Is that slow for coverting float to int ?

[DaviesX]

I have seen a tip that said the standard type convertion of float to in is slow, is that true for gcc ?
If so, how slow it will be ?

They said this code can make the convertion become faster.


#include <stdio.h>
#include <time.h>


inline int FLoatToInt ( float fNumber )
{
int ival = *(int *)&fNumber;

int mantissa = (ival & 0x07fffff) | 0x800000;
int exponent = 150 - ((ival >> 23) & 0xff);

if (exponent < 0)
return mantissa << -exponent;
else
return mantissa >> exponent;
}


int main()
{
int begin = clock ();

float i;
int integer;
for ( i = 0; i < 100000.0f; i ++ )
{
for ( float j = i; j < i + 1.0f; j += 0.01 )
{
integer = FLoatToInt ( j );
// integer = (int)j;
}
}

int end = clock ();

printf ( "%d %d", (int)integer, end - begin );

return 0;
}


if it is faster, why doesn't someone who made compiler use this method ? Is there any downside using the code ?:)

[Bachstelze]

I have seen a tip that said the standard type convertion of float to in is slow, is that true for gcc ?

Where? From whom? How old? CPUs have an instruction to do that nowadays, I highly doubt anything you could tinker up would be faster than that.

[DaviesX]

It's from here,
http://stereopsis.com/sree/fpu2006.html

[DaviesX]

Thanks, You're right :), when I change the test code like this, the standard conversion is obviously faster than any method, after I turned on -march=prescott (Intel Pentium 4 Prescott which support MMX SSE, SSE2, SSE3 instruction set)

The result is:
standard: 18000 (10000 for loop)
so-call fast way: 27000 (10000 for loop)



#include <stdio.h>
#include <time.h>


inline int FLoatToInt ( float fNumber )
{
int ival = *(int *)&fNumber;

int mantissa = (ival & 0x07fffff) | 0x800000;
int exponent = 150 - ((ival >> 23) & 0xff);

if (exponent < 0)
return mantissa << -exponent;
else
return mantissa >> exponent;
}


int main()
{
int begin = clock ();

float i;
int integer;

for ( i = 0; i < 10000000.0f; i ++ )
{
integer = FLoatToInt ( i );
integer = (int)i;
}

int end = clock ();

printf ( "%d %d", (int)integer, end - begin );

return 0;
}

[cgroza]

Still, I seriously doubt that you will ever need to optimize a type conversion...