cudaFree crash in destructor when exit() is called

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi there,

I built a safe thread singleton class, which holds memory that is allocated using cudaMalloc().
The instance of the class is saved in the singleton using a C++ auto_ptr smartpointer.
So that whenever I leave the program, the destructor is automatically called.
In the destructor I make sure the memory is release using cudaFree().

When I call exit(0) the program will crash (EXC_BAD_ACCESS) on cudaFree().

If I manually release my singleton just before the call to exit(0), it will not crash.
The point is that manually releasing the singleton is not acceptable.

Any idea?

2

Can you post the singleton code and the cuda operations you do with it?

also does this happen when only one thread uses the singleton or more than one?

eyal

3

I tried a different strategy that does not involve dynamic allocation of the instance:

class T

{

........

........

........

public:

	static T* GetInstance()

	{

		if (!m_p_singleton)

		{

			pthread_mutex_lock(&m_singleton_mutex);

			if (!m_p_singleton)

			{

				static T instance;

				m_p_singleton = &instance;

				m_p_singleton->Initialise();

			}

			pthread_mutex_unlock(&m_singleton_mutex);

		}

		return m_p_singleton;

	}

	static void Destroy()

	{

		if (m_p_singleton)

		{

			pthread_mutex_lock(&m_singleton_mutex);

			if (m_p_singleton)

			{

				// Release the memory that was dynamically allocated (both on host and device)

				m_p_singleton->ClearMemory();

			}

			pthread_mutex_unlock(&m_singleton_mutex);

		}

	}

........

........

........

private:

	static pthread_mutex_t m_singleton_mutex;

	static T* m_p_singleton;

........

........

........

};

ClearMemory is calling CudaFree (with tests so that I only delete memory that is actually allocated).

It is automatically called within ~T().

And now it works without crash :-)

For information, the version that did not work was:

static T* GetInstance()

	{

		if (!m_p_singleton.get())

		{

			pthread_mutex_lock(&m_singleton_mutex);

			if (!m_p_singleton.get())

			{

				m_p_singleton.reset(new T);

			}

			pthread_mutex_unlock(&m_singleton_mutex);

		}

		return m_p_singleton.get();

	}

	static std::auto_ptr<T> m_p_singleton;

In this version, the program will crash in the destructor ~T (ClearMemory) on cudaFree when exit(0) is called.

The only explanation I imagine is that some kind of cuda context is erased first, then my auto_ptr is deleting my instance, calling the destructor that calls cudaFree.

But it is just a guess.