@Sergey Sharybin (sergey) maybe you have hardware to check this issue? :/

In T49019#384926, @Sergey Sharybin (sergey) wrote:

• Does the issue happen with previous driver versions?

Yes

• Does the issue depend on tile size?

No. I tried with lower tile size and the same happens.

• Does the issue happen with previous Blender releases (2.76, 2.77a) ?

This is render result with 2.76, 2.77a with the same file. It seams that light is not computed right. Render is not the same as with 2.77.3

No, I couldn't reproduce this bug, tested with an R9 380 on Windows.

Thanks. Than it's probably driver issue with AMD 7870.

Occurs on AMD HD7770 (GCN 1.0) on Windows 10 64bit and AMD Crimson drivers 16.8.1, using Blender 2.77a but only when hair is enabled and when the experimental GPU Rendering feature set is also enabled.

Here is a simple scene I can share where the problem easily shows.

I have reported this to AMD's driver team aswell.

If helpful i could provide a Teamviewer Access to a GCN 1.0 System.

OpenCL error (-61): CL_INVALID_BUFFER_SIZE in clCreateBuffer  
OpenCL error: CL_INVALID_MEM_OBJECT in clEnqueueWriteBuffer(cqCommandQueue, CL_MEM_PTR(mem.device_pointer), CL_TRUE, 0, size, (void*)mem.data_pointer, 0, NULL, NULL)

Today it took about 4 minutes to cancel the render.
With 2.76b, Cycles cancels render almost immediately. Scene had about 100 000 faces.

I think it has something to do how driver operates the GPU memory.
Flushing the memory takes ages.

Today's build didn't crash but hair rendering is still fault. I post console log if it helps.
Is it possible to implement test variables for console log to detect this issue's source ?

There is issue with Interpolated mode. In the process of rendering, Blender freezes with Interpolated mode. I captured the screen when it froze.

That's weird, we did not do any explicit fixes for the hair. Could be the commit which changed the internal data layout tho.

For the freezing, did you try setting tiles to some really low value (like 64x64) and rendering few samples (like, 1 sample, if that works 10 samples..) ?

In T49019#390364, @Sergey Sharybin (sergey) wrote:

That's weird, we did not do any explicit fixes for the hair. Could be the commit which changed the internal data layout tho.

Whatever you changed in this release, please keep it.

For example, first 2.78 RC1 build rendered this (Crimson 16.8.2 hotfix 2):

As subdivision level of the mesh goes up, there are more tesselation artifacts if the hair is present on the scene.

Fiji and Polaris cards didn't have this issue, only 7xxx series. Could it be connected with GPU memory and internal data layout which you said you changed?
Or maybe is on behalf of AMD's 16.8.3 drivers.

For the freezing, did you try setting tiles to some really low value (like 64x64) and rendering few samples (like, 1 sample, if that works 10 samples..) ?

Thank you. I'll try.

@Sergey Sharybin (sergey) I did't set up the GPU compute properly.
The issue is still there in the build 9.9.2016 with 16.8.3.

I tried various tiled sizes:
default, 64×60, 512×512 and the same weird shading is present.

I have the same problem with HD7950.
Driver: Crimson 15.12 (сan test with latest version if needed)

blender-2.78-rc2-windows64
GPU:

CPU:

Blender 2.77a (2016-04-05) hash: abf6f08
GPU:

CPU:
identical to GPU

@John (Jns_76) please test newest drivers
http://support.amd.com/en-us/download/desktop?os=Windows+10+-+64
16.9.1

Also, report to AMD http://www.amd.com/report because apparently Fiji and Polaris cards are not affected with this.
AMD is reworking issues from the latest versions of drivers so it's better to give feedback.

Please, test this file https://developer.blender.org/T49371
I think this 2 issues could have some connections because if you have more subdivision levels (standard or new adaptive microdisplacement), there are more tesselation artifacts/render result is darken.

In T49019#392438, @soulfire (vejn) wrote:

@John (Jns_76) please test newest drivers

Installed. Now everything is broken. Great.
whql-win7-64bit-radeon-software-crimson-16.9.1-sep13

blender-2.78-rc2-windows64
sun strength-50

Blender 2.77a (2016-04-05) hash: abf6f08
sun strength-50

Blender both versions
sun strength-1

I don't even sure how to report this...

Testing

//
GPU render with blender-2.78-78c380d-win64 build (you can see weird shading on sphere and in viewport comparing to CPU render)

Still black screen and hair PS issue with blender-2.78-3e460b6-win64.zip and Crimson 10.1 on AMD 7870.

I'm getting the same issue when I try to render hair in 3D view, but it render well in final render.

My system: Rx 480 + Rx 470

It happens in each one alone and with both together with or without Crossfire:

OpenCL error (-61): CL_INVALID_BUFFER_SIZE in clCreateBuffer

@Renato Meneses (renatomeneses) can you please create a new bug report?

Here you go:
https://developer.blender.org/T51755