Am/feat/epbs to experimental

[IceTDrinker]

Brings the extended PBS to the experimental module of core_crypto

No entities for now to make the initial integration easier, it was deemed ok since the new type would only combine a BSK and an extension factor, not worth all the code for something experimental that can be used as is if we want in the shortint dyn atomic pattern

[mayeul-zama]

@mayeul-zama reviewed 11 files and all commit messages, and made 9 comments.
Reviewable status: all files reviewed, 9 unresolved discussions (waiting on IceTDrinker, soonum, and SouchonTheo).

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 77 at r1 (raw file):

    SplitLutCont: ContainerMut<Element = Scalar>,
{
    for (idx, &coeff) in lut.as_ref().iter().enumerate() {

Could add asserts on sizes

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 126 at r1 (raw file):

{
    // extension factor == 1 means classic bootstrap which is already optimized
    if extension_factor.0 == 1 {

Maybe we could simply panic here

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 88 at r1 (raw file):

/// N' = 2^nu * N
/// new_lut_idx = (ai + old_lut_idx) % 2^nu
/// (2^nu + (ai % 2N') - 1 - new_lut_idx)/2^nu seems to work for the multiplication by X^ai

Hard to understand comment
Could detail a bit more

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 154 at r1 (raw file):

    );

    // TODO ? use only a split accumulator and an assign primitive ?

Do we want to keep this TODO?

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 265 at r1 (raw file):

    std::thread::scope(|s| {
        let thread_processing = |id: usize, stack: &mut PodStack| {

Would be easier to read if split into smaller functions or (less ideally) closures

---

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 295 at r1 (raw file):

                        extension_factor,
                        ct_dst_idx,
                    );

A comment to explain how ct_dst_idx and small_monomial_degree are used to "simulate" rotation on the big LUT by doing small rotations and swaps would really help understand here

---

tfhe/src/core_crypto/experimental/commons/parameters.rs line 48 at r1 (raw file):

#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
pub struct LweBootstrapExtensionFactor(pub usize);

Could add a comment to explain what it is

---

tfhe/src/core_crypto/experimental/algorithms/test/lwe_extended_programmable_bootstrapping.rs line 119 at r1 (raw file):

            // Verify our formulas work the way we expect by comparing to the reference not using
            // the trick
            assert_eq!(ref_small_rotated_luts, small_luts_rotated_with_trick);

Putting computation of ref_small_rotated_luts and small_luts_rotated_with_trick in separate functions with extended_lut as input would help follow

---

tfhe/src/core_crypto/experimental/algorithms/test/lwe_extended_programmable_bootstrapping.rs line 245 at r1 (raw file):

    let mut buffers = ComputationBuffers::new();

    // TODO: have req for main thread and for workers ?

Keep TODO?

[mayeul-zama]

tfhe/src/core_crypto/experimental/algorithms/lwe_extended_programmable_bootstrapping.rs line 389 at r1 (raw file):

    }

    GlweCiphertext::from_container(

I think we should return the full extended LUT
And maybe add a function to build the extended glwe secret key and have tests show that we get the initial LUT rotated