torch.nextafter has an incorrect result for bf16 on XPU

[guangyey]

As titled, this is a regression. IPEX could give a correct result for bf16 on XPU.

import torch

torch.manual_seed(0)

a = torch.randn(5, dtype=torch.bfloat16)
b = torch.randn(5, dtype=torch.bfloat16)
print(f"a: {a}")
print(f"b: {b}")
print(torch.nextafter(a, b))

x_a = a.to('xpu')
x_b = b.to('xpu')
print(torch.nextafter(x_a, x_b))

Output:

a: tensor([ 1.5391, -0.2930, -2.1719,  0.5703, -1.0859], dtype=torch.bfloat16)
b: tensor([-1.3984,  0.4043,  0.8398, -0.7188, -0.4043], dtype=torch.bfloat16)
tensor([ 1.5312, -0.2910, -2.1562,  0.5664, -1.0781], dtype=torch.bfloat16)
tensor([ 1.5391, -0.2930, -2.1719,  0.5703, -1.0859], device='xpu:0',
       dtype=torch.bfloat16)

Versions

None

cc @fengyuan14 @xytintel