Question about compute_cycles

[chuckwu2000]

Hi, @hsharma35,
I've reading about the code in accelerator.py, I found I'm confused about the compute_cycles.
In get_compute_cycles function:

"""
Compute instruction
args:
       ic: Input Channels
       oc: Output Channels
       ow: Output Width
       oh: Output Height
       kw: Output Height
       kh: Output Height
       b: Batch Size
       im2col: boolean. If true, we assume the cpu does im2col. Otherwise,
       we do convolutions channel-wise
"""
overhead = 0
if im2col:
    ni = kw * kh * ic
    no = oc
    batch = b * oh * ow
    compute_cycles = batch * ceil_a_by_b(no, self.M) * \
                    (ceil_a_by_b(ni, self.N * self.get_perf_factor(iprec, wprec)) + overhead)
    else:
            compute_cycles = b * ceil_a_by_b(oc, self.M) * \
                    ow * oh * kw * kh * \
                    (ceil_a_by_b(ic, self.N * self.get_perf_factor(iprec, wprec)) + overhead)

    return compute_cycles

My question are:

1. In a systolic array, the partial sums produced by the PEs need to propagate downward to the bottom each cycle. Is the forwarding latency considered (for example, in a 3×3 systolic array, the first output would need to wait three cycles, corresponding to the array’s height)?
2. If the above assumption is correct, does the overhead account for this? If not, what exactly is the purpose of the overhead?