Add cross-basek normalization (#90)

* added cross_basek_normalization

* updated method signatures to take layouts

* fixed cross-base normalization

fix #91
fix #93
This commit is contained in:
Jean-Philippe Bossuat
2025-09-30 14:40:10 +02:00
committed by GitHub
parent 4da790ea6a
commit 37e13b965c
216 changed files with 12481 additions and 7745 deletions

View File

@@ -4,17 +4,20 @@ use poulpy_hal::{
api::{
ScratchAvailable, TakeMatZnx, TakeVecZnx, TakeVecZnxBig, TakeVecZnxDft, TakeVecZnxDftSlice, TakeVecZnxSlice,
VecZnxAddInplace, VecZnxAutomorphismInplace, VecZnxBigAddSmallInplace, VecZnxBigAllocBytes, VecZnxBigAutomorphismInplace,
VecZnxBigNormalize, VecZnxBigNormalizeTmpBytes, VecZnxBigSubSmallBInplace, VecZnxCopy, VecZnxDftAddInplace,
VecZnxBigNormalize, VecZnxBigNormalizeTmpBytes, VecZnxBigSubSmallNegateInplace, VecZnxCopy, VecZnxDftAddInplace,
VecZnxDftAllocBytes, VecZnxDftApply, VecZnxDftCopy, VecZnxIdftApplyConsume, VecZnxIdftApplyTmpA, VecZnxNegateInplace,
VecZnxNormalizeInplace, VecZnxNormalizeTmpBytes, VecZnxRotate, VecZnxRotateInplace, VecZnxRotateInplaceTmpBytes,
VecZnxRshInplace, VecZnxSub, VecZnxSubABInplace, VecZnxSwitchRing, VmpApplyDftToDft, VmpApplyDftToDftAdd,
VmpApplyDftToDftTmpBytes,
VecZnxNormalize, VecZnxNormalizeInplace, VecZnxNormalizeTmpBytes, VecZnxRotate, VecZnxRotateInplace,
VecZnxRotateInplaceTmpBytes, VecZnxRshInplace, VecZnxSub, VecZnxSubInplace, VecZnxSwitchRing, VmpApplyDftToDft,
VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
},
layouts::{Backend, DataMut, DataRef, Module, Scratch, ToOwnedDeep},
oep::{ScratchOwnedAllocImpl, ScratchOwnedBorrowImpl},
};
use poulpy_core::{GLWEOperations, TakeGGLWE, TakeGLWECt, layouts::Infos};
use poulpy_core::{
GLWEOperations, TakeGGLWE, TakeGLWECt,
layouts::{Digits, GGLWECiphertextLayout, GGSWInfos, GLWEInfos, LWEInfos},
};
use poulpy_core::layouts::{GGSWCiphertext, GLWECiphertext, LWECiphertext, prepared::GGLWEAutomorphismKeyPrepared};
@@ -39,7 +42,7 @@ where
+ VecZnxAddInplace
+ VecZnxNegateInplace
+ VecZnxCopy
+ VecZnxSubABInplace
+ VecZnxSubInplace
+ VecZnxDftAllocBytes
+ VmpApplyDftToDftTmpBytes
+ VecZnxBigNormalizeTmpBytes
@@ -50,11 +53,12 @@ where
+ VecZnxBigAddSmallInplace<B>
+ VecZnxBigNormalize<B>
+ VecZnxAutomorphismInplace<B>
+ VecZnxBigSubSmallBInplace<B>
+ VecZnxBigSubSmallNegateInplace<B>
+ VecZnxRotateInplaceTmpBytes
+ VecZnxBigAllocBytes
+ VecZnxDftAddInplace<B>
+ VecZnxRotate,
+ VecZnxRotate
+ VecZnxNormalize<B>,
B: Backend + ScratchOwnedAllocImpl<B> + ScratchOwnedBorrowImpl<B>,
Scratch<B>: TakeVecZnx
+ TakeVecZnxDftSlice<B>
@@ -138,7 +142,7 @@ pub fn circuit_bootstrap_core<DRes, DLwe, DBrk, BRA: BlindRotationAlgo, B>(
+ VecZnxAddInplace
+ VecZnxNegateInplace
+ VecZnxCopy
+ VecZnxSubABInplace
+ VecZnxSubInplace
+ VecZnxDftAllocBytes
+ VmpApplyDftToDftTmpBytes
+ VecZnxBigNormalizeTmpBytes
@@ -149,11 +153,12 @@ pub fn circuit_bootstrap_core<DRes, DLwe, DBrk, BRA: BlindRotationAlgo, B>(
+ VecZnxBigAddSmallInplace<B>
+ VecZnxBigNormalize<B>
+ VecZnxAutomorphismInplace<B>
+ VecZnxBigSubSmallBInplace<B>
+ VecZnxBigSubSmallNegateInplace<B>
+ VecZnxBigAllocBytes
+ VecZnxDftAddInplace<B>
+ VecZnxRotateInplaceTmpBytes
+ VecZnxRotate,
+ VecZnxRotate
+ VecZnxNormalize<B>,
B: Backend + ScratchOwnedAllocImpl<B> + ScratchOwnedBorrowImpl<B>,
Scratch<B>: TakeVecZnxDftSlice<B>
+ TakeVecZnxBig<B>
@@ -166,16 +171,18 @@ pub fn circuit_bootstrap_core<DRes, DLwe, DBrk, BRA: BlindRotationAlgo, B>(
{
#[cfg(debug_assertions)]
{
use poulpy_core::layouts::LWEInfos;
assert_eq!(res.n(), key.brk.n());
assert_eq!(lwe.basek(), key.brk.basek());
assert_eq!(res.basek(), key.brk.basek());
assert_eq!(lwe.base2k(), key.brk.base2k());
assert_eq!(res.base2k(), key.brk.base2k());
}
let n: usize = res.n();
let basek: usize = res.basek();
let rows: usize = res.rows();
let rank: usize = res.rank();
let k: usize = res.k();
let n: usize = res.n().into();
let base2k: usize = res.base2k().into();
let rows: usize = res.rows().into();
let rank: usize = res.rank().into();
let k: usize = res.k().into();
let alpha: usize = rows.next_power_of_two();
@@ -183,27 +190,38 @@ pub fn circuit_bootstrap_core<DRes, DLwe, DBrk, BRA: BlindRotationAlgo, B>(
if to_exponent {
(0..rows).for_each(|i| {
f[i] = 1 << (basek * (rows - 1 - i));
f[i] = 1 << (base2k * (rows - 1 - i));
});
} else {
(0..1 << log_domain).for_each(|j| {
(0..rows).for_each(|i| {
f[j * alpha + i] = j as i64 * (1 << (basek * (rows - 1 - i)));
f[j * alpha + i] = j as i64 * (1 << (base2k * (rows - 1 - i)));
});
});
}
// Lut precision, basically must be able to hold the decomposition power basis of the GGSW
let mut lut: LookUpTable = LookUpTable::alloc(module, basek, basek * rows, extension_factor);
lut.set(module, &f, basek * rows);
let mut lut: LookUpTable = LookUpTable::alloc(module, base2k, base2k * rows, extension_factor);
lut.set(module, &f, base2k * rows);
if to_exponent {
lut.set_rotation_direction(LookUpTableRotationDirection::Right);
}
// TODO: separate GGSW k from output of blind rotation k
let (mut res_glwe, scratch_1) = scratch.take_glwe_ct(n, basek, k, rank);
let (mut tmp_gglwe, scratch_2) = scratch_1.take_gglwe(n, basek, k, rows, 1, rank.max(1), rank);
let (mut res_glwe, scratch_1) = scratch.take_glwe_ct(res);
let gglwe_infos: GGLWECiphertextLayout = GGLWECiphertextLayout {
n: n.into(),
base2k: base2k.into(),
k: k.into(),
rows: rows.into(),
digits: Digits(1),
rank_in: rank.max(1).into(),
rank_out: rank.into(),
};
let (mut tmp_gglwe, scratch_2) = scratch_1.take_gglwe(&gglwe_infos);
key.brk.execute(module, &mut res_glwe, lwe, &lut, scratch_2);
@@ -264,7 +282,7 @@ fn post_process<DataRes, DataA, B: Backend>(
+ VecZnxAddInplace
+ VecZnxNegateInplace
+ VecZnxCopy
+ VecZnxSubABInplace
+ VecZnxSubInplace
+ VecZnxDftAllocBytes
+ VmpApplyDftToDftTmpBytes
+ VecZnxBigNormalizeTmpBytes
@@ -275,8 +293,9 @@ fn post_process<DataRes, DataA, B: Backend>(
+ VecZnxBigAddSmallInplace<B>
+ VecZnxBigNormalize<B>
+ VecZnxAutomorphismInplace<B>
+ VecZnxBigSubSmallBInplace<B>
+ VecZnxRotate,
+ VecZnxBigSubSmallNegateInplace<B>
+ VecZnxRotate
+ VecZnxNormalize<B>,
Scratch<B>: TakeVecZnxDft<B> + ScratchAvailable + TakeVecZnx,
{
let log_n: usize = module.log_n();
@@ -336,7 +355,7 @@ pub fn pack<D: DataMut, B: Backend>(
+ VecZnxAddInplace
+ VecZnxNegateInplace
+ VecZnxCopy
+ VecZnxSubABInplace
+ VecZnxSubInplace
+ VecZnxDftAllocBytes
+ VmpApplyDftToDftTmpBytes
+ VecZnxBigNormalizeTmpBytes
@@ -347,16 +366,13 @@ pub fn pack<D: DataMut, B: Backend>(
+ VecZnxBigAddSmallInplace<B>
+ VecZnxBigNormalize<B>
+ VecZnxAutomorphismInplace<B>
+ VecZnxBigSubSmallBInplace<B>
+ VecZnxRotate,
+ VecZnxBigSubSmallNegateInplace<B>
+ VecZnxRotate
+ VecZnxNormalize<B>,
Scratch<B>: TakeVecZnx + TakeVecZnxDft<B> + ScratchAvailable,
{
let log_n: usize = module.log_n();
let basek: usize = cts.get(&0).unwrap().basek();
let k: usize = cts.get(&0).unwrap().k();
let rank: usize = cts.get(&0).unwrap().rank();
(0..log_n - log_gap_out).for_each(|i| {
let t: usize = 16.min(1 << (log_n - 1 - i));
@@ -370,17 +386,7 @@ pub fn pack<D: DataMut, B: Backend>(
let mut a: Option<GLWECiphertext<D>> = cts.remove(&j);
let mut b: Option<GLWECiphertext<D>> = cts.remove(&(j + t));
combine(
module,
basek,
k,
rank,
a.as_mut(),
b.as_mut(),
i,
auto_key,
scratch,
);
combine(module, a.as_mut(), b.as_mut(), i, auto_key, scratch);
if let Some(a) = a {
cts.insert(j, a);
@@ -394,9 +400,6 @@ pub fn pack<D: DataMut, B: Backend>(
#[allow(clippy::too_many_arguments)]
fn combine<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
module: &Module<B>,
basek: usize,
k: usize,
rank: usize,
a: Option<&mut GLWECiphertext<A>>,
b: Option<&mut GLWECiphertext<D>>,
i: usize,
@@ -415,7 +418,7 @@ fn combine<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
+ VecZnxAddInplace
+ VecZnxNegateInplace
+ VecZnxCopy
+ VecZnxSubABInplace
+ VecZnxSubInplace
+ VecZnxDftAllocBytes
+ VmpApplyDftToDftTmpBytes
+ VecZnxBigNormalizeTmpBytes
@@ -426,8 +429,9 @@ fn combine<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
+ VecZnxBigAddSmallInplace<B>
+ VecZnxBigNormalize<B>
+ VecZnxAutomorphismInplace<B>
+ VecZnxBigSubSmallBInplace<B>
+ VecZnxRotate,
+ VecZnxBigSubSmallNegateInplace<B>
+ VecZnxRotate
+ VecZnxNormalize<B>,
Scratch<B>: TakeVecZnx + TakeVecZnxDft<B> + ScratchAvailable,
{
// Goal is to evaluate: a = a + b*X^t + phi(a - b*X^t))
@@ -441,12 +445,10 @@ fn combine<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
// either mapped to garbage or twice their value which vanishes I(X)
// since 2*(I(X) * Q/2) = I(X) * Q = 0 mod Q.
if let Some(a) = a {
let n: usize = a.n();
let log_n: usize = (u64::BITS - (n - 1).leading_zeros()) as _;
let t: i64 = 1 << (log_n - i - 1);
let t: i64 = 1 << (a.n().log2() - i - 1);
if let Some(b) = b {
let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(n, basek, k, rank);
let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(a);
// a = a * X^-t
a.rotate_inplace(module, -t, scratch_1);
@@ -478,15 +480,13 @@ fn combine<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
a.automorphism_add_inplace(module, auto_key, scratch);
}
} else if let Some(b) = b {
let n: usize = b.n();
let log_n: usize = (u64::BITS - (n - 1).leading_zeros()) as _;
let t: i64 = 1 << (log_n - i - 1);
let t: i64 = 1 << (b.n().log2() - i - 1);
let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(n, basek, k, rank);
let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(b);
tmp_b.rotate(module, t, b);
tmp_b.rsh(module, 1, scratch_1);
// a = (b* X^t - phi(b* X^t))
b.automorphism_sub_ba(module, &tmp_b, auto_key, scratch_1);
b.automorphism_sub_negate(module, &tmp_b, auto_key, scratch_1);
}
}