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
//! Linear algebra operations for Matrix
//!
//! This module provides linear operations:
//! - `transpose()` - Matrix transpose
//! - `matvec()` - Matrix-vector multiplication
//! - `vecmat()` - Vector-matrix multiplication
use crate::{Backend, TruenoError, Vector};
#[cfg(feature = "tracing")]
use tracing::instrument;
/// Backend dispatch macro for dot product - centralizes platform-specific SIMD dispatch
macro_rules! dispatch_dot {
($backend:expr, $a:expr, $b:expr) => {{
#[cfg(target_arch = "x86_64")]
use crate::backends::{avx2::Avx2Backend, sse2::Sse2Backend};
use crate::backends::{scalar::ScalarBackend, VectorBackend};
// SAFETY: CPU features verified at runtime before backend selection
unsafe {
match $backend {
Backend::Scalar => ScalarBackend::dot($a, $b),
#[cfg(target_arch = "x86_64")]
Backend::SSE2 | Backend::AVX => Sse2Backend::dot($a, $b),
#[cfg(target_arch = "x86_64")]
Backend::AVX2 | Backend::AVX512 => Avx2Backend::dot($a, $b),
#[cfg(not(target_arch = "x86_64"))]
Backend::SSE2 | Backend::AVX | Backend::AVX2 | Backend::AVX512 => {
ScalarBackend::dot($a, $b)
}
#[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
Backend::NEON => {
use crate::backends::neon::NeonBackend;
NeonBackend::dot($a, $b)
}
#[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
Backend::NEON => ScalarBackend::dot($a, $b),
#[cfg(target_arch = "wasm32")]
Backend::WasmSIMD => {
use crate::backends::wasm::WasmBackend;
WasmBackend::dot($a, $b)
}
#[cfg(not(target_arch = "wasm32"))]
Backend::WasmSIMD => ScalarBackend::dot($a, $b),
Backend::GPU | Backend::Auto => ScalarBackend::dot($a, $b),
}
}
}};
}
use super::super::Matrix;
impl Matrix<f32> {
/// Transpose this matrix (swap rows and columns)
///
/// Returns a new matrix with dimensions swapped: `self.rows → result.cols`,
/// `self.cols → result.rows`.
///
/// # Performance
///
/// Uses cache-optimized block-wise transpose with 32x32 blocks.
/// Sequential writes for output ensure good cache behavior.
///
/// # Example
///
/// ```
/// use trueno::Matrix;
///
/// let m = Matrix::from_vec(2, 3, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
/// let t = m.transpose();
///
/// // [[1, 2, 3], [[1, 4],
/// // [4, 5, 6]] → [2, 5],
/// // [3, 6]]
/// assert_eq!(t.rows(), 3);
/// assert_eq!(t.cols(), 2);
/// assert_eq!(t.get(0, 0), Some(&1.0));
/// assert_eq!(t.get(0, 1), Some(&4.0));
/// assert_eq!(t.get(1, 0), Some(&2.0));
/// ```
// KAIZEN-040: Delegate to crate::blis::transpose which has AVX2 8×8
// in-register micro-kernel with 64×64 L1-resident tiling and prefetch.
// Previous implementation used scalar 32×32 blocks.
#[cfg_attr(feature = "tracing", instrument(skip(self), fields(dims = %format!("{}x{}", self.rows, self.cols))))]
pub fn transpose(&self) -> Matrix<f32> {
let mut result = Matrix::zeros_with_backend(self.cols, self.rows, self.backend);
// BLIS transpose handles AVX2 dispatch, remainder edges, and shape-adaptive
// loop ordering internally. Dimensions are correct by construction so
// the only possible error (size mismatch) cannot occur.
if let Err(e) =
crate::blis::transpose::transpose(self.rows, self.cols, &self.data, &mut result.data)
{
// Unreachable: result is allocated as cols×rows which matches rows×cols elements.
// If somehow triggered, fall back to scalar element-wise transpose.
debug_assert!(false, "BLIS transpose dimension mismatch: {e}");
for i in 0..self.rows {
for j in 0..self.cols {
result.data[j * self.rows + i] = self.data[i * self.cols + j];
}
}
}
result
}
/// Matrix-vector multiplication (column vector): A × v
///
/// Multiplies this matrix by a column vector, computing `A × v` where the result
/// is a column vector with length equal to the number of rows in `A`.
///
/// # Mathematical Definition
///
/// For an m×n matrix A and an n-dimensional vector v:
/// ```text
/// result[i] = Σ(j=0 to n-1) A[i,j] × v[j]
/// ```
///
/// # Arguments
///
/// * `v` - Column vector with length equal to `self.cols()`
///
/// # Returns
///
/// A new vector with length `self.rows()`
///
/// # Errors
///
/// Returns `InvalidInput` if `v.len() != self.cols()`
///
/// # Example
///
/// ```
/// use trueno::{Matrix, Vector};
///
/// let m = Matrix::from_vec(2, 3, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
/// let v = Vector::from_slice(&[1.0, 2.0, 3.0]);
/// let result = m.matvec(&v).unwrap();
///
/// // [[1, 2, 3] [1] [1×1 + 2×2 + 3×3] [14]
/// // [4, 5, 6]] × [2] = [4×1 + 5×2 + 6×3] = [32]
/// // [3]
/// assert_eq!(result.as_slice(), &[14.0, 32.0]);
/// ```
pub fn matvec(&self, v: &Vector<f32>) -> Result<Vector<f32>, TruenoError> {
if v.len() != self.cols {
return Err(TruenoError::InvalidInput(format!(
"Vector length {} does not match matrix columns {} for matrix-vector multiplication",
v.len(),
self.cols
)));
}
let v_slice = v.as_slice();
let mut result_data = vec![0.0; self.rows];
// Parallel execution for very large matrices (≥4096 rows)
// Note: Thread overhead dominates for smaller matrices
#[cfg(feature = "parallel")]
{
const PARALLEL_THRESHOLD: usize = 4096;
if self.rows >= PARALLEL_THRESHOLD {
use rayon::prelude::*;
use std::sync::atomic::{AtomicPtr, Ordering};
use std::sync::Arc;
let result_ptr = Arc::new(AtomicPtr::new(result_data.as_mut_ptr()));
// Process rows in parallel - each row computes an independent dot product
(0..self.rows).into_par_iter().for_each(|i| {
let row_start = i * self.cols;
let row = &self.data[row_start..(row_start + self.cols)];
let dot_result = dispatch_dot!(self.backend, row, v_slice);
// Write to non-overlapping memory location (thread-safe)
// SAFETY: CPU feature verified at runtime, slices bounds-checked
unsafe {
let ptr = result_ptr.load(Ordering::Relaxed);
*ptr.add(i) = dot_result;
}
});
return Ok(Vector::from_slice(&result_data));
}
}
// SIMD-optimized execution: each row-vector product is a dot product
for (i, result) in result_data.iter_mut().enumerate() {
let row_start = i * self.cols;
let row = &self.data[row_start..(row_start + self.cols)];
// Use SIMD dot product for each row
*result = dispatch_dot!(self.backend, row, v_slice);
}
Ok(Vector::from_slice(&result_data))
}
/// Vector-matrix multiplication (row vector): v^T × A
///
/// Multiplies a row vector by this matrix, computing `v^T × A` where the result
/// is a row vector with length equal to the number of columns in `A`.
///
/// # Mathematical Definition
///
/// For an m-dimensional vector v and an m×n matrix A:
/// ```text
/// result[j] = Σ(i=0 to m-1) v[i] × A[i,j]
/// ```
///
/// # Arguments
///
/// * `v` - Row vector with length equal to `m.rows()`
/// * `m` - Matrix to multiply
///
/// # Returns
///
/// A new vector with length `m.cols()`
///
/// # Errors
///
/// Returns `InvalidInput` if `v.len() != m.rows()`
///
/// # Example
///
/// ```
/// use trueno::{Matrix, Vector};
///
/// let m = Matrix::from_vec(2, 3, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
/// let v = Vector::from_slice(&[1.0, 2.0]);
/// let result = Matrix::vecmat(&v, &m).unwrap();
///
/// // [1, 2] × [[1, 2, 3] = [1×1 + 2×4, 1×2 + 2×5, 1×3 + 2×6]
/// // [4, 5, 6]]
/// // = [9, 12, 15]
/// assert_eq!(result.as_slice(), &[9.0, 12.0, 15.0]);
/// ```
// KAIZEN-041: Uses crate::blis::gemv with AVX2 VFMADD,
// 4-way K-unrolling and N-tiled accumulators.
pub fn vecmat(v: &Vector<f32>, m: &Matrix<f32>) -> Result<Vector<f32>, TruenoError> {
if v.len() != m.rows {
return Err(TruenoError::InvalidInput(format!(
"Vector length {} does not match matrix rows {} for vector-matrix multiplication",
v.len(),
m.rows
)));
}
let mut result_data = vec![0.0f32; m.cols];
crate::blis::gemv::gemv(m.rows, m.cols, v.as_slice(), &m.data, &mut result_data);
Ok(Vector::from_slice(&result_data))
}
}
#[cfg(test)]
mod tests;