XNN Int8 ReLU gives incorrect outputs

[GregoryComer]

🐛 Describe the bug

When running standalone ReLU ops in int8 on XNNPACK, there appears to be some sort of correctness or memory safety issue causing incorrect outputs. Playing around with various inputs, there doesn't seem to be a clear pattern, which makes me suspect that there is an issue handling the tensor memory or dtype somewhere.

Repro:

import torch
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
from executorch.exir import to_edge_transform_and_lower
from executorch.runtime import Runtime

class Model(torch.nn.Module):
    def forward(self, x):
        return torch.nn.ReLU()(x)

inputs = (
    torch.Tensor([-128, -127, -126, -10, 5]).to(torch.int8),

)
et_program = to_edge_transform_and_lower(
    torch.export.export(Model(), inputs),
    partitioner=[XnnpackPartitioner()]
).to_executorch()

print(et_program.exported_program())

runtime = Runtime.get()
program = runtime.load_program(et_program.buffer)
method = program.load_method("forward")
et_out = method.execute(inputs)
ref_out = Model()(*inputs)

print(f"ET:  {et_out}")
print(f"Ref: {ref_out}")

Output:

ET:  [tensor([0, 0, 0, 0, 0], dtype=torch.int8)]
Ref: tensor([0, 0, 0, 0, 5], dtype=torch.int8)

Versions

N/A

cc @digantdesai @mcr229 @cbilgin

[digantdesai]

Can you dump a json from here?

Also do we wire int8 to qu8 with fake scale/zp? And lastly what if you try quantizing fp32 with bogus scale/zp through all ones or something as calibration input?