You can turn trees into numpy-style matrix operations because graphs and matrices are two sides of the same coin. I don't see the code for the binary-tree benchmark in the repo to see how it's written, but there are libraries like graphblas that use the equivalence for optimization.