Altivec vector comparison - GCC AmigaOS4 SDK

6 posts / 0 new
Log in or register to post comments
Last post
Sat, 2018-04-21 16:36
#1
flash
flash's picture
Offline
Last seen: 5 months 4 weeks ago
Joined: 2018-02-24 17:25
Altivec vector comparison - GCC AmigaOS4 SDK

Hello friends,
I'd like to ask you if the following code is ok.
It uses altivec extensions under GCC AmigaOS4 compiler/SDK.

I'm using AmigaOS under Winuae and can't test Altivec code :-(

I'd like to focus you in the following line: "if (vec_all_gt (vec_add (zi2, zr2), maxdist)) break;"

I know vectors comparisions are bit strange and in some others examples in internet I find other more complex ways to make same job using masking tecniques.

Here there's the whole function:

  1. #include "exec/types.h"
  2. #include "FlashMandel.h"
  3.  
  4. #ifdef __ALTIVEC__
  5. #include "altivec.h"
  6.  
  7. __inline WORD MandelnAltivec (LONG Iterazioni, LONG Power, LDouble Cre, LDouble Cim)
  8. {
  9.       REGISTER  WORD Exp;
  10.       const vector float zero = (vector float) {0};      
  11.       const vector float maxdist = (vector float) {4.0};
  12.  
  13.       vector float vCre = (vector float) {(float) Cre};
  14.       vector float vCim = (vector float) {(float) Cim};
  15.       vector float zr = vCre;
  16.       vector float zi = vCim;
  17.       vector float zi2, zr2;
  18.  
  19.      do
  20.      {
  21.         for (Exp = Power; Exp != -1; Exp--)
  22.         {
  23.             zi2 = vec_madd (zi, zi, zero);
  24.             zi = vec_madd (zi, zr, zero);
  25.             zr2 = vec_madd (zr, zr, zero);
  26.             zr = vec_sub (zr2, zi2);
  27.             zi = vec_add (zi, zi);
  28.         }
  29.  
  30.         if (vec_all_gt (vec_add (zi2, zr2), maxdist)) break;
  31.  
  32.         zi = vec_add (zi, vCim);
  33.         zr = vec_add (zr, vCre);
  34.  
  35.         if (--Iterazioni == 0) break;
  36.  
  37.         for (Exp = Power; Exp != -1; Exp--)
  38.         {
  39.             zi2 = vec_madd (zi, zi, zero);
  40.             zi = vec_madd (zi, zr, zero);
  41.             zr2 = vec_madd (zr, zr, zero);
  42.             zr = vec_sub (zr2, zi2);
  43.             zi = vec_add (zi, zi);
  44.         }
  45.  
  46.         if (vec_all_gt (vec_add (zi2, zr2), maxdist)) break;
  47.  
  48.         zi = vec_add (zi, vCim);
  49.         zr = vec_add (zr, vCre);
  50.  
  51.      } while (--Iterazioni);
  52.  
  53.      return 0;
  54. }
  55. #endif /* __ALTIVEC__ */
Top
Sun, 2018-04-29 18:23
flash
flash's picture
Offline
Last seen: 5 months 4 weeks ago
Joined: 2018-02-24 17:25
Re: Altivec vector comparison - GCC AmigaOS4 SDK

Ok friends, i want to update the thread and reply myself because I got the answer.
Due I haven't any real Amiga and Winuae platform doesn't support Altivec extensions I had to try it on my PowerMac G5 Quad under XCode development tool.

..Maybe this is the only existing one mandelbrot program to use altivec ad output a ascii text as result :-)

  1. #include "stdio.h" 
  2. #include "altivec.h"
  3.  
  4. int main() 
  5. { 
  6.     int a, b; 
  7.     vector float cre, cim, zi, zr, zi2, zr2; 
  8.     const int maxiter = 32; 
  9.     int iter; 
  10.  
  11. 	vector bool int exitMask, loopMask = {1}, zeroInt = {0};
  12.  
  13. 	const vector float zero = (vector float) {0};
  14. 	const vector float maxdist = (vector float) {4.0};
  15.  
  16.     for (b = 0; b <= 32; b++) 
  17.     { 
  18. 		cim = (vector float) {((b - 16) / 10.0)}; 
  19.  
  20.         for (a = 0; a <= 85; a++)
  21.         { 
  22. 			iter = 1; 
  23.  
  24.             cre = (vector float) {((a - 50) / 20.0)};    
  25.  
  26. 			zr = cre;
  27. 			zi = cim;
  28.  
  29.             do {                
  30.  
  31. 			zi2 = vec_madd (zi, zi, zero);
  32. 			zi = vec_madd (zi, zr, zero);
  33. 			zr2 = vec_madd (zr, zr, zero);
  34. 			zr = vec_sub (zr2, zi2);
  35. 			zi = vec_add (zi, zi);
  36.  
  37. //			if (vec_all_gt (vec_add (zi2, zr2), maxdist)) break;
  38.  
  39. 			exitMask = vec_and (vec_cmpgt (vec_add (zi2, zr2), maxdist),loopMask);
  40. 			if (vec_any_ne (exitMask,zeroInt))		break;	
  41.  
  42. 			zi = vec_add (zi, cim);
  43. 			zr = vec_add (zr, cre);
  44.  
  45.             iter++; 
  46.             } while (iter < maxiter); 
  47.  
  48.             if (iter == maxiter) 
  49.                 putchar(' '); 
  50.             else 
  51.                 putchar (" .:-;!/>)|&IH%*#"[iter & 15]); 
  52.         } 
  53.         putchar ('\n'); 
  54.     } 
  55.     return 0; 
  56. }

As you can see there's no simple way to compare vectors and I needed to use masking.
If someone knows a faster way is welcome to share.

I'm looking for fortunate amigans with G4 cpu and/or altivec ebabled extensions for testing purposes..

Top
Mon, 2018-04-30 17:04
(Reply to #2)
TSK
TSK's picture
Offline
Last seen: 1 year 1 month ago
Joined: 2011-06-28 02:06
Re: Altivec vector comparison - GCC AmigaOS4 SDK

The code compiles and runs (from the post #2) and returns this on X1000 (with Altivec):

  1. ..........................:::::::::::::::::::::::::::::::::::::::::::::::::...........
  2. ........................:::::::::::::::::::::::::::::::::::::::::::::::::::::.........
  3. ......................:::::::::::::::::::::::::::::::::::::::::::::::::::::::::.......
  4. ....................:::::::::::::::-----------:::::::::::::::::::::::::::::::::::.....
  5. ..................::::::::::-------------------------::::::::::::::::::::::::::::::...
  6. .................:::::::----------------;;;;;!-)!!;;;----:::::::::::::::::::::::::::..
  7. ................:::::----------------;;;;;;!!/>I|I !;;;;----:::::::::::::::::::::::::.
  8. ...............:::----------------;;;;;;;!!!/)| )I)/!!;;;;-----:::::::::::::::::::::::
  9. ..............::----------------;;;;;;;!!//>)*    :)/!!!!;;------:::::::::::::::::::::
  10. .............::--------------;;;;;;!!/>))))|&*    %&|)///:!;;-----::::::::::::::::::::
  11. ............::-------------;;;;!!!!//>H   &          ;H*# #!;------:::::::::::::::::::
  12. ............:-----------;;!!!!!!!///)*H                  ;>/!;------::::::::::::::::::
  13. ...........:-------;;;!!>H>>>>>>>>>)I;                   .|-!;;------:::::::::::::::::
  14. ...........---;;;;;;!!!/>& ;H% %I||I)                     #)/;;-------::::::::::::::::
  15. ...........-;;;;;;!!!!/>)I:       #                       #>!;;;------::::::::::::::::
  16. ...........;;;;;////>)H%*/                                )/!;;;------::::::::::::::::
  17. ..........                                              H)>/!;;;-------:::::::::::::::
  18. ...........;;;;;////>)H%*/                                )/!;;;------::::::::::::::::
  19. ...........-;;;;;;!!!!/>)I:       #                       #>!;;;------::::::::::::::::
  20. ...........---;;;;;;!!!/>& ;H% %I||I)                     #)/;;-------::::::::::::::::
  21. ...........:-------;;;!!>H>>>>>>>>>)I;                   .|-!;;------:::::::::::::::::
  22. ............:-----------;;!!!!!!!///)*H                  ;>/!;------::::::::::::::::::
  23. ............::-------------;;;;!!!!//>H   &          ;H*# #!;------:::::::::::::::::::
  24. .............::--------------;;;;;;!!/>))))|&*    %&|)///:!;;-----::::::::::::::::::::
  25. ..............::----------------;;;;;;;!!//>)*    :)/!!!!;;------:::::::::::::::::::::
  26. ...............:::----------------;;;;;;;!!!/)| )I)/!!;;;;-----:::::::::::::::::::::::
  27. ................:::::----------------;;;;;;!!/>I|I !;;;;----:::::::::::::::::::::::::.
  28. .................:::::::----------------;;;;;!-)!!;;;----:::::::::::::::::::::::::::..
  29. ..................::::::::::-------------------------::::::::::::::::::::::::::::::...
  30. ....................:::::::::::::::-----------:::::::::::::::::::::::::::::::::::.....
  31. ......................:::::::::::::::::::::::::::::::::::::::::::::::::::::::::.......
  32. ........................:::::::::::::::::::::::::::::::::::::::::::::::::::::.........
  33. ..........................:::::::::::::::::::::::::::::::::::::::::::::::::...........
Top
Tue, 2018-05-01 01:45
(Reply to #3)
flash
flash's picture
Offline
Last seen: 5 months 4 weeks ago
Joined: 2018-02-24 17:25
Re: Altivec vector comparison - GCC AmigaOS4 SDK

It was tested on PowerMac g5 quad, I'm very glad to know it runs ok on Amiga X1000 too :-)
I uploaded a version for both, altivec and scalar fpu versions, on os4depot and aminet.
As always C source code is included.

@TSK
I send you a PM

Top
Fri, 2018-11-02 17:21
(Reply to #4)
flash
flash's picture
Offline
Last seen: 5 months 4 weeks ago
Joined: 2018-02-24 17:25
Re: Altivec vector comparison - GCC AmigaOS4 SDK

Well, so I decided to review th Altivec mandelbrot routine because I was sure it could be further optimized.
Than I got 2 vars and one instruction less!
But I got also a big surprise.. the scalar routine is about 30% faster than altivec one!
Maybe using altivec unit just as a regular fpu there isn't any performance improvement, while i guess if it was a spe unit (like Tabor or PS3) it could be faster than fpu..
Is it right?
If you want you can recompile and try.. It's tested under MacOSX Leopard/Xcode 3 and UbuntuMATE 16.04/CodeBlocks (both on PowerMac G5 Quad).
Of course it compiles well also under AmigaOS 4 and in next days I'll update AsciiMandelbrot program on Aminet/Os4Depot adding a benchmark switch.
Meanwhile I have updated FlashMandelNG to v1.8 version to reflect the speed improvements in altivec version.

***

  1. #include <stdio.h>
  2.  
  3. char VERSION_STRING [] = "$VER: AsciiMandelbrot 1.1 by Dino Papararo ©2018\0 ";
  4.  
  5. #define ALTIVEC
  6. #ifdef ALTIVEC
  7. #include <altivec.h>
  8.  
  9. int main()
  10. {
  11.     int a, b;
  12.     vector float cre, cim, zi, zr, zi2, zr2;
  13.     const int maxiter = 800000;
  14.     int iter;
  15.  
  16.     const vector bool int zeroInt = {0};
  17.  
  18. 	const vector float zero = {0.0}, maxdist = {4.0};
  19.  
  20.     for (b = 0; b <= 32; b++)
  21.     {
  22.         cim = (vector float) {((b - 16) / 10.0)};
  23.  
  24.         for (a = 0; a <= 85; a++)
  25.         {
  26.             iter = 1;
  27.  
  28.             cre = (vector float) {((a - 50) / 20.0)};
  29.             zr = cre;
  30.             zi = cim;
  31.  
  32.             do {
  33.                     zi2 = vec_madd (zi, zi, zero);
  34.                     zi = vec_madd (zi, zr, zero);
  35.                     zr2 = vec_madd (zr, zr, zero);
  36.                     zr = vec_sub (zr2, zi2);
  37.                     zi = vec_add (zi, zi);
  38.  
  39.                     if (vec_any_ne (vec_cmpgt (vec_add (zi2, zr2), maxdist),zeroInt)) break;
  40.  
  41.                     zi = vec_add (zi, cim);
  42.                     zr = vec_add (zr, cre);
  43.  
  44.                     iter++;
  45.             } while (iter < maxiter);
  46.  
  47.             if (iter == maxiter) putchar(' ');
  48.             else putchar (" .:-;!/>)|&IH%*#"[iter & 15]);
  49.         }
  50.  
  51.         putchar ('\n');
  52.     }
  53.  
  54.     printf("\nComputed with Altivec unit!\n");
  55.  
  56. //    getchar ();
  57.  
  58.     return 0;
  59. }
  60.  
  61. #else // NOT ALTIVEC
  62.  
  63. int main()
  64. {
  65.     int a, b;
  66.     float cre, cim, zr, zi, zr2, zi2;
  67.     const float maxdist = 4.0;
  68.     const int maxiter = 800000;
  69.     int iter;
  70.  
  71.     for (b = 0; b <= 32; b++)
  72.     {
  73.         cim = (b - 16.0) / 10.0;
  74.  
  75.         for (a = 0; a <= 85; a++)
  76.         {
  77.             cre = (a - 50) / 20.0;
  78.  
  79.             zi = cim;
  80.             zr = cre;
  81.  
  82.             iter = 1;
  83.  
  84.             do {
  85.                     zi2 = zi*zi;
  86.                     zi = zi*zr;
  87.                     zr2 = zr*zr;
  88.                     zr = zr2-zi2;
  89.                     zi += zi;
  90.  
  91.                     if ((zi2+zr2) > maxdist) break;
  92.  
  93.                     zi += cim;
  94.                     zr += cre;
  95.  
  96.                     iter++;
  97.             } while (iter < maxiter);
  98.  
  99.             if (iter == maxiter) putchar(' ');
  100.             else putchar (" .:-;!/>)|&IH%*#"[iter & 15]);
  101.         }
  102.  
  103.         putchar ('\n');
  104.     }
  105.  
  106.     printf ("\nComputed without Altivec unit\n");
  107.  
  108. //    getchar ();
  109.  
  110.     return 0;
  111. }
  112. #endif
Top
Sat, 2018-11-03 20:31
(Reply to #5)
flash
flash's picture
Offline
Last seen: 5 months 4 weeks ago
Joined: 2018-02-24 17:25
Re: Altivec vector comparison - GCC AmigaOS4 SDK

So I have just released a new AsciiMandelbrot V1.4 version on both, Aminet and Os4Depot, just for fun and speedtest purposes.
I have also included timings to track speed on different amiga flavors.
Now there's a single exe for altivec and non altivec routines, of course source code is included.

When/If Tabor will be released, I'd like to add support for it too :-)

Amiga rulez! :-)

Top
Log in or register to post comments