Add `wp.copysign(x, y)` builtin

[c0d1f1ed]

Description

Add wp.copysign(x, y), which returns a value with the magnitude of x and the sign of y. This matches C's copysign and NumPy's np.copysign.

Motivation

The headline use case (and motivation for landing this alongside the work in GH-1376) is forcing a specific sign on a result whose signed-zero behavior is otherwise implementation-defined, e.g. wp.min(-0.0, +0.0) returning either zero per C99's fmin/fmax allowance. With wp.copysign users can pin the sign deterministically:

result = wp.copysign(wp.min(a, b), -1.0)  # always returns -|min|

Requirements

The builtin should be supported for wp.float16 / wp.float32 / wp.float64 / wp.bfloat16. It lowers to a single instruction on every supported target:

• CUDA: libdevice __nv_copysign (one PTX instruction).
• CPU JIT (Clang/LLVM): __builtin_copysignf / __builtin_copysign (compiler intrinsic, no CRT dependency).
• Host build (MSVC fallback): ::copysign from <math.h> via crt.h.

The adjoint adj_copysign routes the gradient on x based on the sign-bit agreement between x and y (+1 when signs agree, -1 otherwise). The gradient on y is 0 almost everywhere — copysign(x, y) = |x| * sign(y) is locally constant in y. Sign comparison goes through copysign(T(1), .) so signed-zero inputs are classified by the IEEE 754 sign bit rather than < 0 (which would misclassify -0 as non-negative).

The builtin is differentiable. The gradient on x is +1 when the signs of x and y agree, -1 otherwise; the gradient on y is 0. Sign agreement is determined by the IEEE 754 sign bit, so ±0 inputs are classified correctly.

Example

import warp as wp


@wp.kernel
def k(
    x: wp.array(dtype=wp.float32),
    y: wp.array(dtype=wp.float32),
    out: wp.array(dtype=wp.float32),
):
    i = wp.tid()
    out[i] = wp.copysign(x[i], y[i])


x = wp.array([3.0, 3.0, -0.0, float("nan")], dtype=wp.float32)
y = wp.array([1.0, -1.0, 1.0, -1.0], dtype=wp.float32)
out = wp.zeros(4, dtype=wp.float32)
wp.launch(k, dim=4, inputs=[x, y], outputs=[out])
print(out.numpy())  # [3.0, -3.0, 0.0, nan]  -- magnitude of x with sign of y

[c0d1f1ed]

Implemented in 2275290.