CODEWITHSUNDEEP
c++optimizationmemory-managementvectorstruct
ParoX
ParoX

Reputation: 5941

Arithmetically getting index in 3D vector

I have a 3d vector of structs. I refer to where the struct is located in the 3d vector by the struct's i j and k. However, when making millions of these, it takes up a bunch of memory because the object stores so much data.

How can I efficiently figure out what i,j,k a particular struct object is without storing that information in the struct itself. Can I do it with some sort of memory arithmetic?

#include <iostream>
#include <vector>
#include <string>

using namespace std;

int main() {
    struct MyStruct {
        size_t i;
        size_t j;
        size_t k;
        bool some_bool;
    };

    vector<vector<vector<MyStruct>>> three_d_struct_v;

    size_t max_i = 1000000;
    size_t max_j = 10;
    size_t max_k = 10;

    for(size_t i = 0; i < max_i; i++) {
        for(size_t j = 0; j < max_j; j++) {
            for(size_t k = 0; k < max_k; k++) {
                three_d_struct_v.emplace_back(MyStruct{i,j,k,false});
            }
        }
    }


    return 0;
}

Upvotes: 0

Views: 1084

Answers (3)

patros
patros

Reputation: 7819

In your case, you can actually figure this out fairly easily by storing a minimal amount of metadata.

Since you have two relatively small vectors, it's possible to store the starting position of all j/k combinations.

size_t max_i = 1000000;
size_t max_j = 10;
size_t max_k = 10;

You would want to restructure your vectors to be stored [k][j][i]. If you store the 100 possible j/k combinations in a std::map, then you can find the j/k values by finding the largest address smaller than the the address of your vector. From there, you calculate the offset in bytes and divide by the size of your struct to figure out i.

It becomes far less practical if max_j and max_k become large.

Upvotes: 0

cmaster - reinstate monica
cmaster - reinstate monica

Reputation: 40665

There is quite a simple way to do this with real arrays. Multilevel std::vector<> won't do, because the memory allocated by all the different line vectors is not contiguous. But with the builtin arrays of the language, this is quite simple:

//Get the memory
bool myData* = new bool[max_i*max_j*max_k];
inline size_t getIndex(size_t i, size_t j, size_t k) { return (i*max_j + j)*max_k + k; }
inline size_t getI(size_t index) { return index/max_j/max_k; }
inline size_t getJ(size_t index) { return (index/max_k)%max_j; }
inline size_t getK(size_t index) { return index%max_k; }

Now you can talk about the indices much in the same way you could talk about pointers to your structs. If you really must do it the C++ way, you can convert references and indices like this:

bool& referenceToElement = myData[anIndex];
size_t recoveredIndex = &referenceToElement - myData;

However, in C you can do much better:

bool (*myData)[max_j][max_k] = malloc(max_i*sizeof(*myData));
myData[i][j][k] = true;    //True 3D array access!

The calculation performed by myData[i][j][k] is precisely the same as the calculation of myData[getIndex(i, j, k)] in the C++ example above. And, as before, you can retrieve the index using pointer arithmetic.

C++ also has multidimensional arrays, but it requires the array dimensions to be compile time constants (and you need to use new instead of malloc()). In C, there is no such restriction, the array sizes may be calculated at runtime.

Upvotes: 1

Matt
Matt

Reputation: 20796

Is this what you are looking for? You can consider m to be an index of all max_i * max_j * max_k structs.

This is untested. You may have to do some casting when calling div on size_t types.

#include <cstdlib> // divmod

size_t max_i = 1000000;
size_t max_j = 10;
size_t max_k = 10;

size_t N = max_i * max_j * max_k; // beware of overflow

for( size_t m=0 ; m<N ; ++m )
{
    div_t q = div( m, max_k );
    size_t k = q.rem;

    q = div( q.quot, max_j );
    size_t j = q.rem;

    q = div( q.quot, max_i );
    size_t i = q.rem;

    // Now i, j, k are set. Do as thou shall.
}

Upvotes: 0

Related Questions