Compiler options & seti #defines for optimiztion

[BenHer]

Here is a little document I wrote...I hope will prove usefull to all

========================================================================
    STATIC LIBRARY : optimize Project Overview
========================================================================

Compiler optimization command line arguments and seti specific DEFINE constants and their functions.


MS VC++ 2003 Compiler options
-----------------------
/arch:sse
   - Its ok for compiler to generate SSE instructions when compiling
/arch:sse2
   - Its ok for compiler to generate SSE2 & SSE instructions when compiling
/arch:sse3
   - Its ok for compiler to generate SSE3 & SSE2 & SSE instructions when compiling

/Oy
   The intel 386 and beyond CPU has 8 floating point stack registers and 8 general
   registers, One, the stack pointer, can't really be used by code...this leaves 7.
   Older code used one of these registers (extended base pointer) to keep track of
   local variables inside of functions...and it was called the "frame pointer"
   /Oy tells compiler to free up the 'ebp' register for regular code use.
   (note - internally Pentium 3 and beyond have many more "registers" but they
    use them behind the scenes when registers get loaded/saved to memory)   

/Ob2 - Inline functions when it seems a good idea
   Tells compiler to "inline" smallish functions when called inside of other functions
   For example the function  add_to_total( int num) {  total += num; } would be a
   perfect candidate for automatic inlining.
   
/Oi - Use processor specific "intrinsic" functions instead of library calls
   example:  the x86 processor has a built in  'sine' opcode
         source             compiled code
   /Oi-   x=sin( num )        ...   call  _library_sine_function
   /Oi    x=sin( num )        ...   fsin

/Ot   - Prefer to make the output code faster rather than making it small

/Og - "global" optmization - Many usefull optimizations within code


Intel C++ Compiler options
-----------------------
/arch:sse
/arch:sse2
/arch:sse3

/Qunroll - Sometimes (when makes code faster) make multiple copies of code inside
   of source loops

/Qipo - Try to do multi-source-file optimizations. 
   Inline small functions from one source file at call points in functions from other files
   Use registers to pass paramenters from some functions to others (rather than stack)
   many other items
   TAKES LONG TIME!! - Uses up to 1.4Gig of RAM + swap file

/QxK - Code is optimized for Intel® Pentium® III and compatible Intel processors.
/QxW - Code is optimized for Intel Pentium 4 and compatible Intel processors.
/QxN - Code is optimized for Intel Pentium 4 and compatible Intel processors
   with Streaming SIMD Extensions 2.
   The resulting code may contain unconditional use of features that are
   not supported on other processors.
   This option also enables new optimizations in addition to Intel
   processor-specific optimizations including advanced data layout
   and code restructuring optimizations to improve memory accesses for Intel processors.

/QxB- Code is optimized for Intel Pentium M and compatible Intel processors.
   This option also enables new optimizations in addition to Intel processor-specific optimizations.

/QxP- Code is optimized for Intel® Core™ Duo processors, Intel® Core™ Solo processors,
   Intel® Pentium® 4 processors with SSE3, and compatible
   Intel processors with SSE3. The resulting code may contain unconditional use
   of features that are not supported on other processors.

/G5 - Intel® Pentium® and Pentium® with MMX™ technology processor

/G6 - Intel® Pentium® Pro, Pentium® II and Pentium® III processors

/G7 - Pentium 4, Xeon, Core Duo, Core solo, Pentium M, SSE3

__INTEL_COMPILER
   -- Automatically created define - lets programmer know what compiler is
     crunching their source code
    

/O2 - Is the same as setting the following flags
   /Og, /Oi-, /Os, /Ob2, /GF (/Qvc7 and above), /Gf (/Qvc6 and below), /Gs, and /Gy

/O3  - Even more optimizations 

Seti@home worker client #define options
--------------------------------
BOINC_APP_GRAPHICS
   - Make code & calls for graphics display of seti WU progress

USE_IPP
   - Fast Fourier Transform functions using the Intel Compiler Primitives
   - an Intel library of fast math functions 

   USE_SSE3
   USE_SSE2
   USE_SSE
      - All only valid when using USE_IPP
      - Determines which version of several IPP functions will be used in FFT
        calculations (which SIMD)

    
USE_FFTWF
   - Fast Fourier Transform functions using 'FFTW3' - 'Fastest Fast Fourier in the West'
   - An open source fast fourier - optimized for all available SIMD processors
     SSE, SSE2, 3Dnow, Altivec     http://www.fftw.org

DO_SMOOTH
   - Call the 'baseline smooth' functions for data in WU processing

[Josef W. Segur]

Here is a little document I wrote...I hope will prove usefull to all
...
Seti@home worker client #define options

USE_IPP
   - Fast Fourier Transform functions using the Intel Compiler Primitives
   - an Intel library of fast math functions 

   USE_SSE3
   USE_SSE2
   USE_SSE
      - All only valid when using USE_IPP
      - Determines which version of several IPP functions will be used in FFT
        calculations (which SIMD)

In the 5.17+ analyzeFuncs_vector.cpp source,
USE_ALTIVEC, USE_SSE3, USE_SSE2,  USE_SSE, USE_3DNOW, and  USE_MMX also control which of some alternative functions are included in the compile for run-time testing.

Only USE_ALTIVEC actully has alternative SIMD versions now, I hope that x86 alternates will be submitted.

If there's a conflict with IPP in that usage, it would be good to get it resolved soon.
                                                                            Joe

[Josef W. Segur]

While looking through the Harold Naparst optimized sources for S@H unenhanced, I noted that Trux had contributed changes which replace all includes of <cmath> with CMATH_LIB, and all <math.h> with MATH_LIB

Then in config.h, there's this block:

#ifdef __INTEL_COMPILER
#define MATH_LIB <mathimf.h>
#define CMATH_LIB <mathimf.h>
#else
#define MATH_LIB <math.h>
#define CMATH_LIB <cmath>
#endif

                                                                          Joe

[Simon]

Er,

I must have missed your post from August, Joe

Just checked our current sources, and no such substitution (math.h/<cmath>->MATH_LIB/CMATH_LIB) seems to be going on, so I'll try and see whether that makes any difference in performance (to my knowledge, IMF is pretty darn quick indeed).

There's also an addendum to Ben's post above, newer ICC versions support a new level of optimization, namely:

/QxT   Optimize for SSSE3-capable CPUs, will NOT run anywhere else
/QaxT  Optimize for SSSE3-capable CPUS, but also compile in a generic code branch (slower, but more compatible, but makes no sense for S@H compilation IMO)

Regards,
Simon.