Optimized ops in Kelvin TFLM

The following table is a list of currently optimized ops in Kelvin TFLM. The relevant source code can be found located here.

Non-Convolutional Ops

OpSupported Data TypeComments
Elementwise Adds8, s16, s32Rescaling with offset and shift, clamping
Leaky ReLUs8, s16
Max Poolings8

Convolutional Ops

OpWeightsActivationBiasComments
Depthwise Conv2ds8s16s64filter size 3x1
Depthwise Conv2ds8s8s64output depth % 32 == 0
Conv2ds8s16s32
Conv2ds8s16s64filter size 1x1, filter depth % 32 == 0
Conv2ds8s16s64filter size 1xn, grouped or ungroups
Conv2ds8s8s32filter size 1x1, output depth % 8 == 0
Conv2ds8s8s32filter depth % 32 == 0
Conv2ds8s8s32filter shape == (48x3x1x48)