examples/autodiff_forward.viv

# Forward-mode automatic differentiation via dual numbers.
#
# A `Dual` carries (value, derivative). Every operation propagates the derivative
# (e.g. d(a*b) = a*db + b*da). Seeding the variable of interest with derivative 1
# (others 0) and evaluating in dual arithmetic yields the exact derivative w.r.t.
# that variable. The gradient of an n-input function is n such passes.
#
# Exact, no finite-difference error. Records make it a plain library (no compiler
# change).

record Dual { v: Num, d: Num }

def dconst(x) -> Dual {
    return Dual { v: x, d: 0.0 }
}
def dvar(x) -> Dual {
    return Dual { v: x, d: 1.0 }
}           # the seeded input
def dadd(a: Dual, b: Dual) -> Dual {
    return Dual { v: a.v + b.v, d: a.d + b.d }
}
def dsub(a: Dual, b: Dual) -> Dual {
    return Dual { v: a.v - b.v, d: a.d - b.d }
}
def dmul(a: Dual, b: Dual) -> Dual {
    return Dual { v: a.v * b.v, d: a.v * b.d + a.d * b.v }
}
def ddiv(a: Dual, b: Dual) -> Dual {
    return Dual { v: a.v / b.v, d: (a.d * b.v - a.v * b.d) / (b.v * b.v) }
}
def dexp(a: Dual) -> Dual {
    e = exp(a.v); return Dual { v: e, d: e * a.d }
}
def dsqrt(a: Dual) -> Dual {
    r = sqrt(a.v); return Dual { v: r, d: a.d / (2.0 * r) }
}

# A scalar function:  f(x) = x^3 - 2x + exp(-x^2)
def f(x: Dual) -> Dual {
    x2 = dmul(x, x)
    x3 = dmul(x2, x)
    lin = dmul(dconst(2.0), x)
    negx2 = dmul(dconst(-1.0), x2)
    return dadd(dsub(x3, lin), dexp(negx2))
}

# Evaluate f and f' at x = 2 in a single pass.  f(2) = 4.018316, f'(2) = 9.926737.
rf = f(dvar(2.0))
f_value = rf.v
f_deriv = rf.d
show f_value
show f_deriv

# A two-variable function:  g(x, y) = x^2 * y + exp(x * y).
def g(x: Dual, y: Dual) -> Dual {
    x2 = dmul(x, x)
    t1 = dmul(x2, y)
    t2 = dexp(dmul(x, y))
    return dadd(t1, t2)
}

# Gradient at (1.5, 0.5): one forward pass per component (seed x, then seed y).
gx = g(dvar(1.5), dconst(0.5))
gy = g(dconst(1.5), dvar(0.5))
g_value = gx.v
dg_dx = gx.d              # 2xy + y e^{xy} = 2.558467
dg_dy = gy.d              # x^2 + x e^{xy} = 5.425401
show g_value
show dg_dx
show dg_dy

# SPDX-License-Identifier: MIT OR Apache-2.0
# SPDX-FileCopyrightText: 2026 The Viv authors