(L) [2019/08/09] [ost
by Philipp] [Maxwell Support for pure Optix?] Wayback!Hello,
thanks for activating my account.
my question:
is there intentionally no support for maxwell GPUs in rendercore_optixrtx_b?
I know that maxwell GPUs do not have RTX cores anyway. So the comment in the code recommends to use rendercore_optixprime_b for better performance. However I am currently in the situation that I would like to develop my software partly on my own laptop, which has no RTX cores. But I would like to also sometimes use computers at my university that do have RTX cores. Especially for benchmarking later. For that reason I would like to use pure OptiX. Because as I understand it pure OptiX will also run on CUDA devices without OptiX. But it will automatically switch to RTX cores when executed on a different PC with an RTX graphics card.
However the function compileToPTX(...) in cudatools.h seems to not support maxwell GPUs in line 142:
Code: [LINK # Select all] sprintf_s(versionString, "compute_%i", cc >= 70 ? 70 : 61 );
const char* compiler_options[] = { "-arch", versionString, "-restrict", "-use_fast_math", "-default-device", "-rdc", "true", "-D__x86_64", 0 };
Is there a reason that this code allows only to use at least compute_61 or is this a bug ? Are there any features needed by lighthouse2 currently that are not available on GPUs with only compute capability 5.0 ?
Actually when I replace the compiler option with compute_50 it seems to run just fine on my laptop with a maxwall GPU (GeForce GTX 960M)
(L) [2019/08/09] [ost
by jbikker] [Maxwell Support for pure Optix?] Wayback!You are right, it should run just fine on Maxwell. In fact, when compiling the .cu files, it should already aim for SM 5.0, like the Optix Prime code. It might even support older devices, but I didn't test it; at some point available GPU memory may become a problem.
(L) [2019/08/13] [ost
by Philipp] [Maxwell Support for pure Optix?] Wayback!Thanks.
But would do you mean "it should already aim for SM 5.0"?
The code in cudatools.h forces it to aim for 6.1 by passing the compiler option compute61 as I understand it.
So this is a bug in the file cudatools.h that should be fixed?