Skip to main content
. 2021 Mar 29;21(7):2364. doi: 10.3390/s21072364
Algorithm 1 Layer-wise PSM
  • 1:

    l.latency: Execution time on CPU or GPU per layer

  • 2:

    l.FPBPdatacopy: Data transfer time on FP and BP per layer

  • 3:

    l.UPdatacopy: Data transfer time on UP per layer

  • 4:

    for i=0net.numberOfLayers do

  • 5:

        l=net.layers[i]

  • 6:

        min=

  • 7:

        if l.updateisTrue then

  • 8:

            l.processor=0

  • 9:

            for j=08 do

  • 10:

               temp=l.latency[i]+l.FPBPdatacopy[j]+l.UPdatacopy[j]

  • 11:

               if min>temp then

  • 12:

                   min=temp

  • 13:

                   net.processors[i]=j

  • 14:

        else

  • 15:

            for j=04 do

  • 16:

               temp=l.latency[i]+l.FPBPdatacopy[j]

  • 17:

               if min>temp then

  • 18:

                   min=temp

  • 19:

                   net.processors[i]=j

  •  

    return net.processors