poulpy/poulpy-schemes/src/tfhe/blind_rotation/cggi_algo.rs
poulpy-schemes/src/tfhe/blind_rotation/cggi_algo.rsRaw
1use itertools::izip;
2use poulpy_hal::{
3 api::{
4 ScratchAvailable, SvpApplyDftToDft, SvpPPolAllocBytes, TakeVecZnx, TakeVecZnxBig, TakeVecZnxDft, TakeVecZnxDftSlice,
5 TakeVecZnxSlice, VecZnxAddInplace, VecZnxBigAddSmallInplace, VecZnxBigAllocBytes, VecZnxBigNormalize,
6 VecZnxBigNormalizeTmpBytes, VecZnxCopy, VecZnxDftAdd, VecZnxDftAddInplace, VecZnxDftAllocBytes, VecZnxDftApply,
7 VecZnxDftSubInplace, VecZnxDftZero, VecZnxIdftApply, VecZnxIdftApplyConsume, VecZnxIdftApplyTmpBytes,
8 VecZnxMulXpMinusOneInplace, VecZnxNormalize, VecZnxNormalizeInplace, VecZnxNormalizeTmpBytes, VecZnxRotate,
9 VecZnxSubInplace, VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
10 },
11 layouts::{Backend, DataMut, DataRef, Module, Scratch, SvpPPol, VecZnx, ZnxView, ZnxZero},
12};
13
14use poulpy_core::{
15 Distribution, GLWEOperations, TakeGLWECt,
16 layouts::{GGSWInfos, GLWE, GLWEInfos, GLWEToMut, LWE, LWECiphertextToRef, LWEInfos},
17};
18
19use crate::tfhe::blind_rotation::{
20 BlincRotationExecute, BlindRotationKeyInfos, BlindRotationKeyPrepared, CGGI, LookUpTable, LookUpTableRotationDirection,
21};
22
23#[allow(clippy::too_many_arguments)]
24pub fn cggi_blind_rotate_scratch_space<B: Backend, OUT, GGSW>(
25 module: &Module<B>,
26 block_size: usize,
27 extension_factor: usize,
28 glwe_infos: &OUT,
29 brk_infos: &GGSW,
30) -> usize
31where
32 OUT: GLWEInfos,
33 GGSW: GGSWInfos,
34 Module<B>: VecZnxDftAllocBytes
35 + VmpApplyDftToDftTmpBytes
36 + VecZnxNormalizeTmpBytes
37 + VecZnxBigAllocBytes
38 + VecZnxIdftApplyTmpBytes
39 + VecZnxBigNormalizeTmpBytes,
40{
41 let brk_size: usize = brk_infos.size();
42
43 if block_size > 1 {
44 let cols: usize = (brk_infos.rank() + 1).into();
45 let dnum: usize = brk_infos.dnum().into();
46 let acc_dft: usize = module.vec_znx_dft_bytes_of(cols, dnum) * extension_factor;
47 let acc_big: usize = module.vec_znx_big_bytes_of(1, brk_size);
48 let vmp_res: usize = module.vec_znx_dft_bytes_of(cols, brk_size) * extension_factor;
49 let vmp_xai: usize = module.vec_znx_dft_bytes_of(1, brk_size);
50 let acc_dft_add: usize = vmp_res;
51 let vmp: usize = module.vmp_apply_dft_to_dft_tmp_bytes(brk_size, dnum, dnum, 2, 2, brk_size); // GGSW product: (1 x 2) x (2 x 2)
52 let acc: usize = if extension_factor > 1 {
53 VecZnx::bytes_of(module.n(), cols, glwe_infos.size()) * extension_factor
54 } else {
55 0
56 };
57
58 acc + acc_dft
59 + acc_dft_add
60 + vmp_res
61 + vmp_xai
62 + (vmp | (acc_big + (module.vec_znx_big_normalize_tmp_bytes() | module.vec_znx_idft_apply_tmp_bytes())))
63 } else {
64 GLWE::bytes_of(glwe_infos) + GLWE::external_product_inplace_scratch_space(module, glwe_infos, brk_infos)
65 }
66}
67
68impl<D: DataRef, B: Backend> BlincRotationExecute<B> for BlindRotationKeyPrepared<D, CGGI, B>
69where
70 Module<B>: VecZnxBigAllocBytes
71 + VecZnxDftAllocBytes
72 + SvpPPolAllocBytes
73 + VmpApplyDftToDftTmpBytes
74 + VecZnxBigNormalizeTmpBytes
75 + VecZnxIdftApplyTmpBytes
76 + VecZnxIdftApply<B>
77 + VecZnxDftAdd<B>
78 + VecZnxDftAddInplace<B>
79 + VecZnxDftApply<B>
80 + VecZnxDftZero<B>
81 + SvpApplyDftToDft<B>
82 + VecZnxDftSubInplace<B>
83 + VecZnxBigAddSmallInplace<B>
84 + VecZnxRotate
85 + VecZnxAddInplace
86 + VecZnxSubInplace
87 + VecZnxNormalize<B>
88 + VecZnxNormalizeInplace<B>
89 + VecZnxCopy
90 + VecZnxMulXpMinusOneInplace<B>
91 + VmpApplyDftToDft<B>
92 + VmpApplyDftToDftAdd<B>
93 + VecZnxIdftApplyConsume<B>
94 + VecZnxBigNormalize<B>
95 + VecZnxNormalizeTmpBytes,
96 Scratch<B>: TakeVecZnxDftSlice<B> + TakeVecZnxDft<B> + TakeVecZnxBig<B> + TakeVecZnxSlice + TakeVecZnx + ScratchAvailable,
97{
98 fn execute<DR: DataMut, DI: DataRef>(
99 &self,
100 module: &Module<B>,
101 res: &mut GLWE<DR>,
102 lwe: &LWE<DI>,
103 lut: &LookUpTable,
104 scratch: &mut Scratch<B>,
105 ) {
106 match self.dist {
107 Distribution::BinaryBlock(_) | Distribution::BinaryFixed(_) | Distribution::BinaryProb(_) | Distribution::ZERO => {
108 if lut.extension_factor() > 1 {
109 execute_block_binary_extended(module, res, lwe, lut, self, scratch)
110 } else if self.block_size() > 1 {
111 execute_block_binary(module, res, lwe, lut, self, scratch);
112 } else {
113 execute_standard(module, res, lwe, lut, self, scratch);
114 }
115 }
116 _ => panic!("invalid CGGI distribution"),
117 }
118 }
119}
120
121fn execute_block_binary_extended<DataRes, DataIn, DataBrk, B: Backend>(
122 module: &Module<B>,
123 res: &mut GLWE<DataRes>,
124 lwe: &LWE<DataIn>,
125 lut: &LookUpTable,
126 brk: &BlindRotationKeyPrepared<DataBrk, CGGI, B>,
127 scratch: &mut Scratch<B>,
128) where
129 DataRes: DataMut,
130 DataIn: DataRef,
131 DataBrk: DataRef,
132 Module<B>: VecZnxBigAllocBytes
133 + VecZnxDftAllocBytes
134 + SvpPPolAllocBytes
135 + VmpApplyDftToDftTmpBytes
136 + VecZnxBigNormalizeTmpBytes
137 + VecZnxIdftApplyTmpBytes
138 + VecZnxIdftApply<B>
139 + VecZnxDftAdd<B>
140 + VecZnxDftAddInplace<B>
141 + VecZnxDftApply<B>
142 + VecZnxDftZero<B>
143 + SvpApplyDftToDft<B>
144 + VecZnxDftSubInplace<B>
145 + VecZnxBigAddSmallInplace<B>
146 + VecZnxRotate
147 + VecZnxAddInplace
148 + VecZnxSubInplace
149 + VecZnxNormalize<B>
150 + VecZnxNormalizeInplace<B>
151 + VecZnxCopy
152 + VecZnxMulXpMinusOneInplace<B>
153 + VecZnxBigNormalize<B>
154 + VmpApplyDftToDft<B>,
155 Scratch<B>: TakeVecZnxDftSlice<B> + TakeVecZnxDft<B> + TakeVecZnxBig<B> + TakeVecZnxSlice + ScratchAvailable + TakeVecZnx,
156{
157 let n_glwe: usize = brk.n_glwe().into();
158 let extension_factor: usize = lut.extension_factor();
159 let base2k: usize = res.base2k().into();
160 let dnum: usize = brk.dnum().into();
161 let cols: usize = (res.rank() + 1).into();
162
163 let (mut acc, scratch_1) = scratch.take_vec_znx_slice(extension_factor, n_glwe, cols, res.size());
164 let (mut acc_dft, scratch_2) = scratch_1.take_vec_znx_dft_slice(extension_factor, n_glwe, cols, dnum);
165 let (mut vmp_res, scratch_3) = scratch_2.take_vec_znx_dft_slice(extension_factor, n_glwe, cols, brk.size());
166 let (mut acc_add_dft, scratch_4) = scratch_3.take_vec_znx_dft_slice(extension_factor, n_glwe, cols, brk.size());
167 let (mut vmp_xai, scratch_5) = scratch_4.take_vec_znx_dft(n_glwe, 1, brk.size());
168
169 (0..extension_factor).for_each(|i| {
170 acc[i].zero();
171 });
172
173 let x_pow_a: &Vec<SvpPPol<Vec<u8>, B>>;
174 if let Some(b) = &brk.x_pow_a {
175 x_pow_a = b
176 } else {
177 panic!("invalid key: x_pow_a has not been initialized")
178 }
179
180 let mut lwe_2n: Vec<i64> = vec![0i64; (lwe.n() + 1).as_usize()]; // TODO: from scratch space
181 let lwe_ref: LWE<&[u8]> = lwe.to_ref();
182
183 let two_n: usize = 2 * n_glwe;
184 let two_n_ext: usize = 2 * lut.domain_size();
185
186 mod_switch_2n(two_n_ext, &mut lwe_2n, &lwe_ref, lut.rotation_direction());
187
188 let a: &[i64] = &lwe_2n[1..];
189 let b_pos: usize = ((lwe_2n[0] + two_n_ext as i64) & (two_n_ext - 1) as i64) as usize;
190
191 let b_hi: usize = b_pos / extension_factor;
192 let b_lo: usize = b_pos & (extension_factor - 1);
193
194 for (i, j) in (0..b_lo).zip(extension_factor - b_lo..extension_factor) {
195 module.vec_znx_rotate(b_hi as i64 + 1, &mut acc[i], 0, &lut.data[j], 0);
196 }
197 for (i, j) in (b_lo..extension_factor).zip(0..extension_factor - b_lo) {
198 module.vec_znx_rotate(b_hi as i64, &mut acc[i], 0, &lut.data[j], 0);
199 }
200
201 let block_size: usize = brk.block_size();
202
203 izip!(
204 a.chunks_exact(block_size),
205 brk.data.chunks_exact(block_size)
206 )
207 .for_each(|(ai, ski)| {
208 (0..extension_factor).for_each(|i| {
209 (0..cols).for_each(|j| {
210 module.vec_znx_dft_apply(1, 0, &mut acc_dft[i], j, &acc[i], j);
211 });
212 module.vec_znx_dft_zero(&mut acc_add_dft[i])
213 });
214
215 // TODO: first & last iterations can be optimized
216 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
217 let ai_pos: usize = ((aii + two_n_ext as i64) & (two_n_ext - 1) as i64) as usize;
218 let ai_hi: usize = ai_pos / extension_factor;
219 let ai_lo: usize = ai_pos & (extension_factor - 1);
220
221 // vmp_res = DFT(acc) * BRK[i]
222 (0..extension_factor).for_each(|i| {
223 module.vmp_apply_dft_to_dft(&mut vmp_res[i], &acc_dft[i], skii.data(), scratch_5);
224 });
225
226 // Trivial case: no rotation between polynomials, we can directly multiply with (X^{-ai} - 1)
227 if ai_lo == 0 {
228 // Sets acc_add_dft[i] = (acc[i] * sk) * X^{-ai} - (acc[i] * sk)
229 if ai_hi != 0 {
230 // DFT X^{-ai}
231 (0..extension_factor).for_each(|j| {
232 (0..cols).for_each(|i| {
233 module.svp_apply_dft_to_dft(&mut vmp_xai, 0, &x_pow_a[ai_hi], 0, &vmp_res[j], i);
234 module.vec_znx_dft_add_inplace(&mut acc_add_dft[j], i, &vmp_xai, 0);
235 module.vec_znx_dft_sub_inplace(&mut acc_add_dft[j], i, &vmp_res[j], i);
236 });
237 });
238 }
239
240 // Non trivial case: rotation between polynomials
241 // In this case we can't directly multiply with (X^{-ai} - 1) because of the
242 // ring homomorphism R^{N} -> prod R^{N/extension_factor}, so we split the
243 // computation in two steps: acc_add_dft = (acc * sk) * (-1) + (acc * sk) * X^{-ai}
244 } else {
245 // Sets acc_add_dft[0..ai_lo] += (acc[extension_factor - ai_lo..extension_factor] * sk) * X^{-ai+1}
246 if (ai_hi + 1) & (two_n - 1) != 0 {
247 for (i, j) in (0..ai_lo).zip(extension_factor - ai_lo..extension_factor) {
248 (0..cols).for_each(|k| {
249 module.svp_apply_dft_to_dft(&mut vmp_xai, 0, &x_pow_a[ai_hi + 1], 0, &vmp_res[j], k);
250 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i], k, &vmp_xai, 0);
251 module.vec_znx_dft_sub_inplace(&mut acc_add_dft[i], k, &vmp_res[i], k);
252 });
253 }
254 }
255
256 // Sets acc_add_dft[ai_lo..extension_factor] += (acc[0..extension_factor - ai_lo] * sk) * X^{-ai}
257 if ai_hi != 0 {
258 // Sets acc_add_dft[ai_lo..extension_factor] += (acc[0..extension_factor - ai_lo] * sk) * X^{-ai}
259 for (i, j) in (ai_lo..extension_factor).zip(0..extension_factor - ai_lo) {
260 (0..cols).for_each(|k| {
261 module.svp_apply_dft_to_dft(&mut vmp_xai, 0, &x_pow_a[ai_hi], 0, &vmp_res[j], k);
262 module.vec_znx_dft_add_inplace(&mut acc_add_dft[i], k, &vmp_xai, 0);
263 module.vec_znx_dft_sub_inplace(&mut acc_add_dft[i], k, &vmp_res[i], k);
264 });
265 }
266 }
267 }
268 });
269
270 {
271 let (mut acc_add_big, scratch7) = scratch_5.take_vec_znx_big(n_glwe, 1, brk.size());
272
273 (0..extension_factor).for_each(|j| {
274 (0..cols).for_each(|i| {
275 module.vec_znx_idft_apply(&mut acc_add_big, 0, &acc_add_dft[j], i, scratch7);
276 module.vec_znx_big_add_small_inplace(&mut acc_add_big, 0, &acc[j], i);
277 module.vec_znx_big_normalize(base2k, &mut acc[j], i, base2k, &acc_add_big, 0, scratch7);
278 });
279 });
280 }
281 });
282
283 (0..cols).for_each(|i| {
284 module.vec_znx_copy(res.data_mut(), i, &acc[0], i);
285 });
286}
287
288fn execute_block_binary<DataRes, DataIn, DataBrk, B: Backend>(
289 module: &Module<B>,
290 res: &mut GLWE<DataRes>,
291 lwe: &LWE<DataIn>,
292 lut: &LookUpTable,
293 brk: &BlindRotationKeyPrepared<DataBrk, CGGI, B>,
294 scratch: &mut Scratch<B>,
295) where
296 DataRes: DataMut,
297 DataIn: DataRef,
298 DataBrk: DataRef,
299 Module<B>: VecZnxBigAllocBytes
300 + VecZnxDftAllocBytes
301 + SvpPPolAllocBytes
302 + VmpApplyDftToDftTmpBytes
303 + VecZnxBigNormalizeTmpBytes
304 + VecZnxIdftApplyTmpBytes
305 + VecZnxIdftApply<B>
306 + VecZnxDftAdd<B>
307 + VecZnxDftAddInplace<B>
308 + VecZnxDftApply<B>
309 + VecZnxDftZero<B>
310 + SvpApplyDftToDft<B>
311 + VecZnxDftSubInplace<B>
312 + VecZnxBigAddSmallInplace<B>
313 + VecZnxRotate
314 + VecZnxAddInplace
315 + VecZnxSubInplace
316 + VecZnxNormalize<B>
317 + VecZnxNormalizeInplace<B>
318 + VecZnxCopy
319 + VecZnxMulXpMinusOneInplace<B>
320 + VmpApplyDftToDft<B>
321 + VecZnxBigNormalize<B>,
322 Scratch<B>: TakeVecZnxDftSlice<B> + TakeVecZnxDft<B> + TakeVecZnxBig<B> + TakeVecZnxSlice + ScratchAvailable + TakeVecZnx,
323{
324 let n_glwe: usize = brk.n_glwe().into();
325 let mut lwe_2n: Vec<i64> = vec![0i64; (lwe.n() + 1).into()]; // TODO: from scratch space
326 let mut out_mut: GLWE<&mut [u8]> = res.to_mut();
327 let lwe_ref: LWE<&[u8]> = lwe.to_ref();
328 let two_n: usize = n_glwe << 1;
329 let base2k: usize = brk.base2k().into();
330 let dnum: usize = brk.dnum().into();
331
332 let cols: usize = (out_mut.rank() + 1).into();
333
334 mod_switch_2n(
335 2 * lut.domain_size(),
336 &mut lwe_2n,
337 &lwe_ref,
338 lut.rotation_direction(),
339 );
340
341 let a: &[i64] = &lwe_2n[1..];
342 let b: i64 = lwe_2n[0];
343
344 out_mut.data_mut().zero();
345
346 // Initialize out to X^{b} * LUT(X)
347 module.vec_znx_rotate(b, out_mut.data_mut(), 0, &lut.data[0], 0);
348
349 let block_size: usize = brk.block_size();
350
351 // ACC + [sum DFT(X^ai -1) * (DFT(ACC) x BRKi)]
352
353 let (mut acc_dft, scratch_1) = scratch.take_vec_znx_dft(n_glwe, cols, dnum);
354 let (mut vmp_res, scratch_2) = scratch_1.take_vec_znx_dft(n_glwe, cols, brk.size());
355 let (mut acc_add_dft, scratch_3) = scratch_2.take_vec_znx_dft(n_glwe, cols, brk.size());
356 let (mut vmp_xai, scratch_4) = scratch_3.take_vec_znx_dft(n_glwe, 1, brk.size());
357
358 let x_pow_a: &Vec<SvpPPol<Vec<u8>, B>>;
359 if let Some(b) = &brk.x_pow_a {
360 x_pow_a = b
361 } else {
362 panic!("invalid key: x_pow_a has not been initialized")
363 }
364
365 izip!(
366 a.chunks_exact(block_size),
367 brk.data.chunks_exact(block_size)
368 )
369 .for_each(|(ai, ski)| {
370 (0..cols).for_each(|j| {
371 module.vec_znx_dft_apply(1, 0, &mut acc_dft, j, out_mut.data_mut(), j);
372 });
373
374 module.vec_znx_dft_zero(&mut acc_add_dft);
375
376 izip!(ai.iter(), ski.iter()).for_each(|(aii, skii)| {
377 let ai_pos: usize = ((aii + two_n as i64) & (two_n - 1) as i64) as usize;
378
379 // vmp_res = DFT(acc) * BRK[i]
380 module.vmp_apply_dft_to_dft(&mut vmp_res, &acc_dft, skii.data(), scratch_4);
381
382 // DFT(X^ai -1) * (DFT(acc) * BRK[i])
383 (0..cols).for_each(|i| {
384 module.svp_apply_dft_to_dft(&mut vmp_xai, 0, &x_pow_a[ai_pos], 0, &vmp_res, i);
385 module.vec_znx_dft_add_inplace(&mut acc_add_dft, i, &vmp_xai, 0);
386 module.vec_znx_dft_sub_inplace(&mut acc_add_dft, i, &vmp_res, i);
387 });
388 });
389
390 {
391 let (mut acc_add_big, scratch_5) = scratch_4.take_vec_znx_big(n_glwe, 1, brk.size());
392
393 (0..cols).for_each(|i| {
394 module.vec_znx_idft_apply(&mut acc_add_big, 0, &acc_add_dft, i, scratch_5);
395 module.vec_znx_big_add_small_inplace(&mut acc_add_big, 0, out_mut.data_mut(), i);
396 module.vec_znx_big_normalize(
397 base2k,
398 out_mut.data_mut(),
399 i,
400 base2k,
401 &acc_add_big,
402 0,
403 scratch_5,
404 );
405 });
406 }
407 });
408}
409
410fn execute_standard<DataRes, DataIn, DataBrk, B: Backend>(
411 module: &Module<B>,
412 res: &mut GLWE<DataRes>,
413 lwe: &LWE<DataIn>,
414 lut: &LookUpTable,
415 brk: &BlindRotationKeyPrepared<DataBrk, CGGI, B>,
416 scratch: &mut Scratch<B>,
417) where
418 DataRes: DataMut,
419 DataIn: DataRef,
420 DataBrk: DataRef,
421 Module<B>: VecZnxBigAllocBytes
422 + VecZnxDftAllocBytes
423 + SvpPPolAllocBytes
424 + VmpApplyDftToDftTmpBytes
425 + VecZnxBigNormalizeTmpBytes
426 + VecZnxIdftApplyTmpBytes
427 + VecZnxIdftApply<B>
428 + VecZnxDftAdd<B>
429 + VecZnxDftAddInplace<B>
430 + VecZnxDftApply<B>
431 + VecZnxDftZero<B>
432 + SvpApplyDftToDft<B>
433 + VecZnxDftSubInplace<B>
434 + VecZnxBigAddSmallInplace<B>
435 + VecZnxRotate
436 + VecZnxAddInplace
437 + VecZnxSubInplace
438 + VecZnxNormalize<B>
439 + VecZnxNormalizeInplace<B>
440 + VecZnxCopy
441 + VecZnxMulXpMinusOneInplace<B>
442 + VmpApplyDftToDft<B>
443 + VmpApplyDftToDftAdd<B>
444 + VecZnxIdftApplyConsume<B>
445 + VecZnxBigNormalize<B>
446 + VecZnxNormalizeTmpBytes,
447 Scratch<B>: TakeVecZnxDftSlice<B> + TakeVecZnxDft<B> + TakeVecZnxBig<B> + TakeVecZnxSlice + ScratchAvailable + TakeVecZnx,
448{
449 #[cfg(debug_assertions)]
450 {
451 assert_eq!(
452 res.n(),
453 brk.n(),
454 "res.n(): {} != brk.n(): {}",
455 res.n(),
456 brk.n()
457 );
458 assert_eq!(
459 lut.domain_size(),
460 brk.n_glwe().as_usize(),
461 "lut.n(): {} != brk.n(): {}",
462 lut.domain_size(),
463 brk.n_glwe().as_usize()
464 );
465 assert_eq!(
466 res.rank(),
467 brk.rank(),
468 "res.rank(): {} != brk.rank(): {}",
469 res.rank(),
470 brk.rank()
471 );
472 assert_eq!(
473 lwe.n(),
474 brk.n_lwe(),
475 "lwe.n(): {} != brk.data.len(): {}",
476 lwe.n(),
477 brk.n_lwe()
478 );
479 }
480
481 let mut lwe_2n: Vec<i64> = vec![0i64; (lwe.n() + 1).into()]; // TODO: from scratch space
482 let mut out_mut: GLWE<&mut [u8]> = res.to_mut();
483 let lwe_ref: LWE<&[u8]> = lwe.to_ref();
484
485 mod_switch_2n(
486 2 * lut.domain_size(),
487 &mut lwe_2n,
488 &lwe_ref,
489 lut.rotation_direction(),
490 );
491
492 let a: &[i64] = &lwe_2n[1..];
493 let b: i64 = lwe_2n[0];
494
495 out_mut.data_mut().zero();
496
497 // Initialize out to X^{b} * LUT(X)
498 module.vec_znx_rotate(b, out_mut.data_mut(), 0, &lut.data[0], 0);
499
500 // ACC + [sum DFT(X^ai -1) * (DFT(ACC) x BRKi)]
501 let (mut acc_tmp, scratch_1) = scratch.take_glwe_ct(&out_mut);
502
503 // TODO: see if faster by skipping normalization in external product and keeping acc in big coeffs
504 // TODO: first iteration can be optimized to be a gglwe product
505 izip!(a.iter(), brk.data.iter()).for_each(|(ai, ski)| {
506 // acc_tmp = sk[i] * acc
507 acc_tmp.external_product(module, &out_mut, ski, scratch_1);
508
509 // acc_tmp = (sk[i] * acc) * (X^{ai} - 1)
510 acc_tmp.mul_xp_minus_one_inplace(module, *ai, scratch_1);
511
512 // acc = acc + (sk[i] * acc) * (X^{ai} - 1)
513 out_mut.add_inplace(module, &acc_tmp);
514 });
515
516 // We can normalize only at the end because we add normalized values in [-2^{base2k-1}, 2^{base2k-1}]
517 // on top of each others, thus ~ 2^{63-base2k} additions are supported before overflow.
518 out_mut.normalize_inplace(module, scratch_1);
519}
520
521pub fn mod_switch_2n(n: usize, res: &mut [i64], lwe: &LWE<&[u8]>, rot_dir: LookUpTableRotationDirection) {
522 let base2k: usize = lwe.base2k().into();
523
524 let log2n: usize = usize::BITS as usize - (n - 1).leading_zeros() as usize + 1;
525
526 res.copy_from_slice(lwe.data().at(0, 0));
527
528 match rot_dir {
529 LookUpTableRotationDirection::Left => {
530 res.iter_mut().for_each(|x| *x = -*x);
531 }
532 LookUpTableRotationDirection::Right => {}
533 }
534
535 if base2k > log2n {
536 let diff: usize = base2k - (log2n - 1); // additional -1 because we map to [-N/2, N/2) instead of [0, N)
537 res.iter_mut().for_each(|x| {
538 *x = div_round_by_pow2(x, diff);
539 })
540 } else {
541 let rem: usize = base2k - (log2n % base2k);
542 let size: usize = log2n.div_ceil(base2k);
543 (1..size).for_each(|i| {
544 if i == size - 1 && rem != base2k {
545 let k_rem: usize = base2k - rem;
546 izip!(lwe.data().at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
547 *y = (*y << k_rem) + (x >> rem);
548 });
549 } else {
550 izip!(lwe.data().at(0, i).iter(), res.iter_mut()).for_each(|(x, y)| {
551 *y = (*y << base2k) + x;
552 });
553 }
554 })
555 }
556}
557
558#[inline(always)]
559fn div_round_by_pow2(x: &i64, k: usize) -> i64 {
560 (x + (1 << (k - 1))) >> k
561}