Speed up For loop with .push_back - c++

Question

I'm wondering if we can speed up this loop using OpenMP or CUDA. Currently, it runs fine with sequential processing but I"m trying to optimize my coding:
for (int curCol = 0; curCol < numRows; ++curCol){ //Long Loop
int lb = csc_colIndices[curCol];
int ub = csc_colIndices[curCol + 1];
// push back the diagonal value to L matrix
vec_L_val[curCol].push_back(1.0f);
vec_L_indices[curCol].push_back(curCol);
for (int curIndex = lb; curIndex < ub; ++curIndex){
int curRow = csc_indices[curIndex];
float curVal = csc_val[curIndex];
if (!Equal(curVal, 0) && curRow <= curCol){// U entry
vec_U_val[curCol].push_back(curVal);
vec_U_indices[curCol].push_back(curRow);
}
else if (!Equal(curVal, 0) && curRow > curCol){// L entry
vec_L_val[curCol].push_back(curVal);
vec_L_indices[curCol].push_back(curRow);
}
}
}
In an effort to parallelise the processing I've tried the following with no effect:
#pragma omp parallel for private(curCol) shared(curIndex)
My suspicion is the use of .push_back but I can be wrong...
How can I improve this code?

Answer 1

first prealloc the whole thing:
for (int curCol = 0; curCol < numRows; ++curCol)
{
vec_L_val[curCol].resize( SIZE_OF_THE_INNER_VECTOR );
vec_L_indices[curCol].resize( SIZE_OF_THE_INNER_VECTOR );
vec_U_val[curCol].resize(SIZE_OF_THE_INNER_VECTOR )
vec_U_indices[curCol].resize(SIZE_OF_INNER_VECTOR )
}
Then your inner loops would probably work faster, as there would not be a need for reallocs inside of the internal vector structures.
for (int curCol = 0; curCol < numRows; ++curCol){ //Long Loop
int lb = csc_colIndices[curCol];
int ub = csc_colIndices[curCol + 1];
// push back the diagonal value to L matrix
vec_L_val[curCol].push_back(1.0f);
vec_L_indices[curCol].push_back(curCol);
for (int curIndex = lb; curIndex < ub; ++curIndex){
int curRow = csc_indices[curIndex];
float curVal = csc_val[curIndex];
if (!Equal(curVal, 0) && curRow <= curCol){// U entry
vec_U_val[curCol].push_back(curVal);
vec_U_indices[curCol].push_back(curRow);
}
else if (!Equal(curVal, 0) && curRow > curCol){// L entry
vec_L_val[curCol].push_back(curVal);
vec_L_indices[curCol].push_back(curRow);
}
}
}