[BUG] Enabling AVX2 makes gather slower

[mechakotik]

Describe the bug

Gather operation is 5 times slower when AVX2 instructions are allowed with -mavx2.

To Reproduce

• Set up Google Benchmark project
• Disable CPU frequency scaling with sudo cpupower frequency-set --governor performance
• Test the following code with -O3 and -O3 -mavx2:

void run(benchmark::State& state) {
    float data[4] = {1, 2, 3, 4};
    for(auto _ : state) {
        eve::wide<float, eve::fixed<4>> vec = eve::gather(data, eve::wide<unsigned char, eve::fixed<4>>{2, 3, 0, 1});
        benchmark::DoNotOptimize(vec);
    }
}

BENCHMARK(run);
BENCHMARK_MAIN();

Without -mavx2:

-----------------------------------------------------
Benchmark           Time             CPU   Iterations
-----------------------------------------------------
run             0.198 ns        0.198 ns   3315565533

With -mavx2:

-----------------------------------------------------
Benchmark           Time             CPU   Iterations
-----------------------------------------------------
run              1.01 ns         1.01 ns    648857193

Setup:

• Compiler: g++ 14.2.1, clang++ 19.1.7
• OS: Gentoo Linux
• CPU: Ryzen 9 7940HS
• Instructions Set used: SSE, AVX2

[jfalcou]

Out of the blue and before checking anything, compile with -DNDEBUG we have a bunch asserts around there.

[mechakotik]

Just tested with -DNDEBUG and no asserts have failed

[jfalcou]

I meant it may have affected performance. I will investigate this week end.