Question

I have a really very basic doubt regarding STL containers.
My requirement is that i want to store double values in the form of multi-dimensional array. I will be performing various algebraic operations directly on them i.e.

myvector[4] = myvector[3] - 2 * myvector[2];

for this I am itterating using for loops & using the [] operator. I am not using STL itterator’s. I found 2 basic approaches here.
I prefer speed over memory efficiency. Since I am accessing these variables frequently I think vector would be slow for me.
So what is your humble opinion on this matter?
I know that the answers would be based on your previous experience, that is why I am asking this question. I am sorry if this question is too basic to be discussed here.

Answer 1

The link you gave listed 2 methods, which creates “real” 2d arrays. In general, 2d arrays are not that efficient, because they require a lot of allocations. Instead, you can use a faked 2d array:

// Array of length L and width W
type* array1 = new type[L * W]; // raw pointers
std::vector<type> array2(L * W); // STL Vector

// Accessing a value. You have to use a convention for indices, and follow it.
// Here the convention is: lines are contiguous (index = x + y * W)
type value = array[x + y * W]; // raw pointer array & vector

Here is a simple benchmark (windows only, except if you change the timer part):

#include <vector>
#include <ctime>
#include <iostream>
#include <stdlib.h>

#include <Windows.h>
typedef LARGE_INTEGER clock_int;

void start_timer(clock_int& v)
{
    QueryPerformanceCounter(&v);
}

void end_timer(clock_int v, const char* str)
{
    clock_int e;
    QueryPerformanceCounter(&e);
    clock_int freq;
    QueryPerformanceFrequency(&freq);
    std::cout << str << 1000.0 * ((double)(e.QuadPart-v.QuadPart) / freq.QuadPart) << " ms\n";
}

void test_2d_vector(unsigned int w, unsigned int h)
{
    std::vector<std::vector<double> > a;
    a.resize(h);
    for(unsigned int t = 0; t < h; t++)
        a[t].resize(w);

    clock_int clock;
    start_timer(clock);
    // Benchmark random write access
    for(unsigned int t = 0; t < w * h; t++)
        a[rand() % h][rand() % w] = 0.0f;
    end_timer(clock,"[2D] Random write (STL) : ");

    start_timer(clock);
    // Benchmark contiguous write access
    for(unsigned int y = 0; y < h; y++)
        for(unsigned int x = 0; x < w; x++)
            a[y][x] = 0.0f;
    end_timer(clock,"[2D] Contiguous write (STL) : ");
}

void test_2d_raw(unsigned int w, unsigned int h)
{
    double** a = new double*[h];
    for(unsigned int t = 0; t < h; t++)
        a[t] = new double[w];

    clock_int clock;
    start_timer(clock);
    // Benchmark random write access
    for(unsigned int t = 0; t < w * h; t++)
        a[rand() % h][rand() % w] = 0.0f;
    end_timer(clock,"[2D] Random write (RAW) : ");

    start_timer(clock);
    // Benchmark contiguous write access
    for(unsigned int y = 0; y < h; y++)
        for(unsigned int x = 0; x < w; x++)
            a[y][x] = 0.0f;
    end_timer(clock,"[2D] Contiguous write (RAW) : ");
}

void test_1d_raw(unsigned int w, unsigned int h)
{
    double* a = new double[h * w];

    clock_int clock;
    start_timer(clock);
    // Benchmark random write access
    for(unsigned int t = 0; t < w * h; t++)
        a[(rand() % h) * w + (rand() % w)] = 0.0f;
    end_timer(clock,"[1D] Random write (RAW) : ");

    start_timer(clock);
    // Benchmark contiguous write access
    for(unsigned int y = 0; y < h; y++)
        for(unsigned int x = 0; x < w; x++)
            a[x + y * w] = 0.0f;
    end_timer(clock,"[1D] Contiguous write (RAW) : ");
}

void test_1d_vector(unsigned int w, unsigned int h)
{
    std::vector<double> a(h * w);

    clock_int clock;
    start_timer(clock);
    // Benchmark random write access
    for(unsigned int t = 0; t < w * h; t++)
        a[(rand() % h) * w + (rand() % w)] = 0.0f;
    end_timer(clock,"[1D] Random write (STL) : ");

    start_timer(clock);
    // Benchmark contiguous write access
    for(unsigned int y = 0; y < h; y++)
        for(unsigned int x = 0; x < w; x++)
            a[x + y * w] = 0.0f;
    end_timer(clock,"[1D] Contiguous write (STL) : ");
}

int main()
{
    int w=1000,h=1000;
    test_2d_vector(w,h);
    test_2d_raw(w,h);
    test_1d_vector(w,h);
    test_1d_raw(w,h);
    system("pause");
    return 0;
}

Compiled with msvc2010, release /Ox /Ot, it outputs for me (Win7 x64, Intel Core i7 2600K):

[2D] Random write (STL) : 32.3436 ms
[2D] Contiguous write (STL) : 0.480035 ms
[2D] Random write (RAW) : 32.3477 ms
[2D] Contiguous write (RAW) : 0.688771 ms
[1D] Random write (STL) : 32.1296 ms
[1D] Contiguous write (STL) : 0.23534 ms
[1D] Random write (RAW) : 32.883 ms
[1D] Contiguous write (RAW) : 0.220138 ms

You can see the STL is equivalent to raw pointers. But 1D is much faster than 2D.