Error when creating array and specifying accelerators

Question

I'm experimenting with C++ AMP and the code below is giving me trouble. It gets compiled to a DLL and called from a C# app. I'm calling the function in a loop (in C#), and that works fine as long as the number of iterations is low. Then I hit some limit and calling the function again will give an error.

What's really strange is that if I create the array without specifying accelerator views (just the extent), it seems to work fine for any number of iterations.

What's going on here?

Thanks.

static array<float, 2> *inputArrayPtr;

extern "C" __declspec ( dllexport ) void _stdcall CopyMyData(
float* input, int m, int n)
{
concurrency::extent<2> inEx(m, n);
accelerator cpuAccelerator(accelerator::cpu_accelerator);
accelerator defaultAccelerator(accelerator::default_accelerator);

delete inputArrayPtr;
inputArrayPtr = NULL;
inputArrayPtr = new array<float, 2>(inEx, cpuAccelerator.default_view, defaultAccelerator.default_view);

copy(input, *inputArrayPtr);
}

Answer 1

Hi tevo,

This seems to be due to a known issue in some GPU drivers  - even on deleting staging arrays, the driver chooses to be lazy about the deletion and eventually runs out of memory. We have addressed this issue in the C++ AMP runtime since Beta.

For now, a potential workaround would be to call "defaultAccelerator.default_view.flush()" after deleting the "inputArrayPtr". Can you try this to see if it helps?

- Amit

---

Amit K Agarwal

Answer 2

Thanks for the info, Amit.

I tried "defaultAccelerator.default_view.flush()" as you suggested, but I still got the error. Then I replaced that call with "defaultAccelerator.default_view.wait()" and now my code runs for any number of iterations.

My original problem is solved, but now I'm wondering why wait() works and flush() does not. Is flush() asynchronous?

Answer 3

Tevo, you are right. This appears to be an issue with NVidia drivers only. Thanks for reporting it - we will bring this issue to their attention. "flush" is expected to synchronously delete any released device resources (staging buffer in your case) as described here.

---

Amit K Agarwal