Created by: sfraczek

PR types

Performance optimization

PR changes

Others

Describe

For Ernie model, adds a fuse that replaces 3 parallel FC ops that accept the same input with a single FC op and a split.

The fuse is applied only when the following attributes have the same values across 3 ops:

• in_num_col_dims,
• Scale_in,
• Scale_out,
• activation_type,
• use_mkldnn,
• padding_weights.

It works by concatenating weights and biases of FCs along dim 1 and adding a Split on FC's output along axis 2 which splits it back into 3 paths.

before

after

padding

If padding_weights is enabled it handles that case also.