Daisyden/stock pt

[daisyden]

Fixes #ISSUE_NUMBER

[EikanWang]

It is better to be a utility function rather than a if-else block.

[daisyden]

updated with this function:
def any_common_cpu_device_one():
return OpDTypes.any_common_cpu_xpu_one if TEST_XPU else OpDTypes.any_common_cpu_cuda_one

[EikanWang]

Abstract the check as a utility function, like has_gpu_device.

def has_gpu_device(devices: List[str]):
    return "cuda" in device or "xpu" in device

[daisyden]

def has_gpu_device(devices: List[str]):
return "cuda" in devices or "xpu" in devices

updated with this.

[EikanWang]

What does it serve?

[daisyden]

removed.

[EikanWang]

It may not make sense to add such a parameter for xpu specific.

[daisyden]

Can we replace it with "skipDevices=[xpu]' ? so that other devices can also use it to skip an op.

[daisyden]

Updated the skip mechanism:
In OpInfo class define two fields:

   # skip the test for a device
    skip_device: Tuple = tuple()

    # enable the test for a device
    enable_skipped_device: Tuple = tuple()

By default skip_device=('xpu', ), if enable_skipped_device=('xpu') is defined in op_db, xpu will be enabled, otherwise xpu will be added to skips.

self.skip_device = ('xpu',)
        for device in (set(self.skip_device).difference(set(self.enable_skipped_device))):
            if self.skips is not None:
                self.skips = (*self.skips, DecorateInfo(unittest.skip, device_type=device, dtypes=None))
            else:
                self.skips = (DecorateInfo(unittest.skip, device_type=device, dtypes=None))

The enable_skipped_device will be passed to OpsInfo __init__ and __post_init__ through kwargs, so no need to pass it manually in all the inherited class like BinaryUfuncInfo.

[daisyden]

Design is changed by using a util function and xpu_op_db.yaml to skip unsupported ops.

[EikanWang]

What does it mean?

[daisyden]

I will remove the comments.

[EikanWang]

I suppose the tests should include XPU but not only for XPU, right?

[daisyden]

yes, the "only_for" is for testing.

[EikanWang]

Regarding this logic, we can refine onlyOn by extending it support device list. The code could be as follows.

class onlyOn:

    def __init__(self, device_type: Union[str, List[str]]):
        self.device_types = []
        if isinstance(device_type, str):
            self.device_types.append(device_type)
        else:
            assert isinstance(device_type, list)
            self.device_types = device_type

    def __call__(self, fn):

        @wraps(fn)
        def only_fn(slf, *args, **kwargs):
            if slf.device_type not in self.device_types:
                reason = f"Only runs on {self.device_types}"
                raise unittest.SkipTest(reason)

            return fn(slf, *args, **kwargs)

        return only_fn

With the refined onlyOn, the above code could be as follows.

def onlyCUDAAndXPU(fn):
    onlyOn(["cuda", "xpu"])(fn)

[daisyden]

Thanks, will do the update.

[EikanWang]

Is it possible to provide a utility function dedicated to xpu to enable some particular operations to be tested on xpu? Otherwise, we have to add more codes with more operations being enabled. And when XPU op coverage is good enough, the code is useless.

[daisyden]

Added enable_skipped_op_dict in common_utils.py to read the xpu_op_db.yaml file. Added enable_skipped_device in OpInfo to skip the unsupported op.

[EikanWang]

The information is per op, why do we need to define two variables?

[daisyden]

Design is changed, enabled xpu test with xpu_op_db.yaml, no new fields are needed in OpInfo.

[daisyden]

Tested on all the classes of test_ops.py, with command:

PYTORCH_TESTING_DEVICE_ONLY_FOR="xpu" PYTORCH_ENABLE_XPU_FALLBACK=1 PYTORCH_TEST_WITH_SLOW=1 pytest -v test_ops.py

The log is attached.
0526_4.log