tremor/misc.h

   1 /********************************************************************
   2  *                                                                  *
   3  * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE.   *
   4  *                                                                  *
   5  * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
   6  * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
   7  * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
   8  *                                                                  *
   9  * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002    *
  10  * BY THE Xiph.Org FOUNDATION http://www.xiph.org/                  *
  11  *                                                                  *
  12  ********************************************************************
  13
  14  function: miscellaneous math and prototypes
  15
  16  ********************************************************************/
  17
  18 #ifndef _V_RANDOM_H_
  19 #define _V_RANDOM_H_
  20 #include "ivorbiscodec.h"
  21 #include "os_types.h"
  22
  23 #include "asm_arm.h"
  24
  25 #ifndef _V_WIDE_MATH
  26 #define _V_WIDE_MATH
  27
  28 #ifndef  _LOW_ACCURACY_
  29 /* 64 bit multiply */
  30
  31 #include <sys/types.h>
  32 #include "config.h"
  33
  34 #ifndef WORDS_BIGENDIAN
  35 union magic {
  36   struct {
  37     ogg_int32_t lo;
  38     ogg_int32_t hi;
  39   } halves;
  40   ogg_int64_t whole;
  41 };
  42 #else
  43 union magic {
  44   struct {
  45     ogg_int32_t hi;
  46     ogg_int32_t lo;
  47   } halves;
  48   ogg_int64_t whole;
  49 };
  50 #endif
  51
  52 static inline ogg_int32_t MULT32(ogg_int32_t x, ogg_int32_t y) {
  53   union magic magic;
  54   magic.whole = (ogg_int64_t)x * y;
  55   return magic.halves.hi;
  56 }
  57
  58 static inline ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
  59   return MULT32(x,y)<<1;
  60 }
  61
  62 static inline ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
  63   union magic magic;
  64   magic.whole  = (ogg_int64_t)x * y;
  65   return ((ogg_uint32_t)(magic.halves.lo)>>15) | ((magic.halves.hi)<<17);
  66 }
  67
  68 #else
  69 /* 32 bit multiply, more portable but less accurate */
  70
  71 /*
  72  * Note: Precision is biased towards the first argument therefore ordering
  73  * is important.  Shift values were chosen for the best sound quality after
  74  * many listening tests.
  75  */
  76
  77 /*
  78  * For MULT32 and MULT31: The second argument is always a lookup table
  79  * value already preshifted from 31 to 8 bits.  We therefore take the
  80  * opportunity to save on text space and use unsigned char for those
  81  * tables in this case.
  82  */
  83
  84 static inline ogg_int32_t MULT32(ogg_int32_t x, ogg_int32_t y) {
  85   return (x >> 9) * y;  /* y preshifted >>23 */
  86 }
  87
  88 static inline ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
  89   return (x >> 8) * y;  /* y preshifted >>23 */
  90 }
  91
  92 static inline ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
  93   return (x >> 6) * y;  /* y preshifted >>9 */
  94 }
  95
  96 #endif
  97
  98 /*
  99  * This should be used as a memory barrier, forcing all cached values in
 100  * registers to wr writen back to memory.  Might or might not be beneficial
 101  * depending on the architecture and compiler.
 102  */
 103 #define MB()
 104
 105 /*
 106  * The XPROD functions are meant to optimize the cross products found all
 107  * over the place in mdct.c by forcing memory operation ordering to avoid
 108  * unnecessary register reloads as soon as memory is being written to.
 109  * However this is only beneficial on CPUs with a sane number of general
 110  * purpose registers which exclude the Intel x86.  On Intel, better let the
 111  * compiler actually reload registers directly from original memory by using
 112  * macros.
 113  */
 114
 115 #ifdef __i386__
 116
 117 #define XPROD32(_a, _b, _t, _v, _x, _y)         \
 118   { *(_x)=MULT32(_a,_t)+MULT32(_b,_v);          \
 119     *(_y)=MULT32(_b,_t)-MULT32(_a,_v); }
 120 #define XPROD31(_a, _b, _t, _v, _x, _y)         \
 121   { *(_x)=MULT31(_a,_t)+MULT31(_b,_v);          \
 122     *(_y)=MULT31(_b,_t)-MULT31(_a,_v); }
 123 #define XNPROD31(_a, _b, _t, _v, _x, _y)        \
 124   { *(_x)=MULT31(_a,_t)-MULT31(_b,_v);          \
 125     *(_y)=MULT31(_b,_t)+MULT31(_a,_v); }
 126
 127 #else
 128
 129 static inline void XPROD32(ogg_int32_t  a, ogg_int32_t  b,
 130                            ogg_int32_t  t, ogg_int32_t  v,
 131                            ogg_int32_t *x, ogg_int32_t *y)
 132 {
 133   *x = MULT32(a, t) + MULT32(b, v);
 134   *y = MULT32(b, t) - MULT32(a, v);
 135 }
 136
 137 static inline void XPROD31(ogg_int32_t  a, ogg_int32_t  b,
 138                            ogg_int32_t  t, ogg_int32_t  v,
 139                            ogg_int32_t *x, ogg_int32_t *y)
 140 {
 141   *x = MULT31(a, t) + MULT31(b, v);
 142   *y = MULT31(b, t) - MULT31(a, v);
 143 }
 144
 145 static inline void XNPROD31(ogg_int32_t  a, ogg_int32_t  b,
 146                             ogg_int32_t  t, ogg_int32_t  v,
 147                             ogg_int32_t *x, ogg_int32_t *y)
 148 {
 149   *x = MULT31(a, t) - MULT31(b, v);
 150   *y = MULT31(b, t) + MULT31(a, v);
 151 }
 152
 153 #endif
 154
 155 #endif
 156
 157 #ifndef _V_CLIP_MATH
 158 #define _V_CLIP_MATH
 159
 160 static inline ogg_int32_t CLIP_TO_15(ogg_int32_t x) {
 161   int ret=x;
 162   ret-= ((x<=32767)-1)&(x-32767);
 163   ret-= ((x>=-32768)-1)&(x+32768);
 164   return(ret);
 165 }
 166
 167 #endif
 168
 169 static inline ogg_int32_t VFLOAT_MULT(ogg_int32_t a,ogg_int32_t ap,
 170                                       ogg_int32_t b,ogg_int32_t bp,
 171                                       ogg_int32_t *p){
 172   if(a && b){
 173 #ifndef _LOW_ACCURACY_
 174     *p=ap+bp+32;
 175     return MULT32(a,b);
 176 #else
 177     *p=ap+bp+31;
 178     return (a>>15)*(b>>16);
 179 #endif
 180   }else
 181     return 0;
 182 }
 183
 184 static inline ogg_int32_t VFLOAT_MULTI(ogg_int32_t a,ogg_int32_t ap,
 185                                       ogg_int32_t i,
 186                                       ogg_int32_t *p){
 187
 188   int ip=_ilog(abs(i))-31;
 189   return VFLOAT_MULT(a,ap,i<<-ip,ip,p);
 190 }
 191
 192 static inline ogg_int32_t VFLOAT_ADD(ogg_int32_t a,ogg_int32_t ap,
 193                                       ogg_int32_t b,ogg_int32_t bp,
 194                                       ogg_int32_t *p){
 195
 196   if(!a){
 197     *p=bp;
 198     return b;
 199   }else if(!b){
 200     *p=ap;
 201     return a;
 202   }
 203
 204   /* yes, this can leak a bit. */
 205   if(ap>bp){
 206     int shift=ap-bp+1;
 207     *p=ap+1;
 208     a>>=1;
 209     if(shift<32){
 210       b=(b+(1<<(shift-1)))>>shift;
 211     }else{
 212       b=0;
 213     }
 214   }else{
 215     int shift=bp-ap+1;
 216     *p=bp+1;
 217     b>>=1;
 218     if(shift<32){
 219       a=(a+(1<<(shift-1)))>>shift;
 220     }else{
 221       a=0;
 222     }
 223   }
 224
 225   a+=b;
 226   if((a&0xc0000000)==0xc0000000 ||
 227      (a&0xc0000000)==0){
 228     a<<=1;
 229     (*p)--;
 230   }
 231   return(a);
 232 }
 233
 234 #endif
 235
 236
 237
 238