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