Convert Float Array to 32-bit Integer Array using C++ SIMD

In performance-critical applications such as scientific computing, graphics, game development, or data processing, converting large arrays of floating-point numbers to integers is a common operation. While a basic scalar loop is simple and effective for small arrays, it quickly becomes a bottleneck when dealing with large datasets. Modern CPUs offer SIMD capabilities, which allow you to process multiple data elements in parallel using a single instruction. This can dramatically speed up such conversions.

The simplest way to convert a float array to int32_t is by looping through each element and casting it individually. Here's what that approach looks like:

#include <iostream>
#include <vector>

void floatint32(const float *data, int32_t *result, const size_t n) {
    for (size_t i = 0; i < n; ++i) {
        result[i] = (int32_t) data[i];
    }
}

int main() {
    std::vector<float> a = {
        -1.4, 0, 1.5, -2.6, 3.4, -4.5, 5.6, -6.4, 7.5,
        -8.6, 9.4, -10.5, 11.6, -12.4, 13.5, -14.6, 15.4, -16.5,
    };
    std::vector<int32_t> result(a.size());

    floatint32(a.data(), result.data(), a.size());
    for (auto value: result) {
        std::cout << value << " ";
    }

    return 0;
}

This method is straightforward and portable, but it doesn't leverage the CPU's vectorization features for improved performance.

Below is the optimized version that takes advantage of AVX2 instructions:

#include <immintrin.h>

void floatint32(const float *data, int32_t *result, const size_t n) {
    size_t i = 0;
    for (; i + 8 <= n; i += 8) {
        __m256 va = _mm256_loadu_ps(&data[i]);
        __m256i vresult = _mm256_cvttps_epi32(va);
        _mm256_storeu_si256((__m256i*) &result[i], vresult);
    }

    for (; i < n; ++i) {
        result[i] = (int32_t) data[i];
    }
}

Explanation:

• _mm256_loadu_ps loads 8 floating-point values from the array.
• _mm256_cvttps_epi32 converts 8 floating-point values to 8 signed integers using truncation, not rounding.
• _mm256_storeu_si256 writes these 8 integers into the result array.

The scalar fallback loop handles any remaining elements at the end when the total count isn't a multiple of 8.