zhanghb97
commented
2 years ago The default lowering pipeline will use the scalar operations, so the poor performance is expected, we can optimize the implementation based on this. For more details, we can discuss at the weekly meeting and then move forward.
Add buddy fir's benchmark, in comparison to kfr's implementation.
Notice that this fir operation is lowered to
linalg.conv_1d, then to affine loops. It's very slow at this moment.