warp.value#

warp.value(x: Float) → Float#
  • Kernel: true
  • Python: true
  • Differentiable: true

Return the value of a scalar jet without its derivatives, or a plain float unchanged.

Experimental

Forward-mode jets are experimental and subject to change without a formal deprecation cycle. See warp.JetSpace().

Comparisons such as a < b are not defined on jets, so a function written generically over plain floats and jets compares wp.value(a) < wp.value(b) instead. A branch chosen this way carries the derivative of the branch taken.

The signatures listed here take plain floats. Each space J = wp.JetSpace(...) or J = wp.JetSpace2(...) adds one that takes a scalar jet of that space, an instance of J.scalar such as J.seed(1.5) (not the space J itself), and returns its value field.

The result carries no forward-mode derivatives, but wp.value() is not a stop-gradient: in reverse mode it is the identity, so a warp.Tape or warp.grad() still differentiates through it.

Parameters:

x – A float16, float32 or float64 scalar, or a scalar jet (an instance of J.scalar for a jet space J).

Returns:

x itself for a plain float, otherwise the jet’s value field, a plain float of the jet’s dtype.

Example

Branch on a value in a function that takes plain floats and scalar jets alike:

J = wp.JetSpace(1)

@wp.func
def ramp_squared(x: Any):
    if wp.value(x) > 0.0:
        return x * x
    return 0.0 * x

@wp.kernel
def ramp_kernel(out: wp.array[float]):
    x = J.seed(1.5)
    y = ramp_squared(x)  # a jet
    out[0] = wp.value(y)  # its value, a plain float
    out[1] = y.grad[0]  # its derivative, d(x**2)/dx = 2x
    out[2] = ramp_squared(1.5)  # the same function on a plain float

out = wp.zeros(3, dtype=float)
wp.launch(ramp_kernel, dim=1, outputs=[out])
print(out.numpy())
[2.25 3.   2.25]