## NVIDIA OpenCL Drivers & News

Discussions related to GPU Acceleration in LuxRender

### Re: NVIDIA OpenCL Drivers & News

Our wet dream comes true : http://barefeats.com/gtx980ti.html

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

### Re: NVIDIA OpenCL Drivers & News

Hello,

are there any positive experiences with latest nvidia drivers under Windows (8.1 64bit is my version)?
I have just updated to the latest drivers (361.43) for my GTX 970, but I'm having problems, like crashes or extremely long kernel compilation times even for very simple scenes.
The previous driver version I had was 347.xx (the one shipped with the card) and it worked fine, even if the enabled OpenCL version was 1.1 (the latest enables OpenCL 1.2).

Thanks,
Ale

Edit: here you can find an extremely simple test scene (just 2 objects, same material, one with a texture).
These are the kernel compilation times reported (first time, then they all drop to 40 - 90 ms):
PathOCL without texture: 875 ms
BiasedPathOCL without texture: 3157 ms
BiasedPathOCL with texture in diffuse channel: 33719 ms
BiasedPathOCL with texture in bump channel: > 300000 ms (more than 5 minutes!)

I think this is a driver problem, as I don't remember having to wait so long with the older version.
I will try what happens with older versions, but it would be good to know others' experiences...

untitled.zip
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v382.05
acasta69

Posts: 299
Joined: Fri Dec 20, 2013 3:18 am

### Re: NVIDIA OpenCL Drivers & News

I confirm insane kernelcompiletimes with procedurals in any tex channel.
Linux 352.63 here, did not checked if this is a driver regression.

I also found the newer drivers are a tad slower in render than 346.xx.

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

### Re: NVIDIA OpenCL Drivers & News

Thanks Jens,
at least I know it's not some strange problem in my system...
I'll revert to the older version for the moment.
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v382.05
acasta69

Posts: 299
Joined: Fri Dec 20, 2013 3:18 am

### Re: NVIDIA OpenCL Drivers & News

acasta69 wrote:Thanks Jens,
at least I know it's not some strange problem in my system...
I'll revert to the older version for the moment.

Please lemme know if 364 reverts to short compiletimes.
Don't wanna fiddle with drivers right now, so will wait for your result.

In the meantime i try to investigate if there is a driverchange, that needs attention aka code tweaking.

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

### Re: NVIDIA OpenCL Drivers & News

jensverwiebe wrote:Please lemme know if 364 reverts to short compiletimes.

364? I don't see it yet on nvidia drive site...
I reverted to 347.88 and compile times are definitely shorter (about 20 seconds for the BiasPathOCL with bump texture), but I'm also having some compile error that I'm sure I did not have before:

Code: Select all
[LuxCore][687.047] [PathOCLBaseRenderThread::0] Compiling kernels[Lux 2016-Jan-04 22:06:58] Rendering aborted: clBuildProgram

Using some of the "overclock" options removes the error, but I did not try all possible combinations. If it can help, let me know and I'll try.
And thanks for looking into this!
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v382.05
acasta69

Posts: 299
Joined: Fri Dec 20, 2013 3:18 am

### Re: NVIDIA OpenCL Drivers & News

346.xx i meant as in former post, this was the most reliable version i ever used. Afaik .96 or so ( with ti support ).

EDIT: Can't help it, it also must be cause of codechanges in luxrays. Now a voronoi bumptex needs 668995ms to compile
where it was 120000 to 180000 before ( even to long either ). Still NVidia 352 driver.
Verified its 3 - 4 times longer after 03283d1

Also tried
Code: Select all
 sudo nvidia-smi -i <gpu_id> -pm 1
to set the persistense mode to keep compiletimes down, does not help too much.

To me it looks as NVidia compiler does not like the new dynamic modes.

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

### Re: NVIDIA OpenCL Drivers & News

Thanks Jens, I am trying older and older drivers, at the moment without success: I still get huge compile times for BiasPathOcl with bump texture.
jensverwiebe wrote:EDIT: Can't help it, it also must be cause of codechanges in luxrays. Now a voronoi bumptex needs 668995ms to compile
where it was 120000 to 180000 before ( even to long either ). Still NVidia 352 driver.
Verified its 3 - 4 times longer after 03283d1

Ok, this can explain why everything seemed to behave differently even with the same driver version as before: I updated to the latest buildbot only a few days ago, but did not use OpenCL modes initially, so I didn't notice anything immediately.
I hope this problem can be fixed... and I'll let know if something more come out from testing.

Edit: at the moment I can only add that the scene attached above does not compile with PathOCL unless I also add the options:
Code: Select all
opencl.kernel.options = -cl-fast-relaxed-math; -cl-mad-enable; -cl-strict-aliasing;

With BiasPathOCL it compiles with and without options, but with the same very long time.
At the moment I am with Nvidia driver 347.25.
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v382.05
acasta69

Posts: 299
Joined: Fri Dec 20, 2013 3:18 am

### Re: NVIDIA OpenCL Drivers & News

Yes, its known that NV compiler is casewise slower that the AMD cl compiler, but now for me 10 min * 3 gpu is unworkable

Lets see if we find the bottleneck.

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

### Re: NVIDIA OpenCL Drivers & News

acasta69 wrote:Thanks Jens, I am trying older and older drivers, at the moment without success: I still get huge compile times for BiasPathOcl with bump texture.
jensverwiebe wrote:EDIT: Can't help it, it also must be cause of codechanges in luxrays. Now a voronoi bumptex needs 668995ms to compile
where it was 120000 to 180000 before ( even to long either ). Still NVidia 352 driver.
Verified its 3 - 4 times longer after 03283d1

Ok, this can explain why everything seemed to behave differently even with the same driver version as before: I updated to the latest buildbot only a few days ago, but did not use OpenCL modes initially, so I didn't notice anything immediately.
I hope this problem can be fixed... and I'll let know if something more come out from testing.

Edit: at the moment I can only add that the scene attached above does not compile with PathOCL unless I also add the options:
Code: Select all
opencl.kernel.options = -cl-fast-relaxed-math; -cl-mad-enable; -cl-strict-aliasing;

With BiasPathOCL it compiles with and without options, but with the same very long time.
At the moment I am with Nvidia driver 347.25.

I now tested your .blend and it works fine here: [LuxCore][25.960] [PathOCLBaseRenderThread::0] Kernels compilation time: 3571ms Both path and biasedpath btw..
Linux Mint 17.3, NV 352.63 driver.
Why the heck this clouds bumpmap compiles so fast i dunno atm.

Jens

jensverwiebe

Posts: 3429
Joined: Wed Apr 02, 2008 4:34 pm

PreviousNext