Conditional Sum

Sum all elements in an array that are strictly less than a given threshold, using AVX2 masking.

This is a classic SIMD predication problem: instead of branching on each element, you compute both paths and use a mask to select which results contribute to the final sum.

Function Signature

#include <immintrin.h>
#include <cstdint>

int32_t conditional_sum(const int32_t* arr, int n, int32_t threshold);

Parameters:

• arr — pointer to an array of n signed 32-bit integers, guaranteed 32-byte aligned
• n — number of elements, guaranteed to be a multiple of 8 and at least 8
• threshold — sum only elements strictly less than this value

Returns: the sum of all elements where arr[i] < threshold (guaranteed to fit in int32_t)

Example

Input:  arr = [1, 5, 3, 8, 2, 7, 4, 6], threshold = 5
Output: 10  (1 + 3 + 2 + 4)

Constraints

• 8 ≤ n ≤ 1,000,000
• n is always a multiple of 8
• All elements are in range [-1000, 1000]
• The conditional sum fits in a 32-bit signed integer
• arr is 32-byte aligned

Notes

AVX2 has no cmplt instruction — use _mm256_cmpgt_epi32 with the arguments swapped: cmpgt(threshold, x) gives you a mask where x < threshold.

When one branch of the predication is zero (as it is here — you either add the element or add nothing), you can use bitwise AND with the mask instead of a blend. This saves a cycle on most microarchitectures.

Useful Intrinsics