Tagging @mhasself here, in case you have opinions on this implementation, since you wrote the G3VectorInt implementation this is copied from.

I haven't profiled this -- @mhasself do you know how efficient the bit_count function is?

I haven't profiled this -- @mhasself do you know how efficient the bit_count function is?

For the record, no I do not.

But it's probably like, 3 assembly instructions (in the loop), and seems like the sort of thing that would pipeline pretty well (despite the ifs). There might be a faster version where you accumulate the +ve and -ve data into separate masks and combine them at the end. An interesting question for a less busy day.

Against my better schedule management ... here is a (relative) profiling. The bit_count function takes 10 ns / sample on my laptop. Tests done with 1k to 1M elements. It seems to take about 10% more time than this simple accumulation function:

int not_bit_count_v2(const std::vector<int64_t> &d) {
    int sum = 0;
    for (auto c: d)
        sum += c;
    return sum;
}

So I think it's close enough to maxed out.

Cool, good to know!