poulpy/poulpy-core/src/keyswitching/glwe.rs
poulpy-core/src/keyswitching/glwe.rsRaw
1use poulpy_hal::{
2 api::{
3 ModuleN, ScratchAvailable, ScratchTakeBasic, VecZnxBigAddSmallInplace, VecZnxBigNormalize, VecZnxBigNormalizeTmpBytes,
4 VecZnxDftApply, VecZnxDftBytesOf, VecZnxDftCopy, VecZnxIdftApplyConsume, VecZnxNormalize, VecZnxNormalizeTmpBytes,
5 VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
6 },
7 layouts::{Backend, DataMut, DataViewMut, Module, Scratch, VecZnxBig, VecZnxDft, VecZnxDftToRef, VmpPMat, ZnxInfos},
8};
9
10use crate::{
11 GLWENormalize, ScratchTakeCore,
12 layouts::{GGLWEInfos, GGLWEPrepared, GGLWEPreparedToRef, GLWE, GLWEInfos, GLWELayout, GLWEToMut, GLWEToRef, LWEInfos},
13};
14
15impl GLWE<Vec<u8>> {
16 pub fn keyswitch_tmp_bytes<M, R, A, B, BE: Backend>(module: &M, res_infos: &R, a_infos: &A, key_infos: &B) -> usize
17 where
18 R: GLWEInfos,
19 A: GLWEInfos,
20 B: GGLWEInfos,
21 M: GLWEKeyswitch<BE>,
22 {
23 module.glwe_keyswitch_tmp_bytes(res_infos, a_infos, key_infos)
24 }
25}
26
27impl<D: DataMut> GLWE<D> {
28 pub fn keyswitch<A, B, M, BE: Backend>(&mut self, module: &M, a: &A, b: &B, scratch: &mut Scratch<BE>)
29 where
30 A: GLWEToRef,
31 B: GGLWEPreparedToRef<BE>,
32 M: GLWEKeyswitch<BE>,
33 Scratch<BE>: ScratchTakeCore<BE>,
34 {
35 module.glwe_keyswitch(self, a, b, scratch);
36 }
37
38 pub fn keyswitch_inplace<A, M, BE: Backend>(&mut self, module: &M, a: &A, scratch: &mut Scratch<BE>)
39 where
40 A: GGLWEPreparedToRef<BE>,
41 M: GLWEKeyswitch<BE>,
42 Scratch<BE>: ScratchTakeCore<BE>,
43 {
44 module.glwe_keyswitch_inplace(self, a, scratch);
45 }
46}
47
48impl<BE: Backend> GLWEKeyswitch<BE> for Module<BE>
49where
50 Self: Sized + GLWEKeySwitchInternal<BE> + VecZnxBigNormalizeTmpBytes + VecZnxBigNormalize<BE> + GLWENormalize<BE>,
51 Scratch<BE>: ScratchTakeCore<BE>,
52{
53 fn glwe_keyswitch_tmp_bytes<R, A, B>(&self, res_infos: &R, a_infos: &A, key_infos: &B) -> usize
54 where
55 R: GLWEInfos,
56 A: GLWEInfos,
57 B: GGLWEInfos,
58 {
59 let cols: usize = res_infos.rank().as_usize() + 1;
60 let size: usize = if a_infos.base2k() != key_infos.base2k() {
61 let a_conv_infos = &GLWELayout {
62 n: a_infos.n(),
63 base2k: key_infos.base2k(),
64 k: a_infos.k(),
65 rank: a_infos.rank(),
66 };
67 self.glwe_keyswitch_internal_tmp_bytes(res_infos, a_conv_infos, key_infos) + GLWE::bytes_of_from_infos(a_conv_infos)
68 } else {
69 self.glwe_keyswitch_internal_tmp_bytes(res_infos, a_infos, key_infos)
70 };
71
72 size.max(self.vec_znx_big_normalize_tmp_bytes()) + self.bytes_of_vec_znx_dft(cols, key_infos.size())
73 }
74
75 fn glwe_keyswitch<R, A, K>(&self, res: &mut R, a: &A, key: &K, scratch: &mut Scratch<BE>)
76 where
77 R: GLWEToMut,
78 A: GLWEToRef,
79 K: GGLWEPreparedToRef<BE>,
80 {
81 let res: &mut GLWE<&mut [u8]> = &mut res.to_mut();
82 let a: &GLWE<&[u8]> = &a.to_ref();
83 let key: &GGLWEPrepared<&[u8], BE> = &key.to_ref();
84
85 assert_eq!(
86 a.rank(),
87 key.rank_in(),
88 "a.rank(): {} != b.rank_in(): {}",
89 a.rank(),
90 key.rank_in()
91 );
92 assert_eq!(
93 res.rank(),
94 key.rank_out(),
95 "res.rank(): {} != b.rank_out(): {}",
96 res.rank(),
97 key.rank_out()
98 );
99
100 assert_eq!(res.n(), self.n() as u32);
101 assert_eq!(a.n(), self.n() as u32);
102 assert_eq!(key.n(), self.n() as u32);
103
104 let scrach_needed: usize = self.glwe_keyswitch_tmp_bytes(res, a, key);
105
106 assert!(
107 scratch.available() >= scrach_needed,
108 "scratch.available()={} < glwe_keyswitch_tmp_bytes={scrach_needed}",
109 scratch.available(),
110 );
111
112 let base2k_a: usize = a.base2k().into();
113 let base2k_key: usize = key.base2k().into();
114 let base2k_res: usize = res.base2k().into();
115
116 let (res_dft, scratch_1) = scratch.take_vec_znx_dft(self, (res.rank() + 1).into(), key.size()); // Todo optimise
117
118 let res_big: VecZnxBig<&mut [u8], BE> = if base2k_a != base2k_key {
119 let (mut a_conv, scratch_2) = scratch_1.take_glwe(&GLWELayout {
120 n: a.n(),
121 base2k: key.base2k(),
122 k: a.k(),
123 rank: a.rank(),
124 });
125 self.glwe_normalize(&mut a_conv, a, scratch_2);
126 self.glwe_keyswitch_internal(res_dft, &a_conv, key, scratch_2)
127 } else {
128 self.glwe_keyswitch_internal(res_dft, a, key, scratch_1)
129 };
130
131 for i in 0..(res.rank() + 1).into() {
132 self.vec_znx_big_normalize(
133 base2k_res,
134 &mut res.data,
135 i,
136 base2k_key,
137 &res_big,
138 i,
139 scratch_1,
140 );
141 }
142 }
143
144 fn glwe_keyswitch_inplace<R, K>(&self, res: &mut R, key: &K, scratch: &mut Scratch<BE>)
145 where
146 R: GLWEToMut,
147 K: GGLWEPreparedToRef<BE>,
148 {
149 let res: &mut GLWE<&mut [u8]> = &mut res.to_mut();
150 let key: &GGLWEPrepared<&[u8], BE> = &key.to_ref();
151
152 assert_eq!(
153 res.rank(),
154 key.rank_in(),
155 "res.rank(): {} != a.rank_in(): {}",
156 res.rank(),
157 key.rank_in()
158 );
159 assert_eq!(
160 res.rank(),
161 key.rank_out(),
162 "res.rank(): {} != b.rank_out(): {}",
163 res.rank(),
164 key.rank_out()
165 );
166
167 assert_eq!(res.n(), self.n() as u32);
168 assert_eq!(key.n(), self.n() as u32);
169
170 let scrach_needed: usize = self.glwe_keyswitch_tmp_bytes(res, res, key);
171
172 assert!(
173 scratch.available() >= scrach_needed,
174 "scratch.available()={} < glwe_keyswitch_tmp_bytes={scrach_needed}",
175 scratch.available(),
176 );
177
178 let base2k_res: usize = res.base2k().as_usize();
179 let base2k_key: usize = key.base2k().as_usize();
180
181 let (res_dft, scratch_1) = scratch.take_vec_znx_dft(self, (res.rank() + 1).into(), key.size()); // Todo optimise
182
183 let res_big: VecZnxBig<&mut [u8], BE> = if base2k_res != base2k_key {
184 let (mut res_conv, scratch_2) = scratch_1.take_glwe(&GLWELayout {
185 n: res.n(),
186 base2k: key.base2k(),
187 k: res.k(),
188 rank: res.rank(),
189 });
190 self.glwe_normalize(&mut res_conv, res, scratch_2);
191
192 self.glwe_keyswitch_internal(res_dft, &res_conv, key, scratch_2)
193 } else {
194 self.glwe_keyswitch_internal(res_dft, res, key, scratch_1)
195 };
196
197 for i in 0..(res.rank() + 1).into() {
198 self.vec_znx_big_normalize(
199 base2k_res,
200 res.data_mut(),
201 i,
202 base2k_key,
203 &res_big,
204 i,
205 scratch_1,
206 );
207 }
208 }
209}
210
211pub trait GLWEKeyswitch<BE: Backend> {
212 fn glwe_keyswitch_tmp_bytes<R, A, B>(&self, res_infos: &R, a_infos: &A, key_infos: &B) -> usize
213 where
214 R: GLWEInfos,
215 A: GLWEInfos,
216 B: GGLWEInfos;
217
218 fn glwe_keyswitch<R, A, K>(&self, res: &mut R, a: &A, key: &K, scratch: &mut Scratch<BE>)
219 where
220 R: GLWEToMut,
221 A: GLWEToRef,
222 K: GGLWEPreparedToRef<BE>;
223
224 fn glwe_keyswitch_inplace<R, K>(&self, res: &mut R, key: &K, scratch: &mut Scratch<BE>)
225 where
226 R: GLWEToMut,
227 K: GGLWEPreparedToRef<BE>;
228}
229
230impl<BE: Backend> GLWEKeySwitchInternal<BE> for Module<BE> where
231 Self: GGLWEProduct<BE>
232 + VecZnxDftApply<BE>
233 + VecZnxNormalize<BE>
234 + VecZnxIdftApplyConsume<BE>
235 + VecZnxBigAddSmallInplace<BE>
236 + VecZnxNormalizeTmpBytes
237{
238}
239
240pub(crate) trait GLWEKeySwitchInternal<BE: Backend>
241where
242 Self: GGLWEProduct<BE>
243 + VecZnxDftApply<BE>
244 + VecZnxNormalize<BE>
245 + VecZnxIdftApplyConsume<BE>
246 + VecZnxBigAddSmallInplace<BE>
247 + VecZnxNormalizeTmpBytes,
248{
249 fn glwe_keyswitch_internal_tmp_bytes<R, A, K>(&self, res_infos: &R, a_infos: &A, key_infos: &K) -> usize
250 where
251 R: GLWEInfos,
252 A: GLWEInfos,
253 K: GGLWEInfos,
254 {
255 let cols: usize = (a_infos.rank() + 1).into();
256 let a_size: usize = a_infos.size();
257 self.gglwe_product_dft_tmp_bytes(res_infos.size(), a_size, key_infos) + self.bytes_of_vec_znx_dft(cols - 1, a_size)
258 }
259
260 fn glwe_keyswitch_internal<DR, A, K>(
261 &self,
262 mut res: VecZnxDft<DR, BE>,
263 a: &A,
264 key: &K,
265 scratch: &mut Scratch<BE>,
266 ) -> VecZnxBig<DR, BE>
267 where
268 DR: DataMut,
269 A: GLWEToRef,
270 K: GGLWEPreparedToRef<BE>,
271 Scratch<BE>: ScratchTakeCore<BE>,
272 {
273 let a: &GLWE<&[u8]> = &a.to_ref();
274 let key: &GGLWEPrepared<&[u8], BE> = &key.to_ref();
275 assert_eq!(a.base2k(), key.base2k());
276 let cols: usize = (a.rank() + 1).into();
277 let a_size: usize = a.size();
278 let (mut a_dft, scratch_1) = scratch.take_vec_znx_dft(self, cols - 1, a_size);
279 for col_i in 0..cols - 1 {
280 self.vec_znx_dft_apply(1, 0, &mut a_dft, col_i, a.data(), col_i + 1);
281 }
282 self.gglwe_product_dft(&mut res, &a_dft, key, scratch_1);
283 let mut res_big: VecZnxBig<DR, BE> = self.vec_znx_idft_apply_consume(res);
284 self.vec_znx_big_add_small_inplace(&mut res_big, 0, a.data(), 0);
285 res_big
286 }
287}
288
289impl<BE: Backend> GGLWEProduct<BE> for Module<BE> where
290 Self: Sized
291 + ModuleN
292 + VecZnxDftBytesOf
293 + VmpApplyDftToDftTmpBytes
294 + VmpApplyDftToDft<BE>
295 + VmpApplyDftToDftAdd<BE>
296 + VecZnxDftCopy<BE>
297{
298}
299
300pub(crate) trait GGLWEProduct<BE: Backend>
301where
302 Self: Sized
303 + ModuleN
304 + VecZnxDftBytesOf
305 + VmpApplyDftToDftTmpBytes
306 + VmpApplyDftToDft<BE>
307 + VmpApplyDftToDftAdd<BE>
308 + VecZnxDftCopy<BE>,
309{
310 fn gglwe_product_dft_tmp_bytes<K>(&self, res_size: usize, a_size: usize, key_infos: &K) -> usize
311 where
312 K: GGLWEInfos,
313 {
314 let dsize: usize = key_infos.dsize().as_usize();
315
316 if dsize == 1 {
317 self.vmp_apply_dft_to_dft_tmp_bytes(
318 res_size,
319 a_size,
320 key_infos.dnum().into(),
321 (key_infos.rank_in()).into(),
322 (key_infos.rank_out() + 1).into(),
323 key_infos.size(),
324 )
325 } else {
326 let dnum: usize = key_infos.dnum().into();
327 let a_size: usize = a_size.div_ceil(dsize).min(dnum);
328 let ai_dft: usize = self.bytes_of_vec_znx_dft(key_infos.rank_in().into(), a_size);
329
330 let vmp: usize = self.vmp_apply_dft_to_dft_tmp_bytes(
331 res_size,
332 a_size,
333 dnum,
334 (key_infos.rank_in()).into(),
335 (key_infos.rank_out() + 1).into(),
336 key_infos.size(),
337 );
338
339 ai_dft + vmp
340 }
341 }
342
343 fn gglwe_product_dft<DR, A, K>(&self, res: &mut VecZnxDft<DR, BE>, a: &A, key: &K, scratch: &mut Scratch<BE>)
344 where
345 DR: DataMut,
346 A: VecZnxDftToRef<BE>,
347 K: GGLWEPreparedToRef<BE>,
348 Scratch<BE>: ScratchTakeCore<BE>,
349 {
350 let a: &VecZnxDft<&[u8], BE> = &a.to_ref();
351 let key: &GGLWEPrepared<&[u8], BE> = &key.to_ref();
352
353 let cols: usize = a.cols();
354 let a_size: usize = a.size();
355 let pmat: &VmpPMat<&[u8], BE> = &key.data;
356
357 // If dsize == 1, then the digit decomposition is equal to Base2K and we can simply
358 // can the vmp API.
359 if key.dsize() == 1 {
360 self.vmp_apply_dft_to_dft(res, a, pmat, scratch);
361 // If dsize != 1, then the digit decomposition is k * Base2K with k > 1.
362 // As such we need to perform a bivariate polynomial convolution in (X, Y) / (X^{N}+1) with Y = 2^-K
363 // (instead of yn univariate one in X).
364 //
365 // Since the basis in Y is small (in practice degree 6-7 max), we perform it naiveley.
366 // To do so, we group the different limbs of ai_dft by their respective degree in Y
367 // which are multiples of the current digit.
368 // For example if dsize = 3, with ai_dft = [a0, a1, a2, a3, a4, a5, a6],
369 // we group them as [[a0, a3, a5], [a1, a4, a6], [a2, a5, 0]]
370 // and evaluate sum(a_di * pmat * 2^{di*Base2k})
371 } else {
372 let dsize: usize = key.dsize().into();
373 let dnum: usize = key.dnum().into();
374
375 // We bound ai_dft size by the number of rows of the matrix
376 let (mut ai_dft, scratch_1) = scratch.take_vec_znx_dft(self, cols, a_size.div_ceil(dsize).min(dnum));
377 ai_dft.data_mut().fill(0);
378
379 for di in 0..dsize {
380 // Sets ai_dft size according to the current digit (if dsize does not divides a_size),
381 // bounded by the number of rows (digits) in the prepared matrix.
382 ai_dft.set_size(((a_size + di) / dsize).min(dnum));
383
384 // Small optimization for dsize > 2
385 // VMP produce some error e, and since we aggregate vmp * 2^{di * Base2k}, then
386 // we also aggregate ei * 2^{di * Base2k}, with the largest error being ei * 2^{(dsize-1) * Base2k}.
387 // As such we can ignore the last dsize-2 limbs safely of the sum of vmp products.
388 // It is possible to further ignore the last dsize-1 limbs, but this introduce
389 // ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
390 // noise is kept with respect to the ideal functionality.
391 res.set_size(pmat.size() - ((dsize - di) as isize - 2).max(0) as usize);
392
393 for j in 0..cols {
394 self.vec_znx_dft_copy(dsize, dsize - di - 1, &mut ai_dft, j, a, j);
395 }
396
397 if di == 0 {
398 // res = pmat * ai_dft
399 self.vmp_apply_dft_to_dft(res, &ai_dft, pmat, scratch_1);
400 } else {
401 // res = (pmat * ai_dft) * 2^{di * Base2k}
402 self.vmp_apply_dft_to_dft_add(res, &ai_dft, pmat, di, scratch_1);
403 }
404 }
405
406 res.set_size(res.max_size());
407 }
408 }
409}