No warning/error when mistakenly cudaMalloc&cudaFree a class which constructor&destructor only declare with __host__ but no __device__?

Accelerated Computing CUDA CUDA Programming and Performance
1

no warning/error when mistakenly cudaMalloc&cudaFree a class which constructor&destructor only declare with __host__ but no __device__?
and what will happen inside CUDA when this happen?
thanks!

as in : https://developer.nvidia.com/blog/separate-compilation-linking-cuda-device-code/
if use this in cpp:

    cudaMalloc(&devPArray, n*sizeof(particle));
    cudaMemcpy(devPArray, pArray, n*sizeof(particle), cudaMemcpyHostToDevice);

should the constructor & desctructor of class particle be declare with nothing or __host__ __device__?

2

If you need to be able to call the constructor and destructor in kernel code and cpu code, you need __host__ __ device__.
However, cudaMalloc and cudaFree do not call constructors or destructors.

3

thanks! then do you know the common reason cause error cudaErrorMemoryAllocation when cudaMalloc or cudaGraphicsGLRegisterImage while there is enough gpu memory there.

4

You may find cudaMemGetInfo useful.

5

used this, that’s why i said there is enough gpu memory there, but got cudaErrorMemoryAllocation…

6

Can you post a minimal reproducer? How much memory is reported free with cudaMemGetInfo, and how much do allocate when it fails?

7

total is 6G, cudaMemGetInfo return available is almost 5G after the returnErrorCode==2(cudaErrorMemoryAllocation) by cudaGraphicsGLRegisterImage or cudaMalloc/cudaMalloca3DArray.
the application delete the window after render and recreate the window which use CUDA-OpenGL-interop, this loops many days, now after loop half a day(eg. 50 thounsands times) got this CUDA error, test a few times and each test not exact 50 thousands but around it, get the error.
thanks very much!