Support for: Fixed-point arithmetic

[GiorgosXou]

I just found out that I can speed up things by using Fixed-point arithmetic, thanks to this video at 3:19 (see also results at 5:00)

• I've also found about this intresting library

I wish i'll find some free time to do something about it, but my fucking life costs money.

[GiorgosXou]

the closest you can get to a (4-byte)-float's precision is SFixed<4,27> or SFixed<3,28>

#include <FixedPoints.h>
#include <FixedPointsCommon.h>

//#define A -9.1233450 <- why this has error 0.0000001 ????
#define A -9.1233459
#define B 0.0753235

void TestS4x27(void)
{
  SFixed<4,27> a = A;
  SFixed<4,27> b = B;
  Serial.println(static_cast<float>(a*b), 7);
  Serial.println(A*B, 7);
}

void setup()
{
  Serial.begin(9600);
  while(!Serial);
  TestS4x27();
}
void loop(){}

Pharap/FixedPointsArduino#18

[GiorgosXou]

ha ha ...... now implement exp log and sqrt to work with fixed point arithmetics 😢

[GiorgosXou]

So here's an exp(x) function for fixed point arithmetics that's really accurate with small numbers. Althought I realised that it won't be as easy as I first thought it would be, because x of exp that's used in most activation functions depeneds upon the bias value [f(x+bias)] which means that the input in the activation function isn't necessarily small at all

[GiorgosXou]

So, should I use casting and use the std one, like this library did?

[GiorgosXou]

Is it even possible... and in what extend?