Test various strategies for creating arrays of random numbers, some with serial dependencies

random_numbers.py

import os
os.environ["NUMBA_LOOP_VECTORIZE"] = "0"
os.environ["NUMBA_SLP_VECTORIZE"] = "0"

from datetime import datetime

from numba import jit, uint32, uint64
import numpy as np


@jit("uint32(uint32)")
def lcg(seed):
    temp = uint64(seed) * 2862933555777941757 + 3037000493
    return temp >> 32


@jit
def generate_random_numbers(n):
    result = np.empty((n,), dtype=np.uint32)
    result[0] = uint32(1)
    for i in range(1, n):
        # 🙁 result[i] has a data dependency on result[i - 1], preventing
        # instruction-level parallelism:
        result[i] = lcg(result[i - 1])
    return result


def to_image(f):
    return f(256 * 256 // 4).view(np.uint8).reshape((256, 256))


@jit
def generate_not_so_random_numbers(n):
    result = np.empty((n,), dtype=np.uint32)
    for i in range(1, n):
        # 🤔 As a temporary experiment, remove all data dependencies that are
        # preventing ILP, allowing us to see how much faster we can go:
        result[i] = lcg(i)
    return result


@jit
def generate_random_numbers_3(n):
    result = np.empty((n,), dtype=np.uint32)

    # Make sure we have 4 sufficiently different starting points:
    result[0] = uint32(1)
    result[1] = lcg(lcg(result[0]) + 1)
    result[2] = lcg(lcg(result[1]) + 1)
    result[3] = lcg(lcg(result[2]) + 1)

    # 😎 Do 4 unrelated calculations that don't share data dependencies with
    # each other, allowing the CPU to run them in parallel:
    for i in range(1, n // 4):
        result[i * 4    ] = lcg(result[(i - 1) * 4    ])
        result[i * 4 + 1] = lcg(result[(i - 1) * 4 + 1])
        result[i * 4 + 2] = lcg(result[(i - 1) * 4 + 2])
        result[i * 4 + 3] = lcg(result[(i - 1) * 4 + 3])

    # Calculate the remaining last few values:
    for i in range(n % 4):
        result[n - (3 - i)] = lcg(result[n - (3 - i) - 1])

    return result


@jit
def generate_random_numbers_4(n):
    result = np.empty((n,), dtype=np.uint32)
    # 😎 Use simple variables to make it faster to access the previous value:
    random_number1 = uint32(1)
    random_number2 = lcg(lcg(random_number1) + 1)
    random_number3 = lcg(lcg(random_number2) + 1)
    random_number4 = lcg(lcg(random_number3) + 1)

    for i in range(n // 4):
        random_number1 = lcg(random_number1)
        random_number2 = lcg(random_number2)
        random_number3 = lcg(random_number3)
        random_number4 = lcg(random_number4)
        # 😎 Less math, no more array reads:
        result[i * 4    ] = random_number1
        result[i * 4 + 1] = random_number2
        result[i * 4 + 2] = random_number3
        result[i * 4 + 3] = random_number4

    for i in range(n % 4):
        result[n - (3 - i)] = lcg(result[n - (3 - i) - 1])

    return result


def main():
    fns = [
        generate_random_numbers,
        generate_not_so_random_numbers,
        generate_random_numbers_3,
        generate_random_numbers_4,
    ]
    for fn in fns:
        print(f"{'-' * 30} {fn.__name__}")
        for _ in range(10):
            s = datetime.now()
            to_image(fn)
            e = datetime.now()
            print(f"Elapsed {e - s}")

if __name__ == '__main__':
    main()

So those final results match the text I think? First is slowest, second is fast (but wrong), third is correct (but in the middle speedwise), fourth is fast again.

Yes, much more like what the text says.