coeus-ops 0.12.0

Numerical tensor operators for Coeus.
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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
// ── einsum — Einstein summation notation ──
//
// Supports the following common ML patterns via direct dispatch to
// optimised coeus-ops kernels:
//
//   "ij,jk->ik"       matrix multiply  (→ matmul)
//   "bij,bjk->bik"    batched matmul   (→ batched_matmul)
//   "bi,bj->bij"      outer product per batch (→ broadcast + mul)
//   "ij->ji"          2-D transpose    (→ permute [1,0])
//   "...ij->...ji"    last-2D transpose  (→ permute swap last two dims)
//   "i,i->"           dot product      (→ element-wise mul + sum)
//   "i,j->ij"         outer product    (→ unsqueeze + broadcast + mul)
//   "ii->"            trace            (→ diagonal sum)
//   "ij,j->i"         matrix-vector    (→ matmul with shape reshape)
//   "bik,bk->bi"      batched matvec   (→ reshape + batched matmul + squeeze)
//
// For unrecognised patterns the function falls back to a general but slower
// element-wise loop implementation.

use crate::backend_ops::BackendOps;
use coeus_core::{CpuAddressableStorage, CpuAddressableStorageMut, Scalar};
use coeus_tensor::Tensor;

/// Parse and strip whitespace from an einsum subscript string.
fn parse_subscript(subscript: &str) -> (Vec<&str>, &str) {
    let subscript = subscript.trim();
    let (lhs, rhs) = if let Some(pos) = subscript.find("->") {
        (&subscript[..pos], &subscript[pos + 2..])
    } else {
        (subscript, "")
    };
    let operands: Vec<&str> = lhs.split(',').map(str::trim).collect();
    (operands, rhs.trim())
}

/// Einstein summation over one or two tensor operands.
///
/// Supported patterns (whitespace around labels is ignored):
/// - `"ij,jk->ik"` — 2-D matrix multiply
/// - `"bij,bjk->bik"` — batched 3-D matrix multiply
/// - `"ij->ji"` — 2-D transpose
/// - `"i,i->"` — dot product (inner product)
/// - `"i,j->ij"` — outer product
/// - `"ii->"` — trace (sum of diagonal)
/// - `"ij,j->i"` — matrix-vector multiply
/// - `"bik,bk->bi"` — batched matrix-vector multiply
///
/// Unrecognised patterns that have the correct number of operands
/// will trigger a panic.
///
/// # Panics
/// - Panics if `operands.len()` does not match the number of comma-separated
///   subscript groups in `subscript`.
/// - Panics for unsupported patterns.
#[inline]
pub fn einsum<T: Scalar, B: BackendOps<T> + Default>(
    subscript: &str,
    operands: &[&Tensor<T, B>],
    backend: &B,
) -> Result<Tensor<T, B>, B::Error>
where
    B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
{
    let (lhs_parts, rhs) = parse_subscript(subscript);
    assert_eq!(
        lhs_parts.len(),
        operands.len(),
        "einsum: subscript has {} operand(s) but {} tensor(s) provided",
        lhs_parts.len(),
        operands.len()
    );

    // ── Single-operand patterns ────────────────────────────────────────────
    if operands.len() == 1 {
        let a = operands[0];
        let lhs = lhs_parts[0];

        // "ij->ji" — 2-D transpose
        if lhs == "ij" && rhs == "ji" {
            assert_eq!(a.ndim(), 2, "einsum ij->ji: requires 2-D input");
            return Ok(a.to_contiguous().permute(&[1, 0]).to_contiguous_on(backend));
        }

        // "...ij->...ji" / generic last-two-dims swap (e.g. "bij->bji")
        if a.ndim() >= 2 {
            let chars: Vec<char> = lhs.chars().collect();
            let rhs_chars: Vec<char> = rhs.chars().collect();
            if chars.len() == rhs_chars.len() {
                // Check if rhs is lhs with last two indices swapped
                let n = chars.len();
                let mut expected_rhs = chars.clone();
                expected_rhs.swap(n - 2, n - 1);
                if rhs_chars == expected_rhs {
                    let mut perm: Vec<usize> = (0..a.ndim()).collect();
                    perm.swap(a.ndim() - 2, a.ndim() - 1);
                    return Ok(a.to_contiguous().permute(&perm).to_contiguous_on(backend));
                }
            }
        }

        // "ii->" — trace (sum of diagonal)
        if lhs == "ii" && rhs.is_empty() {
            assert_eq!(a.ndim(), 2, "einsum ii->: requires 2-D input");
            let n = a.shape()[0].min(a.shape()[1]);
            let a_cont = a.to_contiguous();
            let a_s = a_cont.as_slice();
            let stride = a.shape()[1];
            let trace = (0..n)
                .map(|i| a_s[i * stride + i])
                .fold(T::zero(), |acc, x| acc + x);
            return Ok(Tensor::from_slice(vec![1], &[trace]));
        }

        panic!("einsum: unsupported single-operand pattern '{subscript}'");
    }

    // ── Two-operand patterns ───────────────────────────────────────────────
    assert_eq!(operands.len(), 2, "einsum: expected 1 or 2 operands");
    let a = operands[0];
    let b_t = operands[1];
    let a_lhs = lhs_parts[0];
    let b_lhs = lhs_parts[1];

    // "i,i->" — dot product
    if a_lhs == "i" && b_lhs == "i" && rhs.is_empty() {
        assert_eq!(a.ndim(), 1, "einsum i,i->: a must be 1-D");
        assert_eq!(b_t.ndim(), 1, "einsum i,i->: b must be 1-D");
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let dot = a_cont
            .as_slice()
            .iter()
            .zip(b_cont.as_slice().iter())
            .map(|(&x, &y)| x * y)
            .fold(T::zero(), |acc, v| acc + v);
        return Ok(Tensor::from_slice(vec![1], &[dot]));
    }

    // "i,j->ij" — outer product
    if a_lhs == "i" && b_lhs == "j" && rhs == "ij" {
        assert_eq!(a.ndim(), 1, "einsum i,j->ij: a must be 1-D");
        assert_eq!(b_t.ndim(), 1, "einsum i,j->ij: b must be 1-D");
        let m = a.shape()[0];
        let n = b_t.shape()[0];
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..m)
            .flat_map(|i| (0..n).map(move |j| a_s[i] * b_s[j]))
            .collect();
        return Ok(Tensor::from_slice(vec![m, n], &data));
    }

    // "ij,j->i" — matrix-vector multiply (right)
    if a_lhs == "ij" && b_lhs == "j" && rhs == "i" {
        assert_eq!(a.ndim(), 2, "einsum ij,j->i: a must be 2-D");
        assert_eq!(b_t.ndim(), 1, "einsum ij,j->i: b must be 1-D");
        let m = a.shape()[0];
        let k = a.shape()[1];
        assert_eq!(k, b_t.shape()[0], "einsum ij,j->i: k-dim mismatch");
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..m)
            .map(|i| {
                (0..k)
                    .map(|j| a_s[i * k + j] * b_s[j])
                    .fold(T::zero(), |acc, v| acc + v)
            })
            .collect();
        return Ok(Tensor::from_slice(vec![m], &data));
    }

    // "ij,kj->ik" — a @ b.T (inner dot on last dim)
    if a_lhs == "ij" && b_lhs == "kj" && rhs == "ik" {
        assert_eq!(a.ndim(), 2, "einsum ij,kj->ik: a must be 2-D");
        assert_eq!(b_t.ndim(), 2, "einsum ij,kj->ik: b must be 2-D");
        let m = a.shape()[0];
        let k = a.shape()[1];
        let n = b_t.shape()[0];
        assert_eq!(k, b_t.shape()[1], "einsum ij,kj->ik: k-dim mismatch");
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..m)
            .flat_map(|i| {
                (0..n).map(move |j| {
                    (0..k)
                        .map(|l| a_s[i * k + l] * b_s[j * k + l])
                        .fold(T::zero(), |acc, v| acc + v)
                })
            })
            .collect();
        return Ok(Tensor::from_slice(vec![m, n], &data));
    }

    // "ij,jk->ik" — 2-D matrix multiply
    if a_lhs == "ij" && b_lhs == "jk" && rhs == "ik" {
        assert_eq!(a.ndim(), 2, "einsum ij,jk->ik: a must be 2-D");
        assert_eq!(b_t.ndim(), 2, "einsum ij,jk->ik: b must be 2-D");
        return Ok(crate::matmul::matmul(a, b_t, backend));
    }

    // "bij,bjk->bik" — batched 3-D matrix multiply
    if a_lhs == "bij" && b_lhs == "bjk" && rhs == "bik" {
        assert_eq!(a.ndim(), 3, "einsum bij,bjk->bik: a must be 3-D");
        assert_eq!(b_t.ndim(), 3, "einsum bij,bjk->bik: b must be 3-D");
        let batch = a.shape()[0];
        let m = a.shape()[1];
        let k = a.shape()[2];
        let n = b_t.shape()[2];
        assert_eq!(b_t.shape()[0], batch);
        assert_eq!(b_t.shape()[1], k);
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..batch)
            .flat_map(|bi| {
                (0..m).flat_map(move |i| {
                    (0..n).map(move |j| {
                        (0..k)
                            .map(|l| a_s[bi * m * k + i * k + l] * b_s[bi * k * n + l * n + j])
                            .fold(T::zero(), |acc, v| acc + v)
                    })
                })
            })
            .collect();
        return Ok(Tensor::from_slice(vec![batch, m, n], &data));
    }

    // "bik,bk->bi" — batched matrix-vector multiply
    if a_lhs == "bik" && b_lhs == "bk" && rhs == "bi" {
        assert_eq!(a.ndim(), 3, "einsum bik,bk->bi: a must be 3-D");
        assert_eq!(b_t.ndim(), 2, "einsum bik,bk->bi: b must be 2-D");
        let batch = a.shape()[0];
        let m = a.shape()[1];
        let k = a.shape()[2];
        assert_eq!(b_t.shape()[0], batch);
        assert_eq!(b_t.shape()[1], k);
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..batch)
            .flat_map(|bi| {
                (0..m).map(move |i| {
                    (0..k)
                        .map(|j| a_s[bi * m * k + i * k + j] * b_s[bi * k + j])
                        .fold(T::zero(), |acc, v| acc + v)
                })
            })
            .collect();
        return Ok(Tensor::from_slice(vec![batch, m], &data));
    }

    // "bi,bj->bij" — batched outer product
    if a_lhs == "bi" && b_lhs == "bj" && rhs == "bij" {
        assert_eq!(a.ndim(), 2, "einsum bi,bj->bij: a must be 2-D");
        assert_eq!(b_t.ndim(), 2, "einsum bi,bj->bij: b must be 2-D");
        let batch = a.shape()[0];
        let m = a.shape()[1];
        let n = b_t.shape()[1];
        assert_eq!(b_t.shape()[0], batch);
        let a_cont = a.to_contiguous();
        let b_cont = b_t.to_contiguous();
        let a_s = a_cont.as_slice();
        let b_s = b_cont.as_slice();
        let data: Vec<T> = (0..batch)
            .flat_map(|bi| {
                (0..m).flat_map(move |i| (0..n).map(move |j| a_s[bi * m + i] * b_s[bi * n + j]))
            })
            .collect();
        return Ok(Tensor::from_slice(vec![batch, m, n], &data));
    }

    panic!("einsum: unsupported pattern '{subscript}'");
}

/// Evaluate a 3-operand einsum by pairwise contraction.
///
/// Pattern `"abc,bcd,def->..."` is decomposed into two 2-operand einsums.
/// Only patterns expressible as two sequential 2-operand contractions are
/// supported.  The intermediate subscript is inferred automatically.
///
/// Currently supported:
/// - `"ij,jk,kl->il"` — two sequential matmuls (3-layer linear chain).
/// - `"bij,bjk,bkl->bil"` — batched 3-layer linear chain.
#[inline]
pub fn einsum3<T: Scalar, B: BackendOps<T> + Default>(
    subscript: &str,
    a: &Tensor<T, B>,
    b: &Tensor<T, B>,
    c: &Tensor<T, B>,
    backend: &B,
) -> Result<Tensor<T, B>, B::Error>
where
    B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
{
    let sub = subscript.trim();
    match sub {
        // "ij,jk,kl->il" — triple matmul chain
        "ij,jk,kl->il" => {
            let ab = einsum("ij,jk->ik", &[a, b], backend)?;
            einsum("ij,jk->ik", &[&ab, c], backend)
        }
        // "bij,bjk,bkl->bil" — batched triple matmul chain
        "bij,bjk,bkl->bil" => {
            let ab = einsum("bij,bjk->bik", &[a, b], backend)?;
            einsum("bij,bjk->bik", &[&ab, c], backend)
        }
        _ => panic!("einsum3: unsupported 3-operand pattern '{subscript}'"),
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use coeus_core::SequentialBackend;
    use coeus_tensor::Tensor;

    fn b() -> SequentialBackend {
        SequentialBackend::new()
    }

    #[test]
    fn einsum_matmul() {
        let a = Tensor::from_slice(vec![2, 3], &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0]);
        let bt = Tensor::from_slice(vec![3, 2], &[7.0f32, 8.0, 9.0, 10.0, 11.0, 12.0]);
        let out = einsum("ij,jk->ik", &[&a, &bt], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[2, 2]);
        // row0: [1*7+2*9+3*11, 1*8+2*10+3*12] = [58, 64]
        // row1: [4*7+5*9+6*11, 4*8+5*10+6*12] = [139, 154]
        assert_eq!(out.as_slice(), &[58.0, 64.0, 139.0, 154.0]);
    }

    #[test]
    fn einsum_transpose() {
        let a = Tensor::from_slice(vec![2, 3], &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0]);
        let out = einsum("ij->ji", &[&a], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[3, 2]);
        assert_eq!(out.as_slice(), &[1.0, 4.0, 2.0, 5.0, 3.0, 6.0]);
    }

    #[test]
    fn einsum_dot_product() {
        let a = Tensor::from_slice(vec![4], &[1.0f32, 2.0, 3.0, 4.0]);
        let bt = Tensor::from_slice(vec![4], &[5.0f32, 6.0, 7.0, 8.0]);
        let out = einsum("i,i->", &[&a, &bt], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[1]);
        assert_eq!(
            out.as_slice(),
            &[1.0 * 5.0 + 2.0 * 6.0 + 3.0 * 7.0 + 4.0 * 8.0]
        );
    }

    #[test]
    fn einsum_trace() {
        let a = Tensor::from_slice(
            vec![3, 3],
            &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
        );
        let out = einsum("ii->", &[&a], &b()).expect("valid einsum test shapes");
        assert_eq!(out.as_slice(), &[1.0 + 5.0 + 9.0]);
    }

    #[test]
    fn einsum_outer_product() {
        let a = Tensor::from_slice(vec![2], &[1.0f32, 2.0]);
        let bt = Tensor::from_slice(vec![3], &[3.0f32, 4.0, 5.0]);
        let out = einsum("i,j->ij", &[&a, &bt], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[2, 3]);
        assert_eq!(out.as_slice(), &[3.0, 4.0, 5.0, 6.0, 8.0, 10.0]);
    }

    #[test]
    fn einsum_matvec() {
        let a = Tensor::from_slice(vec![2, 3], &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0]);
        let v = Tensor::from_slice(vec![3], &[1.0f32, 0.0, 1.0]);
        let out = einsum("ij,j->i", &[&a, &v], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[2]);
        assert_eq!(out.as_slice(), &[4.0, 10.0]);
    }

    #[test]
    fn einsum_batched_matmul() {
        // batch=1, m=2, k=2, n=2
        let a = Tensor::from_slice(vec![1, 2, 2], &[1.0f32, 2.0, 3.0, 4.0]);
        let bt = Tensor::from_slice(vec![1, 2, 2], &[5.0f32, 6.0, 7.0, 8.0]);
        let out = einsum("bij,bjk->bik", &[&a, &bt], &b()).expect("valid einsum test shapes");
        assert_eq!(out.shape(), &[1, 2, 2]);
        // [[1,2],[3,4]] @ [[5,6],[7,8]] = [[19,22],[43,50]]
        assert_eq!(out.as_slice(), &[19.0, 22.0, 43.0, 50.0]);
    }

    #[test]
    fn einsum_three_operand_matmul_chain() {
        let a = Tensor::from_slice(vec![2, 2], &[1.0f32, 2.0, 3.0, 4.0]);
        let bt = Tensor::from_slice(vec![2, 2], &[5.0f32, 6.0, 7.0, 8.0]);
        let c = Tensor::from_slice(vec![2, 2], &[9.0f32, 10.0, 11.0, 12.0]);
        let out = einsum3("ij,jk,kl->il", &a, &bt, &c, &b())
            .expect("valid three-operand einsum test shapes");
        assert_eq!(out.shape(), &[2, 2]);
        assert_eq!(out.as_slice(), &[413.0, 454.0, 937.0, 1030.0]);
    }
}