1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219


use strassen::*;

use crate::matrix::*;
use crate::ring::*;

use std::alloc::Allocator;
use std::alloc::Global;

pub mod strassen;

///
/// Trait to allow rings to provide specialized implementations for inner products, i.e.
/// the sums `sum_i a[i] * b[i]`.
/// 
#[stability::unstable(feature = "enable")]
pub trait ComputeInnerProduct: RingBase {

    ///
    /// Computes the inner product `sum_i lhs[i] * rhs[i]`.
    /// 
    fn inner_product_ref<'a, I: Iterator<Item = (&'a Self::Element, &'a Self::Element)>>(&self, els: I) -> Self::Element
        where Self::Element: 'a,
            Self: 'a;

    ///
    /// Computes the inner product `sum_i lhs[i] * rhs[i]`.
    /// 
    fn inner_product_ref_fst<'a, I: Iterator<Item = (&'a Self::Element, Self::Element)>>(&self, els: I) -> Self::Element
        where Self::Element: 'a,
            Self: 'a;

    ///
    /// Computes the inner product `sum_i lhs[i] * rhs[i]`.
    /// 
    fn inner_product<I: Iterator<Item = (Self::Element, Self::Element)>>(&self, els: I) -> Self::Element;
}

impl<R: ?Sized + RingBase> ComputeInnerProduct for R {

    default fn inner_product_ref_fst<'a, I: Iterator<Item = (&'a Self::Element, Self::Element)>>(&self, els: I) -> Self::Element
        where Self::Element: 'a
    {
        self.inner_product(els.map(|(l, r)| (self.clone_el(l), r)))
    }

    default fn inner_product_ref<'a, I: Iterator<Item = (&'a Self::Element, &'a Self::Element)>>(&self, els: I) -> Self::Element
        where Self::Element: 'a
    {
        self.inner_product_ref_fst(els.map(|(l, r)| (l, self.clone_el(r))))
    }

    default fn inner_product<I: Iterator<Item = (Self::Element, Self::Element)>>(&self, els: I) -> Self::Element {
        self.sum(els.map(|(l, r)| self.mul(l, r)))
    }
}

///
/// Trait for objects that can compute a matrix multiplications over a fixed ring.
/// 
#[stability::unstable(feature = "enable")]
pub trait MatmulAlgorithm<R: ?Sized + RingBase> {

    ///
    /// Computes the matrix product of `lhs` and `rhs`, and adds the result to `dst`.
    /// 
    /// This requires that `lhs` is a `nxk` matrix, `rhs` is a `kxm` matrix and `dst` is a `nxm` matrix.
    /// In this case, the function concretely computes `dst[i, j] += sum_l lhs[i, l] * rhs[l, j]` where
    /// `l` runs from `0` to `k - 1`.
    /// 
    fn add_matmul<V1, V2, V3, const T1: bool, const T2: bool, const T3: bool>(&self, lhs: TransposableSubmatrix<V1, R::Element, T1>, rhs: TransposableSubmatrix<V2, R::Element, T2>, dst: TransposableSubmatrixMut<V3, R::Element, T3>, ring: &R)
        where V1: AsPointerToSlice<R::Element>,
            V2: AsPointerToSlice<R::Element>,
            V3: AsPointerToSlice<R::Element>;
         
    ///
    /// Computes the matrix product of `lhs` and `rhs`, and stores the result in `dst`.
    /// 
    /// This requires that `lhs` is a `nxk` matrix, `rhs` is a `kxm` matrix and `dst` is a `nxm` matrix.
    /// In this case, the function concretely computes `dst[i, j] = sum_l lhs[i, l] * rhs[l, j]` where
    /// `l` runs from `0` to `k - 1`.
    ///    
    fn matmul<V1, V2, V3, const T1: bool, const T2: bool, const T3: bool>(&self, lhs: TransposableSubmatrix<V1, R::Element, T1>, rhs: TransposableSubmatrix<V2, R::Element, T2>, mut dst: TransposableSubmatrixMut<V3, R::Element, T3>, ring: &R)
        where V1: AsPointerToSlice<R::Element>,
            V2: AsPointerToSlice<R::Element>,
            V3: AsPointerToSlice<R::Element>
    {
        for i in 0..dst.row_count() {
            for j in 0..dst.col_count() {
                *dst.at_mut(i, j) = ring.zero();
            }
        }
        self.add_matmul(lhs, rhs, dst, ring);
    }
}

///
/// Trait to allow rings to customize the parameters with which [`StrassenAlgorithm`] will
/// compute matrix multiplications.
/// 
#[stability::unstable(feature = "enable")]
pub trait StrassenHint: RingBase {

    ///
    /// Define a threshold from which on [`StrassenAlgorithm`] will use the Strassen algorithm.
    /// 
    /// Concretely, when this returns `k`, [`StrassenAlgorithm`] will reduce the 
    /// matrix multipliction down to `2^k x 2^k` matrices using Strassen's algorithm,
    /// and then use naive matmul for the rest.
    /// 
    /// The value is `0`, but if the considered rings have fast multiplication (compared to addition), 
    /// then setting this higher may result in a performance gain.
    /// 
    fn strassen_threshold(&self) -> usize;
}

impl<R: RingBase + ?Sized> StrassenHint for R {

    default fn strassen_threshold(&self) -> usize {
        0
    }
}

#[stability::unstable(feature = "enable")]
pub const STANDARD_MATMUL: StrassenAlgorithm = StrassenAlgorithm::new(Global);

#[stability::unstable(feature = "enable")]
#[derive(Clone, Copy)]
pub struct StrassenAlgorithm<A: Allocator = Global> {
    allocator: A
}

impl<A: Allocator> StrassenAlgorithm<A> {

    #[stability::unstable(feature = "enable")]
    pub const fn new(allocator: A) -> Self {
        Self { allocator }
    }
}

impl<R: ?Sized + RingBase, A: Allocator> MatmulAlgorithm<R> for StrassenAlgorithm<A> {

    fn add_matmul<V1, V2, V3, const T1: bool, const T2: bool, const T3: bool>(
        &self,
        lhs: TransposableSubmatrix<V1, R::Element, T1>,
        rhs: TransposableSubmatrix<V2, R::Element, T2>,
        dst: TransposableSubmatrixMut<V3, R::Element, T3>,
        ring: &R
    )
        where V1: AsPointerToSlice<R::Element>,
            V2: AsPointerToSlice<R::Element>,
            V3: AsPointerToSlice<R::Element>
    {
        strassen::<_, _, _, _, _, T1, T2, T3>(true, <_ as StrassenHint>::strassen_threshold(ring), lhs, rhs, dst, RingRef::new(ring), &self.allocator)
    }

    fn matmul<V1, V2, V3, const T1: bool, const T2: bool, const T3: bool>(
        &self,
        lhs: TransposableSubmatrix<V1, R::Element, T1>,
        rhs: TransposableSubmatrix<V2, R::Element, T2>,
        dst: TransposableSubmatrixMut<V3, R::Element, T3>,
        ring: &R
    )
        where V1: AsPointerToSlice<R::Element>,
            V2: AsPointerToSlice<R::Element>,
            V3: AsPointerToSlice<R::Element>
    {
        strassen::<_, _, _, _, _, T1, T2, T3>(false, <_ as StrassenHint>::strassen_threshold(ring), lhs, rhs, dst, RingRef::new(ring), &self.allocator)
    }
}

#[cfg(test)]
use test;
#[cfg(test)]
use crate::primitive_int::*;

#[cfg(test)]
const BENCH_SIZE: usize = 128;
#[cfg(test)]
type BenchInt = i64;

#[bench]
fn bench_naive_matmul(bencher: &mut test::Bencher) {
    let lhs = OwnedMatrix::from_fn_in(BENCH_SIZE, BENCH_SIZE, |i, j| std::hint::black_box(i as BenchInt + j as BenchInt), Global);
    let rhs = OwnedMatrix::from_fn_in(BENCH_SIZE, BENCH_SIZE, |i, j| std::hint::black_box(i as BenchInt + j as BenchInt), Global);
    let mut result: OwnedMatrix<BenchInt> = OwnedMatrix::zero(BENCH_SIZE, BENCH_SIZE, StaticRing::<BenchInt>::RING);
    bencher.iter(|| {
        strassen::<_, _, _, _, _, false, false, false>(
            false, 
            100, 
            TransposableSubmatrix::from(lhs.data()), 
            TransposableSubmatrix::from(rhs.data()), 
            TransposableSubmatrixMut::from(result.data_mut()), 
            StaticRing::<BenchInt>::RING, 
            &Global
        );
        assert_eq!((BENCH_SIZE * (BENCH_SIZE + 1) * (BENCH_SIZE * 2 + 1) / 6 - BENCH_SIZE * BENCH_SIZE) as BenchInt, *result.at(0, 0));
    });
}

#[bench]
fn bench_strassen_matmul(bencher: &mut test::Bencher) {
    let threshold_log_2 = 4;
    let lhs = OwnedMatrix::from_fn_in(BENCH_SIZE, BENCH_SIZE, |i, j| std::hint::black_box(i as BenchInt + j as BenchInt), Global);
    let rhs = OwnedMatrix::from_fn_in(BENCH_SIZE, BENCH_SIZE, |i, j| std::hint::black_box(i as BenchInt + j as BenchInt), Global);
    let mut result: OwnedMatrix<BenchInt> = OwnedMatrix::zero(BENCH_SIZE, BENCH_SIZE, StaticRing::<BenchInt>::RING);
    bencher.iter(|| {
        strassen::<_, _, _, _, _, false, false, false>(
            false, 
            threshold_log_2, 
            TransposableSubmatrix::from(lhs.data()), 
            TransposableSubmatrix::from(rhs.data()), 
            TransposableSubmatrixMut::from(result.data_mut()), 
            StaticRing::<BenchInt>::RING, 
            &Global
        );
        assert_eq!((BENCH_SIZE * (BENCH_SIZE + 1) * (BENCH_SIZE * 2 + 1) / 6 - BENCH_SIZE * BENCH_SIZE) as BenchInt, *result.at(0, 0));
    });
}