Lines Matching +full:low +full:- +full:precision
19 * log2(x) = (f - 0.5*f*f + k_log1p(f)) / ln2 + k
20 * in not-quite-routine extra precision.
32 ivln2lo = 1.67517131648865118353e-10; /* 0x3de705fc, 0x2eefa200 */
47 if (hx < 0x00100000) { /* x < 2**-1022 */ in log2()
49 return -two54/vzero; /* log(+-0)=-inf */ in log2()
50 if (hx<0) return (x-x)/zero; /* log(-#) = NaN */ in log2()
51 k -= 54; x *= two54; /* subnormal number, scale up x */ in log2()
57 k += (hx>>20)-1023; in log2()
63 f = x - 1.0; in log2()
68 * f-hfsq must (for args near 1) be evaluated in extra precision in log2()
70 * This is fairly efficient since f-hfsq only depends on f, so can in log2()
73 * so that extra precision is not needed for terms involving R. in log2()
75 * Compiler bugs involving extra precision used to break Dekker's in log2()
76 * theorem for spitting f-hfsq as hi+lo, unless double_t was used in log2()
77 * or the multi-precision calculations were avoided when double_t in log2()
78 * has extra precision. These problems are now automatically in log2()
80 * Dekker splitting step with the clear-low-bits step. in log2()
83 * precision to avoid a very large cancellation when x is very near in log2()
85 * specific to base 2. It is strange that adding +-1 is so much in log2()
86 * harder than adding +-ln2 or +-log10_2. in log2()
94 * The multi-precision calculations for the multiplications are in log2()
97 hi = f - hfsq; in log2()
99 lo = (f - hi) - hfsq + r; in log2()
105 val_lo += (y - w) + val_hi; in log2()