sukker 2.1.3

Linear Algebra and Matrices made easy!
Documentation 
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284

use std::{error::Error, ops::RangeInclusive, str::FromStr};

use rayon::prelude::*;

use crate::{at, Matrix, MatrixElement};

pub fn swap(lhs: &mut usize, rhs: &mut usize) {
    let temp = *lhs;
    *lhs = *rhs;
    *rhs = temp;
}

// simd
impl<'a, T> Matrix<'a, T>
where
    T: MatrixElement + 'a,
    <T as FromStr>::Err: Error + 'static,
    Vec<T>: IntoParallelIterator,
    Vec<&'a T>: IntoParallelRefIterator<'a>,
{
    pub fn determinant_helper(&self) -> T {
        match self.nrows {
            1 => self.at(0, 0),
            2 => Self::det_2x2(self),
            3 => Self::det_3x3(self),
            n => Self::det_nxn(self.data.clone(), n),
        }
    }

    // General helper function calling out to other matmuls based on target architecture
    pub fn matmul_helper(&self, other: &Self) -> Self {
        match (self.shape(), other.shape()) {
            ((1, 2), (2, 1)) => return self.onetwo_by_twoone(other),
            ((2, 2), (2, 1)) => return self.twotwo_by_twoone(other),
            ((1, 2), (2, 2)) => return self.onetwo_by_twotwo(other),
            ((2, 2), (2, 2)) => return self.twotwo_by_twotwo(other),
            _ => {}
        };

        // Target Detection

        // if let Some(result) = optim::get_optimized_matmul(self, other) {
        //     return result;
        // }

        let blck_size = Self::get_block_size(self, other);

        // println!("BS: {}", blck_size);

        if self.shape() == other.shape() {
            // Calculated from lowest possible size where
            // nrows & blck_size == 0.
            // Block size will never be more than 50
            return Self::blocked_matmul(self, other, blck_size);
        }

        Self::optimized_blocked_matmul(self, other, blck_size)
    }

    // Calculate efficient blocksize
    #[inline(always)]
    pub fn get_block_size(&self, other: &Self) -> usize {
        let range = Self::get_range_for_block_size(self, other);

        range
            .collect::<Vec<_>>()
            .into_par_iter()
            .find_last(|b| self.ncols % b == 0 || self.nrows % b == 0 || other.ncols % b == 0)
            .unwrap()
    }

    #[inline(always)]
    pub fn get_range_for_block_size(&self, other: &Self) -> RangeInclusive<usize> {
        if self.nrows < 30 && self.ncols < 30 || other.nrows < 30 && other.ncols < 30 {
            2..=10
        } else if self.nrows < 100 && self.ncols < 100 || other.nrows < 100 && other.ncols < 100 {
            10..=30
        } else {
            30..=50
        }
    }

    // ===================================================
    //           Determinant
    // ===================================================

    #[inline(always)]
    fn det_2x2(&self) -> T {
        self.at(0, 0) * self.at(1, 1) - self.at(0, 1) * self.at(1, 0)
    }

    #[inline(always)]
    fn det_3x3(&self) -> T {
        let a = self.at(0, 0);
        let b = self.at(0, 1);
        let c = self.at(0, 2);
        let d = self.at(1, 0);
        let e = self.at(1, 1);
        let f = self.at(1, 2);
        let g = self.at(2, 0);
        let h = self.at(2, 1);
        let i = self.at(2, 2);

        a * (e * i - f * h) - b * (d * i - f * g) + c * (d * h - e * g)
    }

    fn det_nxn(matrix: Vec<T>, n: usize) -> T {
        if n == 1 {
            return matrix[0];
        }

        let mut det = T::zero();
        let mut sign = T::one();

        for col in 0..n {
            let sub_det = Self::det_nxn(Self::submatrix(matrix.clone(), n, 0, col), n - 1);

            det += sign * matrix[col] * sub_det;

            sign *= -T::one();
        }

        det
    }

    fn submatrix(matrix: Vec<T>, n: usize, row_to_remove: usize, col_to_remove: usize) -> Vec<T> {
        matrix
            .par_iter()
            .enumerate()
            .filter_map(|(i, &value)| {
                let row = i / n;
                let col = i % n;
                if row != row_to_remove && col != col_to_remove {
                    Some(value)
                } else {
                    None
                }
            })
            .collect()
    }

    // ===================================================
    //           Extremely specific optimizations
    // ===================================================

    // 1x2 @ 2x1 matrix
    #[inline(always)]
    fn onetwo_by_twoone(&self, other: &Self) -> Self {
        let a = self.at(0, 0) * other.at(0, 0) + self.at(0, 1) * other.at(1, 0);

        Self::new(vec![a], (1, 1)).unwrap()
    }

    // 2x2 @ 2x1 matrix
    #[inline(always)]
    fn twotwo_by_twoone(&self, other: &Self) -> Self {
        let a = self.at(0, 0) * other.at(0, 0) + self.at(0, 1) * other.at(1, 0);
        let b = self.at(1, 0) * other.at(0, 0) + self.at(1, 1) * other.at(1, 0);

        Self::new(vec![a, b], (2, 1)).unwrap()
    }

    //
    // 1x2 @ 2x2 matrix
    #[inline(always)]
    fn onetwo_by_twotwo(&self, other: &Self) -> Self {
        let a = self.at(0, 0) * other.at(0, 0) + self.at(0, 1) * other.at(1, 0);
        let b = self.at(0, 0) * other.at(1, 0) + self.at(0, 1) * other.at(1, 1);

        Self::new(vec![a, b], (1, 2)).unwrap()
    }

    // 2x2 @ 2x2 matrix
    #[inline(always)]
    fn twotwo_by_twotwo(&self, other: &Self) -> Self {
        let a = self.at(0, 0) * other.at(0, 0) + self.at(1, 0) * other.at(1, 0);
        let b = self.at(0, 0) * other.at(0, 1) + self.at(0, 1) * other.at(1, 1);
        let c = self.at(1, 0) * other.at(0, 0) + self.at(1, 1) * other.at(1, 0);
        let d = self.at(1, 0) * other.at(1, 0) + self.at(1, 1) * other.at(1, 1);

        Self::new(vec![a, b, c, d], (2, 2)).unwrap()
    }

    // ========================================================================
    //
    //    General solutions for matrix multiplication
    //
    // ========================================================================

    /// Naive matmul if you don't have any SIMD intrinsincts
    ///
    /// Also blocked, but doing different than just N
    fn optimized_blocked_matmul(&self, other: &Self, block_size: usize) -> Self {
        let M = self.nrows;
        let N = self.ncols;
        let P = other.ncols;

        let mut data = vec![T::zero(); M * P];

        //let t_other = other.transpose_copy();

        for kk in (0..N).step_by(block_size) {
            for jj in (0..P).step_by(block_size) {
                for ii in (0..M).step_by(block_size) {
                    let block_end_i = (ii + block_size).min(M);
                    let block_end_j = (jj + block_size).min(P);
                    let block_end_k = (kk + block_size).min(N);

                    // Blocking for L0 memory
                    for i in ii..block_end_i {
                        for j in jj..block_end_j {
                            // for k in kk..block_end_k {
                            //     data[at!(i, j, P)] += self.at(i, k) * other.at(k, j);
                            // }
                            data[at!(i, j, P)] = (kk..block_end_k)
                                .into_par_iter()
                                .map(|k| self.at(i, k) * other.at(k, j))
                                .sum();
                        }
                    }
                }
            }
        }
        Self::new(data, (M, P)).unwrap()
    }

    // SUMMA Algorithm
    // https://www.netlib.org/lapack/lawnspdf/lawn96.pdf
    fn summa(&self, other: &Self, block_size: usize) -> Self {
        todo!()
    }

    // The magnum opus of matrix multiply, also known as naive matmul
    // Only optimization is a parallelized innermost summation
    fn naive(&self, other: &Self) -> Self {
        let M = self.nrows;
        let N = self.ncols;
        let P = other.ncols;

        let mut data = vec![T::zero(); M * P];

        for i in 0..M {
            for j in 0..P {
                data[at!(i, j, P)] = (0..N)
                    .into_par_iter()
                    .map(|k| self.at(i, k) * other.at(k, j))
                    .sum();
            }
        }

        Self::new(data, (M, P)).unwrap()
    }

    // Blocked matmul if you don't have any SIMD intrinsincts
    // https://csapp.cs.cmu.edu/public/waside/waside-blocking.pdf
    //
    // Modification involves transposing the B matrix, at the cost
    // of increased space complexity, but better cache hit rate
    //
    // NOTE: Only works for M N @ N M matrices for now
    fn blocked_matmul(&self, other: &Self, block_size: usize) -> Self {
        let n = self.nrows;

        let en = block_size * (n / block_size);

        let mut data = vec![T::zero(); n * n];

        let t_other = other.transpose_copy();

        for kk in (0..n).step_by(en) {
            for jj in (0..n).step_by(en) {
                for i in 0..n {
                    for j in jj..jj + block_size {
                        data[at!(i, j, n)] = (kk..kk + block_size)
                            .into_par_iter()
                            .map(|k| self.at(i, k) * t_other.at(j, k))
                            .sum();
                    }
                }
            }
        }
        Self::new(data, (n, n)).unwrap()
    }
}