Sampling through Algorithmic Diffusion in non-convex Perceptron problems