rand_xoshiro crate = runtime error

[brandonros]

Reproduction:

#[kernel]
#[allow(improper_ctypes_definitions, clippy::missing_safety_doc)]
pub unsafe fn find_vanity_private_key(a: &[u8]) {
    let idx = thread::index_1d() as usize;

    let mut rng_seed: u64 = 42; // TODO: make this unique for each thread?
    let mut num_iterations: u64 = 0;

    for _ in 0..10000 {
        // generate random input for seed
        let mut rng = Xoroshiro128StarStar::seed_from_u64(rng_seed); // if i change this to MockXoroshiro128StarStar it works fine
        let mut input = [0u8; 32];
        rng.fill_bytes(&mut input);
         ...

[jorge-ortega]

Hi @brandonros,

Thanks for the report. I'm not able to reproduce this atm. Can you also share how you launch from the host side?

My kernel currently looks like this:

#[kernel]
#[allow(improper_ctypes_definitions, clippy::missing_safety_doc)]
pub unsafe fn rng_test() {

    let rng_seed: u64 = 42;

    for _ in 0..10000 {
        let mut rng = Xoroshiro128StarStar::seed_from_u64(rng_seed);
        let mut input = [0u8; 32];
        rng.fill_bytes(&mut input);
    }
}

And I launch like this:

unsafe {
    launch!(
        module.rng_test<<<32, 32, 0, stream>>>()
    )?;
}

The path-tracer example uses DefaultRand, which uses Xorohiro128StarStar. The render kernel receives it as a parameter, which the host side provides when launching. Hope that helps.

[brandonros]

I'll report back if I'm able to reproduce. I moved to XorShiftRng for now which doesn't seem to have this issue. Not sure if it is related to resources (too many threads/blocks).

I had the problem trying to launch it with cust and https://github.com/coreylowman/cudarc but I do not want to waste your time tracking it down.

Would a compiled output like .ll help at all?

[brandonros]

brandonros/ed25519-vanity-rs@424ee9d

if you switch Xoroshiro128StarStar and XorShiftRng on this branch it should reproduce

[brandonros]

rng-illegal-address.ptx.zip