lax/layout.rs
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
//! Memory layout of matrices
//!
//! Different from ndarray format which consists of shape and strides,
//! matrix format in LAPACK consists of row or column size and leading dimension.
//!
//! ndarray format and stride
//! --------------------------
//!
//! Let us consider 3-dimensional array for explaining ndarray structure.
//! The address of `(x,y,z)`-element in ndarray satisfies following relation:
//!
//! ```text
//! shape = [Nx, Ny, Nz]
//! where Nx > 0, Ny > 0, Nz > 0
//! stride = [Sx, Sy, Sz]
//!
//! &data[(x, y, z)] = &data[(0, 0, 0)] + Sx*x + Sy*y + Sz*z
//! for x < Nx, y < Ny, z < Nz
//! ```
//!
//! The array is called
//!
//! - C-continuous if `[Sx, Sy, Sz] = [Nz*Ny, Nz, 1]`
//! - F(Fortran)-continuous if `[Sx, Sy, Sz] = [1, Nx, Nx*Ny]`
//!
//! Strides of ndarray `[Sx, Sy, Sz]` take arbitrary value,
//! e.g. it can be non-ordered `Sy > Sx > Sz`, or can be negative `Sx < 0`.
//! If the minimum of `[Sx, Sy, Sz]` equals to `1`,
//! the value of elements fills `data` memory region and called "continuous".
//! Non-continuous ndarray is useful to get sub-array without copying data.
//!
//! Matrix layout for LAPACK
//! -------------------------
//!
//! LAPACK interface focuses on the linear algebra operations for F-continuous 2-dimensional array.
//! Under this restriction, stride becomes far simpler; we only have to consider the case `[1, S]`
//! This `S` for a matrix `A` is called "leading dimension of the array A" in LAPACK document, and denoted by `lda`.
//!
use super::*;
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum MatrixLayout {
C { row: i32, lda: i32 },
F { col: i32, lda: i32 },
}
impl MatrixLayout {
pub fn size(&self) -> (i32, i32) {
match *self {
MatrixLayout::C { row, lda } => (row, lda),
MatrixLayout::F { col, lda } => (lda, col),
}
}
pub fn resized(&self, row: i32, col: i32) -> MatrixLayout {
match *self {
MatrixLayout::C { .. } => MatrixLayout::C { row, lda: col },
MatrixLayout::F { .. } => MatrixLayout::F { col, lda: row },
}
}
pub fn lda(&self) -> i32 {
std::cmp::max(
1,
match *self {
MatrixLayout::C { lda, .. } | MatrixLayout::F { lda, .. } => lda,
},
)
}
pub fn len(&self) -> i32 {
match *self {
MatrixLayout::C { row, .. } => row,
MatrixLayout::F { col, .. } => col,
}
}
pub fn is_empty(&self) -> bool {
self.len() == 0
}
pub fn same_order(&self, other: &MatrixLayout) -> bool {
match (self, other) {
(MatrixLayout::C { .. }, MatrixLayout::C { .. }) => true,
(MatrixLayout::F { .. }, MatrixLayout::F { .. }) => true,
_ => false,
}
}
pub fn toggle_order(&self) -> Self {
match *self {
MatrixLayout::C { row, lda } => MatrixLayout::F { lda: row, col: lda },
MatrixLayout::F { col, lda } => MatrixLayout::C { row: lda, lda: col },
}
}
/// Transpose without changing memory representation
///
/// C-contigious row=2, lda=3
///
/// ```text
/// [[1, 2, 3]
/// [4, 5, 6]]
/// ```
///
/// and F-contigious col=2, lda=3
///
/// ```text
/// [[1, 4]
/// [2, 5]
/// [3, 6]]
/// ```
///
/// have same memory representation `[1, 2, 3, 4, 5, 6]`, and this toggles them.
///
/// ```
/// # use lax::layout::*;
/// let layout = MatrixLayout::C { row: 2, lda: 3 };
/// assert_eq!(layout.t(), MatrixLayout::F { col: 2, lda: 3 });
/// ```
pub fn t(&self) -> Self {
match *self {
MatrixLayout::C { row, lda } => MatrixLayout::F { col: row, lda },
MatrixLayout::F { col, lda } => MatrixLayout::C { row: col, lda },
}
}
}
/// In-place transpose of a square matrix by keeping F/C layout
///
/// Transpose for C-continuous array
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::C { row: 2, lda: 2 };
/// let mut a = vec![1., 2., 3., 4.];
/// square_transpose(layout, &mut a);
/// assert_eq!(a, &[1., 3., 2., 4.]);
/// ```
///
/// Transpose for F-continuous array
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::F { col: 2, lda: 2 };
/// let mut a = vec![1., 3., 2., 4.];
/// square_transpose(layout, &mut a);
/// assert_eq!(a, &[1., 2., 3., 4.]);
/// ```
///
/// Panics
/// ------
/// - If size of `a` and `layout` size mismatch
///
pub fn square_transpose<T: Copy>(layout: MatrixLayout, a: &mut [T]) {
let (m, n) = layout.size();
let n = n as usize;
let m = m as usize;
assert_eq!(a.len(), n * m);
for i in 0..m {
for j in (i + 1)..n {
let a_ij = a[i * n + j];
let a_ji = a[j * m + i];
a[i * n + j] = a_ji;
a[j * m + i] = a_ij;
}
}
}
/// Out-place transpose for general matrix
///
/// Examples
/// ---------
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::C { row: 2, lda: 3 };
/// let a = vec![1., 2., 3., 4., 5., 6.];
/// let (l, b) = transpose(layout, &a);
/// assert_eq!(l, MatrixLayout::F { col: 3, lda: 2 });
/// assert_eq!(b, &[1., 4., 2., 5., 3., 6.]);
/// ```
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::F { col: 2, lda: 3 };
/// let a = vec![1., 2., 3., 4., 5., 6.];
/// let (l, b) = transpose(layout, &a);
/// assert_eq!(l, MatrixLayout::C { row: 3, lda: 2 });
/// assert_eq!(b, &[1., 4., 2., 5., 3., 6.]);
/// ```
///
/// Panics
/// ------
/// - If input array size and `layout` size mismatch
///
pub fn transpose<T: Copy>(layout: MatrixLayout, input: &[T]) -> (MatrixLayout, Vec<T>) {
let (m, n) = layout.size();
let transposed = layout.resized(n, m).t();
let m = m as usize;
let n = n as usize;
assert_eq!(input.len(), m * n);
let mut out: Vec<MaybeUninit<T>> = vec_uninit(m * n);
match layout {
MatrixLayout::C { .. } => {
for i in 0..m {
for j in 0..n {
out[j * m + i].write(input[i * n + j]);
}
}
}
MatrixLayout::F { .. } => {
for i in 0..m {
for j in 0..n {
out[i * n + j].write(input[j * m + i]);
}
}
}
}
(transposed, unsafe { out.assume_init() })
}
/// Out-place transpose for general matrix
///
/// Examples
/// ---------
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::C { row: 2, lda: 3 };
/// let a = vec![1., 2., 3., 4., 5., 6.];
/// let mut b = vec![0.0; a.len()];
/// let l = transpose_over(layout, &a, &mut b);
/// assert_eq!(l, MatrixLayout::F { col: 3, lda: 2 });
/// assert_eq!(b, &[1., 4., 2., 5., 3., 6.]);
/// ```
///
/// ```rust
/// # use lax::layout::*;
/// let layout = MatrixLayout::F { col: 2, lda: 3 };
/// let a = vec![1., 2., 3., 4., 5., 6.];
/// let mut b = vec![0.0; a.len()];
/// let l = transpose_over(layout, &a, &mut b);
/// assert_eq!(l, MatrixLayout::C { row: 3, lda: 2 });
/// assert_eq!(b, &[1., 4., 2., 5., 3., 6.]);
/// ```
///
/// Panics
/// ------
/// - If input array sizes and `layout` size mismatch
///
pub fn transpose_over<T: Copy>(layout: MatrixLayout, from: &[T], to: &mut [T]) -> MatrixLayout {
let (m, n) = layout.size();
let transposed = layout.resized(n, m).t();
let m = m as usize;
let n = n as usize;
assert_eq!(from.len(), m * n);
assert_eq!(to.len(), m * n);
match layout {
MatrixLayout::C { .. } => {
for i in 0..m {
for j in 0..n {
to[j * m + i] = from[i * n + j];
}
}
}
MatrixLayout::F { .. } => {
for i in 0..m {
for j in 0..n {
to[i * n + j] = from[j * m + i];
}
}
}
}
transposed
}