threshold-barrelshift4

4-bit left barrel shifter. Shifts by variable amount 0-3.

Function

barrelshift4(a3, a2, a1, a0, s1, s0) = [a3, a2, a1, a0] << s

where s = 2*s1 + s0 (shift amount 0-3)

Examples

Input	Shift	Output
0001	0	0001
0001	1	0010
0001	2	0100
0001	3	1000
1010	1	0100
1111	2	1100

Architecture

Layer 1: (data_bit AND shift_match) detectors (10 neurons)
  Each neuron fires when a specific data bit is 1 AND shift amount matches.

  a3_s00: a3 AND s=00 (y3 source when s=0)
  a2_s00: a2 AND s=00 (y2 source when s=0)
  a2_s01: a2 AND s=01 (y3 source when s=1)
  a1_s00: a1 AND s=00 (y1 source when s=0)
  a1_s01: a1 AND s=01 (y2 source when s=1)
  a1_s10: a1 AND s=10 (y3 source when s=2)
  a0_s00: a0 AND s=00 (y0 source when s=0)
  a0_s01: a0 AND s=01 (y1 source when s=1)
  a0_s10: a0 AND s=10 (y2 source when s=2)
  a0_s11: a0 AND s=11 (y3 source when s=3)

Layer 2: OR gates combining relevant sources (4 neurons)
  y3 = a3_s00 OR a2_s01 OR a1_s10 OR a0_s11
  y2 = a2_s00 OR a1_s01 OR a0_s10
  y1 = a1_s00 OR a0_s01
  y0 = a0_s00

Parameters

Inputs	6 (4 data + 2 shift)
Outputs	4
Neurons	14
Layers	2
Parameters	114
Magnitude	61

Usage

from safetensors.torch import load_file
# See model.py for full implementation

# barrelshift4(0,0,0,1, 0,1) = [0,0,1,0]  # 0001 << 1 = 0010
# barrelshift4(0,0,0,1, 1,1) = [1,0,0,0]  # 0001 << 3 = 1000

License

MIT