BUILDING libSIMDx86

NOTE: This only builds under GNU C Compiler.


Building with libSIMDx86config, the interactive configuration program.

1. Get the sources and extract them
   This can be done by going to www.sourceforge.net/projects/simdx86. Then extract them to a folder.

2. Compile libSIMDx86config
   libSIMDx86config is written in vanilla ANSI C, compiled with all warnings enabled and under pedantic mode. Pretty much any ANSI C compiler should be able to compile this program without any problems. To compile under the GNU C Compiler, type:
   
   shell$ gcc libSIMDx86config.c -o libSIMDx86config

3. Run libSIMDx86config
Execute libSIMDx86config. On UNIX based systems this is done by:

shell$ ./libSIMDx86config

On Win32 based systems, double-click the program libSIMDx86config.exe
Follow the on-screen instructions through the program.

4. Run the generated build script
Now you need to run the script. On UNIX based systems this is done by:

shell$ sh build.sh

On Win32 based systems, double-click the batch file build.bat




Building without libSIMDx86config

1. Get the sources and extract them.
   This can be done by going to www.sourceforge.net/projects/simdx86. Then extract them to a folder.

2. Decide on instruction set and options
   SIMDx86 can use the following instruction sets:
   1) MMX
   2) 3DNow! (includes MMX)
   3) 3DNow!+/MMX+ (includes MMX)
   4) SSE (includes MMX)
   5) SSE2 (includes MMX/SSE)
   6) SSE3 (includes MMX/SSE/SSE2)
   7) SSSE3 (includes MMX/SSE/SSE2/SSE3)
   
   Additionally, SIMDx86 can optimize for:
   1) Speed (Often for 3D-Games)
   2) Accuracy (Often for non-realtime rendering)
   3) AMD processors rather than Intel
   
   To optimize for speed, do NOT define HIPREC. To optimize for precision, use #define HIPREC. HIPREC tells SIMDx86 not use low precision reciprocal or other various hacks, but instead to do as correct math as possible. Without HIPREC defined, special processor instructions will be used. You can expect precision to at least three decimal places in the worst cases.
   Optimizing for AMD processors hurts performance on Intel processors, and vice versa, but Intel processors take a larger hit from this flag than do AMD processors.
   
   If using 3DNow! or MMX, SIMDx86 can omit '[f]emms' instruction

   Since executing an '[f]emms' instruction can take extra cycles when a context switch is not required, it may be beneficial not to implicitly execute it at the end of each routine. If the symbol "NO_EMMS" is defined, MMX and 3DNow! routines will NOT execute an [f]emms instruction. If you have no idea what is being mentioned here, then ingore this.
   
   For MMX, the preprocessor constant must be: USE_MMX=1
   For 3DNow!, the preprocessor constant must be: USE_3DNOW=1
   For 3DNow!+/MMX+, the preprocessor constant must be: USE_3DNOW=2
   For SSE, the preprocessor constant must be: USE_SSE=1
   For SSE2, the preprocessor constant must be: USE_SSE=2
   For SSE3, the preprocessor constant must be: USE_SSE=3
   For SSSE3, the preprocessor constant must be: USE_SSE=3.5
   
   To omit '[f]emms' instructions, define the symbol NO_EMMS. You must explicitly call 'SIMDx86_femms() or SIMDx86_emms(). Or you can directly add the following: asm("femms"); or asm("emms"); to explictly execute the instructions.
   
   
   

3. Compile it
   This will create libSIMDx86 in the bin/ directory.
   To do a custom build:
   shell$ cd src
   shell$ gcc -c *.c -I../include [-DUSE_XXX=YYY] [-DHIPREC] [-DNO_EMMS]
   shell$ ar cru ../bin/libSIMDx86.a *.o
   shell$ ranlib ../bin/libSIMDx86.a
   
   Note that the [] bracketed arguments are optional. By not defining any USE_XXX constant, libSIMDx86 will be built in x87 mode, meaning no SIMD instructions will be used.
   
   BACK