I'm implementing a quadtree for the spacial partitioning of a very large group of objects (target is 400k and trying to get it as fast as I can (target is no more than 50ms for all 400k). My initial naive approach was only able to handle about 1,000 objects in 27 seconds, but I've since been able to get it down to 10,000 objects in 113ms mostly by paying close attention to unnecessary move, copy, and temporary operations. I'm comfortably certain that there are no more such operations that can be eliminated.

The quadtree is a structure with two std::unordered_set: the first contains the id#'s of all objects inside the tree node aka "owned" objects, and the second contains the first plus any object close enough to the node's border to be included aka "known" objects. It also has a bounding rectangle defined by its top-left and bottom-right points (can be instantiated Rect(x1,y1,x2,y1) to avoid unecessary Point() constructors).

At every update, every known object is inserted into the quadtree. The quadtree is updated with quadtree::insert(const Point& where,const uint64_t& which).
--The method checks first if the quadtree has any children AND if it's depth is < a maximum AND if it has > max objecs within it. If that's true, it creates 4 children and dumps both of it's id containers into the children.
--The method then checks if it has any children - if so, call quadtree::insert for each child. If not, then check if the bounding rect contains "where"
----If the bounding rect contains "where" insert "which" into both containers
----If not, check if the bounding rect collides with a rect formed around where Rect(where.x-width/2,where.y-width/2,where.x+width/2,where.y+width/2). If so, add it to just the "known" container.

Unfortunately, after profiling 10 iterations of 10,000 randomly distributed objects, this leads to about ~408,000 calls to quadtree::insert (~37% execution time), 3,570,000 rect constructors (~6.6%), 7,256,000 point constructors (~10.38%), 3,586,000 stl emplacements (~13.21%).

The only really meaningful way to further optimize this that I can see would be to find a less naive way of splitting the quadtree, so that the top node needn't dump every "known" object into every child, which is what is dramatically multiplying the number of constructors. Any advice in this regard would be seriously appreciated.

Barely Annotated Code:
[source lang="cpp"]
bool LogicTree::insert(const QuadPoint& where,const EntityID& which){
bool within = false;

//check if split needed
if(children.size()==0 && depth < MAX_DEPTH && owned.size()>MAX_ENTITIES){
Coord width = bounds.rightBottom.x-bounds.leftTop.x,height = bounds.rightBottom.y-bounds.leftTop.y;
QuadRect c1(bounds.leftTop.x,bounds.leftTop.y,bounds.leftTop.x+width/2,bounds.leftTop.y+height/2),
c2(bounds.leftTop.x+width/2,bounds.leftTop.y,bounds.leftTop.x+width,bounds.leftTop.y+height/2),
c3(bounds.leftTop.x,bounds.leftTop.y+height/2,bounds.leftTop.x+width/2,bounds.leftTop.y+height),
c4(bounds.leftTop.x+width/2,bounds.leftTop.y+height/2,bounds.rightBottom.x,bounds.rightBottom.y);
children.emplace_back(new LogicTree(c1,depth+1));
children.emplace_back(new LogicTree(c2,depth+1));
children.emplace_back(new LogicTree(c3,depth+1));
children.emplace_back(new LogicTree(c4,depth+1));
for(auto& child : children){
massInsertChild(child);
}
}

//Pass to children or insert
if(children.size()>0){
for(auto& child : children){
if(child->insert(where,which)){
within = true;
}
}
}else{
if(bounds.contains(where)){
owned.emplace(which);
known.emplace(which);
within = true;
}else if(bounds.intersects(QuadRect(where,MAX_RANGE))){
known.emplace(which);
}
}
return within;
}
[/source]P.S>If this comes out as a solid block of text, my appologies - for some reason this forum likes to delete my newline characters.

Edit: Attempt #1 to fix the whitespace nuke. Also fixed a mistake (400k not 3mil, lol).

... new LogicTree ...

A quick, drive by, guess of possible optimization opportunities: be sure allocation uses dedicated pools (the more spatial awareness you can cram into the implementation the better - cache is a bitch too, especially with trees).

Had relatively recently a case where using a pool made an over 100 TIMES (or was it over 1000?) speed difference although allocations were relatively rare compared to work done within the allocated objects - reason being: heap degrades to a standstill extremely fast if one does not use it sparingly/sensibly (LogicTree sounds like massive heap abuse).

edit:

"target is no more than 50ms for all 400k" ... i would say you have about 500 cycles to work with per insert. Way more than needed for the instructions themselves, EXCEPT - one cache miss penalty per one memory access alone can be up to 200-500 (*). Without being cache friendly i do not think your goal is achievable. Other than that, totally doable.

*) Can't remember any typical numbers for x86 of the top of my head, random estimates.

Actually reading from main memory is very slow (~30-100, depends on current clock speed etc)

Cache miss itself (4-60, depending on how many cache levels missed).

TLB miss (15-300, depending on how much of the relevant structures are in cache + how OS implements it, can easily be much slower)