4-bit left barrel shifter, magnitude 61

README.md

@@ -0,0 +1,78 @@
+---
+license: mit
+tags:
+- pytorch
+- safetensors
+- threshold-logic
+- neuromorphic
+---
+
+# 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
+
+```python
+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

config.json

@@ -0,0 +1,9 @@
+{
+  "name": "threshold-barrelshift4",
+  "description": "4-bit left barrel shifter",
+  "inputs": 6,
+  "outputs": 4,
+  "neurons": 14,
+  "layers": 2,
+  "parameters": 114
+}

create_safetensors.py

@@ -0,0 +1,117 @@
+import torch
+from safetensors.torch import save_file
+
+weights = {}
+
+# Input order: [a3, a2, a1, a0, s1, s0] (6 inputs)
+# 4-bit left barrel shift by s (0-3)
+# s=0: [a3, a2, a1, a0], s=1: [a2, a1, a0, 0], s=2: [a1, a0, 0, 0], s=3: [a0, 0, 0, 0]
+
+# Layer 1: (data_bit AND shift_match) neurons
+# Each fires when specific data bit is 1 AND shift amount matches
+
+# a3_s00: a3 AND s=00 (for y3 when s=0)
+weights['a3_s00.weight'] = torch.tensor([[1.0, 0.0, 0.0, 0.0, -1.0, -1.0]], dtype=torch.float32)
+weights['a3_s00.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# a2_s00: a2 AND s=00 (for y2 when s=0)
+weights['a2_s00.weight'] = torch.tensor([[0.0, 1.0, 0.0, 0.0, -1.0, -1.0]], dtype=torch.float32)
+weights['a2_s00.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# a2_s01: a2 AND s=01 (for y3 when s=1)
+weights['a2_s01.weight'] = torch.tensor([[0.0, 1.0, 0.0, 0.0, -1.0, 1.0]], dtype=torch.float32)
+weights['a2_s01.bias'] = torch.tensor([-2.0], dtype=torch.float32)
+
+# a1_s00: a1 AND s=00 (for y1 when s=0)
+weights['a1_s00.weight'] = torch.tensor([[0.0, 0.0, 1.0, 0.0, -1.0, -1.0]], dtype=torch.float32)
+weights['a1_s00.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# a1_s01: a1 AND s=01 (for y2 when s=1)
+weights['a1_s01.weight'] = torch.tensor([[0.0, 0.0, 1.0, 0.0, -1.0, 1.0]], dtype=torch.float32)
+weights['a1_s01.bias'] = torch.tensor([-2.0], dtype=torch.float32)
+
+# a1_s10: a1 AND s=10 (for y3 when s=2)
+weights['a1_s10.weight'] = torch.tensor([[0.0, 0.0, 1.0, 0.0, 1.0, -1.0]], dtype=torch.float32)
+weights['a1_s10.bias'] = torch.tensor([-2.0], dtype=torch.float32)
+
+# a0_s00: a0 AND s=00 (for y0 when s=0)
+weights['a0_s00.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, -1.0, -1.0]], dtype=torch.float32)
+weights['a0_s00.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# a0_s01: a0 AND s=01 (for y1 when s=1)
+weights['a0_s01.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, -1.0, 1.0]], dtype=torch.float32)
+weights['a0_s01.bias'] = torch.tensor([-2.0], dtype=torch.float32)
+
+# a0_s10: a0 AND s=10 (for y2 when s=2)
+weights['a0_s10.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, 1.0, -1.0]], dtype=torch.float32)
+weights['a0_s10.bias'] = torch.tensor([-2.0], dtype=torch.float32)
+
+# a0_s11: a0 AND s=11 (for y3 when s=3)
+weights['a0_s11.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, 1.0, 1.0]], dtype=torch.float32)
+weights['a0_s11.bias'] = torch.tensor([-3.0], dtype=torch.float32)
+
+# Layer 2: OR gates for each output
+# Layer 1 order: [a3_s00, a2_s00, a2_s01, a1_s00, a1_s01, a1_s10, a0_s00, a0_s01, a0_s10, a0_s11]
+#                   0        1       2       3       4       5       6       7       8       9
+
+# y3 = a3_s00 OR a2_s01 OR a1_s10 OR a0_s11 (indices 0, 2, 5, 9)
+weights['y3.weight'] = torch.tensor([[1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]], dtype=torch.float32)
+weights['y3.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# y2 = a2_s00 OR a1_s01 OR a0_s10 (indices 1, 4, 8)
+weights['y2.weight'] = torch.tensor([[0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0]], dtype=torch.float32)
+weights['y2.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# y1 = a1_s00 OR a0_s01 (indices 3, 7)
+weights['y1.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0]], dtype=torch.float32)
+weights['y1.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+# y0 = a0_s00 (index 6)
+weights['y0.weight'] = torch.tensor([[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
+weights['y0.bias'] = torch.tensor([-1.0], dtype=torch.float32)
+
+save_file(weights, 'model.safetensors')
+
+# Verify
+def barrelshift4(a3, a2, a1, a0, s1, s0):
+    inp = torch.tensor([float(a3), float(a2), float(a1), float(a0), float(s1), float(s0)])
+
+    # Layer 1
+    l1_names = ['a3_s00', 'a2_s00', 'a2_s01', 'a1_s00', 'a1_s01', 'a1_s10', 'a0_s00', 'a0_s01', 'a0_s10', 'a0_s11']
+    l1 = []
+    for name in l1_names:
+        v = int((inp @ weights[f'{name}.weight'].T + weights[f'{name}.bias'] >= 0).item())
+        l1.append(float(v))
+    l1_tensor = torch.tensor(l1)
+
+    # Layer 2
+    outputs = []
+    for name in ['y3', 'y2', 'y1', 'y0']:
+        v = int((l1_tensor @ weights[f'{name}.weight'].T + weights[f'{name}.bias'] >= 0).item())
+        outputs.append(v)
+    return outputs
+
+print("Verifying barrelshift4...")
+errors = 0
+for val in range(16):  # all 4-bit values
+    a3, a2, a1, a0 = (val >> 3) & 1, (val >> 2) & 1, (val >> 1) & 1, val & 1
+    for s in range(4):  # all shift amounts
+        s1, s0 = (s >> 1) & 1, s & 1
+        result = barrelshift4(a3, a2, a1, a0, s1, s0)
+
+        # Expected: left shift by s
+        shifted = (val << s) & 0xF
+        expected = [(shifted >> 3) & 1, (shifted >> 2) & 1, (shifted >> 1) & 1, shifted & 1]
+
+        if result != expected:
+            errors += 1
+            if errors <= 5:
+                print(f"ERROR: val={val:04b} s={s} -> {result}, expected {expected}")
+
+if errors == 0:
+    print("All 64 test cases passed!")
+else:
+    print(f"FAILED: {errors} errors")
+
+mag = sum(t.abs().sum().item() for t in weights.values())
+print(f"Magnitude: {mag:.0f}")

model.py

@@ -0,0 +1,40 @@
+import torch
+from safetensors.torch import load_file
+
+def load_model(path='model.safetensors'):
+    return load_file(path)
+
+def barrelshift4(a3, a2, a1, a0, s1, s0, weights):
+    """4-bit left barrel shifter. Shifts left by s = 2*s1 + s0 positions."""
+    inp = torch.tensor([float(a3), float(a2), float(a1), float(a0), float(s1), float(s0)])
+
+    # Layer 1
+    l1_names = ['a3_s00', 'a2_s00', 'a2_s01', 'a1_s00', 'a1_s01', 'a1_s10', 'a0_s00', 'a0_s01', 'a0_s10', 'a0_s11']
+    l1 = []
+    for name in l1_names:
+        v = int((inp @ weights[f'{name}.weight'].T + weights[f'{name}.bias'] >= 0).item())
+        l1.append(float(v))
+    l1_tensor = torch.tensor(l1)
+
+    # Layer 2
+    outputs = []
+    for name in ['y3', 'y2', 'y1', 'y0']:
+        v = int((l1_tensor @ weights[f'{name}.weight'].T + weights[f'{name}.bias'] >= 0).item())
+        outputs.append(v)
+    return outputs
+
+if __name__ == '__main__':
+    w = load_model()
+    print('barrelshift4 examples:')
+    test_cases = [
+        (0, 0, 0, 1, 0, 0),  # 0001 << 0 = 0001
+        (0, 0, 0, 1, 0, 1),  # 0001 << 1 = 0010
+        (0, 0, 0, 1, 1, 0),  # 0001 << 2 = 0100
+        (0, 0, 0, 1, 1, 1),  # 0001 << 3 = 1000
+        (1, 0, 1, 0, 0, 1),  # 1010 << 1 = 0100
+    ]
+    for args in test_cases:
+        a3, a2, a1, a0, s1, s0 = args
+        result = barrelshift4(a3, a2, a1, a0, s1, s0, w)
+        s = s1 * 2 + s0
+        print(f'  {a3}{a2}{a1}{a0} << {s} = {result[0]}{result[1]}{result[2]}{result[3]}')