poulpy/poulpy-core/src/keyswitching/ggsw_ct.rs
poulpy-core/src/keyswitching/ggsw_ct.rsRaw
1use poulpy_hal::{
2 api::{
3 ScratchAvailable, VecZnxBigAddSmallInplace, VecZnxBigBytesOf, VecZnxBigNormalize, VecZnxBigNormalizeTmpBytes, VecZnxCopy,
4 VecZnxDftAddInplace, VecZnxDftApply, VecZnxDftBytesOf, VecZnxDftCopy, VecZnxIdftApplyConsume, VecZnxIdftApplyTmpA,
5 VecZnxNormalize, VecZnxNormalizeTmpBytes, VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
6 },
7 layouts::{Backend, DataMut, DataRef, Module, Scratch, VecZnx, VmpPMat, ZnxInfos},
8};
9
10use crate::{
11 layouts::{
12 GGLWE, GGLWEInfos, GGSW, GGSWInfos, GLWE, GLWEInfos, LWEInfos,
13 prepared::{GLWESwitchingKeyPrepared, TensorKeyPrepared},
14 },
15 operations::GLWEOperations,
16};
17
18impl GGSW<Vec<u8>> {
19 pub(crate) fn expand_row_scratch_space<B: Backend, OUT, TSK>(module: &Module<B>, out_infos: &OUT, tsk_infos: &TSK) -> usize
20 where
21 OUT: GGSWInfos,
22 TSK: GGLWEInfos,
23 Module<B>: VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigBytesOf + VecZnxNormalizeTmpBytes,
24 {
25 let tsk_size: usize = tsk_infos.k().div_ceil(tsk_infos.base2k()) as usize;
26 let size_in: usize = out_infos
27 .k()
28 .div_ceil(tsk_infos.base2k())
29 .div_ceil(tsk_infos.dsize().into()) as usize;
30
31 let tmp_dft_i: usize = module.bytes_of_vec_znx_dft((tsk_infos.rank_out() + 1).into(), tsk_size);
32 let tmp_a: usize = module.bytes_of_vec_znx_dft(1, size_in);
33 let vmp: usize = module.vmp_apply_dft_to_dft_tmp_bytes(
34 tsk_size,
35 size_in,
36 size_in,
37 (tsk_infos.rank_in()).into(), // Verify if rank+1
38 (tsk_infos.rank_out()).into(), // Verify if rank+1
39 tsk_size,
40 );
41 let tmp_idft: usize = module.bytes_of_vec_znx_big(1, tsk_size);
42 let norm: usize = module.vec_znx_normalize_tmp_bytes();
43
44 tmp_dft_i + ((tmp_a + vmp) | (tmp_idft + norm))
45 }
46
47 #[allow(clippy::too_many_arguments)]
48 pub fn keyswitch_scratch_space<B: Backend, OUT, IN, KEY, TSK>(
49 module: &Module<B>,
50 out_infos: &OUT,
51 in_infos: &IN,
52 apply_infos: &KEY,
53 tsk_infos: &TSK,
54 ) -> usize
55 where
56 OUT: GGSWInfos,
57 IN: GGSWInfos,
58 KEY: GGLWEInfos,
59 TSK: GGLWEInfos,
60 Module<B>:
61 VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigBytesOf + VecZnxNormalizeTmpBytes + VecZnxBigNormalizeTmpBytes,
62 {
63 #[cfg(debug_assertions)]
64 {
65 assert_eq!(apply_infos.rank_in(), apply_infos.rank_out());
66 assert_eq!(tsk_infos.rank_in(), tsk_infos.rank_out());
67 assert_eq!(apply_infos.rank_in(), tsk_infos.rank_in());
68 }
69
70 let rank: usize = apply_infos.rank_out().into();
71
72 let size_out: usize = out_infos.k().div_ceil(out_infos.base2k()) as usize;
73 let res_znx: usize = VecZnx::bytes_of(module.n(), rank + 1, size_out);
74 let ci_dft: usize = module.bytes_of_vec_znx_dft(rank + 1, size_out);
75 let ks: usize = GLWE::keyswitch_scratch_space(module, out_infos, in_infos, apply_infos);
76 let expand_rows: usize = GGSW::expand_row_scratch_space(module, out_infos, tsk_infos);
77 let res_dft: usize = module.bytes_of_vec_znx_dft(rank + 1, size_out);
78
79 if in_infos.base2k() == tsk_infos.base2k() {
80 res_znx + ci_dft + (ks | expand_rows | res_dft)
81 } else {
82 let a_conv: usize = VecZnx::bytes_of(
83 module.n(),
84 1,
85 out_infos.k().div_ceil(tsk_infos.base2k()) as usize,
86 ) + module.vec_znx_normalize_tmp_bytes();
87 res_znx + ci_dft + (a_conv | ks | expand_rows | res_dft)
88 }
89 }
90
91 #[allow(clippy::too_many_arguments)]
92 pub fn keyswitch_inplace_scratch_space<B: Backend, OUT, KEY, TSK>(
93 module: &Module<B>,
94 out_infos: &OUT,
95 apply_infos: &KEY,
96 tsk_infos: &TSK,
97 ) -> usize
98 where
99 OUT: GGSWInfos,
100 KEY: GGLWEInfos,
101 TSK: GGLWEInfos,
102 Module<B>:
103 VecZnxDftBytesOf + VmpApplyDftToDftTmpBytes + VecZnxBigBytesOf + VecZnxNormalizeTmpBytes + VecZnxBigNormalizeTmpBytes,
104 {
105 GGSW::keyswitch_scratch_space(module, out_infos, out_infos, apply_infos, tsk_infos)
106 }
107}
108
109impl<DataSelf: DataMut> GGSW<DataSelf> {
110 pub fn from_gglwe<DataA, DataTsk, B: Backend>(
111 &mut self,
112 module: &Module<B>,
113 a: &GGLWE<DataA>,
114 tsk: &TensorKeyPrepared<DataTsk, B>,
115 scratch: &mut Scratch<B>,
116 ) where
117 DataA: DataRef,
118 DataTsk: DataRef,
119 Module<B>: VecZnxCopy
120 + VecZnxDftBytesOf
121 + VmpApplyDftToDftTmpBytes
122 + VecZnxBigBytesOf
123 + VecZnxNormalizeTmpBytes
124 + VecZnxDftApply<B>
125 + VecZnxDftCopy<B>
126 + VmpApplyDftToDft<B>
127 + VmpApplyDftToDftAdd<B>
128 + VecZnxDftAddInplace<B>
129 + VecZnxBigNormalize<B>
130 + VecZnxIdftApplyTmpA<B>
131 + VecZnxNormalize<B>,
132 Scratch<B>: ScratchAvailable,
133 {
134 #[cfg(debug_assertions)]
135 {
136 use crate::layouts::{GLWEInfos, LWEInfos};
137
138 assert_eq!(self.rank(), a.rank_out());
139 assert_eq!(self.dnum(), a.dnum());
140 assert_eq!(self.n(), module.n() as u32);
141 assert_eq!(a.n(), module.n() as u32);
142 assert_eq!(tsk.n(), module.n() as u32);
143 }
144 (0..self.dnum().into()).for_each(|row_i| {
145 self.at_mut(row_i, 0).copy(module, &a.at(row_i, 0));
146 });
147 self.expand_row(module, tsk, scratch);
148 }
149
150 pub fn keyswitch<DataLhs: DataRef, DataKsk: DataRef, DataTsk: DataRef, B: Backend>(
151 &mut self,
152 module: &Module<B>,
153 lhs: &GGSW<DataLhs>,
154 ksk: &GLWESwitchingKeyPrepared<DataKsk, B>,
155 tsk: &TensorKeyPrepared<DataTsk, B>,
156 scratch: &mut Scratch<B>,
157 ) where
158 Module<B>: VecZnxDftBytesOf
159 + VmpApplyDftToDftTmpBytes
160 + VecZnxBigNormalizeTmpBytes
161 + VmpApplyDftToDft<B>
162 + VmpApplyDftToDftAdd<B>
163 + VecZnxDftApply<B>
164 + VecZnxIdftApplyConsume<B>
165 + VecZnxBigAddSmallInplace<B>
166 + VecZnxBigNormalize<B>
167 + VecZnxDftBytesOf
168 + VecZnxBigBytesOf
169 + VecZnxNormalizeTmpBytes
170 + VecZnxDftCopy<B>
171 + VecZnxDftAddInplace<B>
172 + VecZnxIdftApplyTmpA<B>
173 + VecZnxNormalize<B>,
174 Scratch<B>: ScratchAvailable,
175 {
176 (0..lhs.dnum().into()).for_each(|row_i| {
177 // Key-switch column 0, i.e.
178 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0, a1, a2) -> (-(a0s0' + a1s1' + a2s2') + M[i], a0, a1, a2)
179 self.at_mut(row_i, 0)
180 .keyswitch(module, &lhs.at(row_i, 0), ksk, scratch);
181 });
182 self.expand_row(module, tsk, scratch);
183 }
184
185 pub fn keyswitch_inplace<DataKsk: DataRef, DataTsk: DataRef, B: Backend>(
186 &mut self,
187 module: &Module<B>,
188 ksk: &GLWESwitchingKeyPrepared<DataKsk, B>,
189 tsk: &TensorKeyPrepared<DataTsk, B>,
190 scratch: &mut Scratch<B>,
191 ) where
192 Module<B>: VecZnxDftBytesOf
193 + VmpApplyDftToDftTmpBytes
194 + VecZnxBigNormalizeTmpBytes
195 + VmpApplyDftToDft<B>
196 + VmpApplyDftToDftAdd<B>
197 + VecZnxDftApply<B>
198 + VecZnxIdftApplyConsume<B>
199 + VecZnxBigAddSmallInplace<B>
200 + VecZnxBigNormalize<B>
201 + VecZnxDftBytesOf
202 + VecZnxBigBytesOf
203 + VecZnxNormalizeTmpBytes
204 + VecZnxDftCopy<B>
205 + VecZnxDftAddInplace<B>
206 + VecZnxIdftApplyTmpA<B>
207 + VecZnxNormalize<B>,
208 Scratch<B>: ScratchAvailable,
209 {
210 (0..self.dnum().into()).for_each(|row_i| {
211 // Key-switch column 0, i.e.
212 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0, a1, a2) -> (-(a0s0' + a1s1' + a2s2') + M[i], a0, a1, a2)
213 self.at_mut(row_i, 0)
214 .keyswitch_inplace(module, ksk, scratch);
215 });
216 self.expand_row(module, tsk, scratch);
217 }
218
219 pub fn expand_row<DataTsk: DataRef, B: Backend>(
220 &mut self,
221 module: &Module<B>,
222 tsk: &TensorKeyPrepared<DataTsk, B>,
223 scratch: &mut Scratch<B>,
224 ) where
225 Module<B>: VecZnxDftBytesOf
226 + VmpApplyDftToDftTmpBytes
227 + VecZnxBigBytesOf
228 + VecZnxNormalizeTmpBytes
229 + VecZnxDftApply<B>
230 + VecZnxDftCopy<B>
231 + VmpApplyDftToDft<B>
232 + VmpApplyDftToDftAdd<B>
233 + VecZnxDftAddInplace<B>
234 + VecZnxBigNormalize<B>
235 + VecZnxIdftApplyTmpA<B>
236 + VecZnxNormalize<B>,
237 Scratch<B>: ScratchAvailable,
238 {
239 let basek_in: usize = self.base2k().into();
240 let basek_tsk: usize = tsk.base2k().into();
241
242 assert!(scratch.available() >= GGSW::expand_row_scratch_space(module, self, tsk));
243
244 let n: usize = self.n().into();
245 let rank: usize = self.rank().into();
246 let cols: usize = rank + 1;
247
248 let a_size: usize = (self.size() * basek_in).div_ceil(basek_tsk);
249
250 // Keyswitch the j-th row of the col 0
251 for row_i in 0..self.dnum().into() {
252 let a = &self.at(row_i, 0).data;
253
254 // Pre-compute DFT of (a0, a1, a2)
255 let (mut ci_dft, scratch_1) = scratch.take_vec_znx_dft(n, cols, a_size);
256
257 if basek_in == basek_tsk {
258 for i in 0..cols {
259 module.vec_znx_dft_apply(1, 0, &mut ci_dft, i, a, i);
260 }
261 } else {
262 let (mut a_conv, scratch_2) = scratch_1.take_vec_znx(n, 1, a_size);
263 for i in 0..cols {
264 module.vec_znx_normalize(basek_tsk, &mut a_conv, 0, basek_in, a, i, scratch_2);
265 module.vec_znx_dft_apply(1, 0, &mut ci_dft, i, &a_conv, 0);
266 }
267 }
268
269 for col_j in 1..cols {
270 // Example for rank 3:
271 //
272 // Note: M is a vector (m, Bm, B^2m, B^3m, ...), so each column is
273 // actually composed of that many dnum and we focus on a specific row here
274 // implicitely given ci_dft.
275 //
276 // # Input
277 //
278 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0 , a1 , a2 )
279 // col 1: (0, 0, 0, 0)
280 // col 2: (0, 0, 0, 0)
281 // col 3: (0, 0, 0, 0)
282 //
283 // # Output
284 //
285 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0 , a1 , a2 )
286 // col 1: (-(b0s0 + b1s1 + b2s2) , b0 + M[i], b1 , b2 )
287 // col 2: (-(c0s0 + c1s1 + c2s2) , c0 , c1 + M[i], c2 )
288 // col 3: (-(d0s0 + d1s1 + d2s2) , d0 , d1 , d2 + M[i])
289
290 let dsize: usize = tsk.dsize().into();
291
292 let (mut tmp_dft_i, scratch_2) = scratch_1.take_vec_znx_dft(n, cols, tsk.size());
293 let (mut tmp_a, scratch_3) = scratch_2.take_vec_znx_dft(n, 1, ci_dft.size().div_ceil(dsize));
294
295 {
296 // Performs a key-switch for each combination of s[i]*s[j], i.e. for a0, a1, a2
297 //
298 // # Example for col=1
299 //
300 // a0 * (-(f0s0 + f1s1 + f1s2) + s0^2, f0, f1, f2) = (-(a0f0s0 + a0f1s1 + a0f1s2) + a0s0^2, a0f0, a0f1, a0f2)
301 // +
302 // a1 * (-(g0s0 + g1s1 + g1s2) + s0s1, g0, g1, g2) = (-(a1g0s0 + a1g1s1 + a1g1s2) + a1s0s1, a1g0, a1g1, a1g2)
303 // +
304 // a2 * (-(h0s0 + h1s1 + h1s2) + s0s2, h0, h1, h2) = (-(a2h0s0 + a2h1s1 + a2h1s2) + a2s0s2, a2h0, a2h1, a2h2)
305 // =
306 // (-(x0s0 + x1s1 + x2s2) + s0(a0s0 + a1s1 + a2s2), x0, x1, x2)
307 for col_i in 1..cols {
308 let pmat: &VmpPMat<DataTsk, B> = &tsk.at(col_i - 1, col_j - 1).key.data; // Selects Enc(s[i]s[j])
309
310 // Extracts a[i] and multipies with Enc(s[i]s[j])
311 for di in 0..dsize {
312 tmp_a.set_size((ci_dft.size() + di) / dsize);
313
314 // Small optimization for dsize > 2
315 // VMP produce some error e, and since we aggregate vmp * 2^{di * B}, then
316 // we also aggregate ei * 2^{di * B}, with the largest error being ei * 2^{(dsize-1) * B}.
317 // As such we can ignore the last dsize-2 limbs safely of the sum of vmp products.
318 // It is possible to further ignore the last dsize-1 limbs, but this introduce
319 // ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
320 // noise is kept with respect to the ideal functionality.
321 tmp_dft_i.set_size(tsk.size() - ((dsize - di) as isize - 2).max(0) as usize);
322
323 module.vec_znx_dft_copy(dsize, dsize - 1 - di, &mut tmp_a, 0, &ci_dft, col_i);
324 if di == 0 && col_i == 1 {
325 module.vmp_apply_dft_to_dft(&mut tmp_dft_i, &tmp_a, pmat, scratch_3);
326 } else {
327 module.vmp_apply_dft_to_dft_add(&mut tmp_dft_i, &tmp_a, pmat, di, scratch_3);
328 }
329 }
330 }
331 }
332
333 // Adds -(sum a[i] * s[i]) + m) on the i-th column of tmp_idft_i
334 //
335 // (-(x0s0 + x1s1 + x2s2) + a0s0s0 + a1s0s1 + a2s0s2, x0, x1, x2)
336 // +
337 // (0, -(a0s0 + a1s1 + a2s2) + M[i], 0, 0)
338 // =
339 // (-(x0s0 + x1s1 + x2s2) + s0(a0s0 + a1s1 + a2s2), x0 -(a0s0 + a1s1 + a2s2) + M[i], x1, x2)
340 // =
341 // (-(x0s0 + x1s1 + x2s2), x0 + M[i], x1, x2)
342 module.vec_znx_dft_add_inplace(&mut tmp_dft_i, col_j, &ci_dft, 0);
343 let (mut tmp_idft, scratch_3) = scratch_2.take_vec_znx_big(n, 1, tsk.size());
344 for i in 0..cols {
345 module.vec_znx_idft_apply_tmpa(&mut tmp_idft, 0, &mut tmp_dft_i, i);
346 module.vec_znx_big_normalize(
347 basek_in,
348 &mut self.at_mut(row_i, col_j).data,
349 i,
350 basek_tsk,
351 &tmp_idft,
352 0,
353 scratch_3,
354 );
355 }
356 }
357 }
358 }
359}