| /*---------------------------------------------------------------------------+ |
| | poly_l2.c | |
| | | |
| | Compute the base 2 log of a FPU_REG, using a polynomial approximation. | |
| | | |
| | Copyright (C) 1992,1993 | |
| | W. Metzenthen, 22 Parker St, Ormond, Vic 3163, | |
| | Australia. E-mail apm233m@vaxc.cc.monash.edu.au | |
| | | |
| | | |
| +---------------------------------------------------------------------------*/ |
| |
| |
| #include "exception.h" |
| #include "reg_constant.h" |
| #include "fpu_emu.h" |
| #include "control_w.h" |
| |
| |
| |
| #define HIPOWER 9 |
| static unsigned short lterms[HIPOWER][4] = |
| { |
| /* Ideal computation with these coeffs gives about |
| 64.6 bit rel accuracy. */ |
| { 0xe177, 0xb82f, 0x7652, 0x7154 }, |
| { 0xee0f, 0xe80f, 0x2770, 0x7b1c }, |
| { 0x0fc0, 0xbe87, 0xb143, 0x49dd }, |
| { 0x78b9, 0xdadd, 0xec54, 0x34c2 }, |
| { 0x003a, 0x5de9, 0x628b, 0x2909 }, |
| { 0x5588, 0xed16, 0x4abf, 0x2193 }, |
| { 0xb461, 0x85f7, 0x347a, 0x1c6a }, |
| { 0x0975, 0x87b3, 0xd5bf, 0x1876 }, |
| { 0xe85c, 0xcec9, 0x84e7, 0x187d } |
| }; |
| |
| |
| |
| |
| /*--- poly_l2() -------------------------------------------------------------+ |
| | Base 2 logarithm by a polynomial approximation. | |
| +---------------------------------------------------------------------------*/ |
| void poly_l2(FPU_REG *arg, FPU_REG *result) |
| { |
| short exponent; |
| char zero; /* flag for an Xx == 0 */ |
| unsigned short bits, shift; |
| long long Xsq; |
| FPU_REG accum, denom, num, Xx; |
| |
| |
| exponent = arg->exp - EXP_BIAS; |
| |
| accum.tag = TW_Valid; /* set the tags to Valid */ |
| |
| if ( arg->sigh > (unsigned)0xb504f334 ) |
| { |
| /* This is good enough for the computation of the polynomial |
| sum, but actually results in a loss of precision for |
| the computation of Xx. This will matter only if exponent |
| becomes zero. */ |
| exponent++; |
| accum.sign = 1; /* sign to negative */ |
| num.exp = EXP_BIAS; /* needed to prevent errors in div routine */ |
| reg_u_div(&CONST_1, arg, &num, FULL_PRECISION); |
| } |
| else |
| { |
| accum.sign = 0; /* set the sign to positive */ |
| num.sigl = arg->sigl; /* copy the mantissa */ |
| num.sigh = arg->sigh; |
| } |
| |
| |
| /* shift num left, lose the ms bit */ |
| num.sigh <<= 1; |
| if ( num.sigl & 0x80000000 ) num.sigh |= 1; |
| num.sigl <<= 1; |
| |
| denom.sigl = num.sigl; |
| denom.sigh = num.sigh; |
| poly_div4((long long *)&(denom.sigl)); |
| denom.sigh += 0x80000000; /* set the msb */ |
| Xx.exp = EXP_BIAS; /* needed to prevent errors in div routine */ |
| reg_u_div(&num, &denom, &Xx, FULL_PRECISION); |
| |
| zero = !(Xx.sigh | Xx.sigl); |
| |
| mul64((long long *)&Xx.sigl, (long long *)&Xx.sigl, &Xsq); |
| poly_div16(&Xsq); |
| |
| accum.exp = -1; /* exponent of accum */ |
| |
| /* Do the basic fixed point polynomial evaluation */ |
| polynomial((unsigned *)&accum.sigl, (unsigned *)&Xsq, lterms, HIPOWER-1); |
| |
| if ( !exponent ) |
| { |
| /* If the exponent is zero, then we would lose precision by |
| sticking to fixed point computation here */ |
| /* We need to re-compute Xx because of loss of precision. */ |
| FPU_REG lXx; |
| char sign; |
| |
| sign = accum.sign; |
| accum.sign = 0; |
| |
| /* make accum compatible and normalize */ |
| accum.exp = EXP_BIAS + accum.exp; |
| normalize(&accum); |
| |
| if ( zero ) |
| { |
| reg_move(&CONST_Z, result); |
| } |
| else |
| { |
| /* we need to re-compute lXx to better accuracy */ |
| num.tag = TW_Valid; /* set the tags to Vaild */ |
| num.sign = 0; /* set the sign to positive */ |
| num.exp = EXP_BIAS - 1; |
| if ( sign ) |
| { |
| /* The argument is of the form 1-x */ |
| /* Use 1-1/(1-x) = x/(1-x) */ |
| *((long long *)&num.sigl) = - *((long long *)&(arg->sigl)); |
| normalize(&num); |
| reg_div(&num, arg, &num, FULL_PRECISION); |
| } |
| else |
| { |
| normalize(&num); |
| } |
| |
| denom.tag = TW_Valid; /* set the tags to Valid */ |
| denom.sign = SIGN_POS; /* set the sign to positive */ |
| denom.exp = EXP_BIAS; |
| |
| reg_div(&num, &denom, &lXx, FULL_PRECISION); |
| |
| reg_u_mul(&lXx, &accum, &accum, FULL_PRECISION); |
| |
| reg_u_add(&lXx, &accum, result, FULL_PRECISION); |
| |
| normalize(result); |
| } |
| |
| result->sign = sign; |
| return; |
| } |
| |
| mul64((long long *)&accum.sigl, |
| (long long *)&Xx.sigl, (long long *)&accum.sigl); |
| |
| *((long long *)(&accum.sigl)) += *((long long *)(&Xx.sigl)); |
| |
| if ( Xx.sigh > accum.sigh ) |
| { |
| /* There was an overflow */ |
| |
| poly_div2((long long *)&accum.sigl); |
| accum.sigh |= 0x80000000; |
| accum.exp++; |
| } |
| |
| /* When we add the exponent to the accum result later, we will |
| require that their signs are the same. Here we ensure that |
| this is so. */ |
| if ( exponent && ((exponent < 0) ^ (accum.sign)) ) |
| { |
| /* signs are different */ |
| |
| accum.sign = !accum.sign; |
| |
| /* An exceptional case is when accum is zero */ |
| if ( accum.sigl | accum.sigh ) |
| { |
| /* find 1-accum */ |
| /* Shift to get exponent == 0 */ |
| if ( accum.exp < 0 ) |
| { |
| poly_div2((long long *)&accum.sigl); |
| accum.exp++; |
| } |
| /* Just negate, but throw away the sign */ |
| *((long long *)&(accum.sigl)) = - *((long long *)&(accum.sigl)); |
| if ( exponent < 0 ) |
| exponent++; |
| else |
| exponent--; |
| } |
| } |
| |
| shift = exponent >= 0 ? exponent : -exponent ; |
| bits = 0; |
| if ( shift ) |
| { |
| if ( accum.exp ) |
| { |
| accum.exp++; |
| poly_div2((long long *)&accum.sigl); |
| } |
| while ( shift ) |
| { |
| poly_div2((long long *)&accum.sigl); |
| if ( shift & 1) |
| accum.sigh |= 0x80000000; |
| shift >>= 1; |
| bits++; |
| } |
| } |
| |
| /* Convert to 64 bit signed-compatible */ |
| accum.exp += bits + EXP_BIAS - 1; |
| |
| reg_move(&accum, result); |
| normalize(result); |
| |
| return; |
| } |
| |
| |
| /*--- poly_l2p1() -----------------------------------------------------------+ |
| | Base 2 logarithm by a polynomial approximation. | |
| | log2(x+1) | |
| +---------------------------------------------------------------------------*/ |
| int poly_l2p1(FPU_REG *arg, FPU_REG *result) |
| { |
| char sign = 0; |
| long long Xsq; |
| FPU_REG arg_pl1, denom, accum, local_arg, poly_arg; |
| |
| |
| sign = arg->sign; |
| |
| reg_add(arg, &CONST_1, &arg_pl1, FULL_PRECISION); |
| |
| if ( (arg_pl1.sign) | (arg_pl1.tag) ) |
| { /* We need a valid positive number! */ |
| return 1; |
| } |
| |
| reg_add(&CONST_1, &arg_pl1, &denom, FULL_PRECISION); |
| reg_div(arg, &denom, &local_arg, FULL_PRECISION); |
| local_arg.sign = 0; /* Make the sign positive */ |
| |
| /* Now we need to check that |local_arg| is less than |
| 3-2*sqrt(2) = 0.17157.. = .0xafb0ccc0 * 2^-2 */ |
| |
| if ( local_arg.exp >= EXP_BIAS - 3 ) |
| { |
| if ( (local_arg.exp > EXP_BIAS - 3) || |
| (local_arg.sigh > (unsigned)0xafb0ccc0) ) |
| { |
| /* The argument is large */ |
| poly_l2(&arg_pl1, result); return 0; |
| } |
| } |
| |
| /* Make a copy of local_arg */ |
| reg_move(&local_arg, &poly_arg); |
| |
| /* Get poly_arg bits aligned as required */ |
| shrx((unsigned *)&(poly_arg.sigl), -(poly_arg.exp - EXP_BIAS + 3)); |
| |
| mul64((long long *)&(poly_arg.sigl), (long long *)&(poly_arg.sigl), &Xsq); |
| poly_div16(&Xsq); |
| |
| /* Do the basic fixed point polynomial evaluation */ |
| polynomial(&(accum.sigl), (unsigned *)&Xsq, lterms, HIPOWER-1); |
| |
| accum.tag = TW_Valid; /* set the tags to Valid */ |
| accum.sign = SIGN_POS; /* and make accum positive */ |
| |
| /* make accum compatible and normalize */ |
| accum.exp = EXP_BIAS - 1; |
| normalize(&accum); |
| |
| reg_u_mul(&local_arg, &accum, &accum, FULL_PRECISION); |
| |
| reg_u_add(&local_arg, &accum, result, FULL_PRECISION); |
| |
| /* Multiply the result by 2 */ |
| result->exp++; |
| |
| result->sign = sign; |
| |
| return 0; |
| } |