Wrong Floor output on very large input

[nnreduce-bot]

Describe the bug

When floor's input is very large (say 10**10), it returns very strange outputs that are different to PyTorch and Numpy. This seems to be related to shape. When we first tried a random shape [2, 1] it works fine and we need to exactly use [2, 2] as its input tensor shape to reproduce the issue.

Urgency
none.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 20.04
• ONNX Runtime installed from (source or binary): source
• ONNX Runtime version: 1.12
• Python version:3.8

To Reproduce

import torch
import numpy as np

class Net(torch.nn.Module):
    def forward(self, i0):
        return torch.floor(i0),

sh = [2, 2] # sh = [2, 1]  # no bug
inputs = (torch.ones(*sh, dtype=torch.float32) * 10**10.,)
model = Net()
model.eval()
outputs_torch = [o.numpy() for o in model(*inputs)]
iname = [f"i{i}" for i in range(len(inputs))]
oname = [f"o{i}" for i in range(len(outputs_torch))]
torch.onnx.export(model, inputs, "output.onnx",
                  input_names=iname, output_names=oname, opset_version=14)
from onnx import checker
import onnxruntime as ort
checker.check_model("output.onnx", full_check=True)
sess = ort.InferenceSession("output.onnx", providers=["CPUExecutionProvider"])
outputs_ort = sess.run(oname, {k: v.numpy() for k, v in zip(iname, inputs)})
np.testing.assert_allclose(outputs_ort, outputs_torch)

Expected behavior

No exception.

[faxu]

For PyTorch export issues, please file in https://github.com/pytorch/pytorch/issues