cuda Scons fix: remove unnecessary -lcuda which prevents compiling on… #272
Conversation
… cluster head node without cuda drivers
|
Could you explain this a little more? Is it that the code doesn't need to link to the driver directly, and only uses it via cudart, so libcuda is not needed at all? I'm surprised that the node has cudart/cufft/etc but not libcuda, but if this fixes your issue then I think it looks fine to merge. |
|
Yes, the code doesn't need to be linked with libcuda, which comes with the nvidia driver, not CUDA toolkit. On a cluster, the head node typically doesn't have nvidia gpus and therefore they usually don't install the driver on head node, only the toolkit. When the program is compiled on head node, if it looks for libcuda, an error occurs. This only happens to the current scons script. CMake compiles successfully as it has the correct configuration by not including libcuda, e.g., using the (static) cudart only -lcudadevrt -lcudart_static. |
|
Okay, thanks! That makes perfect sense. Changes LGTM so I'll plan to merge this as-is. |
… cluster head node without cuda drivers