AVX2 intrinsics seem to inconsistently take up more processor cycles compared to SSE2 intrinsics (x86_64)

help
1

This post is made after further investigating odd performance behavior I mentioned in a previous post here.

Basically, x86_64 AVX2 vertical addition seems to sometimes take more processor cycles as measured using the rdtsc x86 instruction.

Here's my source code/test case:


use core::mem::transmute;
use core::sync::atomic;
use core::hint::black_box;

#[cfg(target_arch = "x86")]
use core::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use core::arch::x86_64::*;

#[test]
#[inline(never)]
pub fn rdtsc_avx2_vs_sse2() {
    assert!(std::is_x86_feature_detected!("avx2"));

    unsafe {
        let a_256: __m256i = black_box(transmute([0u8; 32]));
        let a_128s: [__m128i; 2] = black_box(transmute(a_256));
        let b_256: __m256i = black_box(transmute([1u8; 32]));
        let b_128s: [__m128i; 2] = black_box(transmute(b_256));
    
        atomic::fence(atomic::Ordering::SeqCst);
        let start_avx2 = _rdtsc();
        atomic::fence(atomic::Ordering::SeqCst);
    
        let result_avx2 = _mm256_add_epi8(a_256, b_256);
    
        atomic::fence(atomic::Ordering::SeqCst);
        let end_avx2 = _rdtsc();
        atomic::fence(atomic::Ordering::SeqCst);
    
        std::println!("AVX2 version got {:?} and took {} CPU cycles", transmute::<_, [u8; 32]>(result_avx2), end_avx2 - start_avx2);
        
        atomic::fence(atomic::Ordering::SeqCst);
        let start_sse2 = _rdtsc();
        atomic::fence(atomic::Ordering::SeqCst);
    
        let result_sse2 = [_mm_add_epi8(a_128s[0], b_128s[0]), _mm_add_epi8(a_128s[1], b_128s[1])];
    
        atomic::fence(atomic::Ordering::SeqCst);
        let end_sse2 = _rdtsc();
        atomic::fence(atomic::Ordering::SeqCst);
    
        std::println!("SSE2 version got {:?} and took {} CPU cycles", transmute::<_, [u8; 32]>(result_sse2), end_sse2 - start_sse2);
    
    }
}

And here's some outputs on my machine (in release mode):

> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 576 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 216 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.05s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 612 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 180 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 216 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 504 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 216 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 576 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 900 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 216 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 252 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 576 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 864 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 216 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

PS C:\Users\alfri\Documents\Projects\x86-simd> cargo test --release rdtsc_avx2_vs_sse2 -- --nocapture
    Finished `release` profile [optimized] target(s) in 0.04s
     Running unittests src\lib.rs (target\release\deps\x86_simd-ef4220bf1b194345.exe)

running 0 tests

test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 8 filtered out; finished in 0.00s

     Running tests\rdtsc_avx2_vs_sse2.rs (target\release\deps\rdtsc_avx2_vs_sse2-d7b94e7d902646c6.exe)

running 1 test
AVX2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 864 CPU cycles
SSE2 version got [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] and took 180 CPU cycles
test rdtsc_avx2_vs_sse2 ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

Is it just my specific chip? I have an AMD Ryzen 5 3600 (in the zen2 line).
It just doesn't make sense to me that a single AVX2 call would seemingly randomly take 800 cycles when 2 SSE2 calls take almost always 200.

2

You are missing #[target_feature(enable = "avx2")] on your function. Without it, LLVM will not use AVX2 instructions inside the function. It will either keep the intrinsic calls outlined, or lower them to SSE instructions, using two (or more) SSE instructions for each AVX2 instruction.

2 Likes
3

You're only doing one operation between the atomic fences and rdtsc. Those will totally overwhelm the cost of the computation itself. What happens if you call the intrinsics thousands of times?