poulpy/poulpy-core/src/glwe_packing.rs
poulpy-core/src/glwe_packing.rsRaw
1use std::collections::HashMap;
2
3use poulpy_hal::{
4 api::{
5 ScratchAvailable, VecZnxAddInplace, VecZnxAutomorphismInplace, VecZnxBigAddSmallInplace, VecZnxBigAutomorphismInplace,
6 VecZnxBigNormalize, VecZnxBigNormalizeTmpBytes, VecZnxBigSubSmallNegateInplace, VecZnxCopy, VecZnxDftApply,
7 VecZnxDftBytesOf, VecZnxDftCopy, VecZnxIdftApplyConsume, VecZnxIdftApplyTmpA, VecZnxNegateInplace, VecZnxNormalize,
8 VecZnxNormalizeInplace, VecZnxNormalizeTmpBytes, VecZnxRotate, VecZnxRotateInplace, VecZnxRshInplace, VecZnxSub,
9 VecZnxSubInplace, VecZnxSwitchRing, VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
10 },
11 layouts::{Backend, DataMut, DataRef, Module, Scratch},
12};
13
14use crate::{
15 GLWEOperations,
16 layouts::{GGLWEInfos, GLWE, GLWEInfos, LWEInfos, prepared::AutomorphismKeyPrepared},
17};
18
19/// [GLWEPacker] enables only the fly GLWE packing
20/// with constant memory of Log(N) ciphertexts.
21/// Main difference with usual GLWE packing is that
22/// the output is bit-reversed.
23pub struct GLWEPacker {
24 accumulators: Vec<Accumulator>,
25 log_batch: usize,
26 counter: usize,
27}
28
29/// [Accumulator] stores intermediate packing result.
30/// There are Log(N) such accumulators in a [GLWEPacker].
31struct Accumulator {
32 data: GLWE<Vec<u8>>,
33 value: bool, // Implicit flag for zero ciphertext
34 control: bool, // Can be combined with incoming value
35}
36
37impl Accumulator {
38 /// Allocates a new [Accumulator].
39 ///
40 /// #Arguments
41 ///
42 /// * `module`: static backend FFT tables.
43 /// * `base2k`: base 2 logarithm of the GLWE ciphertext in memory digit representation.
44 /// * `k`: base 2 precision of the GLWE ciphertext precision over the Torus.
45 /// * `rank`: rank of the GLWE ciphertext.
46 pub fn alloc<A, B: Backend>(module: &Module<B>, infos: &A) -> Self
47 where
48 A: GLWEInfos,
49 {
50 Self {
51 data: GLWE::alloc_from_infos(module, infos),
52 value: false,
53 control: false,
54 }
55 }
56}
57
58impl GLWEPacker {
59 /// Instantiates a new [GLWEPacker].
60 ///
61 /// # Arguments
62 ///
63 /// * `module`: static backend FFT tables.
64 /// * `log_batch`: packs coefficients which are multiples of X^{N/2^log_batch}.
65 /// i.e. with `log_batch=0` only the constant coefficient is packed
66 /// and N GLWE ciphertext can be packed. With `log_batch=2` all coefficients
67 /// which are multiples of X^{N/4} are packed. Meaning that N/4 ciphertexts
68 /// can be packed.
69 pub fn new<A, B: Backend>(module: Module<B>, infos: &A, log_batch: usize) -> Self
70 where
71 A: GLWEInfos,
72 {
73 let mut accumulators: Vec<Accumulator> = Vec::<Accumulator>::new();
74 let log_n: usize = infos.n().log2();
75 (0..log_n - log_batch).for_each(|_| accumulators.push(Accumulator::alloc(module, infos)));
76 Self {
77 accumulators,
78 log_batch,
79 counter: 0,
80 }
81 }
82
83 /// Implicit reset of the internal state (to be called before a new packing procedure).
84 fn reset(&mut self) {
85 for i in 0..self.accumulators.len() {
86 self.accumulators[i].value = false;
87 self.accumulators[i].control = false;
88 }
89 self.counter = 0;
90 }
91
92 /// Number of scratch space bytes required to call [Self::add].
93 pub fn scratch_space<B: Backend, OUT, KEY>(module: &Module<B>, out_infos: &OUT, key_infos: &KEY) -> usize
94 where
95 OUT: GLWEInfos,
96 KEY: GGLWEInfos,
97 Module<B>: VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigNormalizeTmpBytes + VecZnxNormalizeTmpBytes,
98 {
99 pack_core_scratch_space(module, out_infos, key_infos)
100 }
101
102 pub fn galois_elements<B: Backend>(module: &Module<B>) -> Vec<i64> {
103 GLWE::trace_galois_elements(module)
104 }
105
106 /// Adds a GLWE ciphertext to the [GLWEPacker].
107 /// #Arguments
108 ///
109 /// * `module`: static backend FFT tables.
110 /// * `res`: space to append fully packed ciphertext. Only when the number
111 /// of packed ciphertexts reaches N/2^log_batch is a result written.
112 /// * `a`: ciphertext to pack. Can optionally give None to pack a 0 ciphertext.
113 /// * `auto_keys`: a [HashMap] containing the [AutomorphismKeyExec]s.
114 /// * `scratch`: scratch space of size at least [Self::scratch_space].
115 pub fn add<DataA: DataRef, DataAK: DataRef, B: Backend>(
116 &mut self,
117 module: &Module<B>,
118 a: Option<&GLWE<DataA>>,
119 auto_keys: &HashMap<i64, AutomorphismKeyPrepared<DataAK, B>>,
120 scratch: &mut Scratch<B>,
121 ) where
122 Module<B>: VecZnxDftBytesOf
123 + VmpApplyDftToDftTmpBytes
124 + VecZnxBigNormalizeTmpBytes
125 + VmpApplyDftToDft<B>
126 + VmpApplyDftToDftAdd<B>
127 + VecZnxDftApply<B>
128 + VecZnxIdftApplyConsume<B>
129 + VecZnxBigAddSmallInplace<B>
130 + VecZnxBigNormalize<B>
131 + VecZnxCopy
132 + VecZnxRotateInplace<B>
133 + VecZnxSub
134 + VecZnxNegateInplace
135 + VecZnxRshInplace<B>
136 + VecZnxAddInplace
137 + VecZnxNormalizeInplace<B>
138 + VecZnxSubInplace
139 + VecZnxRotate
140 + VecZnxAutomorphismInplace<B>
141 + VecZnxBigSubSmallNegateInplace<B>
142 + VecZnxBigAutomorphismInplace<B>
143 + VecZnxNormalize<B>
144 + VecZnxNormalizeTmpBytes,
145 Scratch<B>: ScratchAvailable,
146 {
147 assert!(
148 (self.counter as u32) < self.accumulators[0].data.n(),
149 "Packing limit of {} reached",
150 self.accumulators[0].data.n().0 as usize >> self.log_batch
151 );
152
153 pack_core(
154 module,
155 a,
156 &mut self.accumulators,
157 self.log_batch,
158 auto_keys,
159 scratch,
160 );
161 self.counter += 1 << self.log_batch;
162 }
163
164 /// Flush result to`res`.
165 pub fn flush<Data: DataMut, B: Backend>(&mut self, module: &Module<B>, res: &mut GLWE<Data>)
166 where
167 Module<B>: VecZnxCopy,
168 {
169 assert!(self.counter as u32 == self.accumulators[0].data.n());
170 // Copy result GLWE into res GLWE
171 res.copy(
172 module,
173 &self.accumulators[module.log_n() - self.log_batch - 1].data,
174 );
175
176 self.reset();
177 }
178}
179
180fn pack_core_scratch_space<B: Backend, OUT, KEY>(module: &Module<B>, out_infos: &OUT, key_infos: &KEY) -> usize
181where
182 OUT: GLWEInfos,
183 KEY: GGLWEInfos,
184 Module<B>: VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigNormalizeTmpBytes + VecZnxNormalizeTmpBytes,
185{
186 combine_scratch_space(module, out_infos, key_infos)
187}
188
189fn pack_core<D: DataRef, DataAK: DataRef, B: Backend>(
190 module: &Module<B>,
191 a: Option<&GLWE<D>>,
192 accumulators: &mut [Accumulator],
193 i: usize,
194 auto_keys: &HashMap<i64, AutomorphismKeyPrepared<DataAK, B>>,
195 scratch: &mut Scratch<B>,
196) where
197 Module<B>: VecZnxDftBytesOf
198 + VmpApplyDftToDftTmpBytes
199 + VecZnxBigNormalizeTmpBytes
200 + VmpApplyDftToDft<B>
201 + VmpApplyDftToDftAdd<B>
202 + VecZnxDftApply<B>
203 + VecZnxIdftApplyConsume<B>
204 + VecZnxBigAddSmallInplace<B>
205 + VecZnxBigNormalize<B>
206 + VecZnxCopy
207 + VecZnxRotateInplace<B>
208 + VecZnxSub
209 + VecZnxNegateInplace
210 + VecZnxRshInplace<B>
211 + VecZnxAddInplace
212 + VecZnxNormalizeInplace<B>
213 + VecZnxSubInplace
214 + VecZnxRotate
215 + VecZnxAutomorphismInplace<B>
216 + VecZnxBigSubSmallNegateInplace<B>
217 + VecZnxBigAutomorphismInplace<B>
218 + VecZnxNormalize<B>
219 + VecZnxNormalizeTmpBytes,
220 Scratch<B>: ScratchAvailable,
221{
222 let log_n: usize = module.log_n();
223
224 if i == log_n {
225 return;
226 }
227
228 // Isolate the first accumulator
229 let (acc_prev, acc_next) = accumulators.split_at_mut(1);
230
231 // Control = true accumlator is free to overide
232 if !acc_prev[0].control {
233 let acc_mut_ref: &mut Accumulator = &mut acc_prev[0]; // from split_at_mut
234
235 // No previous value -> copies and sets flags accordingly
236 if let Some(a_ref) = a {
237 acc_mut_ref.data.copy(module, a_ref);
238 acc_mut_ref.value = true
239 } else {
240 acc_mut_ref.value = false
241 }
242 acc_mut_ref.control = true; // Able to be combined on next call
243 } else {
244 // Compresses acc_prev <- combine(acc_prev, a).
245 combine(module, &mut acc_prev[0], a, i, auto_keys, scratch);
246 acc_prev[0].control = false;
247
248 // Propagates to next accumulator
249 if acc_prev[0].value {
250 pack_core(
251 module,
252 Some(&acc_prev[0].data),
253 acc_next,
254 i + 1,
255 auto_keys,
256 scratch,
257 );
258 } else {
259 pack_core(
260 module,
261 None::<&GLWE<Vec<u8>>>,
262 acc_next,
263 i + 1,
264 auto_keys,
265 scratch,
266 );
267 }
268 }
269}
270
271fn combine_scratch_space<B: Backend, OUT, KEY>(module: &Module<B>, out_infos: &OUT, key_infos: &KEY) -> usize
272where
273 OUT: GLWEInfos,
274 KEY: GGLWEInfos,
275 Module<B>: VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigNormalizeTmpBytes + VecZnxNormalizeTmpBytes,
276{
277 GLWE::bytes_of_from_infos(module, out_infos)
278 + (GLWE::rsh_scratch_space(module.n()) | GLWE::automorphism_inplace_scratch_space(module, out_infos, key_infos))
279}
280
281/// [combine] merges two ciphertexts together.
282fn combine<D: DataRef, DataAK: DataRef, B: Backend>(
283 module: &Module<B>,
284 acc: &mut Accumulator,
285 b: Option<&GLWE<D>>,
286 i: usize,
287 auto_keys: &HashMap<i64, AutomorphismKeyPrepared<DataAK, B>>,
288 scratch: &mut Scratch<B>,
289) where
290 Module<B>: VecZnxDftBytesOf
291 + VmpApplyDftToDftTmpBytes
292 + VecZnxBigNormalizeTmpBytes
293 + VmpApplyDftToDft<B>
294 + VmpApplyDftToDftAdd<B>
295 + VecZnxDftApply<B>
296 + VecZnxIdftApplyConsume<B>
297 + VecZnxBigAddSmallInplace<B>
298 + VecZnxBigNormalize<B>
299 + VecZnxCopy
300 + VecZnxRotateInplace<B>
301 + VecZnxSub
302 + VecZnxNegateInplace
303 + VecZnxRshInplace<B>
304 + VecZnxAddInplace
305 + VecZnxNormalizeInplace<B>
306 + VecZnxSubInplace
307 + VecZnxRotate
308 + VecZnxAutomorphismInplace<B>
309 + VecZnxBigSubSmallNegateInplace<B>
310 + VecZnxBigAutomorphismInplace<B>
311 + VecZnxNormalize<B>
312 + VecZnxNormalizeTmpBytes,
313 Scratch<B>: ScratchAvailable,
314{
315 let log_n: usize = acc.data.n().log2();
316 let a: &mut GLWE<Vec<u8>> = &mut acc.data;
317
318 let gal_el: i64 = if i == 0 {
319 -1
320 } else {
321 module.galois_element(1 << (i - 1))
322 };
323
324 let t: i64 = 1 << (log_n - i - 1);
325
326 // Goal is to evaluate: a = a + b*X^t + phi(a - b*X^t))
327 // We also use the identity: AUTO(a * X^t, g) = -X^t * AUTO(a, g)
328 // where t = 2^(log_n - i - 1) and g = 5^{2^(i - 1)}
329 // Different cases for wether a and/or b are zero.
330 //
331 // Implicite RSH without modulus switch, introduces extra I(X) * Q/2 on decryption.
332 // Necessary so that the scaling of the plaintext remains constant.
333 // It however is ok to do so here because coefficients are eventually
334 // either mapped to garbage or twice their value which vanishes I(X)
335 // since 2*(I(X) * Q/2) = I(X) * Q = 0 mod Q.
336 if acc.value {
337 if let Some(b) = b {
338 let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(a);
339
340 // a = a * X^-t
341 a.rotate_inplace(module, -t, scratch_1);
342
343 // tmp_b = a * X^-t - b
344 tmp_b.sub(module, a, b);
345 tmp_b.rsh(module, 1, scratch_1);
346
347 // a = a * X^-t + b
348 a.add_inplace(module, b);
349 a.rsh(module, 1, scratch_1);
350
351 tmp_b.normalize_inplace(module, scratch_1);
352
353 // tmp_b = phi(a * X^-t - b)
354 if let Some(key) = auto_keys.get(&gal_el) {
355 tmp_b.automorphism_inplace(module, key, scratch_1);
356 } else {
357 panic!("auto_key[{gal_el}] not found");
358 }
359
360 // a = a * X^-t + b - phi(a * X^-t - b)
361 a.sub_inplace_ab(module, &tmp_b);
362 a.normalize_inplace(module, scratch_1);
363
364 // a = a + b * X^t - phi(a * X^-t - b) * X^t
365 // = a + b * X^t - phi(a * X^-t - b) * - phi(X^t)
366 // = a + b * X^t + phi(a - b * X^t)
367 a.rotate_inplace(module, t, scratch_1);
368 } else {
369 a.rsh(module, 1, scratch);
370 // a = a + phi(a)
371 if let Some(key) = auto_keys.get(&gal_el) {
372 a.automorphism_add_inplace(module, key, scratch);
373 } else {
374 panic!("auto_key[{gal_el}] not found");
375 }
376 }
377 } else if let Some(b) = b {
378 let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(a);
379 tmp_b.rotate(module, 1 << (log_n - i - 1), b);
380 tmp_b.rsh(module, 1, scratch_1);
381
382 // a = (b* X^t - phi(b* X^t))
383 if let Some(key) = auto_keys.get(&gal_el) {
384 a.automorphism_sub_negate(module, &tmp_b, key, scratch_1);
385 } else {
386 panic!("auto_key[{gal_el}] not found");
387 }
388
389 acc.value = true;
390 }
391}
392
393/// Packs [x_0: GLWE(m_0), x_1: GLWE(m_1), ..., x_i: GLWE(m_i)]
394/// to [0: GLWE(m_0 * X^x_0 + m_1 * X^x_1 + ... + m_i * X^x_i)]
395pub fn glwe_packing<D: DataMut, ATK, B: Backend>(
396 module: &Module<B>,
397 cts: &mut HashMap<usize, &mut GLWE<D>>,
398 log_gap_out: usize,
399 auto_keys: &HashMap<i64, AutomorphismKeyPrepared<ATK, B>>,
400 scratch: &mut Scratch<B>,
401) where
402 ATK: DataRef,
403 Module<B>: VecZnxRotateInplace<B>
404 + VecZnxNormalizeInplace<B>
405 + VecZnxNormalizeTmpBytes
406 + VecZnxSwitchRing
407 + VecZnxBigAutomorphismInplace<B>
408 + VecZnxRshInplace<B>
409 + VecZnxDftCopy<B>
410 + VecZnxIdftApplyTmpA<B>
411 + VecZnxSub
412 + VecZnxAddInplace
413 + VecZnxNegateInplace
414 + VecZnxCopy
415 + VecZnxSubInplace
416 + VecZnxDftBytesOf
417 + VmpApplyDftToDftTmpBytes
418 + VecZnxBigNormalizeTmpBytes
419 + VmpApplyDftToDft<B>
420 + VmpApplyDftToDftAdd<B>
421 + VecZnxDftApply<B>
422 + VecZnxIdftApplyConsume<B>
423 + VecZnxBigAddSmallInplace<B>
424 + VecZnxBigNormalize<B>
425 + VecZnxAutomorphismInplace<B>
426 + VecZnxBigSubSmallNegateInplace<B>
427 + VecZnxRotate
428 + VecZnxNormalize<B>,
429 Scratch<B>: ScratchAvailable,
430{
431 #[cfg(debug_assertions)]
432 {
433 assert!(*cts.keys().max().unwrap() < module.n())
434 }
435
436 let log_n: usize = module.log_n();
437
438 (0..log_n - log_gap_out).for_each(|i| {
439 let t: usize = (1 << log_n).min(1 << (log_n - 1 - i));
440
441 let auto_key: &AutomorphismKeyPrepared<ATK, B> = if i == 0 {
442 auto_keys.get(&-1).unwrap()
443 } else {
444 auto_keys.get(&module.galois_element(1 << (i - 1))).unwrap()
445 };
446
447 (0..t).for_each(|j| {
448 let mut a: Option<&mut GLWE<D>> = cts.remove(&j);
449 let mut b: Option<&mut GLWE<D>> = cts.remove(&(j + t));
450
451 pack_internal(module, &mut a, &mut b, i, auto_key, scratch);
452
453 if let Some(a) = a {
454 cts.insert(j, a);
455 } else if let Some(b) = b {
456 cts.insert(j, b);
457 }
458 });
459 });
460}
461
462#[allow(clippy::too_many_arguments)]
463fn pack_internal<A: DataMut, D: DataMut, DataAK: DataRef, B: Backend>(
464 module: &Module<B>,
465 a: &mut Option<&mut GLWE<A>>,
466 b: &mut Option<&mut GLWE<D>>,
467 i: usize,
468 auto_key: &AutomorphismKeyPrepared<DataAK, B>,
469 scratch: &mut Scratch<B>,
470) where
471 Module<B>: VecZnxRotateInplace<B>
472 + VecZnxNormalizeInplace<B>
473 + VecZnxNormalizeTmpBytes
474 + VecZnxBigAutomorphismInplace<B>
475 + VecZnxRshInplace<B>
476 + VecZnxDftCopy<B>
477 + VecZnxIdftApplyTmpA<B>
478 + VecZnxSub
479 + VecZnxAddInplace
480 + VecZnxNegateInplace
481 + VecZnxCopy
482 + VecZnxSubInplace
483 + VecZnxDftBytesOf
484 + VmpApplyDftToDftTmpBytes
485 + VecZnxBigNormalizeTmpBytes
486 + VmpApplyDftToDft<B>
487 + VmpApplyDftToDftAdd<B>
488 + VecZnxDftApply<B>
489 + VecZnxIdftApplyConsume<B>
490 + VecZnxBigAddSmallInplace<B>
491 + VecZnxBigNormalize<B>
492 + VecZnxAutomorphismInplace<B>
493 + VecZnxBigSubSmallNegateInplace<B>
494 + VecZnxRotate
495 + VecZnxNormalize<B>,
496 Scratch<B>: ScratchAvailable,
497{
498 // Goal is to evaluate: a = a + b*X^t + phi(a - b*X^t))
499 // We also use the identity: AUTO(a * X^t, g) = -X^t * AUTO(a, g)
500 // where t = 2^(log_n - i - 1) and g = 5^{2^(i - 1)}
501 // Different cases for wether a and/or b are zero.
502 //
503 // Implicite RSH without modulus switch, introduces extra I(X) * Q/2 on decryption.
504 // Necessary so that the scaling of the plaintext remains constant.
505 // It however is ok to do so here because coefficients are eventually
506 // either mapped to garbage or twice their value which vanishes I(X)
507 // since 2*(I(X) * Q/2) = I(X) * Q = 0 mod Q.
508 if let Some(a) = a.as_deref_mut() {
509 let t: i64 = 1 << (a.n().log2() - i - 1);
510
511 if let Some(b) = b.as_deref_mut() {
512 let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(a);
513
514 // a = a * X^-t
515 a.rotate_inplace(module, -t, scratch_1);
516
517 // tmp_b = a * X^-t - b
518 tmp_b.sub(module, a, b);
519 tmp_b.rsh(module, 1, scratch_1);
520
521 // a = a * X^-t + b
522 a.add_inplace(module, b);
523 a.rsh(module, 1, scratch_1);
524
525 tmp_b.normalize_inplace(module, scratch_1);
526
527 // tmp_b = phi(a * X^-t - b)
528 tmp_b.automorphism_inplace(module, auto_key, scratch_1);
529
530 // a = a * X^-t + b - phi(a * X^-t - b)
531 a.sub_inplace_ab(module, &tmp_b);
532 a.normalize_inplace(module, scratch_1);
533
534 // a = a + b * X^t - phi(a * X^-t - b) * X^t
535 // = a + b * X^t - phi(a * X^-t - b) * - phi(X^t)
536 // = a + b * X^t + phi(a - b * X^t)
537 a.rotate_inplace(module, t, scratch_1);
538 } else {
539 a.rsh(module, 1, scratch);
540 // a = a + phi(a)
541 a.automorphism_add_inplace(module, auto_key, scratch);
542 }
543 } else if let Some(b) = b.as_deref_mut() {
544 let t: i64 = 1 << (b.n().log2() - i - 1);
545
546 let (mut tmp_b, scratch_1) = scratch.take_glwe_ct(b);
547 tmp_b.rotate(module, t, b);
548 tmp_b.rsh(module, 1, scratch_1);
549
550 // a = (b* X^t - phi(b* X^t))
551 b.automorphism_sub_negate(module, &tmp_b, auto_key, scratch_1);
552 }
553}