search

pulp-platform / neureka

2-8bit weights, 8-bit activations flexible Neural Processing Engine for PULP clusters
Other
19 stars 2 forks source link

modify the SW to always generate weights aligned #16

Open arpansur opened 1 month ago

arpansur commented 1 month ago

The neureka implementation here relies on TP_IN=32 input channels and is compatible with the highest bandwidth offered by the HCI which is only possible in the aligned access. Thus, force in the stimuli generation to store the weight in the aligned manner. __attribute__ ((aligned(4)))

FrancescoConti commented 1 month ago

This refers to CONV3x3 mode right? not 1x1.