chore(bench): implement throughput benchmarks on core_crypto layer

[soonum]

There is still a bit a of work to do but that won't change much the implementation.
Still to do:

• Check if the GPU devices are fully loaded during benchmarks, adapt elements generation if needed
• Double check usage of streams (seems sub-optimal at the moment)

---

This change is 

[agnesLeroy]

Looks good to me except for the comment I left earlier, and I think we can keep only one workflow for GPU benchmarks for integer+core, wdyt?

[agnesLeroy]

Thanks a lot @soonum! In the end you kept cuda_local_streams and cuda_local_streams_core then?

[IceTDrinker]

@soonum the PR is quite big tell me which part you wanted me to look at in particular that you mentionned

Reviewed 10 of 14 files at r1, all commit messages.
Reviewable status: 10 of 14 files reviewed, 7 unresolved discussions (waiting on @agnesLeroy and @soonum)

---

tfhe/benches/integer/signed_bench.rs line 3071 at r1 (raw file):

fn main() {
    init_bench_type();

can even be a get_bench_type that populates the once lock on demand

---

tfhe/benches/utilities.rs line 703 at r1 (raw file):

    /// Instantiate Cuda server key to each available GPU.
    #[allow(dead_code)]
    #[cfg(feature = "integer")]

pattern to try to make it less annoying with features

if I have somethign for integer

#[cfg(feature = "integer")]
pub mod my_integer_mod {
}

#[cfg(feature = "integer")]
pub use my_integer_mod::*;

avoids to have to repeat the feature and have dead code potentially

---

tfhe/benches/integer/bench.rs line 3329 at r1 (raw file):

fn main() {
    init_bench_type();

can even be a get_bench_type that populates the once lock on demand

---

tfhe/benches/integer/glwe_packing_compression.rs line 368 at r1 (raw file):

fn main() {
    init_bench_type();

can even be a get_bench_type that populates the once lock on demand so no need to manually init

---

tfhe/benches/integer/zk_pke.rs line 440 at r1 (raw file):

fn main() {
    init_bench_type();

same here probably not needed

[soonum]

tfhe/benches/integer/bench.rs line 3329 at r1 (raw file):

Previously, IceTDrinker (Arthur Meyre) wrote…

can even be a get_bench_type that populates the once lock on demand

A getter would mean, to me, returning a value that we should pass around.

[soonum]

tfhe/benches/integer/bench.rs line 3329 at r1 (raw file):

Previously, soonum (David Testé) wrote…

A getter would mean, to me, returning a value that we should pass around.

Unless you mean a get_bench_type called directly in the benchmark functions ?

[IceTDrinker]

tfhe/benches/integer/bench.rs line 3329 at r1 (raw file):

Previously, IceTDrinker (Arthur Meyre) wrote…
A getter would mean, to me, returning a value that we should pass around.

no something like

pub fn get_bench_type() -> BenchType {
    static BENCH_TYPE: OnceLock<BenchType> = OnceLocke::new();

    BENCH_TYPE.get_or_init(|| /* env var logiv*/)
    return bench_type
}

you call it where you need it and the bench type is cacehd in the once lock