Inefficient BRAM utilisation in StreamingFCLayer for int8 models

[haoxi1999]

Hi,

I have noticed the StreamingFCLayer generated by FINN can sometimes be very inefficient in BRAM instantiation. It seems that for 8-bit FC layers with more than 48 neurons or 48 input activations, threshold memories in the StreamingFCLayer are implemented with many instances of BRAMs, with very low utilisation efficiency in each BRAM (e.g. WBitsBanks = 2176 for a BRAM_18 instance). Specifying ram_style = "distributed" doesn't seem to solve this issue.

I really appreciate any help on this issue! Thanks.

Here's an example synthesis report of StreamingFCLayer with input width of 64 and 64 neurons attached.
Synthesis Report for 'Matrix_Vector_Activa'.pdf