I have class, which members are pointers to device space:
In many kernel invocations I will use this memory, but I can determine needed memory space in constructor.
So I call in constructor:
cudaMalloc( (void**) &m_d_ptr, needed_size*sizeof(int));
But in class method on cudaMemcpy I’ve got:
Cuda error in file 'kernel.cu' in line 208 : unknown error.
but if I do cudaMalloc in method just before copy, anythink works fine.
Why? I have no idea what am I doing wrong.
Hard to say, can’t say what’s wrong without seeing your code. How did you pass the variable?
Just a blind guess: a miss-definition of the copy-constructor and/or operator=
I have variable as class member.
Upper class creates my class and then calls from time to time computing function.
Computing function takes pointer from class member and operates on it.
I think that CUDA frees resource, when routine ends in constructor and returns to pure C++ class. Is there any way to tell driver, to do not release this resource, because I will need it later?