asking chatgpt 4 to >write a rust application that will circumvent L1, L2, and L3 and test memory latency, read and write speeds

memtest.rs

#![feature(asm)]
use std::time::Instant;

const BUFFER_SIZE: usize = 1024 * 1024 * 128; // 128 MB

fn main() {
    let mut buffer = vec![0u8; BUFFER_SIZE];

    let start_time = Instant::now();
    unsafe { write_memory_avx(&mut buffer); }
    let write_duration = start_time.elapsed();

    let start_time = Instant::now();
    let sum = unsafe { read_memory_avx(&buffer) };
    let read_duration = start_time.elapsed();

    let read_speed = (BUFFER_SIZE as f64) / read_duration.as_secs_f64() / 1024.0 / 1024.0;
    let write_speed = (BUFFER_SIZE as f64) / write_duration.as_secs_f64() / 1024.0 / 1024.0;

    println!("Read duration: {:?}", read_duration);
    println!("Write duration: {:?}", write_duration);
    println!("Read speed: {:.2} MB/s", read_speed);
    println!("Write speed: {:.2} MB/s", write_speed);
    println!("Sum: {}", sum);
}

#[inline(never)]
unsafe fn read_memory_avx(buffer: &[u8]) -> u64 {
    let mut sum = 0;
    let len = buffer.len() / 32;

    asm!(
        "xor %rax, %rax",
        "mov $0, %rcx",
        "1:",
        "vmovntdqa (%rsi), %ymm0",
        "add $32, %rsi",
        "vpaddq %ymm0, %ymm1, %ymm1",
        "loop 1b",

        "vextracti128 $1, %ymm1, %xmm2",
        "vpaddq %xmm1, %xmm2, %xmm1",
        "movq %xmm1, %rax",

        inout("rsi") buffer.as_ptr() => _,
        inout("rcx") len => _,
        out("rax") sum,
    );

    sum
}

#[inline(never)]
unsafe fn write_memory_avx(buffer: &mut [u8]) {
    let len = buffer.len() / 32;

    asm!(
        "1:",
        "vmovntdq %ymm0, (%rdi)",
        "add $32, %rdi",
        "loop 1b",

        inout("rdi") buffer.as_mut_ptr() => _,
        inout("rcx") len => _,
    );
}