e_exp.c 5.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166
  1. /* @(#)e_exp.c 1.6 04/04/22 */
  2. /*
  3. * ====================================================
  4. * Copyright (C) 2004 by Sun Microsystems, Inc. All rights reserved.
  5. *
  6. * Permission to use, copy, modify, and distribute this
  7. * software is freely granted, provided that this notice
  8. * is preserved.
  9. * ====================================================
  10. */
  11. /* __ieee754_exp(x)
  12. * Returns the exponential of x.
  13. *
  14. * Method
  15. * 1. Argument reduction:
  16. * Reduce x to an r so that |r| <= 0.5*ln2 ~ 0.34658.
  17. * Given x, find r and integer k such that
  18. *
  19. * x = k*ln2 + r, |r| <= 0.5*ln2.
  20. *
  21. * Here r will be represented as r = hi-lo for better
  22. * accuracy.
  23. *
  24. * 2. Approximation of exp(r) by a special rational function on
  25. * the interval [0,0.34658]:
  26. * Write
  27. * R(r**2) = r*(exp(r)+1)/(exp(r)-1) = 2 + r*r/6 - r**4/360 + ...
  28. * We use a special Remes algorithm on [0,0.34658] to generate
  29. * a polynomial of degree 5 to approximate R. The maximum error
  30. * of this polynomial approximation is bounded by 2**-59. In
  31. * other words,
  32. * R(z) ~ 2.0 + P1*z + P2*z**2 + P3*z**3 + P4*z**4 + P5*z**5
  33. * (where z=r*r, and the values of P1 to P5 are listed below)
  34. * and
  35. * | 5 | -59
  36. * | 2.0+P1*z+...+P5*z - R(z) | <= 2
  37. * | |
  38. * The computation of exp(r) thus becomes
  39. * 2*r
  40. * exp(r) = 1 + -------
  41. * R - r
  42. * r*R1(r)
  43. * = 1 + r + ----------- (for better accuracy)
  44. * 2 - R1(r)
  45. * where
  46. * 2 4 10
  47. * R1(r) = r - (P1*r + P2*r + ... + P5*r ).
  48. *
  49. * 3. Scale back to obtain exp(x):
  50. * From step 1, we have
  51. * exp(x) = 2^k * exp(r)
  52. *
  53. * Special cases:
  54. * exp(INF) is INF, exp(NaN) is NaN;
  55. * exp(-INF) is 0, and
  56. * for finite argument, only exp(0)=1 is exact.
  57. *
  58. * Accuracy:
  59. * according to an error analysis, the error is always less than
  60. * 1 ulp (unit in the last place).
  61. *
  62. * Misc. info.
  63. * For IEEE double
  64. * if x > 7.09782712893383973096e+02 then exp(x) overflow
  65. * if x < -7.45133219101941108420e+02 then exp(x) underflow
  66. *
  67. * Constants:
  68. * The hexadecimal values are the intended ones for the following
  69. * constants. The decimal values may be used, provided that the
  70. * compiler will convert from decimal to binary accurately enough
  71. * to produce the hexadecimal values shown.
  72. */
  73. #include "fdlibm.h"
  74. #ifndef _DOUBLE_IS_32BITS
  75. #ifdef __STDC__
  76. static const double
  77. #else
  78. static double
  79. #endif
  80. one = 1.0,
  81. halF[2] = {0.5,-0.5,},
  82. huge = 1.0e+300,
  83. twom1000= 9.33263618503218878990e-302, /* 2**-1000=0x01700000,0*/
  84. o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */
  85. u_threshold= -7.45133219101941108420e+02, /* 0xc0874910, 0xD52D3051 */
  86. ln2HI[2] ={ 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
  87. -6.93147180369123816490e-01,},/* 0xbfe62e42, 0xfee00000 */
  88. ln2LO[2] ={ 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */
  89. -1.90821492927058770002e-10,},/* 0xbdea39ef, 0x35793c76 */
  90. invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
  91. P1 = 1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */
  92. P2 = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */
  93. P3 = 6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */
  94. P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
  95. P5 = 4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */
  96. #ifdef __STDC__
  97. double __ieee754_exp(double x) /* default IEEE double exp */
  98. #else
  99. double __ieee754_exp(x) /* default IEEE double exp */
  100. double x;
  101. #endif
  102. {
  103. double y,hi,lo,c,t;
  104. int32_t k,xsb;
  105. uint32_t hx;
  106. GET_HIGH_WORD(hx,x); /* high word of x */
  107. xsb = (hx>>31)&1; /* sign bit of x */
  108. hx &= 0x7fffffff; /* high word of |x| */
  109. /* filter out non-finite argument */
  110. if(hx >= 0x40862E42) { /* if |x|>=709.78... */
  111. if(hx>=0x7ff00000) {
  112. uint32_t lx;
  113. GET_LOW_WORD(lx,x);
  114. if(((hx&0xfffff)|lx)!=0)
  115. return x+x; /* NaN */
  116. else return (xsb==0)? x:0.0; /* exp(+-inf)={inf,0} */
  117. }
  118. if(x > o_threshold) return huge*huge; /* overflow */
  119. if(x < u_threshold) return twom1000*twom1000; /* underflow */
  120. }
  121. /* argument reduction */
  122. if(hx > 0x3fd62e42) { /* if |x| > 0.5 ln2 */
  123. if(hx < 0x3FF0A2B2) { /* and |x| < 1.5 ln2 */
  124. hi = x-ln2HI[xsb]; lo=ln2LO[xsb]; k = 1-xsb-xsb;
  125. } else {
  126. k = (int32_t)(invln2*x+halF[xsb]);
  127. t = k;
  128. hi = x - t*ln2HI[0]; /* t*ln2HI is exact here */
  129. lo = t*ln2LO[0];
  130. }
  131. x = hi - lo;
  132. }
  133. else if(hx < 0x3e300000) { /* when |x|<2**-28 */
  134. if(huge+x>one) return one+x;/* trigger inexact */
  135. }
  136. else k = 0;
  137. /* x is now in primary range */
  138. t = x*x;
  139. c = x - t*(P1+t*(P2+t*(P3+t*(P4+t*P5))));
  140. if(k==0) return one-((x*c)/(c-2.0)-x);
  141. else y = one-((lo-(x*c)/(2.0-c))-hi);
  142. if(k >= -1021) {
  143. uint32_t hy;
  144. GET_HIGH_WORD(hy, y);
  145. SET_HIGH_WORD(y, hy + (k<<20)); /* add k to y's exponent */
  146. return y;
  147. } else {
  148. uint32_t hy;
  149. GET_HIGH_WORD(hy, y);
  150. SET_HIGH_WORD(y, hy + ((k+1000)<<20));/* add k to y's exponent */
  151. return y*twom1000;
  152. }
  153. }
  154. #endif /* defined(_DOUBLE_IS_32BITS) */