Created by: grygielski

PR types

Performance optimization

PR changes

Others

Describe

This PR adds conv_affine_channel_fuse_pass to MKLDNNPasses and fixes performance of fused convolution by introducing check for denormalized values. It results in increase of oneDNN MASK-RCNN performance ~4.5%.

Data from my i9 developer machine: Native: 0.0547976633974 FPS oneDNN before: 0.0613260459058 FPS oneDNN after: 0.0640848093513 FPS