An explanation

Posted by Astryl on July 28, 2012, 3:47 a.m.

First things first, the last <n>(?) blogs I wrote were written under the

strain of increased fatigue. Figured that out when I nearly blacked out

yesterday. I've been sleeping a lot now, in order to catch up (1-2 hours of

sleep per 24 hours doesn't cut it).

Completely broke my sleeping pattern, trying to get as much done as I could

before I have to start work. :3

Secondly, let's get to what this blog is about: Decisions.

Rob asked a good question in my one blog earlier; one which I failed to

answer coherently. So here we go. Let me try this now that I've had at least ten hours of solid sleep.

Why std::vector instead of std::list?

No reason, they're interchangeable for the most part. I should have

used std::list from the start, but I'm in the bad habit of automatically

using vector.

Linear traversal of a list is definitely faster, whereas vector is better

suited to random access. Of course, in Exile, the only thing I'm really

using Containers for would be the Asset manager.

Before the game can start, all resources are loaded into their respective

vectors, usually with a corresponding nametable.

The Assets class provides the GetImage("name") and GetImageByID(9) static

member functions. These use the nametable and the corresponding index in

the resource table to return a pointer to the image, or throws an exception

if that image was not loaded.

If I was to rework the system, I would have created a pair-list class, with

a basic hash table. And I probably would've done it from scratch.

Of course, I wasn't expecting to get this far with Exile, nor the framework

underlying it; so my all encompassing excuse for the use of vectors where

lists would do better is this: Prototype evolution.

Why my object vectors are insert-sorted

This only takes effect in the 2D mode of the framework, and was only in use

for Exile. It's useless in 3D.

The best way to describe this is as a cheap depth sort.

Of course, a faster method would be to create my own doubly-linked list

class, and use recursion to add new objects.

Then call a recursive function to draw all elements from the tail of the

list.

Things I'm doing now

I created a rather extensive framework thanks to this competition, and I'm

definitely going to be shelving it for further use.

Already it's proven itself useful beyond a doubt, as I started creating

Arbiter with it. And then I moved on to Exile, using the exact same

framework.

There are a lot of optimizations and general rewrites that need to be

performed on the framework, granted. I have marked several locations within

the code that need optimization, mathematical attention, and re-design; but

the 'engine' and the games I created with it work for now.

I'm planning on releasing the cleaned and optimized code to Exile within

the next month or so. I still need to go back and resolve the problems I

'brute-forced' because of the time constraints.

Well, back to whatever it is I was going to do today. Read a book perhaps.

Comments

Rob 12 years, 4 months ago

Quote:
Linear traversal of a list is definitely faster

Whaaaat? No…. it's insertion/deletion from the middle of a list where it's (significantly) faster than vectors…

How would lists be faster for traversal than vectors? Non-continguous+extra structures vs raw contiguous memory implementation.

Are you still just storing everything in a level in a single linear structure though? That seems like it could be awfully slow, especially with large levels, compared to grid-based or even quad-tree based (or kd-trees or whatever the fuck you want to use). Do you iterate through all the things and draw them, regardless of it they're off-screen? For collision detection, do you have a O(n) efficiency since you're pretty much going through every one, rather than only ones that could possibly intersect? What kind of things do you do to speed up collision detection? I'm kinda curious as to what other people do.

JuurianChi 12 years, 4 months ago

Quote:
Read a book perhaps.

I read. What is it that your reading?

Rob 12 years, 3 months ago

Also:

Quote: Mega

Of course, a faster method would be to create my own doubly-linked list class, and use recursion to add new objects. Then call a recursive function to draw all elements from the tail of the list.

Can't this be done iteratively, thus saving the overhead of recursion? (ie function call instructions + stack rape) Why couldn't you do it just how you were doing it with vectors?

Object *obj = new Object(depth or whatever);
for (Node *it = objects.head; it != NULL; it = it->getNext())
{
     if (it->getNext() == NULL || it->getNext()->getObject()->getDepth() > obj->getDepth())
     {
          objects.insert(it, obj); // Assuming insert after iterator
          break;
     }
}

or (for STL lists)

Object *obj = new Object(depth or whatever);
for (std::list<Object*> it = objects.begin(); ; it++)
{
     if (it == objects.end() || (*it)->getDepth() > obj->getDepth())
     {
          objects.insert(it, obj);
          break;
     }
}

or something

Astryl 12 years, 3 months ago

Wow. The line breaks on this one were really stuffed up. I'm using NP++ next time.

@Rob: It's marginally faster. According to the algorithm, they're supposed to be equal, yes. The assembly listings show otherwise.

As for the levels, I just remembered why I used vectors instead of lists. The random access speeds.

Before even entering a loop, I use the current 'view' coordinate to designate a rectangle of indices into the level 'array' (Actually, I used an array later; replaced vectors).

Depending on the resolution of the view, the render time remains constant.

With the levels, in both Exile and Arbiter, the tile data includes a collision field. This is used for basic entity->level collision detection, and in this case, only a tiny rectangle of tiles needs to be checked by the game (Usually player.x-1, player.x+1, player.y-1,player.y+1 to get an encompassing rectangle).

For entity->entity collision I first use an 'in view' test. Then I use a simple distance-squared check to see if two objects are in range. If they are, a simple AABB collision test is performed.

The collision is managed by the RoomManager class in my game, which basically iterates through the Object list, remove 'dead' objects, calls the object->Step() functions, and use double-dispatch on collision between two entities.

Also, in most cases with a no-argument function and recursion, stack overhead is reduced by the compiler optimizer. The difference is negligible on any piece of computer hardware better than an old P2.

Quote:
I read. What is it that your reading?
Tolkien. Robert Jordan. Selected fan fiction. Jules Vern.

EDIT:

The invasion haz begun…

ludamad 12 years, 3 months ago

Quote:
Whaaaat? No…. it's insertion/deletion from the middle of a list where it's (significantly) faster than vectors…

This statement is useless without measurements. Actually, inserting into an std::vector tends to be very very fast due to good cache behaviour.

Please measure your use case with 1000s of items for inserting into std::vector vs whatever. You may very well be surprised.

Rob 12 years, 3 months ago

@Luda

I tried testing it with lots of iterations and the vectors just kept crashing after 0x3FFFF. And at 0x3FFFF it was <1 second for the list, vs 9 seconds for the vector for insertion in the middle of the container. I tried using reserve() to lower than 9 seconds, but it would just crash…

Either way, even when they weren't crashing, std::vectors were slower than std::lists for insertion in the middle.

Also, max_size() was returning 1,073,741,823. And 262,143(what it was crashing when I went over) is a hell of a lot smaller than 1,073,741,823. I always thought vectors used a 32bit uint though, not a 30 bit one…

@Mega

Quote:
@Rob: It's marginally faster. According to the algorithm, they're supposed to be equal, yes. The assembly listings show otherwise.

Wait, lists are faster for traversal?! How? Please explain.

Astryl 12 years, 3 months ago

When iterating you're merely jumping to different addresses. Of course, the [] operator for vector is faster, though access times become equal (In number of operations, so be appeased, ludamad).

Of course, when using assembly operators you have to factor in the number of CPU cycles per operation (Which varies per operation).

My previous statement about it being marginally faster is inaccurate. They're equal. If you want the basic figures, with a vector, chunks of memory are allocated linearly, and traversal is performed using basic pointer arithmetic (I'd say sizeof(stored object), but that's not true with classes that have virtual methods). With a list, the allocation takes place wherever it can (Sometimes in chunks, depending on the compiler implementation), and of course, pointers are used to get to the next/previous element in the list. So by what I can tell, linear access time for reading/writing.

Of course, inserting is a different story. Vector can merely find an empty chunk with enough space.

It gets bitten in the ass when it runs out of space, and has to allocate a new chunk, and copies the old chunk across to the new space.

So list is definitely faster in that respect.

Either way, guess what? I looked at Exile's code again, and guess what I was using all along? std::list. [/facePalm]

ludamad 12 years, 3 months ago

@Rob:

This is more what i had in mind:

Example:

#include <iostream>
#include <list>
#include <ctime>
#include <vector>
using namespace std;

const int testamnt = 100000;
void vector_test() {
    std::vector<int> test;
    for (int i = 0; i < testamnt; i++) {
        std::vector<int>::iterator it = test.begin();
        for (int z = 0; z < i / 2; z++) {
            ++it;
        }
        test.insert(it, 2);
    }
}
void list_test() {
    std::list<int> test;
    for (int i = 0; i < testamnt; i++) {
        std::list<int>::iterator it = test.begin();
        for (int z = 0; z < i / 2; z++) {
            ++it;
        }
        test.insert(it, 2);
    }
}

int main() {
    clock_t c = clock();
    list_test();
//  vector_test();
    cout << double(clock() - c) / CLOCKS_PER_SEC << endl; // prints !!!Hello World!!!
    return 0;
}

11.1s with vector, 15s with list. Without traversal for vector, it is ~ 3s. Insert isn't much extra cost if you have to traverse a list anyway.

Rob 12 years, 3 months ago

But wouldn't you be traversing it anyways if it's an entity list that you're updating every frame already? Traversing the entire thing every single time seems kinda unfair against the list, since I don't think it really applies here.

ludamad 12 years, 3 months ago

Fair enough, but marking all the entries for deletion with a vector and compressing it in one go will still be faster.