Question

is it valid to free device allocated memory from a host function?
I’m writing some C++ class that should be used on host and device. My constructor and destructor are something like this:

class myClass {
public:
__host__ __device__ myClass() {
#if defined(__CUDA_ARCH__)
  data = (char*)malloc(DATA_SIZE);
#else
  cudaMalloc(&data,DATA_SIZE);
#endif
}

__host__ __device__ ~myClass() {
#if defined(__CUDA_ARCH__)
  free(data);
#else
  cudaFree(data);
#endif
}

private:
  char* data;
}

The above code compiles and i didn’t get an error if i construct a class on the device and free it on the host. But this case is not documented in the CUDA developer papers.

Answer 1

For both CUDA 4.2 and CUDA 5.0 RC, the CUDA C Programmer’s guide mentions in section B.17 that: “Memory allocated via malloc() cannot be freed using the runtime (i.e., by calling any
of the free memory functions from Device Memory). ” (This particular text taken from the CUDA 5.0 RC document. In the original doc, Device Memory is a hyperlink to section 3.2.2) The CUDA 4.2 document can be found here (it has similar wording.) I’m wondering if: 1.) Things are actually happening the way you think. Depending on how you allocate the variable in the kernel code, I think it could go out of scope at the completion of the kernel, which would implicitly call your (device-side) destructor. 2.) a call to cudaFree with an invalid pointer (perhaps because it is a device pointer, or perhaps because it has already been freed) simply gets ignored. Without your code to play with, it’s all just speculation. But if you’re checking for errors and not getting one, then it may be getting ignored.