I was trying to compile pcl(point cloud library) with CUDA support.
I got the message
nvcc fatal : Unsupported gpu architecture ‘compute_52’.
CMake Error at pcl_cuda_features_generated_normal_3d.cu.o.cmake:206 (message):
my cuda samples run fine.my setup
ubuntu 14.04, 64bit
graphics card: gt740
First of all
compute_52 code won’t run on/isn’t required for a gt740. But if you want to compile for compute_52 anyway, you’ll need the latest update to CUDA 6.5 which supports this target. As far as I know, at the moment, the compute_52 target only applies to GeForce 980 and 970 products, and their mobile cousins (although presumably more compute_52 GPUs may be added in the future.)
If you want to target a gt740 (a cc3.0 device, i.e. appropriate compile targets are compute_30,sm_30), and you don’t feel like updating your cuda 6.5 to the latest, then you should probably modify the makefile you are using. Since this is CMake, I can’t help you much there, and it may in fact be “easier” just to update your cuda version to the latest:
i already have cuda 6.5, but gt740 doesnt support 5.2. there was a flag for setting compute capablity in pcl. i set it to 3.0 as mentioned in http://en.wikipedia.org/wiki/CUDA#Supported_GPUs
The original cuda 6.5 release did not support compute_52, and it would throw the error you are seeing if you try to use it. The update to CUDA 6.5 that I linked to, does support compute_52. Anyway I’m not sure if you still have a problem. It sounds like an issue in the pcl if you set the desired compute capability to 3.0 and it’s still compiling for compute_52. If you want to make the error go away, you could try installing the update to cuda 6.5 that I linked to.