Event timing strange behaviour

[craig-warren]

Hi,

I'm getting some very strange values when trying to time events. The results are orders of magnitude different from expected. The is an example of a function which is called repeatedly in a for loop:

def update_magnetic(self):
        """Update magnetic field components."""
        event = self.update_magnetic_dev(np.int32(self.grid.nx),
                                         np.int32(self.grid.ny),
                                         np.int32(self.grid.nz),
                                         self.grid.ID_dev,
                                         self.grid.Hx_dev,
                                         self.grid.Hy_dev,
                                         self.grid.Hz_dev,
                                         self.grid.Ex_dev,
                                         self.grid.Ey_dev,
                                         self.grid.Ez_dev)
        event.wait()
        self.compute_time += (event.profile.end - event.profile.start)*1e-9

Multiple functions can update self.compute_time but every function has a event.wait(). When I print self.compute_time at the end it is not what is expected. I'm I doing something stupid here?

Thanks,
Craig

[inducer]

This looks fine from a quick look.

• What ICD is this?
• Is your queue in profiling mode?

[craig-warren]

Thanks @inducer.

• Profiling is enabled on the queue.
• How can I check what ICD is being used? I'm using macOS Monterey (12.2) on a Mac Pro with an AMD Radeon RX580 GPU.

[inducer]

How can I check what ICD is being used?

clinfo, or by printing properties of Device, or by examples/dump-properties.py.

If kernel event profiling is broken on that ICD (it very well might be, the Apple ICDs don't have a stellar reputation), maybe marker-based timing can work? https://documen.tician.de/pyopencl/runtime_queue.html?highlight=enqueue_marker#pyopencl.enqueue_marker

[craig-warren]

Thanks. Definitely Apple ICD being used.

Can you explain a bit more how I could try the marker-based timing?

Do you think its worth me trying the pocl ICD for CPU? - https://documen.tician.de/pyopencl/misc.html#enabling-access-to-cpus-and-gpus-via-py-opencl

[inducer]

Can you explain a bit more how I could try the marker-based timing?

Sure: enqueue a marker before what you'd like to time, enqueue a marker after, wait on both events, look at the time stamp difference to obtain timing.

Do you think its worth me trying the pocl ICD for CPU?

I'm confident it'll work. Folks around me use it all the time. But it'll likely not go much faster than CPU C code. Don't expect it to magically vectorize without substantial help.

[craig-warren]

Marker-based timing appears to work with the Apple ICD on CPU but not GPU - it shows zero time.

I haven't been able to get the pocl ICD working yet. It will install with conda but isn't picked up yet, through either clinfo or pyopencl. Any thoughts?