poulpy/core/src/blind_rotation/ccgi.rs
core/src/blind_rotation/ccgi.rsRaw
1use backend::{
2 FFT64, MatZnxDftOps, MatZnxDftScratch, Module, ScalarZnx, ScalarZnxAlloc, ScalarZnxDft, ScalarZnxDftAlloc, ScalarZnxDftOps,
3 Scratch, VecZnxAlloc, VecZnxBigAlloc, VecZnxBigOps, VecZnxBigScratch, VecZnxDftAlloc, VecZnxDftOps, VecZnxOps, ZnxView,
4 ZnxViewMut, ZnxZero,
5};
6use itertools::izip;
7
8use crate::{
9 GLWECiphertext, GLWECiphertextToMut, Infos, LWECiphertext,
10 blind_rotation::{key::BlindRotationKeyCGGI, lut::LookUpTable},
11 lwe::ciphertext::LWECiphertextToRef,
12};
13
14pub fn cggi_blind_rotate_scratch_space(
15 module: &Module<FFT64>,
16 extension_factor: usize,
17 basek: usize,
18 k_lut: usize,
19 k_brk: usize,
20 rows: usize,
21 rank: usize,
22) -> usize {
23 let cols: usize = rank + 1;
24 let brk_size: usize = k_brk.div_ceil(basek);
25
26 let acc_dft: usize = module.bytes_of_vec_znx_dft(cols, rows) * extension_factor;
27 let acc_big: usize = module.bytes_of_vec_znx_big(1, brk_size);
28 let vmp_res: usize = module.bytes_of_vec_znx_dft(cols, brk_size) * extension_factor;
29 let acc_dft_add: usize = vmp_res;
30 let xai_plus_y: usize = module.bytes_of_scalar_znx(1);
31 let xai_plus_y_dft: usize = module.bytes_of_scalar_znx_dft(1);
32 let vmp: usize = module.vmp_apply_tmp_bytes(brk_size, rows, rows, 2, 2, brk_size); // GGSW product: (1 x 2) x (2 x 2)
33
34 let acc: usize;
35 if extension_factor > 1 {
36 acc = module.bytes_of_vec_znx(cols, k_lut.div_ceil(basek)) * extension_factor;
37 } else {
38 acc = 0;
39 }
40
41 return acc
42 + acc_dft
43 + acc_dft_add
44 + vmp_res
45 + xai_plus_y
46 + xai_plus_y_dft
47 + (vmp | (acc_big + (module.vec_znx_big_normalize_tmp_bytes() | module.vec_znx_idft_tmp_bytes())));
48}
49
50pub fn cggi_blind_rotate<DataRes, DataIn>(
51 module: &Module<FFT64>,
52 res: &mut GLWECiphertext<DataRes>,
53 lwe: &LWECiphertext<DataIn>,
54 lut: &LookUpTable,
55 brk: &BlindRotationKeyCGGI<FFT64>,
56 scratch: &mut Scratch,
57) where
58 DataRes: AsRef<[u8]> + AsMut<[u8]>,
59 DataIn: AsRef<[u8]>,
60{
61 if lut.extension_factor() > 1 {
62 cggi_blind_rotate_block_binary_extended(module, res, lwe, lut, brk, scratch);
63 } else if brk.block_size() > 1 {
64 cggi_blind_rotate_block_binary(module, res, lwe, lut, brk, scratch);
65 } else {
66 todo!("implement this case")
67 }
68}
69
70// TODO: ENSURE DOMAIN EXTENSION AS
71pub(crate) fn cggi_blind_rotate_block_binary_extended<DataRes, DataIn>(
72 module: &Module<FFT64>,
73 res: &mut GLWECiphertext<DataRes>,
74 lwe: &LWECiphertext<DataIn>,
75 lut: &LookUpTable,
76 brk: &BlindRotationKeyCGGI<FFT64>,
77 scratch: &mut Scratch,
78) where
79 DataRes: AsRef<[u8]> + AsMut<[u8]>,
80 DataIn: AsRef<[u8]>,
81{
82 let extension_factor: usize = lut.extension_factor();
83 let basek: usize = res.basek();
84 let rows: usize = brk.rows();
85 let cols: usize = res.rank() + 1;
86
87 let (mut acc, scratch1) = scratch.tmp_slice_vec_znx(extension_factor, module, cols, res.size());
88 let (mut acc_dft, scratch2) = scratch1.tmp_slice_vec_znx_dft(extension_factor, module, cols, rows);
89 let (mut vmp_res, scratch3) = scratch2.tmp_slice_vec_znx_dft(extension_factor, module, cols, brk.size());
90 let (mut acc_add_dft, scratch4) = scratch3.tmp_slice_vec_znx_dft(extension_factor, module, cols, brk.size());
91 let (mut xai_plus_y, scratch5) = scratch4.tmp_scalar_znx(module, 1);
92 let (mut xai_plus_y_dft, scratch6) = scratch5.tmp_scalar_znx_dft(module, 1);
93
94 (0..extension_factor).for_each(|i| {
95 acc[i].zero();
96 });
97
98 let mut lwe_2n: Vec<i64> = vec![0i64; lwe.n() + 1]; // TODO: from scratch space
99 let lwe_ref: LWECiphertext<&[u8]> = lwe.to_ref();
100
101 let two_n_ext: usize = 2 * lut.domain_size();
102
103 negate_and_mod_switch_2n(two_n_ext, &mut lwe_2n, &lwe_ref);
104
105 let a: &[i64] = &lwe_2n[1..];
106 let b_pos: usize = ((lwe_2n[0] + two_n_ext as i64) % two_n_ext as i64) as usize;
107
108 let b_hi: usize = b_pos / extension_factor;
109 let b_lo: usize = b_pos % extension_factor;
110
111 for (i, j) in (0..b_lo).zip(extension_factor - b_lo..extension_factor) {
112 module.vec_znx_rotate(b_hi as i64 + 1, &mut acc[i], 0, &lut.data[j], 0);
113 }
114 for (i, j) in (b_lo..extension_factor).zip(0..extension_factor - b_lo) {
115 module.vec_znx_rotate(b_hi as i64, &mut acc[i], 0, &lut.data[j], 0);
116 }
117
118 let block_size: usize = brk.block_size();
119
120 izip!(
121 a.chunks_exact(block_size),
122 brk.data.chunks_exact(block_size)
123 )
124 .enumerate()
125 .for_each(|(i, (ai, ski))| {
126 (0..extension_factor).for_each(|i| {
127 (0..cols).for_each(|j| {
128 module.vec_znx_dft(1, 0, &mut acc_dft[i], j, &acc[i], j);
129 });
130 acc_add_dft[i].zero();
131 });
132
133 // TODO: first & last iterations can be optimized
134 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
135 let ai_pos: usize = ((aii + two_n_ext as i64) % two_n_ext as i64) as usize;
136 let ai_hi: usize = ai_pos / extension_factor;
137 let ai_lo: usize = ai_pos % extension_factor;
138
139 // vmp_res = DFT(acc) * BRK[i]
140 (0..extension_factor).for_each(|i| {
141 module.vmp_apply(&mut vmp_res[i], &acc_dft[i], &skii.data, scratch6);
142 });
143
144 // Trivial case: no rotation between polynomials, we can directly multiply with (X^{-ai} - 1)
145 if ai_lo == 0 {
146 // DFT X^{-ai}
147 set_xai_plus_y(module, ai_hi, -1, &mut xai_plus_y_dft, &mut xai_plus_y);
148
149 // Sets acc_add_dft[i] = (acc[i] * sk) * (X^{-ai} - 1)
150 (0..extension_factor).for_each(|j| {
151 (0..cols).for_each(|i| {
152 module.svp_apply_inplace(&mut vmp_res[j], i, &xai_plus_y_dft, 0);
153 module.vec_znx_dft_add_inplace(&mut acc_add_dft[j], i, &vmp_res[j], i);
154 });
155 });
156 // Non trivial case: rotation between polynomials
157 // In this case we can't directly multiply with (X^{-ai} - 1) because of the
158 // ring homomorphism R^{N} -> prod R^{N/extension_factor}, so we split the
159 // computation in two steps: acc_add_dft = (acc * sk) * (-1) + (acc * sk) * X^{-ai}
160 } else {
161 // Sets acc_add_dft[i] = acc[i] * sk
162 (0..extension_factor).for_each(|i| {
163 (0..cols).for_each(|k| {
164 module.vec_znx_dft_sub_ab_inplace(&mut acc_add_dft[i], k, &vmp_res[i], k);
165 })
166 });
167
168 // DFT X^{-ai}
169 set_xai_plus_y(module, ai_hi + 1, 0, &mut xai_plus_y_dft, &mut xai_plus_y);
170
171 // Sets acc_add_dft[0..ai_lo] += (acc[extension_factor - ai_lo..extension_factor] * sk) * X^{-ai+1}
172 for (i, j) in (0..ai_lo).zip(extension_factor - ai_lo..extension_factor) {
173 (0..cols).for_each(|k| {
174 module.svp_apply_inplace(&mut vmp_res[j], k, &xai_plus_y_dft, 0);
175 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i], k, &vmp_res[j], k);
176 });
177 }
178
179 // DFT X^{-ai}
180 set_xai_plus_y(module, ai_hi, 0, &mut xai_plus_y_dft, &mut xai_plus_y);
181
182 // Sets acc_add_dft[ai_lo..extension_factor] += (acc[0..extension_factor - ai_lo] * sk) * X^{-ai}
183 for (i, j) in (ai_lo..extension_factor).zip(0..extension_factor - ai_lo) {
184 (0..cols).for_each(|k| {
185 module.svp_apply_inplace(&mut vmp_res[j], k, &xai_plus_y_dft, 0);
186 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i], k, &vmp_res[j], k);
187 });
188 }
189 }
190 });
191
192 {
193 let (mut acc_add_big, scratch7) = scratch6.tmp_vec_znx_big(module, 1, brk.size());
194
195 (0..extension_factor).for_each(|j| {
196 (0..cols).for_each(|i| {
197 module.vec_znx_idft(&mut acc_add_big, 0, &acc_add_dft[j], i, scratch7);
198 module.vec_znx_big_add_small_inplace(&mut acc_add_big, 0, &acc[j], i);
199 module.vec_znx_big_normalize(basek, &mut acc[j], i, &acc_add_big, 0, scratch7);
200 });
201 });
202 }
203 });
204
205 (0..cols).for_each(|i| {
206 module.vec_znx_copy(&mut res.data, i, &acc[0], i);
207 });
208}
209
210fn set_xai_plus_y(
211 module: &Module<FFT64>,
212 ai: usize,
213 y: i64,
214 res: &mut ScalarZnxDft<&mut [u8], FFT64>,
215 buf: &mut ScalarZnx<&mut [u8]>,
216) {
217 let n: usize = module.n();
218
219 {
220 let raw: &mut [i64] = buf.at_mut(0, 0);
221 if ai < n {
222 raw[ai] = 1;
223 } else {
224 raw[(ai - n) & (n - 1)] = -1;
225 }
226 raw[0] += y;
227 }
228
229 module.svp_prepare(res, 0, buf, 0);
230
231 {
232 let raw: &mut [i64] = buf.at_mut(0, 0);
233
234 if ai < n {
235 raw[ai] = 0;
236 } else {
237 raw[(ai - n) & (n - 1)] = 0;
238 }
239 raw[0] = 0;
240 }
241}
242
243pub(crate) fn cggi_blind_rotate_block_binary<DataRes, DataIn>(
244 module: &Module<FFT64>,
245 res: &mut GLWECiphertext<DataRes>,
246 lwe: &LWECiphertext<DataIn>,
247 lut: &LookUpTable,
248 brk: &BlindRotationKeyCGGI<FFT64>,
249 scratch: &mut Scratch,
250) where
251 DataRes: AsRef<[u8]> + AsMut<[u8]>,
252 DataIn: AsRef<[u8]>,
253{
254 let mut lwe_2n: Vec<i64> = vec![0i64; lwe.n() + 1]; // TODO: from scratch space
255 let mut out_mut: GLWECiphertext<&mut [u8]> = res.to_mut();
256 let lwe_ref: LWECiphertext<&[u8]> = lwe.to_ref();
257 let two_n: usize = module.n() << 1;
258 let basek: usize = brk.basek();
259 let rows = brk.rows();
260
261 let cols: usize = out_mut.rank() + 1;
262
263 negate_and_mod_switch_2n(2 * lut.domain_size(), &mut lwe_2n, &lwe_ref);
264
265 let a: &[i64] = &lwe_2n[1..];
266 let b: i64 = lwe_2n[0];
267
268 out_mut.data.zero();
269
270 // Initialize out to X^{b} * LUT(X)
271 module.vec_znx_rotate(b, &mut out_mut.data, 0, &lut.data[0], 0);
272
273 let block_size: usize = brk.block_size();
274
275 // ACC + [sum DFT(X^ai -1) * (DFT(ACC) x BRKi)]
276
277 let (mut acc_dft, scratch1) = scratch.tmp_vec_znx_dft(module, cols, rows);
278 let (mut vmp_res, scratch2) = scratch1.tmp_vec_znx_dft(module, cols, brk.size());
279 let (mut acc_add_dft, scratch3) = scratch2.tmp_vec_znx_dft(module, cols, brk.size());
280 let (mut xai_plus_y, scratch4) = scratch3.tmp_scalar_znx(module, 1);
281 let (mut xai_plus_y_dft, scratch5) = scratch4.tmp_scalar_znx_dft(module, 1);
282
283 izip!(
284 a.chunks_exact(block_size),
285 brk.data.chunks_exact(block_size)
286 )
287 .for_each(|(ai, ski)| {
288 (0..cols).for_each(|j| {
289 module.vec_znx_dft(1, 0, &mut acc_dft, j, &out_mut.data, j);
290 });
291
292 acc_add_dft.zero();
293
294 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
295 let ai_pos: usize = ((aii + two_n as i64) % two_n as i64) as usize;
296
297 // vmp_res = DFT(acc) * BRK[i]
298 module.vmp_apply(&mut vmp_res, &acc_dft, &skii.data, scratch5);
299
300 // DFT(X^ai -1)
301 set_xai_plus_y(module, ai_pos, -1, &mut xai_plus_y_dft, &mut xai_plus_y);
302
303 // DFT(X^ai -1) * (DFT(acc) * BRK[i])
304 (0..cols).for_each(|i| {
305 module.svp_apply_inplace(&mut vmp_res, i, &xai_plus_y_dft, 0);
306 module.vec_znx_dft_add_inplace(&mut acc_add_dft, i, &vmp_res, i);
307 });
308 });
309
310 (0..cols).for_each(|i| {
311 module.vec_znx_dft_add_inplace(&mut acc_dft, i, &acc_add_dft, i);
312 });
313
314 {
315 let (mut acc_add_big, scratch6) = scratch5.tmp_vec_znx_big(module, 1, brk.size());
316
317 (0..cols).for_each(|i| {
318 module.vec_znx_idft(&mut acc_add_big, 0, &acc_add_dft, i, scratch6);
319 module.vec_znx_big_add_small_inplace(&mut acc_add_big, 0, &out_mut.data, i);
320 module.vec_znx_big_normalize(basek, &mut out_mut.data, i, &acc_add_big, 0, scratch6);
321 });
322 }
323 });
324}
325
326pub(crate) fn negate_and_mod_switch_2n(n: usize, res: &mut [i64], lwe: &LWECiphertext<&[u8]>) {
327 let basek: usize = lwe.basek();
328
329 let log2n: usize = usize::BITS as usize - (n - 1).leading_zeros() as usize + 1;
330
331 res.copy_from_slice(&lwe.data.at(0, 0));
332 res.iter_mut().for_each(|x| *x = -*x);
333
334 if basek > log2n {
335 let diff: usize = basek - log2n;
336 res.iter_mut().for_each(|x| {
337 *x = div_ceil_signed_by_pow2(x, diff);
338 })
339 } else {
340 let rem: usize = basek - (log2n % basek);
341 let size: usize = log2n.div_ceil(basek);
342 (1..size).for_each(|i| {
343 if i == size - 1 && rem != basek {
344 let k_rem: usize = basek - rem;
345 izip!(lwe.data.at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
346 *y = (*y << k_rem) + (x >> rem);
347 });
348 } else {
349 izip!(lwe.data.at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
350 *y = (*y << basek) + x;
351 });
352 }
353 })
354 }
355}
356
357#[inline(always)]
358fn div_round_by_pow2(x: &i64, k: usize) -> i64 {
359 if x >= &0 {
360 (x + (1 << (k - 1))) >> k
361 } else {
362 (x + (-1 << (k - 1))) >> k
363 }
364}
365
366// #[inline(always)]
367// fn div_floor_signed_by_pow2(x: &i64, k: usize) -> i64{
368// let bias: i64 = (1 << k) - 1;
369// (x + ((x >> 63) & bias)) >> k
370// }
371
372#[inline(always)]
373fn div_ceil_signed_by_pow2(x: &i64, k: usize) -> i64 {
374 let bias: i64 = (1 << k) - 1;
375 (x + ((x >> 63) & bias)) >> k
376}