[vorbis-dev] optimizing float to int conversions

[Tal]

See remarks in red or with >>> prefix:

-----Original Message-----
From: owner-vorbis-dev at xiph.org [mailto:owner-vorbis-dev at xiph.org] On
Behalf Of illiminable
Sent: Wednesday, May 05, 2004 6:59 PM
To: vorbis-dev at xiph.org
Subject: Re: [vorbis-dev] optimizing float to int conversions

<p>I gave this a test compiling with VS2002 (VC7.0)... optimised the build
settings as much as possible.

My test is below (using C++)... To alternate tests just comment out one
test or the other.
>>> What is the running time of the source libvorbis and what was the
running time after using the macro?

>>> Could you check that the output binaries are the same?
On my Athlon 2500 your macro runs in just over 50% of the time of teh
cast... ie just under 50% improvement with the most heavily optimised
compiler settings for each test.

>>> that is great :)

I just made another chagne to your macro and got it up to a 55%
improvement... Probably a bit more tweaking can happen too. See where
you had the if (d>31)
  MyInt = 0;
>>> On intel platform you are shifting by x modulo 32. We've seen the
bug and this is why we added the code.

Assuming your ints are of size 32 bits (I'm pretty sure this is a safe
assumption in ogg), this is not necessary if you shift any 32 bit int by
32 or more bits in either direction you get zero anyway (is there some
case i'm not thinking of that this is not true ?)... so you can remove a
a comparison and jump there. All bits left of the 24th bit would be zero
anyway.

But either way... it's a whole bunch faster than casting !

<p>Zen...

Test follows.

#define CONVERT_FLT2INT(MyFloat,MyInt) \
{ \
   int FltInt = *(int *)&MyFloat; \
\
   int mantissa = (FltInt & 0x07fffff) | 0x800000; \
   int exponent = ((FltInt >> 23) & 0xff) - 0x7f; \
   int d = 23-exponent;\
   if ((d) < 0) \
   {\
     MyInt = (mantissa << -(d)); \
  d = -d;\
   }\
   else \
     MyInt = (mantissa >> (d)); \
\
   if (FltInt & 0x80000000) \
     MyInt = -MyInt; \
}

<p>int _tmain(int argc, _TCHAR* argv[])
{
 const ASIZE = 500000000;
 float* fa = new float[ASIZE];
 int* ia = new int[ASIZE];

//Create some numbers
 for (int i = 0; i < ASIZE; i++) {
  int a = rand();
  fa[i] = (float)a;
  if (i % 2 == 1) fa[i] = -fa[i];
 }

   clock_t start, finish;

   /* Measure the duration of an event. */
   start = clock();

   //Test 1
   for (int i = 0; i < ASIZE; i++) {
    ia[i] = (float)fa[i];
   }

   //Test 2
   //for (int i = 0; i < ASIZE; i++) {
     //CONVERT_FLT2INT(fa[i], ia[i]);
   //}
   finish = clock();

   cout << "Duration of "<<ASIZE<<" conversions is "<<finish -
start<<endl;

 return 0;
}
----- Original Message -----
From: "dean gaudet" <dean-list-vorbis-dev at arctic.org>
To: <vorbis-dev at xiph.org>
Sent: Thursday, May 06, 2004 12:29 AM
Subject: Re: [vorbis-dev] optimizing float to int conversions

<p>> On Wed, 5 May 2004 shabi at t2.technion.ac.il wrote:
>
> > Hi,
> > We compiled libvorbis with vc6 enviroment.
> > We've discovered that changing float to int conversion results in a
> > 4%
speedup
> > on a benchmark we've come with on pentium4 3GHz.
>
> i'm curious where the real gains are coming from here -- is your
> compiler generating cmov instructions after your patches?  it looks
> like most of the potential for gains are in parallelising the code in
> and around where you've made these float to int conversions -- p4 is a
> lot happier parallizing cmov than it is fcmov.
>
> i gather vc6 is too old to generate sse code instead of x87 code? 
> because the places where you're changing things are vectorizable.
>
> also -- you might want to try out this change to psy.c as well:
>
> /* this is for per-channel noise normalization */
> static int apsort(const void *a, const void *b){
> #if 0
>   float f1=fabs(**(float**)a);
>   float f2=fabs(**(float**)b);
>   return (f1<f2)-(f1>f2);
> #else
>   uint32_t u1 = **(uint32_t **)a;
>   uint32_t u2 = **(uint32_t **)b;
>   u1 *= 2; // clear sign bit
>   u2 *= 2;
>   //return (u1<u2)-(u1>u2);
>   return (int32_t)u2 - (int32_t)u1;
> #endif
> }
>
> -dean
> --- >8 ----
> List archives:  http://www.xiph.org/archives/
> Ogg project homepage: http://www.xiph.org/ogg/
> To unsubscribe from this list, send a message to
'vorbis-dev-request at xiph.org'
> containing only the word 'unsubscribe' in the body.  No subject is
> needed. Unsubscribe messages sent to the list will be
> ignored/filtered.
>
>
>

<p>--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request at xiph.org' containing only the word 'unsubscribe' in
the body.  No subject is needed. Unsubscribe messages sent to the list
will be ignored/filtered.

<p><p><p>--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-dev-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.