chore: add another implementation of multiversion

[usamoi]

I noticed that it's hard to test multiversions of a function and the code can be written incorrectly easily.

Also, dispatching manual versions is not efficient.

So I reimplemented multiversion.

Examples

// It generate x86_64/v4, x86_64/v3, x86_64/v2, aarch64/neon and fallback versions of this function
#[detect::multiversion(v4, v3, v2, neon, fallback)]
fn f() {}

#[cfg(any(target_arch = "x86_64", doc))]
#[doc(cfg(target_arch = "x86_64"))]
#[detect::target_cpu(enable = "v4")]
unsafe fn g_v4() {}

// It generate x86_64/v3, x86_64/v2, aarch64/neon and fallback versions of this function
// It takes advantage of `g_v4` as x86_64/v4 version of this function
// It exposes the fallback version with the name "g_fallback"
#[detect::multiversion(v4 = import, v3, v2, neon, fallback = export)]
fn g() {}

#[cfg(test)]
fn g_test() {
    #[cfg(target_arch = "x86_64")]
    if detect::v4::detect() {
        assert_eq!(unsafe { g_v4() }, unsafe { g_fallback() });
    } else {
        println!("skipped: v4");
    }
}

[cutecutecat]

I have a question, why it's hard to test multiversions of a function? Could you explain it more detailed?

Could we use same name for different version like:

#[multiversion(targets("x86_64+avx"))]
#[cfg(test)]
fn g() {
...
}

#[multiversion(targets("aarch64+neon"))]
#[cfg(test)]
fn g() {
...
}

#[cfg(test)]
#[multiversion(targets("aarch64+neon", "x86_64+avx"))]
fn g_test() {
   g();
}

Or a flattened test:

#[multiversion(targets("x86_64+avx"))]
#[cfg(test)]
fn xxx_x86_64_avx_test() {
g();
}

#[multiversion(targets("aarch64+neon"))]
#[cfg(test)]
fn xxx_aarch64_neon_test() {
g_aarch64_neon();
}

[usamoi]

I have a question, why it's hard to test multiversions of a function? Could you explain it more detailed?

It can be written incorrectly. Some cases:

• detect::init() is not called in the tests of svector.
• test() is called, instead of detect() is called in the implementation of veci8.
• Versions of a function is written everywhere and hard to track.

Could we use same name for different version

I do not understand what do you mean. detect::multiversion is for implementation, not for tests. It just forces name conversion and module level, and allows you write less code. Thus we have a standard way to write tests and view them by cargo doc.

[usamoi]

@silver-ymz

base::vector::svecf32::sl2_v4_test (former base::vector::svecf32::test_sl2_svector) failed (https://github.com/tensorchord/pgvecto.rs/actions/runs/8434503680/job/23098303512?pr=446#step:9:608) in emulator (you can set up local environment following https://github.com/usamoi/pgvecto.rs/blob/multiversion/.github/workflows/rust.yml#L167).

Can you help me with it?

Edit:

1. you have permissions to push commits to usamoi:multiversion
2. this test fails with a probability

[cutecutecat]

Need a comments for multiversion, other LGTM.

[VoVAllen]

sparse vector should not have precision problem I think. 1e-5 should be enough, the failure might mean something is wrong

[silver-ymz]

silver-ymz@5f0bac3 can fix base::vector::svecf32::sl2_v4_test. And about precision, I ran 100 times, the max difference is 2e-4.

[silver-ymz]

Also, in my tests. vector::vecf16::dot_v4_avx512fp16_test failed once.

---- vector::vecf16::dot_v4_avx512fp16_test stdout ----
thread 'vector::vecf16::dot_v4_avx512fp16_test' panicked at crates/base/src/vector/vecf16.rs:229:5:
specialized = 1024, fallback = 1025.4672.

[usamoi]

silver-ymz@5f0bac3 can fix base::vector::svecf32::sl2_v4_test. And about precision, I ran 100 times, the max difference is 2e-4.

100 times evaluation on emulator shows EPS should be larger than 7.0138855.

Are you testing on a real machine? Is the difference caused by emulating?

[usamoi]

I found if I just copy the code in dot_v4 and evaluate $\Sigma x^2 + \Sigma y^2 - 2 D$, the smallest EPS in 10000 times should be 0.00032043457.

Also, in my tests. vector::vecf16::dot_v4_avx512fp16_test failed once.

All tests should be stressful to get real EPS...

[silver-ymz]

Are you testing on a real machine?

Yes, I test on real machine.

Is the difference caused by emulating?

I'm not sure. Could you provide the vectors which caused large EPS to reproduce?

[usamoi]

I think I know why now. I should copy code in left zone to my editor.