various bug fixes

This commit is contained in:
Jean-Philippe Bossuat
2025-06-10 12:55:34 +02:00
parent 113231da55
commit a673b84047
8 changed files with 95 additions and 57 deletions

View File

@@ -344,7 +344,12 @@ impl<DataSelf: AsMut<[u8]> + AsRef<[u8]>> GGSWCiphertext<DataSelf, FFT64> {
// col 2: (-(c0s0 + c1s1 + c2s2) , c0 , c1 + M[i], c2 )
// col 3: (-(d0s0 + d1s1 + d2s2) , d0 , d1 , d2 + M[i])
let digits: usize = tsk.digits();
let (mut tmp_dft_i, scratch1) = scratch.tmp_vec_znx_dft(module, cols, tsk.size());
let (mut tmp_a, scratch2) = scratch1.tmp_vec_znx_dft(module, 1, (ci_dft.size() + digits - 1) / digits);
let res_size: usize = res.to_mut().size();
{
// Performs a key-switch for each combination of s[i]*s[j], i.e. for a0, a1, a2
//
@@ -358,28 +363,30 @@ impl<DataSelf: AsMut<[u8]> + AsRef<[u8]>> GGSWCiphertext<DataSelf, FFT64> {
// =
// (-(x0s0 + x1s1 + x2s2) + s0(a0s0 + a1s1 + a2s2), x0, x1, x2)
(1..cols).for_each(|col_i| {
let digits: usize = tsk.digits();
let pmat: &MatZnxDft<DataTsk, FFT64> = &tsk.at(col_i - 1, col_j - 1).0.data; // Selects Enc(s[i]s[j])
// Extracts a[i] and multipies with Enc(s[i]s[j])
if col_i == 1 {
(0..digits).for_each(|di| {
let (mut tmp_a, scratch2) = scratch1.tmp_vec_znx_dft(module, 1, (ci_dft.size() + di) / digits);
module.vec_znx_dft_copy(digits, digits - 1 - di, &mut tmp_a, 0, ci_dft, col_i);
if di == 0 {
module.vmp_apply(&mut tmp_dft_i, &tmp_a, pmat, scratch2);
} else {
module.vmp_apply_add(&mut tmp_dft_i, &tmp_a, pmat, di, scratch2);
}
});
} else {
(0..digits).for_each(|di| {
let (mut tmp_a, scratch2) = scratch1.tmp_vec_znx_dft(module, 1, (ci_dft.size() + di) / digits);
module.vec_znx_dft_copy(digits, digits - 1 - di, &mut tmp_a, 0, ci_dft, col_i);
(0..digits).for_each(|di| {
tmp_a.set_size((ci_dft.size() + di) / digits);
// Small optimization for digits > 2
// VMP produce some error e, and since we aggregate vmp * 2^{di * B}, then
// we also aggregate ei * 2^{di * B}, with the largest error being ei * 2^{(digits-1) * B}.
// As such we can ignore the last digits-2 limbs safely of the sum of vmp products.
// It is possible to further ignore the last digits-1 limbs, but this introduce
// ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
// noise is kept with respect to the ideal functionality.
//tmp_dft_i.set_size(res_size - ((digits - di) as isize - 2).max(0) as usize);
module.vec_znx_dft_copy(digits, digits - 1 - di, &mut tmp_a, 0, ci_dft, col_i);
if di == 0 && col_i == 1 {
module.vmp_apply(&mut tmp_dft_i, &tmp_a, pmat, scratch2);
} else {
module.vmp_apply_add(&mut tmp_dft_i, &tmp_a, pmat, di, scratch2);
});
}
}
});
});
}

View File

@@ -1,7 +1,5 @@
use backend::{
AddNormal, Backend, FFT64, FillUniform, MatZnxDftOps, MatZnxDftScratch, Module, ScalarZnxAlloc, ScalarZnxDftAlloc,
ScalarZnxDftOps, Scratch, VecZnx, VecZnxAlloc, VecZnxBig, VecZnxBigAlloc, VecZnxBigOps, VecZnxBigScratch, VecZnxDftAlloc,
VecZnxDftOps, VecZnxOps, VecZnxToMut, VecZnxToRef, ZnxZero,
AddNormal, Backend, FillUniform, MatZnxDftOps, MatZnxDftScratch, Module, ScalarZnxAlloc, ScalarZnxDftAlloc, ScalarZnxDftOps, Scratch, VecZnx, VecZnxAlloc, VecZnxBig, VecZnxBigAlloc, VecZnxBigOps, VecZnxBigScratch, VecZnxDftAlloc, VecZnxDftOps, VecZnxOps, VecZnxToMut, VecZnxToRef, ZnxZero, FFT64
};
use sampling::source::Source;
@@ -94,14 +92,15 @@ impl GLWECiphertext<Vec<u8>> {
rank_in: usize,
rank_out: usize,
) -> usize {
let res_dft: usize = GLWECiphertextFourier::bytes_of(module, basek, k_out, rank_out);
let res_dft: usize = GLWECiphertextFourier::bytes_of(module, basek, k_out, rank_out + 1);
let in_size: usize = div_ceil(div_ceil(k_in, basek), digits);
let out_size: usize = div_ceil(k_out, basek);
let ksk_size: usize = div_ceil(k_ksk, basek);
let ai_dft: usize = module.bytes_of_vec_znx_dft(rank_in, in_size);
let vmp: usize = module.vmp_apply_tmp_bytes(out_size, in_size, in_size, rank_in, rank_out + 1, ksk_size)
+ module.bytes_of_vec_znx_dft(rank_in, in_size);
let normalize: usize = module.vec_znx_big_normalize_tmp_bytes();
return res_dft + (vmp | normalize);
return res_dft + ((ai_dft + vmp) | normalize);
}
pub fn keyswitch_from_fourier_scratch_space(
@@ -494,20 +493,29 @@ impl<DataSelf: AsRef<[u8]> + AsMut<[u8]>> GLWECiphertext<DataSelf> {
let cols_in: usize = rhs.rank_in();
let cols_out: usize = rhs.rank_out() + 1;
let digits: usize = rhs.digits();
let (mut res_dft, scratch1) = scratch.tmp_vec_znx_dft(module, cols_out, rhs.size()); // Todo optimise
let (mut ai_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols_in, (lhs.size() + digits - 1) / digits);
ai_dft.zero();
{
let digits = rhs.digits();
(0..digits).for_each(|di| {
// (lhs.size() + di) / digits = (a - (digit - di - 1) + digit - 1) / digits
let (mut ai_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols_in, (lhs.size() + di) / digits);
ai_dft.set_size((lhs.size() + di) / digits);
// Small optimization for digits > 2
// VMP produce some error e, and since we aggregate vmp * 2^{di * B}, then
// we also aggregate ei * 2^{di * B}, with the largest error being ei * 2^{(digits-1) * B}.
// As such we can ignore the last digits-2 limbs safely of the sum of vmp products.
// It is possible to further ignore the last digits-1 limbs, but this introduce
// ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
// noise is kept with respect to the ideal functionality.
//res_dft.set_size(rhs.size() - ((digits - di) as isize - 2).max(0) as usize);
(0..cols_in).for_each(|col_i| {
module.vec_znx_dft(
digits,
digits - 1 - di,
digits - di - 1,
&mut ai_dft,
col_i,
&lhs.data,
@@ -587,15 +595,24 @@ impl<DataSelf: AsRef<[u8]> + AsMut<[u8]>> GLWECiphertext<DataSelf> {
}
let cols: usize = rhs.rank() + 1;
let digits: usize = rhs.digits();
let (mut res_dft, scratch1) = scratch.tmp_vec_znx_dft(module, cols, rhs.size()); // Todo optimise
let (mut a_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols, (lhs.size() + digits-1) / digits);
{
let digits = rhs.digits();
(0..digits).for_each(|di| {
// (lhs.size() + di) / digits = (a - (digit - di - 1) + digit - 1) / digits
let (mut a_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols, (lhs.size() + di) / digits);
a_dft.set_size((lhs.size() + di) / digits);
// Small optimization for digits > 2
// VMP produce some error e, and since we aggregate vmp * 2^{di * B}, then
// we also aggregate ei * 2^{di * B}, with the largest error being ei * 2^{(digits-1) * B}.
// As such we can ignore the last digits-2 limbs safely of the sum of vmp products.
// It is possible to further ignore the last digits-1 limbs, but this introduce
// ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
// noise is kept with respect to the ideal functionality.
//res_dft.set_size(rhs.size() - ((digits - di) as isize - 2).max(0) as usize);
(0..cols).for_each(|col_i| {
module.vec_znx_dft(digits, digits - 1 - di, &mut a_dft, col_i, &lhs.data, col_i);

View File

@@ -96,7 +96,7 @@ impl GLWECiphertextFourier<Vec<u8>, FFT64> {
pub fn external_product_scratch_space(
module: &Module<FFT64>,
basek: usize,
k_out: usize,
_k_out: usize,
k_in: usize,
k_ggsw: usize,
digits: usize,
@@ -197,17 +197,19 @@ impl<DataSelf: AsMut<[u8]> + AsRef<[u8]>> GLWECiphertextFourier<DataSelf, FFT64>
}
let cols: usize = rhs.rank() + 1;
let digits = rhs.digits();
// Space for VMP result in DFT domain and high precision
let (mut res_dft, scratch1) = scratch.tmp_vec_znx_dft(module, cols, rhs.size());
let (mut a_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols, (lhs.size() + digits - 1) / digits);
{
let digits = rhs.digits();
(0..digits).for_each(|di| {
// (lhs.size() + di) / digits = (a - (digit - di - 1) + digit - 1) / digits
let (mut a_dft, scratch2) = scratch1.tmp_vec_znx_dft(module, cols, (lhs.size() + di) / digits);
a_dft.set_size((lhs.size() + di) / digits);
res_dft.set_size(rhs.size() - (digits - di - 1));
(0..cols).for_each(|col_i| {
module.vec_znx_dft_copy(digits, digits - 1 - di, &mut a_dft, col_i, &lhs.data, col_i);
});

View File

@@ -557,7 +557,7 @@ fn test_automorphism(
);
assert!(
(noise_have - noise_want).abs() <= 0.5,
noise_have <= noise_want + 1.0,
"{} {}",
noise_have,
noise_want