[Performance] onnxruntime 3 times slower than pytorch on cpu

Nobody has claimed this yet.

### Describe the issue

I converted a xlnet model to onnx, and the model is 3 times slower compared to pytorch in centos 7. I tried add_session_config_entry("session.set_denormal_as_zero", "1") based on previous issues. But still same performance.

It is interesting that on mac for the same model I am getting 2 times faster compared to pytorch.

### To reproduce

self.sessopts = onnxruntime.SessionOptions()
self.sessopts.graph_optimization_level = (onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL )
self.sessopts.add_session_config_entry("session.set_denormal_as_zero", "1")
self.sessopts.intra_op_num_threads=4
self.session = onnxruntime.InferenceSession(model_filename, sess_options =self.sessopts, providers=[available_provider]
)

### Urgency

we have a release soon, we were hoping to deploy onnx model

### Platform

Linux

### OS Version

CentOS Linux 7 (Core)

### ONNX Runtime Installation

Built from Source

### ONNX Runtime Version or Commit ID

1.13.1

### ONNX Runtime API

Python

### Architecture

X86

### Execution Provider

Default CPU

### Execution Provider Library Version

_No response_

### Model File

_No response_

### Is this a quantized model?

No

Start with the Python reproduction snippet and the reported environment: ONNX Runtime 1.13.1, CentOS 7, x86, and the default CPU provider. The model file and provider library version are not supplied, so first obtain a reproducible XLNet model and benchmark it against PyTorch on the stated platform. Done means identifying the source of the slowdown or documenting the missing conditions needed to investigate it.

Written by the indexing model from the issue text.