alpha06 bugs

[jensverwiebe]

Hi all,

I've made a test with a scene I'm working on two different computers, one with an gforce gtx 1070 and another with a radeon r9 280x.
After a while I'm getting 1.6M with the gforce and 1.3M with the radeon. Only opencl.

Given radeon is on a much older computer, is this normal?

Thanks.

Typically on light scenes NVidia lacks a bit, the performance jumps in more significant on rich scenes.
Anyway, you did not provided much information to really judge.

From my HW:
R9 280x is somewhat a 7970 tahiti which gives me 2.8M max on Hotel scene.
GTX 1070 should be a tad slower than a 1080 i have, which gives 4.14M on Hotel.

So the gtx speed looks too low compared to 280x.
Then again its all os/driver/scene dependent.

[frank_yifei]

About NVIDIA performance, I'm unable to replicate the problem on Linux with a 980 GTX.

This is likely to be related to Windows terrible thread scheduler and hybrid rendering: the CPU may be busy and unable to "feed" the GPUs fast enough.

Try to compare alpha5 and alpha6 disabling hybrid rendering by setting the number of native threads to 0. Once hybrid rendering is disabled, you should obtain exactly the same performance.

If this is true, you can than try to re-enable hybrid rendering leaving, at least, 1 core free for each GPU you are using (i.e. 4xCore+1xGPU => 3 native threads, 4xCore+2xGPU => 2 native threads, etc.).

P.S. NVIDIA compilation times are yet another topic.

I have tried blender cycles, amd prorender and luxrender and compared the opencl cache file that generated during the rendering, I found a big difference in number of files. cycles and prorender generates many bin files, luxrender generates only a few but much larger.
On my system, windows 10 gtx 1060, luxrender rendering the blender benchmark bmw scene generates only one 14.4mb ocl file. Prorender generates 19 bin files and in total 6.24 mb. cycles generates 3 files in total 5.3 mb if choose single program in debug option, not checking this option results in 19 bin files and in total 10mb.
are the differenct ways they generate bin files affect the compiling performance? for luxcorerender can the bin file be reused if I use same material in other files ?

[Dade]

I have tried blender cycles, amd prorender and luxrender and compared the opencl cache file that generated during the rendering, I found a big difference in number of files. cycles and prorender generates many bin files, luxrender generates only a few but much larger.
On my system, windows 10 gtx 1060, luxrender rendering the blender benchmark bmw scene generates only one 14.4mb ocl file. Prorender generates 19 bin files and in total 6.24 mb. cycles generates 3 files in total 5.3 mb if choose single program in debug option, not checking this option results in 19 bin files and in total 10mb.
are the differenct ways they generate bin files affect the compiling performance?

It is a specific problem of NVIDIA compiler trying to inline every single function (and probably unrolling loops at the same time). This lead to an exponential compilation time. AMD suffered of a similar problem some year ago but they fixed this sick compiler behavior.

LuxCore compile multiple kernels in a single call. This should be faster because shared code is compiled only once.

for luxcorerender can the bin file be reused if I use same material in other files ?

Yes, LuxCore has a kernel cache. Just try to run a rendering 2 times. The second time will take only few millisecond to compile because the cached kernel will be used.

[frank_yifei]

I tried another blender benchmark scene classroom, blendlux just can't passed the compiling and then crashed. I suspect that split kernel is better than megakernel in this case for complex scene. Also I noticed that when I change one material in the scene, the megakernel is then rebuilt and takes a long time.

[Dade]

I tried another blender benchmark scene classroom, blendlux just can't passed the compiling and then crashed.

It is the NVIDIA compiler to generate broken code. The same scene will work with AMD OpenCL compiler for any GPU and with Intel/AMD compiler for any CPU.

I suspect that split kernel is better than megakernel in this case for complex scene.
Also I noticed that when I change one material in the scene, the megakernel is then rebuilt and takes a long time.

LuxCore has only micro-kernels and dynamically generates the code for evaluating recursive texture and material definitions (like mix, scale, etc.). So if you use a combination of material types for the first time (i.e. not cached), you will trigger a new compilation.

[Stem]

I tried another blender benchmark scene classroom, blendlux just can't passed the compiling and then crashed.

I am not seeing that problem.

Compile time approx 9sec, renders OK.

Win7 64/ gtx1080(x2)

[frank_yifei]

I tried another blender benchmark scene classroom, blendlux just can't passed the compiling and then crashed.

I am not seeing that problem.

Compile time approx 9sec, renders OK.

Win7 64/ gtx1080(x2)

windows 10 might be the problem. will give linux a try. some argues it is faster

[frank_yifei]

I tried another blender benchmark scene classroom, blendlux just can't passed the compiling and then crashed.

It is the NVIDIA compiler to generate broken code. The same scene will work with AMD OpenCL compiler for any GPU and with Intel/AMD compiler for any CPU.

I suspect that split kernel is better than megakernel in this case for complex scene.
Also I noticed that when I change one material in the scene, the megakernel is then rebuilt and takes a long time.

LuxCore has only micro-kernels and dynamically generates the code for evaluating recursive texture and material definitions (like mix, scale, etc.). So if you use a combination of material types for the first time (i.e. not cached), you will trigger a new compilation.

I used lots of mix of material.

[Sprocket]

Im guessing this problem is limited to me: When trying to load cornell.cfg in luxcoreui I get the error:
Unable to open properties file: scenes/cornell/cornell.scn

[Dade]

Im guessing this problem is limited to me: When trying to load cornell.cfg in luxcoreui I get the error:
Unable to open properties file: scenes/cornell/cornell.scn

This is only a path problem, you have to run luxcoreui from command line (from the root directory) with something like:

Code: Select all

david@i7-3930k:~/tmp/luxcorerender-v2.0alpha6-linux64$ ./luxcoreui scenes/cornell/cornell.cfg 

Otherwise the luxcoreui will not be able to find the .scn file.