Why is there a large drop in performance between 2.0.5 and 2.0.6?

[technoblogy]

Board

ESP32

Device Description

Adafruit ESP32 Feather

Hardware Configuration

No.

Version

v2.0.7

IDE Name

Arduino IDE

Operating System

macOS 10.13.6

Flash frequency

80 MHz

PSRAM enabled

no

Upload speed

921600

Description

I've compared benchmarks for three application programs running on the ESP32 version of uLisp, my Lisp interpreter for microcontrollers, and there's a big drop in performance between arduino-esp32 version 2.0.5 and 2.0.6 (the figures are execution times, so shorter times are better):

Version	2.0.5	2.0.6	2.0.7
Tak	8.1 s	14.2 s	13.8 s
Q2	21.2 s	38.1 s	37.2 s
FFT	241 ms	416 ms	408 ms

As you can see from the figures, 2.0.7 is a slight improvement, but it's still about 70% slower than 2.0.5.

Is there a reason for these differences, and is there anything I can do about it?

Sketch

For details of the benchmarks see [Benchmarks](http://www.ulisp.com/show?1EO1).

Debug Message

There is no error.

Other Steps to Reproduce

No response

I have checked existing issues, online documentation and the Troubleshooting Guide

• I confirm I have checked existing issues, online documentation and Troubleshooting guide.

[mrengineer7777]

That's interesting. Not sure what your benchmark is doing, but I'm guessing it's processor bound work. The differences are probably due to updates in the underlying IDF.

[mrengineer7777]

It's also possible the compiler optimization flags have changed.

[technoblogy]

Is there some way I can check that - I assume it's in the platform.txt file?

[mrengineer7777]

The IDF version is noted in release notes. It's also a #define and there's a function call to retrieve the string version. I believe compile flags are set in arduino-lib-builder.

…

On Sat, Apr 1, 2023, 12:09 PM David Johnson-Davies ***@***.***> wrote: Is there some way I can check that - I assume it's in the platform.txt file? — Reply to this email directly, view it on GitHub <#8019 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKQB624XEAJXNWM7GBBHOGTW7BOMVANCNFSM6AAAAAAWPURHYQ> . You are receiving this because you commented.Message ID: ***@***.***>

[technoblogy]

Sorry, why is this closed? Is there no plan to investigate what's causing it?

[mrengineer7777]

If you can specify which library is causing the performance issue and can provide a simple code example to replicate it, then we can investigate. While interesting, your issue is not a bug and would take a lot of time to research.

[technoblogy]

I've tried a pure C implementation of the first benchmark I used for the above comparison:

int tak (int x, int y, int z) {
  if (!(y < x)) return z;
  else return tak(tak(x-1,y,z), tak(y-1, z,x), tak(z-1,x,y));
} 

The calculation of tak(24, 16, 8) took exactly the same time (170ms) on 2.0.5, 2.0.6, and 2.0.7.

I'll have to think about what my Lisp interpreter might be doing that's affected by the different versions of the ESP32 core.

[mrengineer7777]

Sounds good. FYI I'm not an Espressif employee. I volunteered to help with managing their issues because they are way behind on issues and PRs. I have even fixed some issues by submitting PRs.

[dragoncoder047]

I'll have to think about what my Lisp interpreter might be doing that's affected by the different versions of the ESP32 core.

My guess from a user standpoint is that the arduino core runs freeRTOS background tasks and the uLisp evaluator calls yield() somehow (although I can't figure out where) allowing other non-uLisp stuff to run in the middle of uLisp execution. The C example clearly doesn't call yield() so no tasks switch.

@technoblogy if you want to get rid of this issue, I would maybe add a flag NOYIELD, and turn it on when in (time), and then only ever yield() if the NOYIELD flag is cleared -- similar to calling testescape() and the NOESC flag.

[technoblogy]

@dragoncoder047 thanks for the suggestions, but I think I've tracked down what the problem is. My uLisp interpreter periodically calls Serial.read() to allow the user to escape from a runaway program by entering the '~' character.

Here's a test program:

void testescape () {
  if (Serial.read() == '~') { Serial.println("escape!"); for(;;); }
}

void setup() {
  Serial.begin(9600);
  Serial.println();
  Serial.println("Start");
}

void loop() {
  unsigned long start = millis();
  for (int i=0; i<10000; i++) testescape();
  Serial.print(millis()-start); Serial.println("ms");
  for(;;);
}

and here are the timings (the figures are execution times, so shorter times are better):

Version	2.0.5	2.0.6	2.0.7
Time	71 ms	133 ms	133 ms

So the change in performance between versions of the ESP32 Arduino core could have something to do with the execution of Serial.read().

Changing the testescape() routine to:

void testescape () {
  if (Serial.available() && Serial.read() == '~') { Serial.println("escape!"); for(;;); }
}

makes the execution time for the test program 68 ms for all three versions; ie eliminates the difference.

Apologies that I reported this as a more general scare before fully tracking down what was causing it.