LuxCoreRender Forums

acasta69 wrote: ↑Wed Apr 29, 2020 6:51 am

zuljin3d wrote: ↑Tue Apr 28, 2020 11:56 pm

mib2berlin wrote: ↑Tue Apr 28, 2020 9:28 pm Hi, I test only small scenes but does anybody have the same performance boost like me?
It is about 1.8 times faster than OCL on my stone age GTX 760.

I have a similar boost result on my old (gtx960 4gb)

neo2068 wrote: ↑Mon Apr 27, 2020 6:21 pm I get errors with latest build and OpenCL and the food demo scene, too, e.g. PATHOCL:clFinish(CL_INVALID_COMMAND_QUEUE) or TILEPATHOCL: clEnqueueReadBuffer(CL_OUT_OF_RESOURCES). With cuda my GTX 780 is 1.5 times faster than the GTX 1080. Is it normal behaviour?

acasta69 wrote: ↑Tue Apr 21, 2020 5:36 pm Now, without native threads and samples/sec as reference, I have 12% faster performance with OpenCL [than with Cuda on GTX970]

It seems we have a wide variety of results, mostly improvements but not all of them.
Also a GTX780 faster than a 1080 is quite strange, isn't it? The 1080 has more Cuda cores and nearly double clock speed and memory bandwitdh than the 780.

may be GTX780 faster than a 1080 for next reason gtx780 have GK110(Kepler) stone and 1080 GP104-400 Pascal (architecture) and Cuda cores are not the same.
Such case was happend earlier
1 core Fermi = 3 cores Kepler
this info from
https://milovsky.ru/nvidia-6xx-cuda-problem/

It seems we have a wide variety of results, mostly improvements but not all of them.
Also a GTX780 faster than a 1080 is quite strange, isn't it? The 1080 has more Cuda cores and nearly double clock speed and memory bandwitdh than the 780.

more Cuda cores and nearly double clock speed and memory bandwitdh
Sure something is wrong. it is not the 980 but the 1080.

I should have fixed the OpenCL problem: it was a mess of alignment and size of a 64bit field (CPU, OpenCL and CUDA work in different ways ).

Dade wrote: ↑Wed Apr 29, 2020 11:18 am I should have fixed the OpenCL problem: it was a mess of alignment and size of a 64bit field (CPU, OpenCL and CUDA work in different ways ).

Well Done ! thanks

Dade wrote: ↑Wed Apr 29, 2020 11:18 am I should have fixed the OpenCL problem

Yep, confirmed with Benchmark scene on Win10PC: 4.6 vs 3.3 megasamples/s CUDA vs OCL.

Great work, big thanks!

epilectrolytics wrote: ↑Wed Apr 29, 2020 1:01 pm Yep, confirmed with Benchmark scene on Win10PC: 4.6 vs 3.3 megasamples/s CUDA vs OCL.

Apparently, it is even faster with the new under construction cache friendly Sobol sampler

Yes, the results are really strange. I get very low cuda performance in general with my builds compared to the azure build, factor 4 slower for the 1080. The OpenCL kernel compilation ended with the mentioned error messages. I tested various versions of cuda SDK (8.0, 10.1 and 10.2) and the results were all the same. An old build of LuxCoreUI from 14.04. before the cuda merge show a normal OpenCL behaviour.

Ok, I've been roughly reading this over the last week. Now I have time again andcould finally try myself

Win10 + RTX2070
I installed CUDA SDK 10.2 and the latest studio driver 442.92
Then the latest BlendLuxCore and LuxCore_CUDA build, and I have 2.3 as well.
Didn't try compiling myself yet.

In the daily build, both the LuxBall and Canelle&Fromage test scenes show a 25% improvement with CUDA over OpenCl. (CPU disabled)

Then I tried OCl in V2.3 with the LuxBall scene. And indeed, I get the same performance as with CUDA in 2.4

Latest daily is broken...

You have posted an empty file. Anyway I'm fixing it.