Thread: Fastest float to int conversion

  1. #1
    Registered User
    Join Date
    Oct 2004
    Posts
    36

    Fastest float to int conversion

    Hi guys,

    This is probably more of an ASM question. Once again I am using ARM, no hardware floats (all software). Does anyone have a fast technique for converting float to int?

    I don't mind if its the same as 'round' , 'ceil' or 'floor', as long as its consistant, and FAST

    Cheers,
    Pea

  2. #2
    ATH0 quzah's Avatar
    Join Date
    Oct 2001
    Posts
    14,826
    Code:
    float f = 10.0;
    int i = 0;
    
    i = (int)f;
    I doubt you'll find anything faster.

    Quzah.
    Hope is the first step on the road to disappointment.

  3. #3
    Software Developer jverkoey's Avatar
    Join Date
    Feb 2003
    Location
    New York
    Posts
    1,905
    I believe some compilers sometimes use the ftol asm command which is a slow, rounding assembly call. There is a way to disable this and make it chop instead of round...*runs off to get book*

    K, in Andre LaMothe's Tricks of the 3D Game Programming Gurus:
    When assigning a float to and int, the compiler will many times call the internal function, _ftol(), which can slow things down to a crawl. For example:

    float f=10.5;
    int i=f;

    To avoid this, use inline assembly and FPU instructions, such as fistp/fst:

    _asm
    {
    fld f;
    fistp i;
    }

    Or use the compiler flag /QIfist to force the rounding mode to chop, and no call to _ftol() will be made.

  4. #4
    ATH0 quzah's Avatar
    Join Date
    Oct 2001
    Posts
    14,826
    Code:
    union convert
    {
        int i;
        float f;
    };
    That'll beat your assembly method hands down for speed. This is found in the "Tips of the Wise-Ass Gurus".

    Quzah.
    Hope is the first step on the road to disappointment.

  5. #5
    Registered User
    Join Date
    Oct 2004
    Posts
    36
    Thanks guys. I'll give it a go. I'll have to look up some of those assembler instructions thought

    I knew there was a faster way than a cast

    BTW, whats 'union convert' ?

  6. #6
    Registered User
    Join Date
    Sep 2004
    Location
    California
    Posts
    3,268
    BTW, whats 'union convert' ?
    It's a joke

  7. #7
    ATH0 quzah's Avatar
    Join Date
    Oct 2001
    Posts
    14,826
    Code:
    #include <stdio.h>
    int main( void )
    {
            union convert
            {
                    float f;
                    int i;
            } data;
    
            data.f = 10.0;
            printf("Using a union to convert %f to an int is fun: %d!\n",
                            data.f, data.i );
            return 0;
    }
    
    /*
        Using a union to convert 10.000000 to an int is fun: 1092616192!
    */
    Nothing was ever said about accuracy. Just speed.

    Quzah.
    Last edited by quzah; 01-24-2005 at 02:42 AM.
    Hope is the first step on the road to disappointment.

  8. #8
    Registered User
    Join Date
    Oct 2004
    Posts
    36

  9. #9
    and the hat of int overfl Salem's Avatar
    Join Date
    Aug 2001
    Location
    The edge of the known universe
    Posts
    39,660
    > no hardware floats (all software)
    Then you're already hosed.
    The speed of converting a float to an int is inconsequential to all the work which gets done as soon as you start using floating point.

    Consider your need carefully, and consider whether fixed-point arithmetic will work just as well, because it can be a hell of a lot quicker than true floating point.
    If you dance barefoot on the broken glass of undefined behaviour, you've got to expect the occasional cut.
    If at first you don't succeed, try writing your phone number on the exam paper.

  10. #10
    The C-er
    Join Date
    Mar 2004
    Posts
    192
    Yes, I would agree with Salem. Years ago I wrote a mandelbrot program in C (actuallly my first real C program) at college on a 386 (no coprocessor).It ran pretty slowly using the library float routines, so I recoded it to use fixed point integer, needless to say the performance difference was dramatic.

    Remember too that that Doom /Doom2 is all integer, so you can do quite a lot this way.

  11. #11
    UT2004 Addict Kleid-0's Avatar
    Join Date
    Dec 2004
    Posts
    656
    Quote Originally Posted by quzah
    Nothing was ever said about accuracy. Just speed.
    That's worse than my random milliseconds answer I gave awhile back! :P

  12. #12
    Registered /usr
    Join Date
    Aug 2001
    Location
    Newport, South Wales, UK
    Posts
    1,273
    No FPU eh? Gawd, that takes me back to the days when Intel FPUs were optional (8086-386). Back then, you had to spend serious time configuring the compiler otherwise your program would crash regardless of the code. (Those who are still using Borland/Turbo C 3 will know what I mean )

    For FPU support then, you had 3 options:- Hardware (i.e. you knew that the x87 was installed on the motherboard), emulation or none. Am I right in thinking that FPU emulation was a fixed-point approximation (otherwise these 16 MHz boxes would've been there for hours)?

    Perhaps there's a similar option on your ARM compiler.
    Last edited by SMurf; 01-25-2005 at 07:48 AM. Reason: Historically inaccurate :)

  13. #13
    Registered User
    Join Date
    Oct 2004
    Posts
    36
    Thanks guys, but the argument to float or not to float is a different one. Just after the quickest method...

  14. #14
    and the hat of int overfl Salem's Avatar
    Join Date
    Aug 2001
    Location
    The edge of the known universe
    Posts
    39,660
    > but the argument to float or not to float is a different one
    Except you seem to consider tying your feet together to be a good start, then asking what's the best way to win a 100M race is.

    > Just after the quickest method.
    Hire someone with a clue would be my next suggestion, from reading your other "performance" post
    If you dance barefoot on the broken glass of undefined behaviour, you've got to expect the occasional cut.
    If at first you don't succeed, try writing your phone number on the exam paper.

  15. #15
    The C-er
    Join Date
    Mar 2004
    Posts
    192
    Salem has a point. Although we could probably help more if we knew more about what the problem space is. What does your program actually do?

Popular pages Recent additions subscribe to a feed

Similar Threads

  1. Replies: 48
    Last Post: 09-26-2008, 03:45 AM
  2. memory leak
    By aruna1 in forum C++ Programming
    Replies: 3
    Last Post: 08-17-2008, 10:28 PM
  3. Replies: 1
    Last Post: 10-27-2006, 01:21 PM
  4. My graphics library
    By stupid_mutt in forum C Programming
    Replies: 3
    Last Post: 11-26-2001, 06:05 PM
  5. A Simple (?) Problem
    By Unregistered in forum C++ Programming
    Replies: 8
    Last Post: 10-12-2001, 04:28 AM