Created by: mozga-intel

Problem with the MKLDNN performance after this change : commit

Description and solution The mkldnn performance of the newest Paddle is a worse than the code before the changes: nearly 50%, The mkldnn version of code need to have all space of the memory to work correctly. During the tests with htop the loss of power of the threads is significant and we can not keep the fluency between the threads.

The code shows the solution to this problem. Summary: this is the fall back to the previous version of the code + MKLDNN flag.