As a benchmarking fan, I was looking for a minimal code to run able to give us the difference between CPU and GPU calculations. I found it on this linuxhint article. At the bottom, you have two links for further reading: GPU programming with C++ and GPU programming with python. I’ve tried both on my CentOS 7.9 and a Quadro K5200. First step is to install the default CUDA:
yum install cuda
In my case, the above command installed cuda-11.7. I previoulsy had the drivers for the GPU installed. Unfortunately that seems not to be enough. You need the nvidia compiler, nvcc. Very easy to install if you don’t have it:
yum install nvcc
The compiler lands on /usr/local/cuda/bin/nvcc. You may want to modify your PATH variable to point to that folder also. I didn’t. Time to copy the sample code given above. I save everything in my project folder and type make as suggested.
gputest ## > make
/usr/local/cuda/bin/nvcc
-std=c++11 gpu-example.cpp -o gpu-example
gpu-example.cpp: In function
'void vector_add_gpu(int*, int*, int*, int)':
gpu-example.cpp:24:13: error: 'threadIdx'
was not declared in this scope
int i = threadIdx.x;
^
gpu-example.cpp: In function 'int main()':
gpu-example.cpp:64:22: error: expected primary-expression
before '<' token
vector_add_gpu <<<1, ITER>>>
(gpu_a, gpu_b, gpu_c, ITER);
^
gpu-example.cpp:64:32: error: expected primary-expression
before '>' token
vector_add_gpu <<<1, ITER>>>
(gpu_a, gpu_b, gpu_c, ITER);
^
make: *** [all] Error 1
In case it’s not clear, the code copied as-is didn’t work. I check the nvcc compiler is there (it is!) so it’s something else. I guess from this nvidia forum post the problem may be in the convention I’m using. I rename my file and adjust the Makefile accordingly
mv gpu-example.cpp gpu-example.cu
After that, make works and I get a binary that I can run. This is my output.
gputest ## > make
/usr/local/cuda/bin/nvcc -std=c++11 gpu-example.cu -o gpu-example
gputest ## > ls
gpu-example* gpu-example.cu Makefile
gputest ## > ./gpu-example
vector_add_cpu: 432218 nanoseconds.
vector_add_gpu: 19573 nanoseconds.
Later runs give me different numbers but in the same order: the GPU calculation is 20 times faster than the CPU one. Further analysis is needed 🧐
Bits and Dragons 