如何用x86 Intrinsic快速交错两个__mm256浮点向量？

Given your __mm256 vectors of separate real (a) and imaginary (b) components, the most efficient AVX2 implementation to produce the interleaved vectors c and d leverages 256-bit unpack instructions combined with lane extraction—this minimizes operations compared to splitting into 128-bit vectors first.

Optimal Code Implementation #

#include <immintrin.h>

// Note: _mm256_set_ps takes values in reverse order (highest index first)
__mm256 a = _mm256_set_ps(7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 0.0f);
__mm256 b = _mm256_set_ps(17.0f, 16.0f, 15.0f, 14.0f, 13.0f, 12.0f, 11.0f, 10.0f);

// Step 1: Unpack low and high pairs within each 128-bit lane
__mm256 temp_low = _mm256_unpacklo_ps(a, b); // Lower lane: [a0,b0,a1,b1], Upper lane: [a4,b4,a5,b5]
__mm256 temp_high = _mm256_unpackhi_ps(a, b); // Lower lane: [a2,b2,a3,b3], Upper lane: [a6,b6,a7,b7]

// Step 2: Combine lanes to form target vectors
__mm256 c = _mm256_setr_m128(_mm256_castps256_ps128(temp_low), _mm256_castps256_ps128(temp_high));
__mm256 d = _mm256_setr_m128(_mm256_extractf128_ps(temp_low, 1), _mm256_extractf128_ps(temp_high, 1));

How This Works #

1. 256-bit Unpack Instructions:

   • _mm256_unpacklo_ps(a, b) operates on each 128-bit lane independently: it interleaves the first two elements of a's lane with the first two of b's lane.
   • _mm256_unpackhi_ps(a, b) does the same for the last two elements of each lane.

2. Lane Combination:

   • c concatenates the lower lanes of both temp vectors to get the first four interleaved real/imaginary pairs.
   • d uses the upper lanes of both temp vectors to form the last four interleaved pairs.

Efficiency Advantage #

Your initial plan of splitting into 128-bit vectors first adds extra extract/cast steps. This approach uses only 6 total instructions (2 unpack, 2 extract, 2 lane concatenation), with unpack operations being single uops on modern x86 CPUs—making it the most efficient AVX2 solution for this task.

Context Note #

Since FFT libraries like Intel IPP and FFTW require interleaved complex data, this method provides minimal-overhead conversion from your separate real/imaginary vectors to the required format.

内容的提问来源于stack exchange，提问作者Zvi Vered