DNNl MaxPool can't get nan result when input's value is nan

[XiaobingSuper]

DNNL MaxPool should get nan value when input values are all nan.

test case:

import torch
import intel_pytorch_extension
from torch._six import inf, nan
intel_pytorch_extension.core.enable_auto_dnnl()

input = torch.full([1, 1] + 2 * [3], nan).to('dpcpp')
output = F.max_pool2d(x, 3, 3)

[EikanWang]

@pinzhenx , If a tensor cotains nan data, what's the beahavior of dnnl?

[pinzhenx]

If a tensor cotains nan data, what's the beahavior of dnnl?

UB. The observed behavior from DNNL is that it will return some values out of nowhere

[XiaobingSuper]

It will always return the minimum number of float data type(-3.4028E+38).

[EikanWang]

It will always return the minimum number of float data type(-3.4028E+38).

Could I say that DNNL treats NAN as zero?