perf: Elide the total order wrapper for non-(float/option) types

[nameexhaustion]

This should be a slight perf improvement for hashing primitive types, but more importantly will revert the change in the hash output values for integer types introduced by #14617 (see #14617 (comment)).

Benchmark

running 1 test
n_iterations: 70000000
(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): 2449ms vs 2157ms
(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): 2411ms vs 2151ms
(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): 2400ms vs 2122ms
(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): 2437ms vs 2145ms
(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): 2440ms vs 2166ms
(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2916ms vs 3357ms
(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 3051ms vs 3350ms
(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2921ms vs 3279ms
(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2961ms vs 3262ms
(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2989ms vs 3253ms
(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2399ms vs 2584ms
test total_ord::test::test has been running for over 60 seconds
(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2432ms vs 2665ms
(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2403ms vs 2665ms
(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2425ms vs 2578ms
(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): 2411ms vs 2681ms

mod test {
    #[test]
    fn test() {
        use std::time::SystemTime;

        use super::*;
        let rs = ahash::RandomState::new();

        let n_iterations = 70_000_000;

        println!("n_iterations: {}", n_iterations);

        fn time_it<F: Fn() -> T, T>(f: F, n: u32) -> u128 {
            let start = SystemTime::now();
            for _ in 0..n {
                f();
            }
            let end = SystemTime::now();
            let duration = end.duration_since(start).unwrap();
            return duration.as_millis();
        }

        let x: i32 = 10;

        for _ in 0..5 {
            let hash_one_wrap_t = time_it(|| rs.clone().hash_one(TotalOrdWrap(x)), n_iterations);
            let hash_one_to_tot_t = time_it(|| rs.clone().hash_one(x.to_total_ord()), n_iterations);
            println!(
                "(10 as i32) hash_one(TotalOrdWrap(x)) vs hash_one(x.to_total_ord()): {}ms vs {}ms",
                hash_one_wrap_t, hash_one_to_tot_t
            );
        }

        let x: Option<i32> = Some(10);

        for _ in 0..5 {
            let hash_one_t = time_it(|| rs.clone().hash_one(x), n_iterations);
            let hash_one_to_tot_t = time_it(|| rs.clone().hash_one(x.to_total_ord()), n_iterations);
            println!(
                "(Some(10) as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): {}ms vs {}ms",
                hash_one_t, hash_one_to_tot_t
            );
        }

        let x: Option<i32> = None;

        for _ in 0..5 {
            let hash_one_t = time_it(|| rs.clone().hash_one(x), n_iterations);
            let hash_one_to_tot_t = time_it(|| rs.clone().hash_one(x.to_total_ord()), n_iterations);
            println!(
                "(None as Option<i32>) hash_one(x) vs hash_one(x.to_total_ord()): {}ms vs {}ms",
                hash_one_t, hash_one_to_tot_t
            );
        }
    }
}

[codecov]

Codecov Report

Attention: Patch coverage is 86.74033% with 24 lines in your changes are missing coverage. Please review.

Project coverage is 80.77%. Comparing base (9560c34) to head (7f12860).
Report is 2 commits behind head on main.

Files	Patch %	Lines
crates/polars-utils/src/total_ord.rs	58.33%	15 Missing ⚠️
crates/polars-core/src/frame/group_by/hashing.rs	66.66%	4 Missing ⚠️
crates/polars-core/src/datatypes/any_value.rs	25.00%	3 Missing ⚠️
crates/polars-core/src/hashing/vector_hasher.rs	80.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #14648      +/-   ##
==========================================
- Coverage   80.78%   80.77%   -0.02%     
==========================================
  Files        1326     1326              
  Lines      173157   173227      +70     
  Branches     2453     2453              
==========================================
+ Hits       139891   139923      +32     
- Misses      32792    32830      +38     
  Partials      474      474              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nameexhaustion]

wasn't straightforward to add ToTotalOrd trait for vec_hash on PolarsObject so juts hash directly for now, since I don't think it's needed

[ritchie46]

Indeed. Objects are 🤷

[nameexhaustion]

Because the build_tables function now uses .to_total_ord() this is now just a plain BytesHash instead of a TotalOrdWrap<BytesHash>, so we don't need to call .to_total_ord() here (this function is asof_join_by_binary).

[ritchie46]

Thank you @nameexhaustion.