QNN Avg and Max pool with ceil_mode=True appear to have incorrect outputs

🐛 Describe the bug

When running avg and max pool operators with ceil_mode=True, many output elements are zero. There appears to be an error in the computation.

Example output:

Eager: tensor([[[[ 0.2050, -0.1577,  0.7572],
          [-0.1823,  0.8233, -0.7934],
          [ 0.6227, -0.0854, -1.4602]],

         [[-0.0229, -0.0853,  0.0022],
          [-0.0799,  0.1280, -0.8410],
          [ 0.0817, -0.0908, -0.5292]]]])
QNN: tensor([[[[ 0.2048, -0.1576,  0.0000],
          [-0.1823,  0.8232,  0.0000],
          [ 0.0000,  0.0000,  0.0000]],

         [[-0.0229, -0.0851,  0.0000],
          [-0.0798,  0.1281,  0.0000],
          [ 0.0000,  0.0000,  0.0000]]]])

Repro:

import torch

from executorch.backends.qualcomm.utils.utils import (
    generate_htp_compiler_spec,
    generate_qnn_executorch_compiler_spec,
    get_soc_to_chipset_map,
)
from executorch.backends.qualcomm.utils.utils import QcomChipset
from executorch.backends.qualcomm.utils.utils import (
    to_edge_transform_and_lower_to_qnn,
)
from executorch.exir import ExecutorchBackendConfig
from executorch.runtime import Runtime
from typing import Optional, Tuple, List, Union

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.pool = torch.nn.AvgPool2d(3, ceil_mode=True)

    def forward(self, x):
        return self.pool(x)

inputs = (torch.randn(1,2,7,7),)
model = Model()

# HTP Compiler Configuration
backend_options = generate_htp_compiler_spec(
    use_fp16=True,
)

# QNN Compiler Spec
compile_spec = generate_qnn_executorch_compiler_spec(
    soc_model=QcomChipset.SM8650,  # Your target SoC
    backend_options=backend_options,
)


# Lower to QNN backend
delegated_program = to_edge_transform_and_lower_to_qnn(
    model,
    inputs,
    compile_spec
)

# Export to ExecuTorch format
executorch_program = delegated_program.to_executorch(
    config=ExecutorchBackendConfig(extract_delegate_segments=False)
)

runtime = Runtime.get()
program = runtime.load_program(executorch_program.buffer)
method = program.load_method("forward")
output: List[torch.Tensor] = method.execute([*inputs])

print(f"Eager: {model(*inputs)}")
print(f"QNN: {output[0]}")

Versions

Commit 335de46fa76866a76599def5e34296f0ee5f6106

cc @cccclai @cbilgin