test host device about omptest HOT 2 CLOSED

omptest/HybridOMP.C

Lines 35 to 50 in 403e28d

	for (int i=0; i<num_devices_visible; i++)
	{
	int A[1] = {-1};
	#pragma omp target
	{
	A[0] = omp_is_initial_device();
	}
	if (A[0]==0)
	num_gpus++;
	else
	printf("A[0]=%d\n",A[0]);
	}
	if (num_gpus == 0)
	printf("HybridOMP: Able to use offloading: %s, #devices visible: %d #devices available: %d\n", num_gpus>0?"yes":"NO", num_devices_visible, num_gpus);
	return num_gpus;

• Let say you have N devices available. The loop just says that you offload the code N times, but you don't specify to which device to offload, so probably it will be offloaded always to the same device. Am I right?

Yes, you're right, the code is incorrect. Line 38 should be

#pragma omp target device(i)

• Specification says that omp_is_initial_device() returns true (despite it is int) if the task is executed at a host device. You increment num_gpus if the function returns 0 (false), and then at line 49 you say that offloading is possible if the num_gpus > 0 .

I think this works as it should. If omp_is_initial_device() returns false, i.e. A[0]==0, we know the code was executed on a GPU, and num_gpus is incremented.

• The condition at line 48 is the same as the condition at the following line, so there is a dead code.

I'm not exactly sure what you mean here, but line 48 should probably be removed, so that the status message is always printed.

I hope you get my point.

When executing your test application omp_is_initial_device() always returns 0. ldd confirms, that the binary is linked with the libraries from the GCC (GCC-10.2.0 or GCC-7.3.0 + CUDA/10.2.89) I have compiled (libstdc++.so, libgomp.so, libgcc_s.so). Do you have an idea could be wrong?

Not sure, maybe it is picking the wrong libraries (or library versions)? Maybe try checking with nvprof whether any code is executed on a device. Also, make sure you can run CUDA code correctly.

from omptest.

My problem was in misunderstanding the return value of the omp_is_initial_device(). The code works fine now, thanks!

from omptest.