poulpy/poulpy-core/src/conversion/gglwe_to_ggsw.rs
poulpy-core/src/conversion/gglwe_to_ggsw.rsRaw
1use poulpy_hal::{
2 api::{
3 ModuleN, ScratchAvailable, ScratchTakeBasic, VecZnxBigBytesOf, VecZnxBigNormalize, VecZnxCopy, VecZnxDftAddInplace,
4 VecZnxDftApply, VecZnxDftBytesOf, VecZnxDftCopy, VecZnxIdftApplyTmpA, VecZnxNormalize, VecZnxNormalizeTmpBytes,
5 VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftTmpBytes,
6 },
7 layouts::{Backend, DataMut, Module, Scratch, VmpPMat, ZnxInfos},
8};
9
10use crate::{
11 ScratchTakeCore,
12 layouts::{
13 GGLWE, GGLWEInfos, GGLWEToRef, GGSW, GGSWInfos, GGSWToMut, GLWEInfos, LWEInfos,
14 prepared::{TensorKeyPrepared, TensorKeyPreparedToRef},
15 },
16 operations::GLWEOperations,
17};
18
19impl GGLWE<Vec<u8>> {
20 pub fn from_gglw_tmp_bytes<R, A, M, BE: Backend>(module: &M, res_infos: &R, tsk_infos: &A) -> usize
21 where
22 M: GGSWFromGGLWE<BE>,
23 R: GGSWInfos,
24 A: GGLWEInfos,
25 {
26 module.ggsw_from_gglwe_tmp_bytes(res_infos, tsk_infos)
27 }
28}
29
30impl<D: DataMut> GGSW<D> {
31 pub fn from_gglwe<G, M, T, BE: Backend>(&mut self, module: &M, gglwe: &G, tsk: &T, scratch: &mut Scratch<BE>)
32 where
33 M: GGSWFromGGLWE<BE>,
34 G: GGLWEToRef,
35 T: TensorKeyPreparedToRef<BE>,
36 Scratch<BE>: ScratchTakeCore<BE>,
37 {
38 module.ggsw_from_gglwe(self, gglwe, tsk, scratch);
39 }
40}
41
42impl<BE: Backend> GGSWFromGGLWE<BE> for Module<BE> where Self: GGSWExpandRows<BE> + VecZnxCopy {}
43
44pub trait GGSWFromGGLWE<BE: Backend>
45where
46 Self: GGSWExpandRows<BE> + VecZnxCopy,
47{
48 fn ggsw_from_gglwe_tmp_bytes<R, A>(&self, res_infos: &R, tsk_infos: &A) -> usize
49 where
50 R: GGSWInfos,
51 A: GGLWEInfos,
52 {
53 self.ggsw_expand_rows_tmp_bytes(res_infos, tsk_infos)
54 }
55
56 fn ggsw_from_gglwe<R, A, T>(&self, res: &mut R, a: &A, tsk: &T, scratch: &mut Scratch<BE>)
57 where
58 R: GGSWToMut,
59 A: GGLWEToRef,
60 T: TensorKeyPreparedToRef<BE>,
61 Scratch<BE>: ScratchTakeCore<BE>,
62 {
63 let res: &mut GGSW<&mut [u8]> = &mut res.to_mut();
64 let a: &GGLWE<&[u8]> = &a.to_ref();
65 let tsk: &TensorKeyPrepared<&[u8], BE> = &tsk.to_ref();
66
67 assert_eq!(res.rank(), a.rank_out());
68 assert_eq!(res.dnum(), a.dnum());
69 assert_eq!(res.n(), self.n() as u32);
70 assert_eq!(a.n(), self.n() as u32);
71 assert_eq!(tsk.n(), self.n() as u32);
72
73 for row in 0..res.dnum().into() {
74 res.at_mut(row, 0).copy(self, &a.at(row, 0));
75 }
76
77 self.ggsw_expand_row(res, tsk, scratch);
78 }
79}
80
81impl<BE: Backend> GGSWExpandRows<BE> for Module<BE> where
82 Self: Sized
83 + ModuleN
84 + VecZnxDftBytesOf
85 + VmpApplyDftToDftTmpBytes
86 + VecZnxBigBytesOf
87 + VecZnxNormalizeTmpBytes
88 + VecZnxDftBytesOf
89 + VmpApplyDftToDftTmpBytes
90 + VecZnxBigBytesOf
91 + VecZnxNormalizeTmpBytes
92 + VecZnxDftApply<BE>
93 + VecZnxDftCopy<BE>
94 + VmpApplyDftToDft<BE>
95 + VmpApplyDftToDftAdd<BE>
96 + VecZnxDftAddInplace<BE>
97 + VecZnxBigNormalize<BE>
98 + VecZnxIdftApplyTmpA<BE>
99 + VecZnxNormalize<BE>
100{
101}
102
103pub(crate) trait GGSWExpandRows<BE: Backend>
104where
105 Self: Sized
106 + ModuleN
107 + VecZnxDftBytesOf
108 + VmpApplyDftToDftTmpBytes
109 + VecZnxBigBytesOf
110 + VecZnxNormalizeTmpBytes
111 + VecZnxDftBytesOf
112 + VmpApplyDftToDftTmpBytes
113 + VecZnxBigBytesOf
114 + VecZnxNormalizeTmpBytes
115 + VecZnxDftApply<BE>
116 + VecZnxDftCopy<BE>
117 + VmpApplyDftToDft<BE>
118 + VmpApplyDftToDftAdd<BE>
119 + VecZnxDftAddInplace<BE>
120 + VecZnxBigNormalize<BE>
121 + VecZnxIdftApplyTmpA<BE>
122 + VecZnxNormalize<BE>,
123{
124 fn ggsw_expand_rows_tmp_bytes<R, A>(&self, res_infos: &R, tsk_infos: &A) -> usize
125 where
126 R: GGSWInfos,
127 A: GGLWEInfos,
128 {
129 let tsk_size: usize = tsk_infos.k().div_ceil(tsk_infos.base2k()) as usize;
130 let size_in: usize = res_infos
131 .k()
132 .div_ceil(tsk_infos.base2k())
133 .div_ceil(tsk_infos.dsize().into()) as usize;
134
135 let tmp_dft_i: usize = self.bytes_of_vec_znx_dft((tsk_infos.rank_out() + 1).into(), tsk_size);
136 let tmp_a: usize = self.bytes_of_vec_znx_dft(1, size_in);
137 let vmp: usize = self.vmp_apply_dft_to_dft_tmp_bytes(
138 tsk_size,
139 size_in,
140 size_in,
141 (tsk_infos.rank_in()).into(), // Verify if rank+1
142 (tsk_infos.rank_out()).into(), // Verify if rank+1
143 tsk_size,
144 );
145 let tmp_idft: usize = self.bytes_of_vec_znx_big(1, tsk_size);
146 let norm: usize = self.vec_znx_normalize_tmp_bytes();
147
148 tmp_dft_i + ((tmp_a + vmp) | (tmp_idft + norm))
149 }
150
151 fn ggsw_expand_row<R, T>(&self, res: &mut R, tsk: &T, scratch: &mut Scratch<BE>)
152 where
153 R: GGSWToMut,
154 T: TensorKeyPreparedToRef<BE>,
155 Scratch<BE>: ScratchTakeCore<BE>,
156 {
157 let res: &mut GGSW<&mut [u8]> = &mut res.to_mut();
158 let tsk: &TensorKeyPrepared<&[u8], BE> = &tsk.to_ref();
159
160 let basek_in: usize = res.base2k().into();
161 let basek_tsk: usize = tsk.base2k().into();
162
163 assert!(scratch.available() >= self.ggsw_expand_rows_tmp_bytes(res, tsk));
164
165 let rank: usize = res.rank().into();
166 let cols: usize = rank + 1;
167
168 let a_size: usize = (res.size() * basek_in).div_ceil(basek_tsk);
169
170 // Keyswitch the j-th row of the col 0
171 for row_i in 0..res.dnum().into() {
172 let a = &res.at(row_i, 0).data;
173
174 // Pre-compute DFT of (a0, a1, a2)
175 let (mut ci_dft, scratch_1) = scratch.take_vec_znx_dft(self, cols, a_size);
176
177 if basek_in == basek_tsk {
178 for i in 0..cols {
179 self.vec_znx_dft_apply(1, 0, &mut ci_dft, i, a, i);
180 }
181 } else {
182 let (mut a_conv, scratch_2) = scratch_1.take_vec_znx(self, 1, a_size);
183 for i in 0..cols {
184 self.vec_znx_normalize(basek_tsk, &mut a_conv, 0, basek_in, a, i, scratch_2);
185 self.vec_znx_dft_apply(1, 0, &mut ci_dft, i, &a_conv, 0);
186 }
187 }
188
189 for col_j in 1..cols {
190 // Example for rank 3:
191 //
192 // Note: M is a vector (m, Bm, B^2m, B^3m, ...), so each column is
193 // actually composed of that many dnum and we focus on a specific row here
194 // implicitely given ci_dft.
195 //
196 // # Input
197 //
198 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0 , a1 , a2 )
199 // col 1: (0, 0, 0, 0)
200 // col 2: (0, 0, 0, 0)
201 // col 3: (0, 0, 0, 0)
202 //
203 // # Output
204 //
205 // col 0: (-(a0s0 + a1s1 + a2s2) + M[i], a0 , a1 , a2 )
206 // col 1: (-(b0s0 + b1s1 + b2s2) , b0 + M[i], b1 , b2 )
207 // col 2: (-(c0s0 + c1s1 + c2s2) , c0 , c1 + M[i], c2 )
208 // col 3: (-(d0s0 + d1s1 + d2s2) , d0 , d1 , d2 + M[i])
209
210 let dsize: usize = tsk.dsize().into();
211
212 let (mut tmp_dft_i, scratch_2) = scratch_1.take_vec_znx_dft(self, cols, tsk.size());
213 let (mut tmp_a, scratch_3) = scratch_2.take_vec_znx_dft(self, 1, ci_dft.size().div_ceil(dsize));
214
215 {
216 // Performs a key-switch for each combination of s[i]*s[j], i.e. for a0, a1, a2
217 //
218 // # Example for col=1
219 //
220 // a0 * (-(f0s0 + f1s1 + f1s2) + s0^2, f0, f1, f2) = (-(a0f0s0 + a0f1s1 + a0f1s2) + a0s0^2, a0f0, a0f1, a0f2)
221 // +
222 // a1 * (-(g0s0 + g1s1 + g1s2) + s0s1, g0, g1, g2) = (-(a1g0s0 + a1g1s1 + a1g1s2) + a1s0s1, a1g0, a1g1, a1g2)
223 // +
224 // a2 * (-(h0s0 + h1s1 + h1s2) + s0s2, h0, h1, h2) = (-(a2h0s0 + a2h1s1 + a2h1s2) + a2s0s2, a2h0, a2h1, a2h2)
225 // =
226 // (-(x0s0 + x1s1 + x2s2) + s0(a0s0 + a1s1 + a2s2), x0, x1, x2)
227 for col_i in 1..cols {
228 let pmat: &VmpPMat<&[u8], BE> = &tsk.at(col_i - 1, col_j - 1).key.data; // Selects Enc(s[i]s[j])
229
230 // Extracts a[i] and multipies with Enc(s[i]s[j])
231 for di in 0..dsize {
232 tmp_a.set_size((ci_dft.size() + di) / dsize);
233
234 // Small optimization for dsize > 2
235 // VMP produce some error e, and since we aggregate vmp * 2^{di * B}, then
236 // we also aggregate ei * 2^{di * B}, with the largest error being ei * 2^{(dsize-1) * B}.
237 // As such we can ignore the last dsize-2 limbs safely of the sum of vmp products.
238 // It is possible to further ignore the last dsize-1 limbs, but this introduce
239 // ~0.5 to 1 bit of additional noise, and thus not chosen here to ensure that the same
240 // noise is kept with respect to the ideal functionality.
241 tmp_dft_i.set_size(tsk.size() - ((dsize - di) as isize - 2).max(0) as usize);
242
243 self.vec_znx_dft_copy(dsize, dsize - 1 - di, &mut tmp_a, 0, &ci_dft, col_i);
244 if di == 0 && col_i == 1 {
245 self.vmp_apply_dft_to_dft(&mut tmp_dft_i, &tmp_a, pmat, scratch_3);
246 } else {
247 self.vmp_apply_dft_to_dft_add(&mut tmp_dft_i, &tmp_a, pmat, di, scratch_3);
248 }
249 }
250 }
251 }
252
253 // Adds -(sum a[i] * s[i]) + m) on the i-th column of tmp_idft_i
254 //
255 // (-(x0s0 + x1s1 + x2s2) + a0s0s0 + a1s0s1 + a2s0s2, x0, x1, x2)
256 // +
257 // (0, -(a0s0 + a1s1 + a2s2) + M[i], 0, 0)
258 // =
259 // (-(x0s0 + x1s1 + x2s2) + s0(a0s0 + a1s1 + a2s2), x0 -(a0s0 + a1s1 + a2s2) + M[i], x1, x2)
260 // =
261 // (-(x0s0 + x1s1 + x2s2), x0 + M[i], x1, x2)
262 self.vec_znx_dft_add_inplace(&mut tmp_dft_i, col_j, &ci_dft, 0);
263 let (mut tmp_idft, scratch_3) = scratch_2.take_vec_znx_big(self, 1, tsk.size());
264 for i in 0..cols {
265 self.vec_znx_idft_apply_tmpa(&mut tmp_idft, 0, &mut tmp_dft_i, i);
266 self.vec_znx_big_normalize(
267 basek_in,
268 &mut res.at_mut(row_i, col_j).data,
269 i,
270 basek_tsk,
271 &tmp_idft,
272 0,
273 scratch_3,
274 );
275 }
276 }
277 }
278 }
279}