Sycl ext oneapi current device test plan

[dklochkov-emb]

Test plan to ext oneapi device feature. See docs

[gmlueck]

Some specific comments below. At a higher level, I think there are three things to test:

1. Call set_current_device(d) then call get_current_device() from the same thread. Make sure it returns d.

2. Do not call set_current_device. Call get_current_device() and make sure it returns the default device.

3. Create multiple threads. Each thread calls the following in a loop: set_current_device(d) followed by get_current_device(). Each thread uses a different device d (if the system has more than one device). Make sure the call to get_device returns the same device that was set in set_current_device. This tests that each thread is keeping a separate copy of the "current device".

[gmlueck]

I think you mean that the test will do:

set_current_device(dev);
// ...
d = get_current_devic();
assert(d == dev);

Correct?

What do you mean by "check thread ID"?

[gmlueck]

We don't usually write tests in the CTS to detect compilation failures for illegal programs. In fact, we should not do this in the CTS because there is no requirement for a SYCL implementation to diagnose a compilation error in this case.

[gmlueck]

What does "async task" mean? Is this a SYCL host task?

Assuming you mean host task, we should not test this case. There is no requirement for the compiler to diagnose a compile-time error. In fact, I don't think DPC++ would generate a compilation error in this case because we can't tell at compile time whether the call comes from a host task.

[gmlueck]

Sorry, I can't follow this description. Is this test different from the one above for get_current_device?

[dklochkov-emb]

Some specific comments below. At a higher level, I think there are three things to test:

1. Call set_current_device(d) then call get_current_device() from the same thread. Make sure it returns d.
2. Do not call set_current_device. Call get_current_device() and make sure it returns the default device.
3. Create multiple threads. Each thread calls the following in a loop: set_current_device(d) followed by get_current_device(). Each thread uses a different device d (if the system has more than one device). Make sure the call to get_device returns the same device that was set in set_current_device. This tests that each thread is keeping a separate copy of the "current device".

@gmlueck What do you think about additional test: we should check if one thread sets device and another reads it? I.e. compare value(device name) read from thread 2 and value(device name) set earlier in thread 1.

[gmlueck]

What do you think about additional test: we should check if one thread sets device and another reads it? I.e. compare value(device name) read from thread 2 and value(device name) set earlier in thread 1.

This sounds like a wrong test. The spec for set_current_device says:

Sets the current default device to dev for the calling host thread.

Therefore, setting the current device in one thread should not be visible in another thread.

[gmlueck]

You can get the default device just by default-constructing a device object:

void test() {
  sycl::device d;  // d is the default device
}

There is no need to use an environment variable.

[gmlueck]

I was thinking about this more. A better test would be:

• Thread 1 does:
  • Call set_current_device(d1)
  • Signal thread 2
  • Wait for thread 2 to signal
  • Call get_current_device() and make sure it returns d1.
• Thread 2 does:
  • Call set_current_device(d2)
  • Signal thread 1
  • Wait for thread 1 to signal
  • Call get_current_device() and make sure it returns d2.

Here d1 and d2 should be different devices, if the system has more than one device. There is no need to run this in a loop.