char _cosminus1ltPiDiv4( floatnum x, int digits) { floatstruct tmp; int reductions; if (float_iszero(x)) return 1; float_abs(x); reductions = 0; while(float_getexponent(x) >= -2) { float_mul(x, x, &c1Div2, digits+1); ++reductions; } if (!cosminus1near0(x, digits) && reductions == 0) return !float_iszero(x); float_create(&tmp); for(; reductions-- > 0;) { float_mul(&tmp, x, x, digits); float_add(x, x, x, digits+2); float_add(x, x, &tmp, digits+2); float_add(x, x, x, digits+2); } float_free(&tmp); return 1; }
static char _lngamma_prim( floatnum x, floatnum revfactor, int* infinity, int digits) { floatstruct tmp; char result; char odd; *infinity = 0; if (float_getsign(x) > 0) return _lngamma_prim_xgt0(x, revfactor, digits); float_copy(revfactor, x, digits + 2); float_sub(x, &c1, x, digits+2); float_create(&tmp); result = _lngamma_prim_xgt0(x, &tmp, digits); if (result) { float_neg(x); odd = float_isodd(revfactor); _sinpix(revfactor, digits); if (float_iszero(revfactor)) { *infinity = 1; float_setinteger(revfactor, odd? -1 : 1); } else float_mul(&tmp, &tmp, &cPi, digits+2); float_div(revfactor, revfactor, &tmp, digits+2); } float_free(&tmp); return result; }
Error float_out( p_otokens tokens, floatnum x, int scale, signed char base, char outmode) { t_number_desc n; _emptytokens(tokens); /* do some sanity checks first */ if (!_validmode(outmode) || scale < 0 || !_isvalidbase(base)) return InvalidParam; _clearnumber(&n); if (float_iszero(x)) n.prefix.base = IO_BASE_ZERO; else if (!float_isnan(x)) n.prefix.base = base; if (!_isvalidbase(n.prefix.base)) /* NaN and 0 are handled here */ return desc2str(tokens, &n, 0); n.prefix.sign = float_getsign(x); float_abs(x); switch (outmode) { case IO_MODE_FIXPOINT: return _outfixp(tokens, x, &n, scale); case IO_MODE_ENG: return _outeng(tokens, x, &n, scale); case IO_MODE_COMPLEMENT: return _outcompl(tokens, x, &n, 0); default: return _outsci(tokens, x, &n, scale); } }
char _floatnum2logic( t_longint* longint, cfloatnum x) { floatstruct tmp; int digits; digits = float_getexponent(x)+1; if (float_iszero(x) || digits <= 0) { longint->length = 1; longint->value[0] = 0; } else { if (digits > MATHPRECISION) return 0; float_create(&tmp); /* floatnum2longint rounds, we have to truncate first */ float_copy(&tmp, x, digits); if (float_getsign(x) < 0) float_add(&tmp, &tmp, &c1, EXACT); _floatnum2longint(longint, &tmp); float_free(&tmp); if (_bitlength(longint) > LOGICRANGE) return 0; } _zeroextend(longint); if (float_getsign(x) < 0) _not(longint); return 1; }
char erfseries( floatnum x, int digits) { floatstruct xsqr, smd, pwr; int i, workprec, expx; expx = float_getexponent(x); workprec = digits + 2*expx + 2; if (workprec <= 0 || float_iszero(x)) /* for tiny arguments approx. == x */ return 1; float_create(&xsqr); float_create(&smd); float_create(&pwr); float_mul(&xsqr, x, x, workprec + 1); workprec = digits + float_getexponent(&xsqr) + 1; float_copy(&pwr, x, workprec + 1); i = 1; while (workprec > 0) { float_mul(&pwr, &pwr, &xsqr, workprec + 1); float_divi(&pwr, &pwr, -i, workprec + 1); float_divi(&smd, &pwr, 2 * i++ + 1, workprec); float_add(x, x, &smd, digits + 3); workprec = digits + float_getexponent(&smd) + expx + 2; } float_free(&pwr); float_free(&smd); float_free(&xsqr); return 1; }
char _doshift( floatnum dest, cfloatnum x, cfloatnum shift, char right) { int ishift; t_longint lx; if (float_isnan(shift)) return _seterror(dest, NoOperand); if (!float_isinteger(shift)) return _seterror(dest, OutOfDomain); if(!_cvtlogic(&lx, x)) return 0; if (float_iszero(shift)) { float_copy(dest, x, EXACT); return 1; } ishift = float_asinteger(shift); if (ishift == 0) ishift = (3*LOGICRANGE) * float_getsign(shift); if (!right) ishift = -ishift; if (ishift > 0) _shr(&lx, ishift); else _shl(&lx, -ishift); _logic2floatnum(dest, &lx); return 1; }
char float_raisei( floatnum power, cfloatnum base, int exponent, int digits) { if (digits <= 0 || digits > maxdigits) return _seterror(power, InvalidPrecision); if (float_isnan(base)) return _seterror(power, NoOperand); if (float_iszero(base)) { if (exponent == 0) return _seterror(power, OutOfDomain); if (exponent < 0) return _seterror(power, ZeroDivide); return _setzero(power); } digits += 14; if (digits > maxdigits) digits = maxdigits; float_copy(power, base, digits); if (!_raisei(power, exponent, digits) || !float_isvalidexp(float_getexponent(power))) { if (float_getexponent(base) < 0) return _seterror(power, Underflow); return _seterror(power, Overflow); } return 1; }
/* evaluates arctan x for |x| <= 1 relative error for a 100 digit result is 6e-100 */ void _arctanlt1( floatnum x, int digits) { floatstruct tmp; int reductions; if (float_iszero(x)) return; float_create(&tmp); reductions = 0; while(float_getexponent(x) >= -2) { float_mul(&tmp, x, x, digits); float_add(&tmp, &tmp, &c1, digits+2); float_sqrt(&tmp, digits); float_add(&tmp, &tmp, &c1, digits+1); float_div(x, x, &tmp, digits); ++reductions; } arctannear0(x, digits); for (;reductions-- > 0;) float_add(x, x, x, digits+1); float_free(&tmp); }
/* evaluates tan x for |x| <= pi. A return value of 0 indicates that x = +/- pi/2 within small tolerances, so that tan x cannot be reliable computed */ char _tan( floatnum x, int digits) { signed char sgn; sgn = float_getsign(x); float_abs(x); if (float_cmp(x, &cPiDiv2) > 0) { float_sub(x, &cPi, x, digits+1); sgn = -sgn; } if (float_cmp(x, &cPiDiv4) <= 0) _tanltPiDiv4(x, digits); else { float_sub(x, &cPiDiv2, x, digits+1); if (float_iszero(x) || float_getexponent(x) < -digits) return 0; _tanltPiDiv4(x, digits); float_reciprocal(x, digits); } float_setsign(x, sgn); return 1; }
/* series expansion of cos/cosh - 1 used for small x, |x| <= 0.01. The function returns 0, if an underflow occurs. The relative error seems to be less than 5e-100 for a 100-digit calculation with |x| < 0.01 */ char cosminus1series( floatnum x, int digits, char alternating) { floatstruct sum, smd; int expsqrx, pwrsz, addsz, i; expsqrx = 2 * float_getexponent(x); float_setexponent(x, 0); float_mul(x, x, x, digits+1); float_mul(x, x, &c1Div2, digits+1); float_setsign(x, alternating? -1 : 1); expsqrx += float_getexponent(x); if (float_iszero(x) || expsqrx < EXPMIN) { /* underflow */ float_setzero(x); return expsqrx == 0; } float_setexponent(x, expsqrx); pwrsz = digits + expsqrx + 2; if (pwrsz <= 0) /* for very small x, cos/cosh(x) - 1 = (-/+)0.5*x*x */ return 1; addsz = pwrsz; float_create(&sum); float_create(&smd); float_copy(&smd, x, pwrsz); float_setzero(&sum); i = 2; while (pwrsz > 0) { float_mul(&smd, &smd, x, pwrsz+1); float_divi(&smd, &smd, i*(2*i-1), pwrsz); float_add(&sum, &sum, &smd, addsz); ++i; pwrsz = digits + float_getexponent(&smd); } float_add(x, x, &sum, digits+1); float_free(&sum); float_free(&smd); return 1; }
char float_raise( floatnum power, cfloatnum base, cfloatnum exponent, int digits) { signed char sgn; if (float_isnan(exponent) || float_isnan(base)) return _seterror(power, NoOperand); if (digits <= 0 || digits > MATHPRECISION) return _seterror(power, InvalidPrecision); if (float_iszero(base)) { switch(float_getsign(exponent)) { case 0: return _seterror(power, OutOfDomain); case -1: return _seterror(power, ZeroDivide); } return _setzero(power); } sgn = float_getsign(base); if (sgn < 0) { if (!float_isinteger(exponent)) return _seterror(power, OutOfDomain); if ((float_getdigit(exponent, float_getexponent(exponent)) & 1) == 0) sgn = 1; } float_copy(power, base, digits+1); float_abs(power); if (!_raise(power, exponent, digits)) { float_seterror(Overflow); if (float_getexponent(base) * float_getsign(exponent) < 0) float_seterror(Underflow); return _setnan(power); } float_setsign(power, sgn); return 1; }
char float_tanhminus1( floatnum x, int digits) { if (!chckmathparam(x, digits)) return 0; if (float_cmp(x, &c1Div2) >= 0) return _tanhminus1gt0(x, digits)? 1 : _seterror(x, Underflow); if (!float_iszero(x)) { if (float_abscmp(x, &c1Div2) <= 0) _tanhlt0_5(x, digits); else { float_setsign(x, 1); _tanhgt0_5(x, digits); float_setsign(x, -1); } } return float_sub(x, x, &c1, digits); }
static char _pochhammer_i( floatnum x, cfloatnum n, int digits) { /* do not use the expensive Gamma function when a few multiplications do the same */ /* pre: n is an integer */ int ni; signed char result; if (float_iszero(n)) return float_copy(x, &c1, EXACT); if (float_isinteger(x)) { result = -1; float_neg((floatnum)n); if (float_getsign(x) <= 0 && float_cmp(x, n) > 0) /* x and x+n have opposite signs, meaning 0 is among the factors */ result = _setzero(x); else if (float_getsign(x) > 0 && float_cmp(x, n) <= 0) /* x and x+n have opposite signs, meaning at one point you have to divide by 0 */ result = _seterror(x, ZeroDivide); float_neg((floatnum)n); if (result >= 0) return result; } if (float_getexponent(x) < EXPMAX/100) { ni = float_asinteger(n); if (ni != 0 && ni < 50 && ni > -50) return _pochhammer_si(x, ni, digits+2); } return _pochhammer_g(x, n, digits); }
/* the Taylor series of arctan/arctanh x at x == 0. For small |x| < 0.01 this series converges very fast, yielding 4 or more digits of the result with every summand. The working precision is adjusted, so that the relative error for 100-digit arguments is around 5.0e-100. This means, the error is 1 in the 100-th place (or less) */ void arctanseries( floatnum x, int digits, char alternating) { int expx; int expsqrx; int pwrsz; int addsz; int i; floatstruct xsqr; floatstruct pwr; floatstruct smd; floatstruct sum; /* upper limit of log(x) and log(result) */ expx = float_getexponent(x)+1; /* the summands of the series from the second on are bounded by x^(2*i-1)/3. So the summation yields a result bounded by (x^3/(1-x*x))/3. For x < sqrt(1/3) approx.= 0.5, this is less than 0.5*x^3. We need to sum up only, if the first <digits> places of the result (roughly x) are touched. Ignoring the effect of a possile carry, this is only the case, if x*x >= 2*10^(-digits) > 10^(-digits) Example: for x = 9e-51, a 100-digits result covers the decimal places from 1e-51 to 1e-150. x^3/3 is roughly 3e-151, and so is the sum of the series. So we can ignore the sum, but we couldn't for x = 9e-50 */ if (float_iszero(x) || 2*expx < -digits) /* for very tiny arguments arctan/arctanh x is approx.== x */ return; float_create(&xsqr); float_create(&pwr); float_create(&smd); float_create(&sum); /* we adapt the working precision to the decreasing summands, saving time when multiplying. Unfortunately, there is no error bound given for the operations of bc_num. Tests show, that the last digit in an incomplete multiplication is usually not correct up to 5 ULP's. */ pwrsz = digits + 2*expx + 1; /* the precision of the addition must not decrease, of course */ addsz = pwrsz; i = 3; float_mul(&xsqr, x, x, pwrsz); float_setsign(&xsqr, alternating? -1 : 1); expsqrx = float_getexponent(&xsqr); float_copy(&pwr, x, pwrsz); float_setzero(&sum); for(; pwrsz > 0; ) { /* x^i */ float_mul(&pwr, &pwr, &xsqr, pwrsz+1); /* x^i/i */ float_divi(&smd, &pwr, i, pwrsz); /* The addition virtually does not introduce errors */ float_add(&sum, &sum, &smd, addsz); /* reduce the working precision according to the decreasing powers */ pwrsz = digits - expx + float_getexponent(&smd) + expsqrx + 3; i += 2; } /* add the first summand */ float_add(x, x, &sum, digits+1); float_free(&xsqr); float_free(&pwr); float_free(&smd); float_free(&sum); }