Created by: jczaja
Changes discussed here are are introducing Reusing concept for MKL-DNN primitive descriptors (computational ones) for fp32 mkl-dnn ops (conv, conv transpose, softmax and batch norm).
Motivation:
- To be independent from creation time of primitive descriptors, that may vary from version of MKL-DNN.
- Performance improvement
Performance improvements (SKX8180, ResNet50):
- Significant boost for small workload when lots of computing resources ( BS=1, num threads=28) : ~10%
- Tiny improvement for more often use scenarios (BS=1, num_threads=1) : ~0.5%