Performances, la NumWorks détrône la HP Prime ! :o

[critor]

Merci, édité. Oui simple erreur de recopie, sinon ça n'aurait pas marché.

[Lionel Debroux]

Je pense que Numworks a raison d'utiliser double pour les nombres flottants Python, et j'imagine qu'à moyen terme, pour des raisons de place disponible dans la Flash (du premier modèle), ils devront arrêter cette coûteuse duplication de code entre float et double dans le moteur de calcul. Même si les performances en souffriront.

[zardam]

J'ai finalement isolé la source de la différence de performance. En fait, cela provient de l'implémentation softfloat de la liba qui est moins rapide que celle de la libgcc. En utilisant la libgcc, j'ai des performances cohérentes avec celles de la pyboard. La dèrnière version du mandelbrot du forum s'exécute en environ 38 secondes (avec 320, 222, 10 en argument). Cerise sur le gâteau, utiliser la libgcc libère quelques ko de flash

Le code est ici : https://github.com/zardam/epsilon/tree/libgcc et il faut compiler avec l'option "USE_LIBGCC=1" pour l'activer.

[compsystems]

What is the time used by the following code ?, Since I do not have a real hp-prime.

Please can someone make a comparison with ti-nspire & micropython

Code: Select all

EXPORT fractal(W,H,N)
BEGIN
LOCAL x,y,z,c,j,w,h,n;
w:=2.7/(W-1);
h:=1.87/(H-1);
n:=255.0/N;
FOR y FROM 0 TO H-1 DO
  c:=-2.1-i*(h*y-.935);
  FOR x FROM 0 TO W-1 DO
   z:=0;
   FOR j FROM 0 to N-1 DO
    IF ABS(z:=z²+c)>2 THEN
     break;
    END;
   END;
   PIXON_P(x,y,j*255*20);
   c:=c+w;
  END;
END;
FREEZE;
//WAIT(0);
END;

[critor]

Unfortunately, we can't compare with TI-Nspire+MicroPython, since the Nspire graphic module for MicroPython won't let you write to the screen directly.

Your code takes 1min28.64s.

[compsystems]

This implies that you must port a library or graphic engine or develop a new one?

[critor]

If we want to include the TI-Nspire in the comparison, we need to add instructions writing directly to the screen to a MicroPython module.

Currently :
- TI-Basic won't let you output anything else than characters to the screen
- Lua is buffering screen outputs, performing them all at once when it wants
- MicroPython and DukTape are forcing you to use offscreen layers (called Textures)

So none of them can be used.

The latter was a very bad choice from an educational point of view (not that it matters with the exam mode anyway nowadays, but the MicroPython port was supposed to target education if I remember the first posts about it well enough).
Since programming beginners with their simple graphic programs won't accept/understand why they have to define, dim and deal with all those big offscreens just to turn on a few pixels.
And you can't turn them into advanced programmers accepting and understanding this without showing them why it's needed for more advanced graphic programs.