tensorflow: tf.mul after tf.split + tf.sigmoid produces wrong numerical results with MKL enabled

Issue Type

Bug

Have you reproduced the bug with TF nightly?

No

Source

binary

Tensorflow Version

intel-tensorflow 2.8 - 2.12

Custom Code

No

OS Platform and Distribution

Ubuntu 22.04

Mobile device

No response

Python version

3.10

Bazel version

No response

GCC/Compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current Behaviour?

When running with an MKL enabled tensorflow (e.g. intel-tensorflow from pypi) (or self-compiled with --config=mkl). Starting with tensorflow 2.8.0 up until 2.12.0 The attached code produces the wrong numerical result. (1.7615 vs expected 2.6439).

If line 25 is changed to m = sig * (b + 0.0) one can get the correct result. This issue does not occur if installing “vanilla” tesorflow from pip with pip install tensorflow.

This issue also does not occur if one uses tf.exp or tf.log instead of tf.sigmoid.

Standalone code to reproduce the issue

#!/usr/bin/env python3

import math

import tensorflow as tf
import numpy as np

def sigmoid(x):
  return 1 / (1 + math.exp(-x))

data = [[[2.0, 3.0]]]

tf.compat.v1.disable_eager_execution()

s = tf.compat.v1.Session()
p = tf.compat.v1.placeholder(dtype=tf.float32)
a, b = tf.split(p, 2, axis=2)
sig = tf.sigmoid(a)
m = sig * b

out = s.run([p, m], feed_dict={p: data})
print(out)

print('computed: ', out[-1][0,0,0])
print('expected: ', sigmoid(data[0][0][0]) * data[0][0][1])

Relevant log output

2023-06-21 18:37:35.713027: I tensorflow/core/platform/cpu_feature_guard.cc:151] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-06-21 18:37:35.715264: I tensorflow/core/common_runtime/process_util.cc:146] Creating new thread pool with default inter op setting: 
[array([[[2., 3.]]], dtype=float32), array([[[1.7615942]]], dtype=float32)]
computed:  1.7615942
expected:  2.642391233933647