Below is largely based on following the CUDA tutorial on readthedocs.io run on Ubuntu laptop with an NVIDA P2000 Max-Q.
Example code to add vectors of values is in vector_add.c, similar to the tutorial's
example.
nvcc vector_add.cu -o vector_add
Runing is done same as a normal binary. The CUDA-specific calls allocate memory on the GPU, copy data and run the vector adding kernel.
time ./vector_add
out[0] = 3.000000
PASSED
./vector_add 1.92s user 0.29s system 94% cpu 2.341 total
Here are the profiling results using nvprof.
sudo nvprof ./vector_add
==970283== NVPROF is profiling process 970283, command: ./vector_add
out[0] = 3.000000
PASSED
==970283== Profiling application: ./vector_add
==970283== Profiling result:
Type Time(%) Time Calls Avg Min Max Name
GPU activities: 97.82% 1.58634s 1 1.58634s 1.58634s 1.58634s vector_add(float*, float*, float*, int)
1.18% 19.082ms 1 19.082ms 19.082ms 19.082ms [CUDA memcpy DtoH]
1.00% 16.238ms 2 8.1188ms 7.0825ms 9.1550ms [CUDA memcpy HtoD]
API calls: 83.29% 1.62249s 3 540.83ms 7.1400ms 1.60596s cudaMemcpy
16.51% 321.60ms 3 107.20ms 152.94us 321.29ms cudaMalloc
0.15% 2.8922ms 3 964.08us 216.46us 1.3731ms cudaFree
0.02% 471.06us 97 4.8560us 424ns 214.95us cuDeviceGetAttribute
0.02% 338.44us 1 338.44us 338.44us 338.44us cuDeviceTotalMem
0.00% 75.495us 1 75.495us 75.495us 75.495us cuDeviceGetName
0.00% 58.404us 1 58.404us 58.404us 58.404us cudaLaunchKernel
0.00% 31.653us 1 31.653us 31.653us 31.653us cuDeviceGetPCIBusId
0.00% 5.6920us 3 1.8970us 454ns 3.1350us cuDeviceGetCount
0.00% 2.6850us 2 1.3420us 666ns 2.0190us cuDeviceGet
0.00% 730ns 1 730ns 730ns 730ns cuDeviceGetUuid