Speed optimizations to gpio.py

[Gadgetoid]

These are discussed and shared by @sheffieldnick in #11

I have commited these changes for evaluation since they differ slightly to those in PR #12

[Gadgetoid]

I ran some tests on alternate methods of file IO and general pokery, boiling down to the following findings.

This is an absolute best case approximation of the current method of reading a pin 10 million times:

with open("/sys/class/gpio/gpio10/value", "w+") as f:
    for _ in range(10 * 1000 * 1000):
        f.seek(0)
        val = int(f.read())

This takes roughly 4min 30sec on my Pi 4.

Compare to using unbuffered bytes IO and some oldschool bytes-to-int conversion magc:

with open("/sys/class/gpio/gpio10/value", "rb", buffering=0) as f:
    for _ in range(10 * 1000 * 1000):
        f.seek(0)
        val = f.read()[0] - 48

This takes 1min 41sec on my Pi 4. Roughly 2.6x faster.

This is pretty reasonable. Avoiding a costly conversion from a "string" to an int should be faster - you'd think - but it's actually the switch from text to bytes IO that makes all the difference. Let's eliminate the conversion from whatever-to-int from the equation and just read the value. I'll do 1m iterations here because waiting is dull.

Here's regular ol' text mode:

    for _ in range(1 * 1000 * 1000):
        f.seek(0)
        val = f.read()

Which takes 25.15s.

And here's in bytes IO:

with open("/sys/class/gpio/gpio10/value", "rb", buffering=0) as f:
    for _ in range(1 * 1000 * 1000):
        f.seek(0)
        val = f.read()

Which took 9.78s.

Again the bytes method is roughly 2.6x faster and anything Python does to the values is largely inconsequential.

I applied these findings both to the read and set methods of this library (ignoring for now that the method used for writing breaks Python 2.x support) and confirmed on my oscilloscope that the result is - again - roughly 2.6x faster than this optimized version of the library.

With my own rewrite I achieve a 130KHz pin toggle frequency:

Running the same code against this change set yields roughly 53KHz:

My measured results are roughly 2.4x faster at - currently - the cost of Python 2.x support.

[Gadgetoid]

Replaced by #22