reproduce the error** Switch mode to cycles render. Add Suzanne smart uv unwrap her. Add image texture in node editor, and create and use a texture. Switch device to GPU compute in render properties. Switch Bake type to Ambient Occlusion or Normals.

Delo Lat commented

2015-09-22 12:00:12 +02:00

Changed status to: 'Open'

Delo Lat commented

2015-09-22 12:00:12 +02:00

Added subscriber: @edisto

blender-admin commented

2015-09-22 12:00:12 +02:00

#48130 was marked as duplicate of this issue

Delo Lat commented

2015-09-22 20:18:34 +02:00

I realized after posting that techs might not be familiar with baking. So I attached a premade file that just requires you to hit bake within the render properties under bake tab.

bakeTest2.blend

I'm guessing this bug probably has something to do with my Radeon Fury X.

Edit
The moment I posted that comment I remembered I could go to system and use a single card. After checking It does it on both cards. So the Radeon Fury X and Firepro w8100.

I realized after posting that techs might not be familiar with baking. So I attached a premade file that just requires you to hit bake within the render properties under bake tab. [bakeTest2.blend](https://archive.blender.org/developer/F237500/bakeTest2.blend) I'm guessing this bug probably has something to do with my Radeon Fury X. **Edit** The moment I posted that comment I remembered I could go to system and use a single card. After checking It does it on both cards. So the Radeon Fury X and Firepro w8100.

Sergey Sharybin commented

2015-09-23 12:18:50 +02:00

Added subscriber: @Sergey

Sergey Sharybin commented

2015-09-23 12:18:50 +02:00

Can confirm OpemCL is producing totally wrong result. Not totally sure yet why, will need to have a closer look.

Sergey Sharybin self-assigned this 2015-09-23 12:19:23 +02:00

Brecht Van Lommel commented

2016-02-14 20:44:58 +01:00

Added subscriber: @brecht

Brecht Van Lommel commented

2016-02-14 20:44:58 +01:00

Findings so far:

Baking with AA sample 1 works ok, only with multiple samples there is a problem.
Adding clFinish() between baking samples does not solve the issue (though we should add it)
Marking the output memory as MEM_READ_WRITE does not solve the issue (though we should do that)

Commenting out some parts of the code in the kernel seems to make it work, though I'm not sure if that's avoiding some kind of race condition, working around a compiler bug, or something else.

Findings so far: * Baking with AA sample 1 works ok, only with multiple samples there is a problem. * Adding `clFinish()` between baking samples does not solve the issue (though we should add it) * Marking the output memory as `MEM_READ_WRITE` does not solve the issue (though we should do that) Commenting out some parts of the code in the kernel seems to make it work, though I'm not sure if that's avoiding some kind of race condition, working around a compiler bug, or something else.

Brecht Van Lommel commented

2016-04-13 01:42:04 +02:00

Added subscriber: @VeloFX