cudaMemcpyToSymbol error while running "hello_world" program in carlsim

[arnab roy]

Hi ,

I am trying to install CARLsim in my GPU system. I have downloaded CARLsim 3.,  set up the the appropriate environment variable.
After that I have executed make and make install command (which got executed successfully).
However when I am trying to run the example project "hello_world" program, inside projects/hello_world directory I am getting the following error,

CUDA error at carlsim/kernel/src/snn_gpu.cu:320 code=13(cudaErrorInvalidSymbol) "cudaMemcpyToSymbol(loadBufferCount, &bufferCnt, sizeof(int), 0, cudaMemcpyHostToDevice)"

Could you please throw some light on this issue?


Platform information:
nvcc version : 5.5
CUDA device: Tesla M2070
CUDA Compute Capability (CC) : 2.0

.bashrc information:

export CUDA_INSTALL_PATH=/usr/local/cuda-5.5
export CUDAHOME=/usr/local/cuda-5.5
export CARLSIM_LIB_DIR=~/CARL/CARLsim
export CARLSIM_CUDAVER=6
export CUDA_MAJOR_NUM=3
export CUDA_MINOR_NUM=5
export CPU_ONLY=0

os information:
Linux 2.6.32


Regards,
Arnab

[Michael Beyeler]

Hi Arnab,

This is weird, especially since installing CARLsim didn't give any warnings or errors.

One possibility is: The .bashrc variables define the CUDA compile flags, and if they don't match your hardware, sometimes weird errors pop up.

Can you try updating your .bashrc variables to reflect the compute capability of your card?

export CUDA_MAJOR_NUM=2
export CUDA_MINOR_NUM=0

Also (although it shouldn't matter too much):

export CARLSIM_CUDAVER=5

Then source your .bashrc to make sure changes take effect:

$ source ~/.bashrc

Then re-install CARLsim

$ make clean

$ make -j4

$ sudo make install

Then re-run the project

$ make clean && make

$ ./hello_world

...and see if anything changed?

Best,

Michael

[arnab roy]

Yes it seems the compile flags are not supported by the hardware. Because when I am running the the same hello_world program in "Tesla K40c" machine it's working fine. Thanks for the help.