From 61dc29e68b5bfd11253a3e3b8d3993c2134d39ea Mon Sep 17 00:00:00 2001
From: Nicolas Kruse <nicolas@nonan.net>
Date: Wed, 3 Dec 2025 17:30:38 +0100
Subject: [PATCH] added auto grad feature "grad(..)"

---
 src/copapy/__init__.py  |   2 +
 src/copapy/_autograd.py | 102 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 104 insertions(+)
 create mode 100644 src/copapy/_autograd.py

diff --git a/src/copapy/__init__.py b/src/copapy/__init__.py
index 3948a7e..b8216c6 100644
--- a/src/copapy/__init__.py
+++ b/src/copapy/__init__.py
@@ -3,6 +3,7 @@ from ._basic_types import NumLike, variable, generic_sdb, iif
 from ._vectors import vector, distance, scalar_projection, angle_between, rotate_vector, vector_projection
 from ._matrices import matrix, identity, zeros, ones, diagonal
 from ._math import sqrt, abs, sign, sin, cos, tan, asin, acos, atan, atan2, log, exp, pow, get_42, clamp, min, max, relu
+from ._autograd import grad
 
 __all__ = [
     "Target",
@@ -39,4 +40,5 @@ __all__ = [
     "angle_between",
     "rotate_vector",
     "vector_projection",
+    "grad"
 ]
diff --git a/src/copapy/_autograd.py b/src/copapy/_autograd.py
new file mode 100644
index 0000000..613256d
--- /dev/null
+++ b/src/copapy/_autograd.py
@@ -0,0 +1,102 @@
+from . import variable, vector
+import copapy.backend as cpb
+from typing import Any, Sequence, overload
+import copapy as cp
+from ._basic_types import Net, unifloat
+
+
+@overload
+def grad(var: variable[Any], to: variable[Any]) -> unifloat: ...
+@overload
+def grad(var: variable[Any], to: Sequence[variable[Any]]) -> Sequence[unifloat]: ...
+@overload
+def grad(var: variable[Any], to: vector[Any]) -> vector[float]: ...
+def grad(var: variable[Any], to: variable[Any] | Sequence[variable[Any]] | vector[Any]) -> unifloat | Sequence[unifloat] | vector[float]:
+    edges = cpb.get_all_dag_edges([var.source])
+    ordered_ops = cpb.stable_toposort(edges)
+
+    net_lookup = {net.source: net for node in ordered_ops for net in node.args}
+    grad_dict: dict[Net, unifloat] = dict()
+
+    def add_grad(val: variable[Any], gradient_value: unifloat) -> None:
+        grad_dict[val] = grad_dict.get(val, 0.0) + gradient_value
+
+    for node in reversed(ordered_ops):
+        print(f"-->   {'x' if node in net_lookup else ' '}", node, f"{net_lookup.get(node)}")
+        if node.args:
+            args: Sequence[Any] = [v for v in node.args]
+            g = 1.0 if node is var.source else grad_dict[net_lookup[node]]
+            opn = node.name.split('_')[0]
+            x: variable[Any] = args[0] 
+            y: variable[Any] = args[1] if len(args) > 1 else x
+
+            if opn in ['ge', 'gt', 'eq', 'ne']:
+                pass  # Derivative is 0
+
+            elif opn == 'add':
+                add_grad(x, g)
+                add_grad(y, g)
+
+            elif opn == 'sub':
+                add_grad(x, g)
+                add_grad(y, -g)
+
+            elif opn == 'mul':
+                add_grad(x, y * g)
+                add_grad(y, x * g)
+
+            elif opn == 'div':  
+                add_grad(x, g / y)
+                add_grad(y, -x * g / (y**2))
+
+            elif opn == 'pow':
+                add_grad(x, (y * (x ** (y - 1))) * g)
+                add_grad(y, (x ** y * cp.log(x)) * g)
+
+            elif opn == 'sqrt':
+                add_grad(x, g * (0.5 / cp.sqrt(x)))
+
+            elif opn == 'abs':
+                add_grad(x, g * cp.sign(x))
+
+            elif opn == 'sin':
+                add_grad(x, g * cp.cos(x))
+
+            elif opn == 'cos':
+                add_grad(x, g * -cp.sin(x))
+
+            elif opn == 'tan':
+                add_grad(x, g * (1 / cp.cos(x) ** 2))
+
+            elif opn == 'asin':
+                add_grad(x, g * (1 / cp.sqrt(1 - x**2)))
+
+            elif opn == 'acos':
+                add_grad(x, g * (-1 / cp.sqrt(1 - x**2)))
+
+            elif opn == 'atan':
+                add_grad(x, g * (1 / (1 + x**2)))
+
+            elif opn == 'atan2':
+                denom = x**2 + y**2
+                add_grad(x, g * (-y / denom))
+                add_grad(y, g * ( x / denom))
+
+            elif opn == 'log':
+                add_grad(x, g / x)
+
+            elif opn == 'exp':
+                add_grad(x, g * cp.exp(x))
+
+            elif opn == 'gt':
+                add_grad(x, g)
+                add_grad(y, -g)
+
+            else:
+                raise ValueError(f"Operation {opn} not yet supported for auto diff.")
+
+    if isinstance(to, variable):
+        return grad_dict[to]
+    if isinstance(to, vector):
+        return vector(grad_dict[dvar] for dvar in to)
+    return [grad_dict[dvar] for dvar in to]
\ No newline at end of file