understood, but, inside the code, we use CUDA methods to get the memory info, so, would it be correct to select ‘None’ for CUDA Runtime Library?
Anyway, I tried it out and I obtained the following error:
main.c:(.text+0x16): undefined reference to `cudaMemGetInfo'
/tmp/tmpxft_00000cc6_00000000-4_NTTest_dlink.o: In function `__cudaUnregisterBinaryUtil':
link.stub:(.text+0x10): undefined reference to `__cudaUnregisterFatBinary'
collect2: error: ld returned 1 exit status
There is some information need to be correct.
Please follow this step to compile your program:
- copy the library from TK1 back to x86-host.
scp -r /usr/local/cuda-6.5/targets/armv7-linux-gnueabihf/ [Your ID]:/tmp/
Include and link the library in Nsight:
NVCC Linker -> CUDA Runtime Library: Static
NVCC Linker -> Library search path:
Thanks and sorry for any inconvenience.
I can do the steps you mentioned above, but the libs, that
contains, were previously on the board, and they are not the new libs contained into the Release 21.6.
So, even if I’ll be able to build my source code with those libs, I would be using old libs, is this right?
I want to point out that the main goal of this thread is to update the firmware version to R21.6.
I already applied new binaries and flashed the board. I was only wondering if I need to compile my source code (using Nsight) with the new libs. If so, I’m not able to find libcudadevrt.so (as nvcc linker says), also into /usr/local/cuda-6.5/targets/armv7-linux-gnueabihf/lib.
If not (please confirm), i.e. it’s enough to copy new libs from R21.6 to the board, not using new libs to build the code, I think we can close the topic.
Yes. CUDA libraries come from JetPack3.1, not from rel-21.6.