TBCI Numerical high perf. C++ Library: perf

00001 
00008 #ifndef TBCI_PERF_OPT_H
00009 #define TBCI_PERF_OPT_H
00010 
00011 /* gcc-4:
00012  * The user should use -funroll-loops -fvectorize-trees 
00013  * and -fprefetch-loop-arrays as needed.
00014  */
00015 #if defined(__GNUC__) && __GNUC__ >= 4 && !defined(USE_PLAIN_VEC_KERNELS) \
00016                 && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00017 # define USE_PLAIN_VEC_KERNELS
00018 #endif
00019 
00020 /* DEC Alpha architecture */
00021 #ifdef __alpha__
00022 # define DEF_CACHELINE_SZ 32
00023 # ifdef __GNUC__
00024 #  if __GNUC__ <= 2
00025 #   if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00026 //#    define USE_PLAIN_VEC_KERNELS
00027 #    define USE_UNR_VEC_KERNELS
00028 #   endif
00029 #   define DEF_UNROLL_DEPTH 8
00030 #   define DEF_PREFETCH_AHEAD 8
00031 #  else /* gcc 3 */
00032 #   if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00033 #    define USE_UNR_VEC_KERNELS
00034 #   endif
00035 #   define DEF_UNROLL_DEPTH 8
00036 #   define DEF_PREFETCH_AHEAD 8
00037 #  endif
00038 # else /* DEC/Compaq/HP compiled */
00039 #  if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00040 #   define USE_UNR_VEC_KERNELS
00041 #  endif
00042 #  define DEF_UNROLL_DEPTH 1
00043 #  define DEF_PREFETCH_AHEAD 4
00044 # endif
00045 #endif
00046 
00047 /* iA32 arch */
00048 #if defined(__i386__) || defined(__x86_64__)
00049 # if defined(OPT_PENTIUM4) || defined(OPT_CORE2) || defined(__x86_64__)
00050 #  define DEF_CACHELINE_SZ 64
00051 # else
00052 #  define DEF_CACHELINE_SZ 32
00053 # endif
00054 # ifdef __GNUC__
00055 #  ifdef OPT_PENTIUM4
00056 #   if __GNUC__ <= 2
00057 #    if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00058 #     define USE_UNR_VEC_KERNELS
00059 #    endif
00060 #    define DEF_UNROLL_DEPTH 1  /* Hopefully -funroll-loops is enabled */
00061 #    define DEF_PREFETCH_AHEAD 4
00062 #   else /* gcc >= 3 */
00063 #    if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00064 #     define USE_PLAIN_VEC_KERNELS      /* P4 does HW prefetch */
00065 #    endif
00066 #    define DEF_UNROLL_DEPTH 4
00067 #    define DEF_PREFETCH_AHEAD 16
00068 #    if !defined(FORCE_PREFETCH) && !defined(NO_PREFETCH)
00069 #     define NO_PREFETCH        /* P4 does HW prefetch */
00070 #    endif
00071 #   endif
00072 #  elif defined(OPT_PENTIUM3) || defined(OPT_CORE2) || defined(SSE_PREFETCH)    /* Pentium M / 3 */
00073 #   if !defined(FORCE_PREFETCH) && !defined(NO_PREFETCH)
00074 #    define NO_PREFETCH /* Pentium-M does HW prefetch, Pentium-3 does not ... */
00075 #   endif
00076 #   if !defined(SSE_PREFETCH) && defined(OPT_ARCH_PENTIUM3) &&!defined(NO_PREFETCH)
00077 #    define SSE_PREFETCH
00078 #   endif
00079 #   if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00080 #    define USE_UNR_VEC_KERNELS
00081 #   endif
00082 #   define DEF_UNROLL_DEPTH 4
00083 #   define DEF_PREFETCH_AHEAD 8
00084 #  elif defined(OPT_ATHLON) || defined(AMD_PREFETCH)    /* Athlon / Opteron */
00085 #   if !defined(FORCE_PREFETCH) && !defined(NO_PREFETCH) && defined(__x86_64__)
00086 #    define NO_PREFETCH /* AMD64 does HW prefetch */
00087 #   endif
00088 #   if !defined(AMD_PREFETCH) && defined(OPT_ARCH_ATHLON) && !defined(NO_PREFETCH) 
00089 #    define AMD_PREFETCH
00090 #   endif
00091 #   if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00092 #    define USE_UNR_VEC_KERNELS
00093 #   endif
00094 #   define DEF_UNROLL_DEPTH 1   /* Hopefully -funroll-loops is enabled */
00095 #   define DEF_PREFETCH_AHEAD 8
00096 #  else /* deflt proc */
00097 #   if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00098 #    define USE_UNR_VEC_KERNELS
00099 #   endif
00100 #   define DEF_UNROLL_DEPTH 8
00101 #   define DEF_PREFETCH_AHEAD 4
00102 #  endif
00103 # else /* ! __GNUC__ */
00104 #  if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00105 #   define USE_UNR_VEC_KERNELS
00106 #  endif
00107 #  define DEF_UNROLL_DEPTH 8
00108 #  define DEF_PREFETCH_AHEAD 8
00109 # endif
00110 #endif /* __i386__ || __x86_64__*/
00111 
00112 /* defaults */
00113 #if !defined(USE_PLAIN_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS) && !defined(USE_UNR_VEC_KERNELS2)
00114 # define USE_PLAIN_VEC_KERNELS
00115 #endif
00116 #ifndef DEF_UNROLL_DEPTH
00117 # define DEF_UNROLL_DEPTH 4
00118 #endif
00119 #ifndef DEF_PREFETCH_AHEAD
00120 # define DEF_PREFETCH_AHEAD 4
00121 #endif
00122 #ifndef DEF_CACHELINE_SZ
00123 # define DEF_CACHELINE_SZ 32
00124 #endif
00125 
00128 #define DEF_CACHE_LOC_READ  2
00129 #define DEF_CACHE_LOC_WRITE 3
00130 
00131 /* Apply defaults */
00132 
00135 #ifndef PREFETCH_AHEAD
00136 # define PREFETCH_AHEAD DEF_PREFETCH_AHEAD
00137 #endif
00141 #ifndef UNROLL_DEPTH
00142 # define UNROLL_DEPTH DEF_UNROLL_DEPTH
00143 #endif
00147 #ifndef CACHELINE_SZ
00148 # define CACHELINE_SZ DEF_CACHELINE_SZ
00149 #endif
00150 
00160 #ifndef CACHE_LOC_READ
00161 # define CACHE_LOC_READ  DEF_CACHE_LOC_READ
00162 #endif
00163 #ifndef CACHE_LOC_WRITE
00164 # define CACHE_LOC_WRITE DEF_CACHE_LOC_WRITE
00165 #endif
00166 
00167 // Those should be evaluated at compile time
00168 #define EL_PER_CL(T) (signed)((CACHELINE_SZ/sizeof( T ))?(CACHELINE_SZ/sizeof( T )):1)
00169 #define PREF_OFFS(T) (EL_PER_CL(T)*PREFETCH_AHEAD)
00170 
00171 
00172 #endif  /* TBCI_PERF_OPT_H */
perf_opt.h
