poulpy/core/src/blind_rotation/ccgi.rs
core/src/blind_rotation/ccgi.rsRaw
1use std::time::Instant;
2
3use backend::{
4 FFT64, MatZnxDftOps, MatZnxDftScratch, Module, ScalarZnx, ScalarZnxAlloc, ScalarZnxDft, ScalarZnxDftAlloc, ScalarZnxDftOps,
5 Scratch, VecZnxBigAlloc, VecZnxBigOps, VecZnxDftOps, VecZnxOps, ZnxInfos, ZnxView, ZnxViewMut, ZnxZero,
6};
7use itertools::izip;
8
9use crate::{
10 FourierGLWECiphertext, FourierGLWESecret, GLWECiphertext, GLWECiphertextToMut, GLWEPlaintext, Infos, LWECiphertext,
11 ScratchCore,
12 blind_rotation::{key::BlindRotationKeyCGGI, lut::LookUpTable},
13 lwe::ciphertext::LWECiphertextToRef,
14};
15
16pub fn cggi_blind_rotate_scratch_space(
17 module: &Module<FFT64>,
18 extension_factor: usize,
19 basek: usize,
20 k_lut: usize,
21 k_brk: usize,
22 rows: usize,
23 rank: usize,
24) -> usize {
25 let lut_size: usize = k_lut.div_ceil(basek);
26 let brk_size: usize = k_brk.div_ceil(basek);
27
28 let acc_dft: usize = FourierGLWECiphertext::bytes_of(module, basek, k_brk, rank) * extension_factor;
29 let acc_big: usize = module.bytes_of_vec_znx_big(rank + 1, brk_size);
30 let acc_dft_add: usize = FourierGLWECiphertext::bytes_of(module, basek, k_brk, rank) * extension_factor;
31 let vmp_res: usize = FourierGLWECiphertext::bytes_of(module, basek, k_brk, rank) * extension_factor;
32 let xai_plus_y: usize = module.bytes_of_scalar_znx(1);
33 let xai_plus_y_dft: usize = module.bytes_of_scalar_znx_dft(1);
34 let vmp: usize = module.vmp_apply_tmp_bytes(lut_size, lut_size, rows, 2, 2, brk_size); // GGSW product: (1 x 2) x (2 x 2)
35
36 let acc: usize;
37 if extension_factor > 1 {
38 acc = GLWECiphertext::bytes_of(module, basek, k_lut, rank) * extension_factor;
39 } else {
40 acc = 0;
41 }
42
43 return acc + acc_dft + acc_dft_add + vmp_res + xai_plus_y + xai_plus_y_dft + (vmp | acc_big);
44}
45
46pub fn cggi_blind_rotate<DataRes, DataIn>(
47 module: &Module<FFT64>,
48 res: &mut GLWECiphertext<DataRes>,
49 lwe: &LWECiphertext<DataIn>,
50 lut: &LookUpTable,
51 brk: &BlindRotationKeyCGGI<FFT64>,
52 scratch: &mut Scratch,
53) where
54 DataRes: AsRef<[u8]> + AsMut<[u8]>,
55 DataIn: AsRef<[u8]>,
56{
57 if lut.extension_factor() > 1 {
58 cggi_blind_rotate_block_binary_extended(module, res, lwe, lut, brk, scratch);
59 } else if brk.block_size() > 1 {
60 cggi_blind_rotate_block_binary(module, res, lwe, lut, brk, scratch);
61 } else {
62 todo!("implement this case")
63 }
64}
65
66pub(crate) fn cggi_blind_rotate_block_binary_extended<DataRes, DataIn>(
67 module: &Module<FFT64>,
68 res: &mut GLWECiphertext<DataRes>,
69 lwe: &LWECiphertext<DataIn>,
70 lut: &LookUpTable,
71 brk: &BlindRotationKeyCGGI<FFT64>,
72 scratch: &mut Scratch,
73) where
74 DataRes: AsRef<[u8]> + AsMut<[u8]>,
75 DataIn: AsRef<[u8]>,
76{
77 let extension_factor: usize = lut.extension_factor();
78 let basek: usize = res.basek();
79
80 let (mut acc, scratch1) = scratch.tmp_vec_glwe_ct(extension_factor, module, basek, res.k(), res.rank());
81 let (mut acc_dft, scratch2) = scratch1.tmp_vec_fourier_glwe_ct(extension_factor, module, basek, brk.k(), res.rank());
82 let (mut vmp_res, scratch3) = scratch2.tmp_vec_fourier_glwe_ct(extension_factor, module, basek, brk.k(), res.rank());
83 let (mut acc_add_dft, scratch4) = scratch3.tmp_vec_fourier_glwe_ct(extension_factor, module, basek, brk.k(), res.rank());
84
85 (0..extension_factor).for_each(|i| {
86 acc[i].data.zero();
87 });
88
89 let (mut xai_plus_y, scratch5) = scratch4.tmp_scalar_znx(module, 1);
90 let (mut xai_plus_y_dft, scratch6) = scratch5.tmp_scalar_znx_dft(module, 1);
91 let (mut acc_add_big, scratch7) = scratch6.tmp_vec_znx_big(module, 1, brk.size());
92
93 let mut lwe_2n: Vec<i64> = vec![0i64; lwe.n() + 1]; // TODO: from scratch space
94 let lwe_ref: LWECiphertext<&[u8]> = lwe.to_ref();
95
96 let two_n_ext: usize = 2 * lut.domain_size();
97
98 let cols: usize = res.rank() + 1;
99
100 negate_and_mod_switch_2n(two_n_ext, &mut lwe_2n, &lwe_ref);
101
102 let a: &[i64] = &lwe_2n[1..];
103 let b_pos: usize = ((lwe_2n[0] + two_n_ext as i64) % two_n_ext as i64) as usize;
104
105 let b_hi: usize = b_pos / extension_factor;
106 let b_lo: usize = b_pos % extension_factor;
107
108 for (i, j) in (0..b_lo).zip(extension_factor - b_lo..extension_factor) {
109 module.vec_znx_rotate(b_hi as i64 + 1, &mut acc[i].data, 0, &lut.data[j], 0);
110 }
111 for (i, j) in (b_lo..extension_factor).zip(0..extension_factor - b_lo) {
112 module.vec_znx_rotate(b_hi as i64, &mut acc[i].data, 0, &lut.data[j], 0);
113 }
114
115 let block_size: usize = brk.block_size();
116
117 izip!(
118 a.chunks_exact(block_size),
119 brk.data.chunks_exact(block_size)
120 )
121 .enumerate()
122 .for_each(|(i, (ai, ski))| {
123 (0..extension_factor).for_each(|i| {
124 (0..cols).for_each(|j| {
125 module.vec_znx_dft(1, 0, &mut acc_dft[i].data, j, &acc[i].data, j);
126 });
127 acc_add_dft[i].data.zero();
128 });
129
130 // TODO: first & last iterations can be optimized
131 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
132 let ai_pos: usize = ((aii + two_n_ext as i64) % two_n_ext as i64) as usize;
133 let ai_hi: usize = ai_pos / extension_factor;
134 let ai_lo: usize = ai_pos % extension_factor;
135
136 // vmp_res = DFT(acc) * BRK[i]
137 (0..extension_factor).for_each(|i| {
138 module.vmp_apply(&mut vmp_res[i].data, &acc_dft[i].data, &skii.data, scratch7);
139 });
140
141 // Trivial case: no rotation between polynomials, we can directly multiply with (X^{-ai} - 1)
142 if ai_lo == 0 {
143 // DFT X^{-ai}
144 set_xai_plus_y(
145 module,
146 ai_hi as i64,
147 -1,
148 &mut xai_plus_y_dft,
149 &mut xai_plus_y,
150 );
151
152 // Sets acc_add_dft[i] = (acc[i] * sk) * (X^{-ai} - 1)
153 (0..extension_factor).for_each(|j| {
154 (0..cols).for_each(|i| {
155 module.svp_apply_inplace(&mut vmp_res[j].data, i, &xai_plus_y_dft, 0);
156 module.vec_znx_dft_add_inplace(&mut acc_add_dft[j].data, i, &vmp_res[j].data, i);
157 });
158 });
159 // Non trivial case: rotation between polynomials
160 // In this case we can't directly multiply with (X^{-ai} - 1) because of the
161 // ring homomorphism R^{N} -> prod R^{N/extension_factor}, so we split the
162 // computation in two steps: acc_add_dft = (acc * sk) * (-1) + (acc * sk) * X^{-ai}
163 } else {
164 // Sets acc_add_dft[i] = acc[i] * sk
165 (0..extension_factor).for_each(|i| {
166 (0..cols).for_each(|k| {
167 module.vec_znx_dft_sub_ab_inplace(&mut acc_add_dft[i].data, k, &vmp_res[i].data, k);
168 })
169 });
170
171 // DFT X^{-ai+1}
172 set_xai_plus_y(
173 module,
174 ai_hi as i64 + 1,
175 0,
176 &mut xai_plus_y_dft,
177 &mut xai_plus_y,
178 );
179
180 // Sets acc_add_dft[0..ai_lo] += (acc[extension_factor - ai_lo..extension_factor] * sk) * X^{-ai+1}
181 for (i, j) in (0..ai_lo).zip(extension_factor - ai_lo..extension_factor) {
182 module.vec_znx_rotate(b_hi as i64 + 1, &mut acc[i].data, 0, &lut.data[j], 0);
183 (0..cols).for_each(|k| {
184 module.svp_apply_inplace(&mut vmp_res[j].data, k, &xai_plus_y_dft, 0);
185 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i].data, k, &vmp_res[j].data, k);
186 });
187 }
188
189 // DFT X^{-ai}
190 set_xai_plus_y(
191 module,
192 ai_hi as i64,
193 0,
194 &mut xai_plus_y_dft,
195 &mut xai_plus_y,
196 );
197
198 // Sets acc_add_dft[ai_lo..extension_factor] += (acc[0..extension_factor - ai_lo] * sk) * X^{-ai}
199 for (i, j) in (ai_lo..extension_factor).zip(0..extension_factor - ai_lo) {
200 module.vec_znx_rotate(b_hi as i64, &mut acc[i].data, 0, &lut.data[j], 0);
201 (0..cols).for_each(|k| {
202 module.svp_apply_inplace(&mut vmp_res[j].data, k, &xai_plus_y_dft, 0);
203 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i].data, k, &vmp_res[j].data, k);
204 });
205 }
206 }
207 });
208
209 (0..extension_factor).for_each(|j| {
210 (0..cols).for_each(|i| {
211 module.vec_znx_dft_add_inplace(&mut acc_dft[j].data, i, &acc_add_dft[j].data, i);
212 module.vec_znx_idft(&mut acc_add_big, 0, &acc_dft[j].data, i, scratch7);
213 module.vec_znx_big_normalize(basek, &mut acc[j].data, i, &acc_add_big, 0, scratch7);
214 });
215 });
216 });
217
218 (0..cols).for_each(|i| {
219 module.vec_znx_copy(&mut res.data, i, &acc[0].data, i);
220 });
221}
222
223fn set_xai_plus_y(
224 module: &Module<FFT64>,
225 k: i64,
226 y: i64,
227 res: &mut ScalarZnxDft<&mut [u8], FFT64>,
228 buf: &mut ScalarZnx<&mut [u8]>,
229) {
230 buf.zero();
231 buf.at_mut(0, 0)[0] = 1;
232 module.vec_znx_rotate_inplace(k, buf, 0);
233 buf.at_mut(0, 0)[0] += y;
234 module.svp_prepare(res, 0, buf, 0);
235}
236
237pub(crate) fn cggi_blind_rotate_block_binary<DataRes, DataIn>(
238 module: &Module<FFT64>,
239 res: &mut GLWECiphertext<DataRes>,
240 lwe: &LWECiphertext<DataIn>,
241 lut: &LookUpTable,
242 brk: &BlindRotationKeyCGGI<FFT64>,
243 scratch: &mut Scratch,
244) where
245 DataRes: AsRef<[u8]> + AsMut<[u8]>,
246 DataIn: AsRef<[u8]>,
247{
248 let basek: usize = res.basek();
249
250 let mut lwe_2n: Vec<i64> = vec![0i64; lwe.n() + 1]; // TODO: from scratch space
251 let mut out_mut: GLWECiphertext<&mut [u8]> = res.to_mut();
252 let lwe_ref: LWECiphertext<&[u8]> = lwe.to_ref();
253
254 let cols: usize = out_mut.rank() + 1;
255
256 negate_and_mod_switch_2n(2 * lut.domain_size(), &mut lwe_2n, &lwe_ref);
257
258 let a: &[i64] = &lwe_2n[1..];
259 let b: i64 = lwe_2n[0];
260
261 out_mut.data.zero();
262
263 // Initialize out to X^{b} * LUT(X)
264 module.vec_znx_rotate(-b, &mut out_mut.data, 0, &lut.data[0], 0);
265
266 let block_size: usize = brk.block_size();
267
268 // ACC + [sum DFT(X^ai -1) * (DFT(ACC) x BRKi)]
269
270 let (mut acc_dft, scratch1) = scratch.tmp_fourier_glwe_ct(module, brk.basek(), out_mut.k(), out_mut.rank());
271 let (mut acc_add_dft, scratch2) = scratch1.tmp_fourier_glwe_ct(module, brk.basek(), out_mut.k(), out_mut.rank());
272 let (mut vmp_res, scratch3) = scratch2.tmp_fourier_glwe_ct(module, basek, out_mut.k(), out_mut.rank());
273 let (mut xai_plus_y, scratch4) = scratch3.tmp_scalar_znx(module, 1);
274 let (mut xai_plus_y_dft, scratch5) = scratch4.tmp_scalar_znx_dft(module, 1);
275
276 let start: Instant = Instant::now();
277 izip!(
278 a.chunks_exact(block_size),
279 brk.data.chunks_exact(block_size)
280 )
281 .for_each(|(ai, ski)| {
282 out_mut.dft(module, &mut acc_dft);
283 acc_add_dft.data.zero();
284
285 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
286 // vmp_res = DFT(acc) * BRK[i]
287 module.vmp_apply(&mut vmp_res.data, &acc_dft.data, &skii.data, scratch5);
288
289 // DFT(X^ai -1)
290 set_xai_plus_y(module, *aii, -1, &mut xai_plus_y_dft, &mut xai_plus_y);
291
292 // DFT(X^ai -1) * (DFT(acc) * BRK[i])
293 (0..cols).for_each(|i| {
294 module.svp_apply_inplace(&mut vmp_res.data, i, &xai_plus_y_dft, 0);
295 module.vec_znx_dft_add_inplace(&mut acc_add_dft.data, i, &vmp_res.data, i);
296 });
297 });
298
299 (0..cols).for_each(|i| {
300 module.vec_znx_dft_add_inplace(&mut acc_dft.data, i, &acc_add_dft.data, i);
301 });
302
303 acc_dft.idft(module, &mut out_mut, scratch5);
304 });
305 let duration: std::time::Duration = start.elapsed();
306}
307
308pub(crate) fn negate_and_mod_switch_2n(n: usize, res: &mut [i64], lwe: &LWECiphertext<&[u8]>) {
309 let basek: usize = lwe.basek();
310
311 let log2n: usize = usize::BITS as usize - (n - 1).leading_zeros() as usize + 1;
312
313 res.copy_from_slice(&lwe.data.at(0, 0));
314 res.iter_mut().for_each(|x| *x = -*x);
315
316 if basek > log2n {
317 let diff: usize = basek - log2n;
318 res.iter_mut().for_each(|x| {
319 *x = div_signed_by_pow2(x, diff);
320 })
321 } else {
322 let rem: usize = basek - (log2n % basek);
323 let size: usize = log2n.div_ceil(basek);
324 (1..size).for_each(|i| {
325 if i == size - 1 && rem != basek {
326 let k_rem: usize = basek - rem;
327 izip!(lwe.data.at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
328 *y = (*y << k_rem) + (x >> rem);
329 });
330 } else {
331 izip!(lwe.data.at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
332 *y = (*y << basek) + x;
333 });
334 }
335 })
336 }
337}
338
339#[inline(always)]
340fn div_signed_by_pow2(x: &i64, k: usize) -> i64 {
341 let bias: i64 = (1 << k) - 1;
342 (x + ((x >> 63) & bias)) >> k
343}