Vectorization
- C++ course
- Expert
- Lesson 27/47
- 5 min
Learn automatic parallelization with SIMD vectorization.
Vectorization
Vectorization is the compiler turning a loop into SIMD (Single Instruction Multiple Data) instructions. A single instruction processes multiple array elements; this gives a speedup.
On modern CPUs SSE/AVX registers process 4-16 ints at once. The compiler auto-vectorizes with -O3. Obstacles: pointer aliasing (two pointers may point to the same memory), conditions inside the loop, complex dependencies. The __restrict__ keyword says there is no aliasing.
Helping vectorization
For the compiler to vectorize a loop: the loop must be simple (single block, no condition), the trip count must be known in advance (constant or n), and pointers must be disjoint (no aliasing). Writing int* __restrict__ a declares no aliasing.
Example: void add(int* __restrict__ a, int* b, int n) { for (int i = 0; i < n; ++i) a[i] += b[i]; }. Here a and b are disjoint, so the compiler can safely process each step with SIMD. Without vectorization each step is processed one by one; that is slower.
Practice Write a vectorizable sum
Write int sum(int* a, int n). Body: int s = 0; for (int i = 0; i < n; ++i) s += a[i]; return s;. In main define int a[4] = {1, 2, 3, 4}; and write std::cout << sum(a, 4);.
Starter code
#include <iostream>
// Define the sum function here
int main() {
// Define the a array, call sum and print it
return 0;
}Sample solution
#include <iostream>
int sum(int* a, int n) {
int s = 0;
for (int i = 0; i < n; ++i) s += a[i];
return s;
}
int main() {
int a[4] = {1, 2, 3, 4};
std::cout << sum(a, 4);
return 0;
}Expected output
10
Questions and answers
Review the lesson questions together with their answers. To test yourself, solve the lesson interactively.
1 Match each piece to its role
| Term | Meaning |
|---|---|
SIMD |
Single instruction multiple data |
#pragma omp simd |
OpenMP vectorization |
__restrict__ |
Pointer disjointness |
-O3 |
Vectorization flag |
2 Complete the restrict keyword
void add(int* ________ a, int* b, int n) { for (int i = 0; i < n; ++i) a[i] += b[i]; }void add(int* __restrict__ a, int* b, int n) { for (int i = 0; i < n; ++i) a[i] += b[i]; }Explanation: __restrict__ says there is no aliasing.
3 What optimization does -O3 enable?
g++ -O3 main.cpp- Nothing
- Only warnings
- Aggressive vectorization and loop unrolling Correct answer
- The same thing
Explanation: -O3 adds aggressive vectorization and loop unrolling on top of -O2.