search

TENNLab-UTK / fpga

FPGA neuromorphic elements, networks, processors, tooling, and software interfaces.
Mozilla Public License 2.0
1 stars 0 forks source link

RISP neuron model SUM_WIDTH is incorrectly calculated #14

Closed BGull00 closed 4 weeks ago

BGull00 commented 1 month ago

The SUM_WIDTH field in risp_neuron.sv is being calculated incorrectly for the following neuron case:

The incorrect calculation uses this line of code: localparam SUM_WIDTH = $clog2(NUM_INP * ((1 << (CHARGE_WIDTH - 2)) - 1) + POTENTIAL_ABS_MAX) + 1;

In this case, SUM_WIDTH results in 1. However, the incoming synapses to the neurons have weights of 1, 1, and -1, respectively. The correct SUM_WIDTH should be 2 in this case.

keegandent commented 1 month ago

I think it might be best to rework this so it is based on the negative extrema of signed integers instead of constantly dealing with the -1 shenanigans of the positive end.

keegandent commented 4 weeks ago

@BGull00, I did a fair amount of spreadsheet math for this, and put the RTL changes in. The neuron is not "aware" of the synapse weights, so it wouldn't be able to optimize SUM_WIDTH based on them; it assumes it can receive any signed integer of CHARGE_WIDTH on any input port at any time. The synthesis tool is obviously free to optimize further, but I think to preserve design abstraction and modularity, I don't want to add that level of awareness to the neuron. As a result, SUM_WIDTH values may be overly generous.