LuxMark v3.0: LuxRender propaganda with OpenCL

Discussions related to GPU Acceleration in LuxRender

Moderators: Dade, jromang, tomb, coordinators

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby SATtva » Tue Jun 30, 2015 7:53 am

Yes to both.
Linux builds packager
聞くのは一時の恥、聞かぬのは一生の恥
User avatar
SATtva
Developer
 
Posts: 7162
Joined: Tue Apr 07, 2009 12:19 pm
Location: from Siberia with love

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby LuxMarked One » Thu Jul 09, 2015 8:52 am

Hello,

I have some questions regarding LuxMark 3.0 and the Error-Correcting Code support on the AMD FirePro W8100 GPU.


My last four results are (ECC off vs. on):

Hotel lobby

Result: 2413 vs. 259

Image Validation OK (144749 different pixels, 24.54% vs. 337870 different pixels, 57.28%)


LuxBall HDR

Result: 13387 vs. 1611

Image Validation OK (964 different pixels, 0.15% vs. 22578 different pixels, 3.53%)



I know that the available video memory size, and bandwidth will be reduced when enabling ECC, but I doubt that this kind of performance loss is normal (In 'contrast', the LuxMark 2.0 Room results are somewhat better (1201 vs. 810)

The image validation concerns me more, how can be the image even less accurate with enabled ECC?
Is this a driver issue, or something wrong with the video card?
LuxMarked One
 
Posts: 2
Joined: Thu Jul 09, 2015 7:32 am

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby Dade » Thu Jul 09, 2015 9:03 am

LuxMarked One wrote:The image validation concerns me more, how can be the image even less accurate with enabled ECC?


The image validation tell you how far you are from the reference (i.e. noise free) image. So, slower is the rendering and larger is the error. This is way you get a larger error by enabling ECC, it is 10 times slower and so, at the end of the test, there is far more noise.

About the ECC slowdown: it doesn't look normal, you are supposed to loose some GPU ram space but there is something running really wrong if you run 10x times slower.

Are you using Linux ? I have also a FirePro w8100 and I have observed some odd result on Linux (i.e. quite slower than expected).

LuxMarked One wrote:Is this a driver issue, or something wrong with the video card?


It looks like a software and/or hardware problem. Try to run some other OpenCL benchmark to cross check if you have the same kind of slowdown.
User avatar
Dade
Developer
 
Posts: 8357
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby LuxMarked One » Thu Jul 09, 2015 12:25 pm

Thanks for the quick reply.


I don't have Linux, only using Windows 7 64 bit SP1, and the latest 14.502.1019 FirePro WHQL driver.



Some more benchmarks (ECC on vs. off):



AIDA64 GPGPU: No difference except of the memory bandwidth. (187487 MB /s vs. 239135 MB / s)


-----------------------------------------------------------------------------------------
CompuBench 1.5.8 x64


Computer Vision
Face detection 79.169 MPixels/s vs. 83.348 MPixels / s
TV-L1 Optical Flow 22.504 MPixels/s vs. 25.047 MPixels / s

Physics
Ocean Surface Simulation 1770.5 Frames/s vs. 2055.9 Frames/s
Particle Simulation - 64k 110.62 MInteractions/s vs. 596.51 MInteractions/s


Graphics
T-Rex 4.8139 Frames/s vs. 8.2121 Frames/s


Video Processing
Video Composition 81.8 Frames/s vs. 84.771 Frames/s

Throughput
Bitcoin Mining 459.75 MHash / s vs. 475.93 MHash / s


-----------------------------------------------------------------------------------------
LuxMark v2.0

Room 810 vs. 1201
Sala 1266 vs. 2336
LuxBall HDR 5282 vs. 18324


-----------------------------------------------------------------------------------------
LuxMark v3.0

Hotel room 259 vs. 2413
Neumann TLM-102-SE 1008 vs. 5335
LuxBall HDR 1611 vs. 13387

----------------------------------------------------------------------------------------------------------
Blender 2.75 Cycles BMW27 (tiles: 1024x1024) 1 min 25 sec vs. N/A (freezes after 3 min)



Something definitely not right with these numbers.

It would be good if there will be some more LuxMark 3 results with a W8100-9100, or even with an NVIDIA Quadro or Tesla to see
the performance difference with and without ECC, just for comparsion.
LuxMarked One
 
Posts: 2
Joined: Thu Jul 09, 2015 7:32 am

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby acasta69 » Sun Jul 19, 2015 2:27 pm

Hello,

how does the pure CPU performance compare to the OpenCL CPU?
I ran some tests with LuxMark on my PC and got curious about this...

These are the results I get:
Luxball scene:
Native C++: 2190
OpenCL CPUs: 1690

Hotel lobby scene:
Native C++: 562
OpenCL CPUs: 375

Are these results expected?
Is it correct to say that OpenCL code is much less efficient than native C++ code on CPU (even if it's useful to be able to run it on CPU also)?

Ale
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v376.54
acasta69
 
Posts: 297
Joined: Fri Dec 20, 2013 3:18 am

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby SATtva » Mon Jul 20, 2015 4:01 am

Before Embree (native CPU geometry accelerator) integration it was the other way around. OpenCL is efficient on CPU, but Embree is plain f***ing damn efficient.
Linux builds packager
聞くのは一時の恥、聞かぬのは一生の恥
User avatar
SATtva
Developer
 
Posts: 7162
Joined: Tue Apr 07, 2009 12:19 pm
Location: from Siberia with love

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby acasta69 » Mon Jul 20, 2015 6:02 am

OK, I see.
Thanks. :)
Windows 10 64 bits, i7-4770 3.4 GHz, RAM 16 GB, GTX 970 4GB v376.54
acasta69
 
Posts: 297
Joined: Fri Dec 20, 2013 3:18 am

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby cwizou » Wed Jul 29, 2015 9:26 am

Hi folks, I'm a hardware reviewer at www.hardware.fr and we are currently using LuxMark 2.0 in our benchmarks (done with Windows 8.1 64 bit).

We wanted to use V3.0 (delighted to see native CPU back !) but we've had a few issues so far, mostly with Intel drivers who couldn't run it.

I wanted to try V3.0 again for an upcoming review with Intel's newest drivers but I ran in another weird issue that probably isn't driver related, since it's a boost path error :

2015-07-29 15:59:58 - [LuxCore] [7.109] [PathOCLBaseRenderThread::0] Compiling kernels
2015-07-29 16:00:03 - RUNTIME ERROR: boost::filesystem::create_directories

Weirdly this ONLY happens in OpenCL CPU/OpenCL CPU+GPU modes, Native and OpenCL GPU runs fine.

I've tried disabling UAC, running as administrator, compatibility modes, even the folder where v3.0 sits, didn't change a thing. I've also tried 3.1beta1 and hit the exact same issue. To recap, this happens only when OpenCL CPU mode is involved, with any of the three scenes.

Config for this test :
AMD A8-7600 (running with integrated graphics)
Asus A88XM-Plus
8 GB RAM
Windows 8.1 64bit
AMD Catalyst 15.7

Any idea on what I could try to fix this ? Thanks a lot for providing this benchmark, we're glad to use it.
cwizou
 
Posts: 7
Joined: Wed Jul 29, 2015 9:12 am

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby Dade » Wed Jul 29, 2015 9:50 am

cwizou wrote:Any idea on what I could try to fix this ?


LuxMark (and LuxRender) use a file system cache of compiled OpenCL kernels (i.e. they are compiled once and than saved on the file system for later usage). On my Linux box, the cache directory is created under "/tmp" with the "kernel_cache":

Code: Select all
david@i7-3930k:~/projects/luxrender-dev/luxrays$ ls -l /tmp/kernel_cache/
total 4
drwxrwxr-x 3 david david 4096 Jul 29 16:47 LUXCORE_1.5RC2
david@i7-3930k:~/projects/luxrender-dev/luxrays$ find /tmp/kernel_cache/
/tmp/kernel_cache/
/tmp/kernel_cache/LUXCORE_1.5RC2
/tmp/kernel_cache/LUXCORE_1.5RC2/Advanced Micro Devices, Inc.
/tmp/kernel_cache/LUXCORE_1.5RC2/Advanced Micro Devices, Inc./Hawaii
/tmp/kernel_cache/LUXCORE_1.5RC2/Advanced Micro Devices, Inc./Hawaii/44
/tmp/kernel_cache/LUXCORE_1.5RC2/Advanced Micro Devices, Inc./Hawaii/44/7d43c95e-68e03a33.ocl


It looks like LuxMark is unable to create the "kernel_cache" directory in your case. I'm not sure of the exact directory cache name on Windows but according Boost documentation (http://www.boost.org/doc/libs/1_58_0/li ... ctory_path):

Window's GetTempPath uses similar logic to the POSIX version: the first environment variable in the list TMP, TEMP, USERPROFILE. If none of these are found, it returns the Windows directory.


So you may have some unusual environment variable setting for TMP, TEMP, USERPROFILE and/or you may lack the right to write there.

A simple test would be to install an AMD or NVIDIA GPU on the Intel test system and check if you still get the same error (you should and you would be sure it is not a problem related to Intel drivers).
User avatar
Dade
Developer
 
Posts: 8357
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: LuxMark v3.0: LuxRender propaganda with OpenCL

Postby Dade » Wed Jul 29, 2015 9:54 am

cwizou wrote:Weirdly this ONLY happens in OpenCL CPU/OpenCL CPU+GPU modes, Native and OpenCL GPU runs fine.


But are you using only the AMD OpenCL drivers (they include both CPU and GPU OpenCL device support) or AMD + Intel OpenCL drivers ? I'm a bit confused :?:
User avatar
Dade
Developer
 
Posts: 8357
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

PreviousNext

Return to GPU Acceleration

Who is online

Users browsing this forum: No registered users and 2 guests