projectgus
commented
5 years ago Hi @dsyleixa ,
Thanks for an interesting bug report! We don't have an exact answer for you, but I do have an observation.
I modified the code to plain .cpp so I could build it in IDF without Arduino, and also so I could build it on my host PC. Code here: lunar.cpp.txt
On my ESP32 with ESP-IDF, I get the same final value reported here: 54196.625000.
Building it on my 64-bit Linux PC (g++ -o lunar lunar.cpp), I get the same final result (54196.625000) as on the ESP32. Note that there is nothing ESP-specific about how I'm building it on my PC, that's treating it vanilla C++. You can see the same result online, here: http://www.cpp.sh/2xwry
I also tried with -ffast-math on my local PC and get the same final result.
Here's the different output:
Output at each step is the same for PC vs ESP32. Maybe there is some small difference in the way I modified the Arduino code to make it C++ compatible, although I don't see what it would be.
I expect this is probably something subtle with with different floating point math implementations and how they handle errors, or trade off acceptable error vs performance. The errors are accumulating differently and leading to these very different results.
As long as g++ on Linux is producing the same results, I'm inclined to think it's not a bug in ESP-IDF or the ESP32 as such. But I'm curious if you find out exactly where the root cause is...
(NB: If you want to avoid any kind of floating point errors, the only surefire way is to go to fixed point math. Or integer math where you "think big" with all your units, which is basically fixed point math.)
dsyleixa
igrr
Arduino IDE 1.8.8 different ESP32 MCUs (Adafruit Feather ESP32, ESP32 Dev Board, WROOM32) vs. different ARM Cortex boards (Adafruit Feather M4, Arduino Zero (M0), Arduino Due (ATSAM3X8E) Arduino AVR Mega2560
In the following code all ARM Cortex Boards and the Arduino Mega evalutate to 32.843750, whilst all ESP32 evaluate to 54196.625
Ref.: "HaWe" (ESP32, M3, M4, Mega2560): https://www.roboternetz.de/community/threads/73360-ESP32-berechnet-floats-falsch-im-Vergleich-zu-M4-und-Due-M3
"Ceos" (ESP32): https://www.roboternetz.de/community/threads/73360-ESP32-berechnet-floats-falsch-im-Vergleich-zu-M4-und-Due-M3?p=651789&viewfull=1#post651789
"BDL" (M4): https://forums.adafruit.com/viewtopic.php?f=57&t=151070&p=746220#p746147
"MartinL" (ESP32, M0): http://forum.arduino.cc/index.php?topic=613077.msg4157601#msg4157601
(to be continued...)
application source code (issue coincidentally detected): the questionable value is the last one in either Serial output line (TBase.m)
changing float (fp32) to double (fp64) makes no difference basically.