View page source

warp.sample_cdf#

warp.sample_cdf(state: uint32, cdf: Array[float32]) → int#

  • Kernel

Sample a discrete distribution by inverse-transform sampling of a CDF.

Draws a uniform value u in [0.0, 1.0) and returns the index of the first entry of cdf greater than or equal to u via binary search. cdf must be a 1D, monotonically non-decreasing array normalized so its last element is 1.0 (for example, a cumulative sum of non-negative weights divided by their total). The returned index lies in [0, len(cdf) - 1] and selects the sampled bin.

Unlike the other random built-ins, this function is only callable from within kernels, where it advances state in place (see rand_init()).

Build a normalized CDF in Python, then sample it inside a kernel:

@wp.kernel
def sample_bins(seed: int, cdf: wp.array[float], out: wp.array[int]):
    i = wp.tid()
    rng = wp.rand_init(seed, i)
    out[i] = wp.sample_cdf(rng, cdf)  # index in [0, len(cdf) - 1]

weights = np.array([0.1, 0.4, 0.5], dtype=np.float32)
cdf = wp.array(np.cumsum(weights) / weights.sum(), dtype=float)
out = wp.empty(1000, dtype=int)
wp.launch(sample_bins, dim=out.shape, inputs=[42, cdf], outputs=[out])
Parameters:

  • state – RNG state, advanced in place (see rand_init()).

  • cdf – Normalized, non-decreasing cumulative distribution (1D float array).

Returns:

The sampled index into cdf.