Document the log table generation method

Add comments with enough detail so the log lookup tables can be recreated.
This commit is contained in:
Szabolcs Nagy 2018-09-05 12:15:55 +01:00 committed by Corinna Vinschen
parent 7283d2513c
commit f92a4c5d2d
3 changed files with 74 additions and 0 deletions

View File

@ -66,6 +66,32 @@ const struct log2_data __log2_data = {
0x1.a6225e117f92ep-3,
#endif
},
/* Algorithm:
x = 2^k z
log2(x) = k + log2(c) + log2(z/c)
log2(z/c) = poly(z/c - 1)
where z is in [1.6p-1; 1.6p0] which is split into N subintervals and z falls
into the ith one, then table entries are computed as
tab[i].invc = 1/c
tab[i].logc = (double)log2(c)
tab2[i].chi = (double)c
tab2[i].clo = (double)(c - (double)c)
where c is near the center of the subinterval and is chosen by trying +-2^29
floating point invc candidates around 1/center and selecting one for which
1) the rounding error in 0x1.8p10 + logc is 0,
2) the rounding error in z - chi - clo is < 0x1p-64 and
3) the rounding error in (double)log2(c) is minimized (< 0x1p-68).
Note: 1) ensures that k + logc can be computed without rounding error, 2)
ensures that z/c - 1 can be computed as (z - chi - clo)*invc with close to a
single rounding error when there is no fast fma for z*invc - 1, 3) ensures
that logc + poly(z/c - 1) has small error, however near x == 1 when
|log2(x)| < 0x1p-4, this is not enough so that is special cased. */
.tab = {
#if N == 64
{0x1.724286bb1acf8p+0, -0x1.1095feecdb000p-1},

View File

@ -110,6 +110,32 @@ const struct log_data __log_data = {
0x1.2493c29331a5cp-3,
#endif
},
/* Algorithm:
x = 2^k z
log(x) = k ln2 + log(c) + log(z/c)
log(z/c) = poly(z/c - 1)
where z is in [1.6p-1; 1.6p0] which is split into N subintervals and z falls
into the ith one, then table entries are computed as
tab[i].invc = 1/c
tab[i].logc = (double)log(c)
tab2[i].chi = (double)c
tab2[i].clo = (double)(c - (double)c)
where c is near the center of the subinterval and is chosen by trying +-2^29
floating point invc candidates around 1/center and selecting one for which
1) the rounding error in 0x1.8p9 + logc is 0,
2) the rounding error in z - chi - clo is < 0x1p-66 and
3) the rounding error in (double)log(c) is minimized (< 0x1p-66).
Note: 1) ensures that k*ln2hi + logc can be computed without rounding error,
2) ensures that z/c - 1 can be computed as (z - chi - clo)*invc with close to
a single rounding error when there is no fast fma for z*invc - 1, 3) ensures
that logc + poly(z/c - 1) has small error, however near x == 1 when
|log(x)| < 0x1p-4, this is not enough so that is special cased. */
.tab = {
#if N == 64
{0x1.7242886495cd8p+0, -0x1.79e267bdfe000p-2},

View File

@ -50,6 +50,28 @@ const struct pow_log_data __pow_log_data = {
-0x1.0002b8b263fc3p-3 * -8,
#endif
},
/* Algorithm:
x = 2^k z
log(x) = k ln2 + log(c) + log(z/c)
log(z/c) = poly(z/c - 1)
where z is in [0x1.69555p-1; 0x1.69555p0] which is split into N subintervals
and z falls into the ith one, then table entries are computed as
tab[i].invc = 1/c
tab[i].logc = round(0x1p43*log(c))/0x1p43
tab[i].logctail = (double)(log(c) - logc)
where c is chosen near the center of the subinterval such that 1/c has only a
few precision bits so z/c - 1 is exactly representible as double:
1/c = center < 1 ? round(N/center)/N : round(2*N/center)/N/2
Note: |z/c - 1| < 1/N for the chosen c, |log(c) - logc - logctail| < 0x1p-97,
the last few bits of logc are rounded away so k*ln2hi + logc has no rounding
error and the interval for z is selected such that near x == 1, where log(x)
is tiny, large cancellation error is avoided in logc + poly(z/c - 1). */
.tab = {
#if N == 128
#define A(a,b,c) {a,0,b,c},