static void SampleRule(This *t, ccount iregion)
{
SAMPLERDEFS;
Set *first = samples->rule->first;
Set *last = samples->rule->last;
Set *s;
creal *errcoeff = samples->rule->errcoeff;
count comp, rul, sn;
csize_t setsize = SetSize;
for( s = first; s <= last; NextSet(s) )
if( s->n ) x = ExpandFS(t, b, s->gen, x);
DoSample(t, n, samples->x, f);
for( comp = 0; comp < t->ncomp; ++comp ) {
real sum[nrules];
creal *f1 = f++;
Zap(sum);
for( s = first; s <= last; NextSet(s) )
for( sn = s->n; sn; --sn ) {
creal fun = *f1;
f1 += t->ncomp;
for( rul = 0; rul < nrules; ++rul )
sum[rul] += fun*s->weight[rul];
}
/* Search for the null rule, in the linear space spanned by two
successive null rules in our sequence, which gives the greatest
error estimate among all normalized (1-norm) null rules in this
space. */
for( rul = 1; rul < nrules - 1; ++rul ) {
real maxerr = 0;
for( s = first; s <= last; NextSet(s) )
maxerr = Max(maxerr,
fabsx(sum[rul + 1] + s->scale[rul]*sum[rul])*s->norm[rul]);
sum[rul] = maxerr;
}
res[comp].avg = region->vol*sum[0];
res[comp].err = region->vol*(
(errcoeff[0]*sum[1] <= sum[2] && errcoeff[0]*sum[2] <= sum[3]) ?
errcoeff[1]*sum[1] :
errcoeff[2]*Max(Max(sum[1], sum[2]), sum[3]) );
}
}
static void Sample(This *t, void *voidregion)
{
TYPEDEFREGION;
TYPEDEFSET;
Region *const region = (Region *)voidregion;
creal vol = ldexp(1., -region->div);
real *x = t->rule.x, *f = t->rule.f;
Set *first = (Set *)t->rule.first, *last = (Set *)t->rule.last, *s;
creal *errcoeff = t->rule.errcoeff;
creal ratio = Sq(first[2].gen[0]/first[1].gen[0]);
ccount offset = 2*t->ndim*t->ncomp;
count dim, comp, rul, n, maxdim = 0;
real maxrange = 0;
for( dim = 0; dim < t->ndim; ++dim ) {
cBounds *b = ®ion->bounds[dim];
creal range = b->upper - b->lower;
if( range > maxrange ) {
maxrange = range;
maxdim = dim;
}
}
for( s = first; s <= last; ++s )
if( s->n ) x = ExpandFS(t, region->bounds, s->gen, x);
DoSample(t, t->rule.n, t->rule.x, f);
for( comp = 0; comp < t->ncomp; ++comp ) {
Result *r = ®ion->result[comp];
real sum[nrules];
creal *f1 = f;
creal base = *f1*2*(1 - ratio);
real maxdiff = 0;
count bisectdim = maxdim;
for( dim = 0; dim < t->ndim; ++dim ) {
creal *fp = f1 + t->ncomp;
creal *fm = fp + t->ncomp;
creal fourthdiff = fabs(base +
ratio*(fp[0] + fm[0]) - (fp[offset] + fm[offset]));
f1 = fm;
if( fourthdiff > maxdiff ) {
maxdiff = fourthdiff;
bisectdim = dim;
}
}
r->bisectdim = bisectdim;
f1 = f++;
Zap(sum);
for( s = first; s <= last; ++s )
for( n = s->n; n; --n ) {
creal fun = *f1;
f1 += t->ncomp;
for( rul = 0; rul < nrules; ++rul )
sum[rul] += fun*s->weight[rul];
}
/* Search for the null rule, in the linear space spanned by two
successive null rules in our sequence, which gives the greatest
error estimate among all normalized (1-norm) null rules in this
space. */
for( rul = 1; rul < nrules - 1; ++rul ) {
real maxerr = 0;
for( s = first; s <= last; ++s )
maxerr = Max(maxerr,
fabs(sum[rul + 1] + s->scale[rul]*sum[rul])*s->norm[rul]);
sum[rul] = maxerr;
}
r->avg = vol*sum[0];
r->err = vol*(
(errcoeff[0]*sum[1] <= sum[2] && errcoeff[0]*sum[2] <= sum[3]) ?
errcoeff[1]*sum[1] :
errcoeff[2]*Max(Max(sum[1], sum[2]), sum[3]) );
}
if( VERBOSE > 2 ) {
char s[64*NDIM + 128*NCOMP], *p = s;
for( dim = 0; dim < t->ndim; ++dim ) {
cBounds *b = ®ion->bounds[dim];
p += sprintf(p,
(dim == 0) ? "\nRegion (" REALF ") - (" REALF ")" :
"\n (" REALF ") - (" REALF ")",
b->lower, b->upper);
}
for( comp = 0; comp < t->ncomp; ++comp ) {
cResult *r = ®ion->result[comp];
p += sprintf(p, "\n[" COUNT "] "
REAL " +- " REAL, comp + 1, r->avg, r->err);
}
Print(s);
}
}
static void Sample(This *t, Region *region)
{
csize_t setsize = SetSize;
creal vol = ldexp(1., -region->div);
real *x = t->frame, *f = x + t->rule.n*t->ndim;
Set *first = t->rule.first, *last = t->rule.last, *s;
Bounds *b, *B = region->bounds + t->ndim;
Result *result = RegionResult(region), *res, *Res = result + t->ncomp;
creal *errcoeff = t->rule.errcoeff;
creal ratio = Sq(IndexSet(first,2)->gen[0]/
IndexSet(first,1)->gen[0]);
ccount offset = 2*t->ndim*t->ncomp;
count dim, rul, n, maxdim = 0;
real maxrange = 0;
for( b = region->bounds, dim = 0; b < B; ++b, ++dim ) {
creal range = b->upper - b->lower;
if( range > maxrange ) {
maxrange = range;
maxdim = dim;
}
}
for( s = first; s <= last; NextSet(s) )
if( s->n ) x = ExpandFS(t, region->bounds, s->gen, x);
DoSample(t, t->rule.n, t->frame, f);
for( res = result; res < Res; ++res ) {
real sum[nrules];
creal *f1 = f;
creal base = *f1*2*(1 - ratio);
real maxdiff = 0;
count bisectdim = maxdim;
for( dim = 0; dim < t->ndim; ++dim ) {
creal *fp = f1 + t->ncomp;
creal *fm = fp + t->ncomp;
creal fourthdiff = fabs(base +
ratio*(fp[0] + fm[0]) - (fp[offset] + fm[offset]));
f1 = fm;
if( fourthdiff > maxdiff ) {
maxdiff = fourthdiff;
bisectdim = dim;
}
}
res->bisectdim = bisectdim;
f1 = f++;
Zap(sum);
for( s = first; s <= last; NextSet(s) )
for( n = s->n; n; --n ) {
creal fun = *f1;
f1 += t->ncomp;
for( rul = 0; rul < nrules; ++rul )
sum[rul] += fun*s->weight[rul];
}
/* Search for the null rule, in the linear space spanned by two
successive null rules in our sequence, which gives the greatest
error estimate among all normalized (1-norm) null rules in this
space. */
for( rul = 1; rul < nrules - 1; ++rul ) {
real maxerr = 0;
for( s = first; s <= last; NextSet(s) )
maxerr = Max(maxerr,
fabs(sum[rul + 1] + s->scale[rul]*sum[rul])*s->norm[rul]);
sum[rul] = maxerr;
}
res->avg = vol*sum[0];
res->err = vol*(
(errcoeff[0]*sum[1] <= sum[2] && errcoeff[0]*sum[2] <= sum[3]) ?
errcoeff[1]*sum[1] :
errcoeff[2]*Max(Max(sum[1], sum[2]), sum[3]) );
}
if( VERBOSE > 2 ) {
Vector(char, out, 64*NDIM + 128*NCOMP);
char *oe = out;
count comp;
cchar *msg = "\nRegion (" REALF ") - (" REALF ")";
for( b = region->bounds; b < B; ++b ) {
oe += sprintf(oe, msg, b->lower, b->upper);
msg = "\n (" REALF ") - (" REALF ")";
}
for( res = result, comp = 0; res < Res; ++res )
oe += sprintf(oe, "\n[" COUNT "] "
REAL " +- " REAL, ++comp, res->avg, res->err);
Print(out);
}
