tensorflow: MKL no longer works with tensorflow 1.15

---

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): NO

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Centos 7

• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: N/A

• TensorFlow installed from (source or binary): source

• TensorFlow version (use command below): 1.15.2

• Python version: N/A

• Bazel version (if compiling from source): 0.24.1

• GCC/Compiler version (if compiling from source): gcc-6 (devtoolset-6 on centos 7)

• CUDA/cuDNN version: N/A

• GPU model and memory: N/A

• Exact command to reproduce:

 bazel build -c opt --copt=-msse4.2 --copt=-mavx \
       --copt=-O3 --config=mkl --linkopt -ldl \
       --copt=-march=x86-64 \
       //tensorflow/tools/pip_package:build_pip_package \
       //tensorflow/tools/lib_package:libtensorflow_jni \
       //tensorflow/tools/lib_package:libtensorflow \
       //tensorflow/tools/lib_package:libtensorflow_proto \

Describe the problem

libtensorflow_framework.so build this way does not have any symbols from MKL. When trying to import tensorflow from java scala, it fails with symbol not found for tensorflow::DisableMKL()

The number of MKL symbols found in libtensorflow_framework.so for 1.15 are also significantly lower than those found in 1.14.

Source code / logs

Code used to import tensorflow in scala

import org.tensorflow.Tensorflow

Note: We have to ensure libiomp5.so and libmklml_intel.so are available on library load path.

The simplest solution we found was to load the libraries manually in order. The code snippet can be seen here: https://gist.github.com/pavanky/ea6e71e3e7e52c013db844b715723be0

Error

libtensorflow_jni.so: undefined symbol: _ZN10tensorflow10DisableMKLEv

Looking at the number of symbols related to MKL:

 $ nm -D org/tensorflow/native/linux-x86_64/libtensorflow_framework.so.1 | grep -i mkl | wc -l
1
 $ nm -D org/tensorflow/native/linux-x86_64/libtensorflow_jni.so | grep -i mkl | wc -l
9
 $ nm -D org/tensorflow/native/linux-x86_64/libtensorflow_framework.so.1 | grep -i mkl
0000000000e127b0 T _ZN10tensorflow12IsMklEnabledEv

For reference, 1.14 had a lot more

 $ nm -D org/tensorflow/native/linux-x86_64/libtensorflow_framework.so.1 | grep -i mkl | wc -l
11388
 $ nm -D org/tensorflow/native/linux-x86_64/libtensorflow_jni.so | grep -i mkl | wc -l
8

---

MKL is also not available in the wheel built by the command mentioned above.

>>> import tensorflow as tf
>>> tf.python.pywrap_tensorflow.IsMklEnabled()
False