OpenMP-Accelerators: interoperability

[JPoznanovic]

Hello,

How does OpenMP (for accelerator targets) provide interoperability between itself and another programming language or numerical library? (e.g. keeping data persistent on the device between OpenMP accelerator regions and an external accelerator-targeted library call)

Seems like there should be a way to expose and/or use device data addresses on the host side. Without this, it seems like it would be difficult to allow device data persistence between uses of OpenMP and alternative approaches.

Please let me know your plans on this topic and/or if I've overlooked something.

Thanks,

Jeff

[RaviNarayanaswamy]

The OpenMP-Accelerator sub-committee understands the need for being able to pass around a device pointer address.
We are discussing 2 alternatives 1) specify "ptr" attribute to a pointer to indicate the value of the pointer should be sent/received as show in Example 1
or 2) if you use “p” instead of p[:] in map clause then the pointer value is sent instead of what it points to


Example 1
foo()
{
int *p;

#pragma target mapfrom(ptr:p) {
p = malloc(100*sizeof(int));
initialize(p);
}

#pragma target mapto(ptr:p)
{
Use(p);
free(p);
}
}

[JPoznanovic]

Dear Ravi,

I think this makes sense, as it appears to work in two key situations:

1) If the device data was allocated by the alternative approach and then used by OpenMP

2) If the device data was allocated by OpenMP and then used by the alternative approach

Thanks for the response.

Regards,

Jeff

[wapeter]

Dear Ravi,

I think this makes sense, as it appears to work in two key situations:

1) If the device data was allocated by the alternative approach and then used by OpenMP

2) If the device data was allocated by OpenMP and then used by the alternative approach

Thanks for the response.

Regards,

Jeff

Ravi which one did you end up going for? Excuse my newbieness but I'm just curious