Userspace is slow? Especially on microcontrollers

0-slow.py

import json

import os

try:
    sysname = os.uname().sysname
except AttributeError:
    sysname = "micropython-darwin"

if sysname in ["esp32", "micropython-darwin"]:
    # micropython
    print(f"micropython {sysname}")

    from random import getrandbits
    from utime import ticks_us, ticks_diff
    from time import sleep

    def rand8():
        getrandbits(8)


else:
    # cpython
    print(f"cpython: {sysname}")

    from random import random
    from math import floor
    from time import time, sleep

    def rand8():
        return floor(random() * 255)

    def ticks_us():
        return round(time() * 1000 * 1000)

    def ticks_diff(t2, t1):
        return t2 - t1


pixels = []
for p in range(255):
    pix = []
    for i in range(3):
        pix.append(rand8())
    pixels.append(pix)

# before = ticks_us()
# sleep(1)
# print(f"{ticks_diff(ticks_us(), before)} usec")


def handroll(pixels):
    ret = "pixels = ["
    for p in pixels:
        ret += f"[{p[0]}, {p[1]}, {p[2]}],"


def jsonroll(pixels):
    ret = "pixels = ["
    ret += json.dumps(pixels)


before = ticks_us()
handroll(pixels)
print(f"handroll time: {ticks_diff(ticks_us(), before)} usec")

before = ticks_us()
jsonroll(pixels)
print(f"jsonroll time: {ticks_diff(ticks_us(), before)} usec")

1-outputs.txt

cpython: Darwin
handroll time: 149 usec
jsonroll time: 151 usec

----

micropython micropython-darwin
handroll time: 2320 usec
jsonroll time: 1257 usec

----

micropython esp32
handroll time: 2183061 usec
jsonroll time: 22771 usec

Almost 100x faster for 256 elements on micropython esp32!
(cpython is just about the same speed, and micropython-darwin is ~2x)

FWIW:

micropython esp32 with @micropython.native decorator on handroll
handroll time: 2155378 usec
jsonroll time: 17526 usec

micropython esp32 with @micropython.viper decorator on handroll
handroll time: 2170409 usec
jsonroll time: 16919 usec

I bet this is actually a memory allocation problem, based on this maximizing speed document.

There's probably some hinting we can do with e.g. bytearray and memoryview instead of string reallocation, but I think I'm going to push the heavy lifting out to another system (the visiting browser user-agent in this case).