## NVIDIA OpenCL Drivers & News

Discussions related to GPU Acceleration in LuxRender

### Re: NVIDIA OpenCL Drivers & News

So far i had no succes in trying to control optimizations in parts of the kernel.
There is only the #pragma unroll right now, but this only gave me insignificant
benefit for procedural textures. It seems there is no way to say let just a part with
lot of switch statements unoptimized ( where optimized would not mean anything
but longer compile of kernel anyway ).

It looks as one have to make those data parallel by for example split kernels here again ?
If anyone else has another idea lemme know.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News + WARNING

Actually tried cuda8 + sdk and got also new drivers in the installation.
367.48-0ubuntu1 is a real catastrophy. I get crashes i never had, also in luxmark.

My new friend is:
Code: Select all
[LuxCore][75.921] [PathOCLBaseRenderThread::1] Compiling kernels [LuxCore][75.954] [PathOCLRenderThread::0] Rendering thread ERROR: clFlush(-9999)terminate called after throwing an instance of 'cl::Error'  what():  clEnqueueReadBuffer

Stay away from this 367.44 and .48 drivers.

BTW: The cuda 8rc included 361.xx drivers where okay.

EDIT: after lot of testing went back to 352 drivers, faster and more reliable.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

Hi
I ordered a gtx 1080 for testing now ( this time NV does not seem to have interest in us )
It will ariive tomorrow hopefully.
I had to install cuda8 and pascal capable drivers again and 'am now @ 367.48 again.
Unfortunately the luxmark mic scene again always crashes on my maxwell cards and i dunno how to track this down atm..
Hotel and luxball work okay.

In last post there is the error occuring when using luxmark scene directly, here additionally with "LuxVR"
Code: Select all
2016-10-14 14:20:23 - [LuxCore][75.016] [PathOCLBaseRenderThread::1] Compiling kernels 2016-10-14 14:20:23 - [LuxCore][75.055] [RTPathOCLRenderThread::0] Rendering thread ERROR: clFinish(CL_INVALID_COMMAND_QUEUE)2016-10-14 14:20:23 - terminate called after throwing an instance of 'cl::Error'2016-10-14 14:20:23 -   what():  clFinish2016-10-14 14:20:24 - Aborted

Perhaps someone has an idea ? Or again waiting for a driver fix ?
Why the heck must driver updates always break older ( older ? ) gpu ?

EDIT: for cuda in blender cycles, cuda8 and new drivers are a win so far

I would like to find out if this is solvable somehow, else the judgment would be: "mixing pascal and maxwell impossible"

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

GTX 1080 arrived today and its a dud with actual drivers so far

In Luxmark it is only slightly faster as the gtx 980ti in hotel, in other scenes slower !
Notice: i found in original LM the micro scene does not crash, so thats something in our code i fear.

Actually the 367.48 driver seems to not like procedurals at all, goes into the above issues all the time.
Also with actual luxrays the 1080 is on par if not slower !!!! than the 980 ( no ti ) most of the time.

In cuda the 1080 is only 10% faster than my 980ti ( both reference cards ), so also no halleluja there.

All in all it looks as some work is to be done to adapt to pascal in existing cl and cuda code. Not really convincing.

This for now ... Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

Update:

The opencl behavoiur of the gtx 1080 is still odd, it seems also loadbalancing is not good with mixed maxwell/pascal.
Then also it seems the NVidia compiler again ignored inline attribute, played a bit with __forceinline and it got faster. Hmmmm....

The actual code still tends to crash with certain scenes a lot ( also mic bench scene ) when used along with 367.48 driver.

Did not yet tried beta 370 yet.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

On Linux 367.57 is just out which restores former speeds nealy but does not fix the mircrophone ( procedurals ? ) crash.

This is the distro driver. I also tested all from NVidia itself up to 375.10 to no avai.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

Seems like NVIDIA is finally updating their driver to support OpenCL 2.0.
Although the driver is early and LuxMark benchmarks shows that the OpenCL 2.0 with "evaluation support" is slower than the older 1.2 driver.

http://www.anandtech.com/show/11180/the ... -review/14
"The OpenCL situation for NVIDIA right now is a bit weird. The company is in the middle of rolling out OpenCL 2.0 support to their video cards – something that I had actually given up hope on until it happened – and as a result their OpenCL drivers are in a state of flux as company continues to refine their updated driver. The end result is that OpenCL performance has dipped a bit compared to where the GTX 1080 launched at back in May, with said card dropping from 4138 points to 3648 points. Not that the GTX 1080 Ti is too fazed, mind you – it’s still king of the hill by a good degree – but the point is that once NVIDIA gets their drivers sorted out, there’s every reason to believe that NVIDIA can improve their OpenCL performance."

Latest NVIDIA driver here, should be 378.66 or higher: http://www.geforce.com/drivers
HDD

Posts: 48
Joined: Sat May 08, 2010 11:33 am

### Re: NVIDIA OpenCL Drivers & News

In case nobody noticed, i found a solution to get rid of long kernel compile times with NVidia.
As LuxRender is now almost stalled here, i offer linux testbuilds for LuxCoreRender ( the Lux reboot )
which are 100 to 300 times faster in kernelcompile, while having negligible tradeoff in renderspeed.
Especially the nasty procedural textures used as bump are down from > 9min to 1.5s etc.

It makes not much sense to backport anything to old luxcore(1) here, so in case you are interested,
visit the relaunched site.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

Dade took now over this workaround into luxcore2 master, thus its basically an improvement for all OS now.
Linux users can still test also my profiled version, which goes a bit further.

NVidia is informed since a while and i hope a future compiler handles this better without tweaking.

Jens

jensverwiebe

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

### Re: NVIDIA OpenCL Drivers & News

Although i don't use macOS for CG anymore, made a a test with actual sources ( luxcorerender alpha7 ff ) and found
the situation with Apple odd, but solvable.

So far had no luck with NVidia on 10.10, 10.11 but made a testinstall 10.13 where i got NVidia to finally work as expected.
This is with webdriver 387.10.10.10.25.161. and a gtx 980 as well as 1080. 10.12 might work, did skip.
This might bee a good sign for eGPU users later. Tbh installing a linux on that mac was way more efficient

Ppl can proof with luxmark if their system version works so far, but the actual development stage behaves a bit different again.

My profiled linux builds still did not show problems and compiles kernels 2 to 4 times faster compared to the master implementation. ( my profile never was requested )
The OSX builds behave across the board as far as i checked some scenes. Not surprising AMD gpu's need less tweaks.

Jens

jensverwiebe

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

Previous