I'd like to improve the performance of PickPotatoes
in the below code by using move instead of copy, but I can't figure out how to do that with insert
and a boost::variant
. In my actual use case, parsing the data takes about 75% of the time, and the real version of PickPotatoes takes about 25%, due to some slow copies. By improving PickPotatoes
I should be able to get that down. Is it possible to move something out of a boost::variant
and improve PickPotatoes
?
#include <map>
#include "boost/variant.hpp"
#include <string>
#include <vector>
#include <functional>
struct tuber
{
int z;
std::vector<double> r;
};
int getZ(const tuber& t)
{
return t.z;
}
boost::variant<std::string, tuber> GrowPotato()
{
int z = std::rand() / (RAND_MAX / 10);
if (z < 2)
{
return "BAD POTATO";
}
else
{
tuber ret;
ret.z = z;
ret.r.resize(10000);
for (int i = 0;i < 10000;++i)
{
ret.r[i] = std::rand() / (RAND_MAX / 50);
}
return ret;
}
}
std::vector<boost::variant<std::string,tuber>> GrowPotatoes(int n)
{
std::vector<boost::variant<std::string, tuber>> ret;
ret.resize(n);
for (int i = 0; i < n; ++i)
{
ret[i] = GrowPotato();
}
return ret;
}
//could make this more efficient.
std::pair<std::vector<std::string>,std::multimap<int, tuber>> PickPotatoes(std::vector <boost::variant<std::string, tuber>> result)
{
std::pair<std::vector<std::string>,std::multimap<int,tuber>> ret;
int numTypTwo = 0;
for (const auto& item : result)
{
numTypTwo += item.which();
}
ret.first.resize(result.size() - numTypTwo);
int fstSpot = 0;
for (int i = 0; i < result.size();++i)
{
if (result[i].which())
{
ret.second.insert(std::pair<int, tuber>(getZ(boost::get<tuber>(result[i])), boost::get<tuber>(result[i])));
}
else
{
ret.first[fstSpot++] = std::move(boost::get<std::string>(result[i]));
}
}
return ret;
}
int main()
{
std::srand(0);
std::vector<boost::variant<std::string, tuber>> q= GrowPotatoes(5000);
std::pair<std::vector<std::string>, std::multimap<int, tuber>> z = PickPotatoes(q);
return 0;
}
The simplest win would be to move the parameter value:
std::pair<std::vector<std::string>, std::multimap<int, tuber>> z = PickPotatoes(std::move(q));
Indeed, it wins 14% of performance, roughly on my benchmarks. The rest heavily depends on what it all means, how it's going to be used.
Focus on reducing allocations (use a non-nodebased container if you can, e.g. boost::flat_multimap
, sort explicitly, use string_view, parse into the desired datastructure instead of intermediate).
I was able to shave off about 30% using:
std::pair<std::vector<std::string>, std::multimap<int, tuber> >
PickPotatoes(std::vector<boost::variant<std::string, tuber> >&& result) {
std::pair<std::vector<std::string>, std::multimap<int, tuber> > ret;
ret.first.reserve(result.size());
struct Vis {
using result_type = void;
void operator()(std::string& s) const {
first.emplace_back(std::move(s));
}
void operator()(tuber& tbr) const {
second.emplace(tbr.z, std::move(tbr));
}
std::vector<std::string>& first;
std::multimap<int, tuber>& second;
} visitor { ret.first, ret.second };
for (auto& element : result) {
boost::apply_visitor(visitor, element);
}
return ret;
}
Using emplace, avoiding repeated get<>
, avoiding the loop to get the first
size etc.