Indigo 3.0 - OpenCL error

[Bosseye]

Ok, Indigo 3.0 trial:

If I enable GPU acceleration then I only get a black screen, no image, SPP does not increase despite render time increasing. Works fine if GPU acceleration is off.

CUDA-Z is reporting CUDA compute capability of my card (Nvidia Quadro 4000) as 2.0, so should work considering Indigo apparently needs at least compute 1.1

[ENSLAVER]

Check the render log, try with an extremely simple scene.

[Bosseye]

Render log returns this using a basic scene containing a cube only - presumably the last line is the most important:


Detecting available GPU(s)...
Device number: 0
Name: Quadro 4000
Memory: 2014 MB
Type: OpenCL

Scene file path: 'C:\Users\adown\AppData\Roaming/Indigo Renderer/default.igs'
Sky zenith luminance: 3387.095215 cdm^-2
Rendering with GPU acceleration
Forcing bidirectional=false, hybrid=false, batched=true
Light luminous flux (geometry name=EnvSphereGeometry): 5.746E+009 lm
Building Object Tree...
1 objects.
calcing root AABB.
AABB: (0.236831, 0.292659, 0.000000), (3.618678, 5.213458, 2.038000)
max tree depth: 2
reserving N nodes: 1(8 B)
total nodes used: 1 (8 B)
total leafgeom size: 1 (8 B)
Finished building tree.
AutoFocus: setting camera focus distance to 5.76820 m.
Num buffer layers: 1
Master buffer size: 20.960 MB
Auxiliary buffer size: 48.307 MB
Settings:
Image width: 958 px
Image height: 466 px
Internal image width: 1932 px
Internal image height: 948 px
Verbose: false
Display period: 10.00000 s
Image save period: 60.00000 s
Frame upload period: 40.00000 s
Splat filter: FastBox
Downsize filter: mn_cubic, blur=0.33330, ring=0.33330, support=4.00000px
Render region: false
Supersample factor: 2
Metropolis: false
Bidirectional: false
Hybrid: false
Aperture diffraction: false
Post-process diffraction: true
Auto setting number of threads to 8.
Finished initialisation (Time Taken: 0.07843 s)
Time to render thread start: 0.07898 s.
Starting threads... (num threads:
OpenCL build log: :5: error: ilegal address space on image or sampler type
__constant sampler_t l1lll11l1l1 = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST; inline float4 ll1l1l11lll(read_only image2d_t texture, const uint ll1l1lllll1) { const uint y = ll1l1lllll1 / OMTX; const uint x = ll1l1lllll1 % OMTX; return read_imagef(texture, l1lll11l1l1, (int2)(x, y)); }

Error initialising GPU ray tracer: Build failed

Not sure what that means...

[Bosseye]

Updated Nvidia Drivers to latest version 275.36 - works like a charm now. Also in the GPU dialogue box its showing as Quadro 4000 (CUDA) rather than Quadro 4000 (OpenCL).

Much much faster too

Thanks for your help.

[Zom-B]

Updated Nvidia Drivers to latest version 275.36 - works like a charm now. Also in the GPU dialogue box its showing as Quadro 4000 (CUDA) rather than Quadro 4000 (OpenCL).

Latest Drivers are 275.50 but don't let us get picky

Does your GPU also has a OpenCL option in the dropdown? It should have one! Does that work now?

[Bosseye]

Pshaw, whats in a couple of numbers eh?

Hmmm, it does have the OpenCL listed as an option in the drop down box - but it doesn't work.

OpenCL mode renders a strangely distorted version of the model, but does so at a crazy level, sometimes around 12000k compared to around 4000k for the same scene in CUDA mode. See below.

For the black background scenes as an example:
CUDA - 2 minute render, 370spp, 4097k
OpenCL - 40 second render, 370spp, 11900k

[Bosseye]

Anyone got any thoughts on the OpenCL issue?