c++ optimization of image resize array - c++

Question

Trying to increase the width of an image array to return to an opencv mat. The problem is speed when the temp_mat array needs to be shifted by a certain amount as the image increases in size. See function below:
This line will run with good speed:
//temp_mat[height][width] = in_mat[i][j];
But the speed decreases by a lot when changed to:
temp_mat[height][width + int(((width - middle_point) * -1) * FLOAT_HERE)] = in_mat[i][j];
The loop takes many milliseconds longer to run. Here is the complete function, variable names have been changed.
#define D_HEIGHT 1000
#define D_WIDTH 1200
int DEFAULT_HEIGHT = 1000;
int DEFAULT_WIDTH = 1200;
float FLOAT_HERE = .04;
static int temp_mat[D_HEIGHT][D_WIDTH];
cv::Mat get_mat(int in_mat[D_HEIGHT][300]){
int height = 0;
int width = 0;
int middle_point = DEFAULT_WIDTH/2;
for(int i=0;i < DEFAULT_HEIGHT;i++){
width = 0;
for(int j =0;j < DEFAULT_WIDTH / 4;j++){
for(int il = 0; il < DEFAULT_WIDTH / (DEFAULT_WIDTH/4); il++){
//This is to slow, but what I need
temp_mat[height][width + int(((width - middle_point) * -1) * FLOAT_HERE)] = in_mat[i][j];
//This is ok
//temp_mat[height][width] = in_mat[i][j];
width++;
}
}
height++;
}
return cv::Mat(D_HEIGHT,D_WIDTH,CV_8UC4,temp_mat);
}
Any ideas to make it faster are welcome. I am hoping to avoid a new thread.

Answer 1

You are doing that wrong just use Affine Transformation and OpenCV will do this in fastest possible way.

Answer 2

Even though DEFAULT_WIDTH is not declared const it appears to be used as a constant, and the naming of the variable suggests it as well. You should probably make it constant, even though that in it self will not improve performance. I say this because you are calculating a middle_point that is then also constant, and can be pre calculated. The same goes for the FLOAT_HERE, which also appears to be constant.
Having made those constant the only variable in the calculation, which you make multiple times is the width variable. Since you are always looping the same number of iterations, you might consider pre-calculating the different values, simply creating a cache of values instead of calculating on the fly.
For each value of width you can create a corresponding calculated value, you can store this in an array where the index is the width, and the value is what is calculated:
int width_cache[DEFAULT_WIDTH];
...
for (int i = 0; i < DEFAULT_WIDTH; ++i) {
width_cache[i] = i + int(((i - middle_point) * -1) * FLOAT_HERE);
}
In your loop, you could then do:
temp_mat[height][width_cache[width]] = in_mat[i][j];