This recommendation is part of the open catalog of best practice rules for performance that is automatically detected and reported by Codee.
Issue
Performance can be increased by using the highest possible trip count in the vectorized loop.
Relevance
Vectorization takes advantage of having as high a trip count (ie. number of iterations) as possible. When loops are perfectly nested and they can be safely interchanged, making the loop with the highest trip count the innermost should increase vectorization performance.
Actions
Interchange loops so that the loop with the highest trip count becomes the innermost loop.
Notes
If the loop interchange introduces non-sequential memory accesses, the runtime can get slower because of the inefficient memory access pattern.
Code example
The following code shows two nested loops, where the outer one has a larger trip count than the inner one:
for (int i = 0; i < n; i++) {
for (int j = margin; j < n - margin; j++) {
bb[i][j] = 0.0;
for (int k = -margin; k < margin; k++) {
bb[i][j] += aa[i][j + k];
}
}
}
The value of margin
is not known at compile time, but it is typically low. We can increase the loop trip count of the innermost loop by performing loop interchange. To do loop interchange, the loop over j
and the loop over k
need to be perfectly nested. We can make them perfectly nested by moving the initialization bb[i][j] = 0.0
into a separate loop.
for (int i = 0; i < n; i++) {
for (int j = margin; j < n - margin; j++) {
bb[i][j] = 0.0;
}
for (int k = -margin; k < margin; k++) {
for (int j = margin; j < LEN_2D - margin; j++) {
bb[i][j] += aa[i][j + k];
}
}
}

Building performance into the code from day one with Codee