1use poulpy_hal::{
  2    api::{VmpPMatAlloc, VmpPMatBytesOf, VmpPrepare, VmpPrepareTmpBytes},
  3    layouts::{Backend, Data, DataMut, DataRef, Module, Scratch, VmpPMat, VmpPMatToMut, VmpPMatToRef, ZnxInfos},
  4};
  5
  6use crate::layouts::{
  7    Base2K, Dnum, Dsize, GGSW, GGSWInfos, GGSWToRef, GLWEInfos, GetRingDegree, LWEInfos, Rank, RingDegree, TorusPrecision,
  8};
  9
 10#[derive(PartialEq, Eq)]
 11pub struct GGSWPrepared<D: Data, B: Backend> {
 12    pub(crate) data: VmpPMat<D, B>,
 13    pub(crate) k: TorusPrecision,
 14    pub(crate) base2k: Base2K,
 15    pub(crate) dsize: Dsize,
 16}
 17
 18impl<D: Data, B: Backend> LWEInfos for GGSWPrepared<D, B> {
 19    fn n(&self) -> RingDegree {
 20        RingDegree(self.data.n() as u32)
 21    }
 22
 23    fn base2k(&self) -> Base2K {
 24        self.base2k
 25    }
 26
 27    fn k(&self) -> TorusPrecision {
 28        self.k
 29    }
 30
 31    fn size(&self) -> usize {
 32        self.data.size()
 33    }
 34}
 35
 36impl<D: Data, B: Backend> GLWEInfos for GGSWPrepared<D, B> {
 37    fn rank(&self) -> Rank {
 38        Rank(self.data.cols_out() as u32 - 1)
 39    }
 40}
 41
 42impl<D: Data, B: Backend> GGSWInfos for GGSWPrepared<D, B> {
 43    fn dsize(&self) -> Dsize {
 44        self.dsize
 45    }
 46
 47    fn dnum(&self) -> Dnum {
 48        Dnum(self.data.rows() as u32)
 49    }
 50}
 51
 52pub trait GGSWPreparedAlloc<B: Backend>
 53where
 54    Self: GetRingDegree + VmpPMatAlloc<B> + VmpPMatBytesOf,
 55{
 56    fn alloc_ggsw_prepared(
 57        &self,
 58        base2k: Base2K,
 59        k: TorusPrecision,
 60        dnum: Dnum,
 61        dsize: Dsize,
 62        rank: Rank,
 63    ) -> GGSWPrepared<Vec<u8>, B> {
 64        let size: usize = k.0.div_ceil(base2k.0) as usize;
 65        debug_assert!(
 66            size as u32 > dsize.0,
 67            "invalid ggsw: ceil(k/base2k): {size} <= dsize: {}",
 68            dsize.0
 69        );
 70
 71        assert!(
 72            dnum.0 * dsize.0 <= size as u32,
 73            "invalid ggsw: dnum: {} * dsize:{} > ceil(k/base2k): {size}",
 74            dnum.0,
 75            dsize.0,
 76        );
 77
 78        GGSWPrepared {
 79            data: self.vmp_pmat_alloc(
 80                dnum.into(),
 81                (rank + 1).into(),
 82                (rank + 1).into(),
 83                k.0.div_ceil(base2k.0) as usize,
 84            ),
 85            k,
 86            base2k,
 87            dsize,
 88        }
 89    }
 90
 91    fn alloc_ggsw_prepared_from_infos<A>(&self, infos: &A) -> GGSWPrepared<Vec<u8>, B>
 92    where
 93        A: GGSWInfos,
 94    {
 95        assert_eq!(self.ring_degree(), infos.n());
 96        self.alloc_ggsw_prepared(
 97            infos.base2k(),
 98            infos.k(),
 99            infos.dnum(),
100            infos.dsize(),
101            infos.rank(),
102        )
103    }
104
105    fn bytes_of_ggsw_prepared(&self, base2k: Base2K, k: TorusPrecision, dnum: Dnum, dsize: Dsize, rank: Rank) -> usize {
106        let size: usize = k.0.div_ceil(base2k.0) as usize;
107        debug_assert!(
108            size as u32 > dsize.0,
109            "invalid ggsw: ceil(k/base2k): {size} <= dsize: {}",
110            dsize.0
111        );
112
113        assert!(
114            dnum.0 * dsize.0 <= size as u32,
115            "invalid ggsw: dnum: {} * dsize:{} > ceil(k/base2k): {size}",
116            dnum.0,
117            dsize.0,
118        );
119
120        self.bytes_of_vmp_pmat(dnum.into(), (rank + 1).into(), (rank + 1).into(), size)
121    }
122
123    fn bytes_of_ggsw_prepared_from_infos<A>(&self, infos: &A) -> usize
124    where
125        A: GGSWInfos,
126    {
127        assert_eq!(self.ring_degree(), infos.n());
128        self.bytes_of_ggsw_prepared(
129            infos.base2k(),
130            infos.k(),
131            infos.dnum(),
132            infos.dsize(),
133            infos.rank(),
134        )
135    }
136}
137
138impl<B: Backend> GGSWPreparedAlloc<B> for Module<B> where Self: GetRingDegree + VmpPMatAlloc<B> + VmpPMatBytesOf {}
139
140impl<B: Backend> GGSWPrepared<Vec<u8>, B> {
141    pub fn alloc_from_infos<A, M>(module: &M, infos: &A) -> Self
142    where
143        A: GGSWInfos,
144        M: GGSWPreparedAlloc<B>,
145    {
146        module.alloc_ggsw_prepared_from_infos(infos)
147    }
148
149    pub fn alloc<M>(module: &M, base2k: Base2K, k: TorusPrecision, dnum: Dnum, dsize: Dsize, rank: Rank) -> Self
150    where
151        M: GGSWPreparedAlloc<B>,
152    {
153        module.alloc_ggsw_prepared(base2k, k, dnum, dsize, rank)
154    }
155
156    pub fn bytes_of_from_infos<A, M>(module: &M, infos: &A) -> usize
157    where
158        A: GGSWInfos,
159        M: GGSWPreparedAlloc<B>,
160    {
161        module.bytes_of_ggsw_prepared_from_infos(infos)
162    }
163
164    pub fn bytes_of<M>(module: &M, base2k: Base2K, k: TorusPrecision, dnum: Dnum, dsize: Dsize, rank: Rank) -> usize
165    where
166        M: GGSWPreparedAlloc<B>,
167    {
168        module.bytes_of_ggsw_prepared(base2k, k, dnum, dsize, rank)
169    }
170}
171
172impl<D: DataRef, B: Backend> GGSWPrepared<D, B> {
173    pub fn data(&self) -> &VmpPMat<D, B> {
174        &self.data
175    }
176}
177
178pub trait GGSWPrepare<B: Backend>
179where
180    Self: GetRingDegree + VmpPrepareTmpBytes + VmpPrepare<B>,
181{
182    fn ggsw_prepare_tmp_bytes<A>(&self, infos: &A) -> usize
183    where
184        A: GGSWInfos,
185    {
186        assert_eq!(self.ring_degree(), infos.n());
187        self.vmp_prepare_tmp_bytes(
188            infos.dnum().into(),
189            (infos.rank() + 1).into(),
190            (infos.rank() + 1).into(),
191            infos.size(),
192        )
193    }
194    fn ggsw_prepare<R, O>(&self, res: &mut R, other: &O, scratch: &mut Scratch<B>)
195    where
196        R: GGSWPreparedToMut<B>,
197        O: GGSWToRef,
198    {
199        let mut res: GGSWPrepared<&mut [u8], B> = res.to_mut();
200        let other: GGSW<&[u8]> = other.to_ref();
201        assert_eq!(res.n(), self.ring_degree());
202        assert_eq!(other.n(), self.ring_degree());
203        assert_eq!(res.k, other.k);
204        assert_eq!(res.base2k, other.base2k);
205        assert_eq!(res.dsize, other.dsize);
206        self.vmp_prepare(&mut res.data, &other.data, scratch);
207    }
208}
209
210impl<B: Backend> GGSWPrepare<B> for Module<B> where Self: GetRingDegree + VmpPrepareTmpBytes + VmpPrepare<B> {}
211
212impl<B: Backend> GGSWPrepared<Vec<u8>, B> {
213    pub fn prepare_tmp_bytes<A, M>(&self, module: &M, infos: &A) -> usize
214    where
215        A: GGSWInfos,
216        M: GGSWPrepare<B>,
217    {
218        module.ggsw_prepare_tmp_bytes(infos)
219    }
220}
221
222impl<D: DataMut, B: Backend> GGSWPrepared<D, B> {
223    pub fn prepare<O, M>(&mut self, module: &M, other: &O, scratch: &mut Scratch<B>)
224    where
225        O: GGSWToRef,
226        M: GGSWPrepare<B>,
227    {
228        module.ggsw_prepare(self, other, scratch);
229    }
230}
231
232pub trait GGSWPreparedToMut<B: Backend> {
233    fn to_mut(&mut self) -> GGSWPrepared<&mut [u8], B>;
234}
235
236impl<D: DataMut, B: Backend> GGSWPreparedToMut<B> for GGSWPrepared<D, B> {
237    fn to_mut(&mut self) -> GGSWPrepared<&mut [u8], B> {
238        GGSWPrepared {
239            base2k: self.base2k,
240            k: self.k,
241            dsize: self.dsize,
242            data: self.data.to_mut(),
243        }
244    }
245}
246
247pub trait GGSWPreparedToRef<B: Backend> {
248    fn to_ref(&self) -> GGSWPrepared<&[u8], B>;
249}
250
251impl<D: DataRef, B: Backend> GGSWPreparedToRef<B> for GGSWPrepared<D, B> {
252    fn to_ref(&self) -> GGSWPrepared<&[u8], B> {
253        GGSWPrepared {
254            base2k: self.base2k,
255            k: self.k,
256            dsize: self.dsize,
257            data: self.data.to_ref(),
258        }
259    }
260}