static void updateLineX(HOOD_NEW& hoodNew, int indexEnd, HOOD_OLD& hoodOld, unsigned /* nanoStep */)
{
for (int i = hoodOld.index(); i < indexEnd; ++i, ++hoodOld) {
ShortVec tmp;
tmp.load_aligned(&hoodNew->sum() + i * C);
for (const auto& j: hoodOld.weights(0)) {
ShortVec weights;
ShortVec values;
weights.load_aligned(j.second());
values.gather(&hoodOld->value(), j.first());
tmp += values * weights;
}
tmp.store_aligned(&hoodNew->sum() + i * C);
}
}
static void updateLineX(HOOD_NEW& hoodNew, int indexEnd, HOOD_OLD& hoodOld, unsigned /* nanoStep */)
{
for (int i = hoodOld.index(); i < indexEnd; ++i, ++hoodOld) {
ShortVec tmp;
tmp.load_aligned(&hoodNew->sum() + i * C);
for (const auto& j: hoodOld.weights(0)) {
ShortVec weights;
ShortVec values;
weights.load_aligned(j.second());
// fixme: is this gahter actually correct? shouldn't we use offset 0 for the gather? see also hpxperformancetests/main.cpp
values.gather(&hoodOld->value(), j.first());
tmp += values * weights;
}
tmp.store_aligned(&hoodNew->sum() + i * C);
}
}