Adding permute_axes method for Tensor2D, Tensor3D, Tensor4D

src/devices/mod.rs

@@ -5,6 +5,7 @@ mod broadcast;
 mod fill;
 mod foreach;
 mod matmul;
+mod permute;
 mod reduce_all;
 mod reduce_axis;
 mod select;
@@ -14,6 +15,7 @@ pub use broadcast::*;
 pub use fill::*;
 pub use foreach::*;
 pub use matmul::*;
+pub use permute::*;
 pub use reduce_all::*;
 pub use reduce_axis::*;
 pub use select::*;

src/devices/permute.rs

@@ -0,0 +1,273 @@
+//! Permutation implementation is basically just looping over
+//! all the elements and reordering them.
+//!
+//! This implementation heavily relies on macros to expand on the
+//! possible versions. In the future it may be possible to refactor
+//! this to use const generics, however the attempt at this on
+//! initial implementation was too verbose and hard to understand.
+//!
+//! - `permutations!` expands all the possible permutations of axes
+//! - `impl_permute!` does the actual implementation.
+//! - [permuted_loop2], [permuted_loop3], [permuted_loop4], and [const_idx]
+//!   are used to do the permutations.
+//!
+//! [permuted_loop2] takes in a function that receives the unpermuted set of
+//! indices, and the permuted set of indices. This type of function enables
+//! only specifying the looping & indexing logic once. Both
+//! [DevicePermute2D::permute] and [DevicePermute2D::inverse_permute] share
+//! this looping logic, but only differ in what they do with the indices.
+
+use super::Cpu;
+
+/// Permutes axes of `A` resulting in `B`.
+pub trait DevicePermute2D<A, B, const I: isize, const J: isize> {
+    fn permute(a: &A, b: &mut B);
+    fn inverse_permute(a: &mut A, b: &B);
+}
+
+/// Permutes axes of `A` resulting in `B`.
+pub trait DevicePermute3D<A, B, const I: isize, const J: isize, const K: isize> {
+    fn permute(a: &A, b: &mut B);
+    fn inverse_permute(a: &mut A, b: &B);
+}
+
+/// Permutes axes of `A` resulting in `B`.
+pub trait DevicePermute4D<A, B, const I: isize, const J: isize, const K: isize, const L: isize> {
+    fn permute(a: &A, b: &mut B);
+    fn inverse_permute(a: &mut A, b: &B);
+}
+
+/// Expands to the const generic for a specific axis. This is purely convention only.
+#[rustfmt::skip]
+macro_rules! axis { (0) => { M }; (1) => { N }; (2) => { O }; (3) => { P }; }
+
+/// Expands to a array type using the axes passed in.
+/// E.g. `array!(2, 0, 1)` expands to `[[[f32; N]; M]; O]`
+#[rustfmt::skip]
+macro_rules! array {
+    ($Ax0:tt) => { [f32; axis!($Ax0)] };
+    ($Ax0:tt, $Ax1:tt) => { [[f32; axis!($Ax1)]; axis!($Ax0)] };
+    ($Ax0:tt, $Ax1:tt, $Ax2:tt) => { [[[f32; axis!($Ax2)]; axis!($Ax1)]; axis!($Ax0)] };
+    ($Ax0:tt, $Ax1:tt, $Ax2:tt, $Ax3:tt) => { [[[[f32; axis!($Ax3)]; axis!($Ax2)]; axis!($Ax1)]; axis!($Ax0)] };
+}
+
+/// Concrete implementations for the permute and inverse permute functions.
+#[rustfmt::skip]
+macro_rules! impl_permute {
+    ($Ax0:tt, $Ax1:tt) => {
+impl<const M: usize, const N: usize>
+    DevicePermute2D<array!(0, 1), array!($Ax0, $Ax1), $Ax0, $Ax1> for Cpu
+{
+    fn permute(a: &array!(0, 1), b: &mut array!($Ax0, $Ax1)) {
+        permuted_loop2::<M, N, $Ax0, $Ax1, _>(&mut |[m, n], [i, j]| {
+            b[i][j] = a[m][n];
+        });
+    }
+    fn inverse_permute(a: &mut array!(0, 1), b: &array!($Ax0, $Ax1)) {
+        permuted_loop2::<M, N, $Ax0, $Ax1, _>(&mut |[m, n], [i, j]| {
+            a[m][n] = b[i][j];
+        });
+    }
+}
+    };
+    ($Ax0:tt, $Ax1:tt, $Ax2:tt) => {
+impl<const M: usize, const N: usize, const O: usize>
+    DevicePermute3D<array!(0, 1, 2), array!($Ax0, $Ax1, $Ax2), $Ax0, $Ax1, $Ax2> for Cpu
+{
+    fn permute(a: &array!(0, 1, 2), b: &mut array!($Ax0, $Ax1, $Ax2)) {
+        permuted_loop3::<M, N, O, $Ax0, $Ax1, $Ax2, _>(&mut |[m, n, o], [i, j, k]| {
+            b[i][j][k] = a[m][n][o];
+        });
+    }
+    fn inverse_permute(a: &mut array!(0, 1, 2), b: &array!($Ax0, $Ax1, $Ax2)) {
+        permuted_loop3::<M, N, O, $Ax0, $Ax1, $Ax2, _>(&mut |[m, n, o], [i, j, k]| {
+            a[m][n][o] = b[i][j][k];
+        });
+    }
+}
+    };
+    ($Ax0:tt, $Ax1:tt, $Ax2:tt, $Ax3:tt) => {
+impl<const M: usize, const N: usize, const O: usize, const P: usize>
+    DevicePermute4D<array!(0,1,2,3), array!($Ax0,$Ax1,$Ax2,$Ax3), $Ax0,$Ax1,$Ax2,$Ax3> for Cpu
+{
+    fn permute(a: &array!(0, 1, 2, 3), b: &mut array!($Ax0, $Ax1, $Ax2, $Ax3)) {
+        permuted_loop4::<M, N, O, P, $Ax0, $Ax1, $Ax2, $Ax3, _>(
+            &mut |[m, n, o, p], [i, j, k, l]| {
+                b[i][j][k][l] = a[m][n][o][p];
+            },
+        );
+    }
+    fn inverse_permute(a: &mut array!(0, 1, 2, 3), b: &array!($Ax0, $Ax1, $Ax2, $Ax3)) {
+        permuted_loop4::<M, N, O, P, $Ax0, $Ax1, $Ax2, $Ax3, _>(
+            &mut |[m, n, o, p], [i, j, k, l]| {
+                a[m][n][o][p] = b[i][j][k][l];
+            },
+        );
+    }
+}
+    };
+}
+
+/// Index into `indices` using the const `I`. If `I` < 0 then use `N - I`.
+fn const_idx<const I: isize, const N: usize>(indices: &[usize; N]) -> usize {
+    if I < 0 {
+        indices[(N as isize - I) as usize]
+    } else {
+        indices[I as usize]
+    }
+}
+
+/// Apply a function `f` to two sets of 2d indices.
+fn permuted_loop2<const M: usize, const N: usize, const I: isize, const J: isize, F>(f: &mut F)
+where
+    F: FnMut([usize; 2], [usize; 2]),
+{
+    for m in 0..M {
+        for n in 0..N {
+            let indices = [m, n];
+            let i = const_idx::<I, 2>(&indices);
+            let j = const_idx::<J, 2>(&indices);
+            f(indices, [i, j]);
+        }
+    }
+}
+
+/// Apply a function `f` to two sets of 2d indices.
+fn permuted_loop3<
+    const M: usize,
+    const N: usize,
+    const O: usize,
+    const I: isize,
+    const J: isize,
+    const K: isize,
+    F: FnMut([usize; 3], [usize; 3]),
+>(
+    f: &mut F,
+) {
+    for m in 0..M {
+        for n in 0..N {
+            for o in 0..O {
+                let indices = [m, n, o];
+                let i = const_idx::<I, 3>(&indices);
+                let j = const_idx::<J, 3>(&indices);
+                let k = const_idx::<K, 3>(&indices);
+                f(indices, [i, j, k]);
+            }
+        }
+    }
+}
+
+/// Apply a function `f` to two sets of 2d indices.
+fn permuted_loop4<
+    const M: usize,
+    const N: usize,
+    const O: usize,
+    const P: usize,
+    const I: isize,
+    const J: isize,
+    const K: isize,
+    const L: isize,
+    F: FnMut([usize; 4], [usize; 4]),
+>(
+    f: &mut F,
+) {
+    for m in 0..M {
+        for n in 0..N {
+            for o in 0..O {
+                for p in 0..P {
+                    let indices = [m, n, o, p];
+                    let i = const_idx::<I, 4>(&indices);
+                    let j = const_idx::<J, 4>(&indices);
+                    let k = const_idx::<K, 4>(&indices);
+                    let l = const_idx::<L, 4>(&indices);
+                    f(indices, [i, j, k, l]);
+                }
+            }
+        }
+    }
+}
+
+/// Expand out all the possible permutations for 2-4d
+macro_rules! permutations {
+    ([$Ax0:tt, $Ax1:tt]) => {
+        impl_permute!($Ax0, $Ax1);
+        impl_permute!($Ax1, $Ax0);
+    };
+
+    ([$Ax0:tt, $Ax1:tt, $Ax2:tt]) => {
+        permutations!($Ax0, [$Ax1, $Ax2]);
+        permutations!($Ax1, [$Ax0, $Ax2]);
+        permutations!($Ax2, [$Ax0, $Ax1]);
+    };
+    ($Ax0:tt, [$Ax1:tt, $Ax2:tt]) => {
+        impl_permute!($Ax0, $Ax1, $Ax2);
+        impl_permute!($Ax0, $Ax2, $Ax1);
+    };
+
+    ([$Ax0:tt, $Ax1:tt, $Ax2:tt, $Ax3:tt]) => {
+        permutations!($Ax0, [$Ax1, $Ax2, $Ax3]);
+        permutations!($Ax1, [$Ax0, $Ax2, $Ax3]);
+        permutations!($Ax2, [$Ax0, $Ax1, $Ax3]);
+        permutations!($Ax3, [$Ax0, $Ax1, $Ax2]);
+    };
+    ($Ax0:tt, [$Ax1:tt, $Ax2:tt, $Ax3:tt]) => {
+        permutations!($Ax0, $Ax1, [$Ax2, $Ax3]);
+        permutations!($Ax0, $Ax2, [$Ax1, $Ax3]);
+        permutations!($Ax0, $Ax3, [$Ax1, $Ax2]);
+    };
+    ($Ax0:tt, $Ax1:tt, [$Ax2:tt, $Ax3:tt]) => {
+        impl_permute!($Ax0, $Ax1, $Ax2, $Ax3);
+        impl_permute!($Ax0, $Ax1, $Ax3, $Ax2);
+    };
+}
+
+permutations!([0, 1]);
+permutations!([0, 1, 2]);
+permutations!([0, 1, 2, 3]);
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::prelude::FillElements;
+    use rand::{thread_rng, Rng};
+
+    #[test]
+    fn test_2d_permute() {
+        let a = [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]];
+        let mut b = [[0.0; 2]; 3];
+        <Cpu as DevicePermute2D<_, _, 1, 0>>::permute(&a, &mut b);
+        assert_eq!(b, [[1.0, 4.0], [2.0, 5.0], [3.0, 6.0]]);
+
+        let mut c = [[0.0; 3]; 2];
+        <Cpu as DevicePermute2D<_, _, 1, 0>>::inverse_permute(&mut c, &b);
+        assert_eq!(a, c);
+    }
+
+    #[test]
+    fn test_3d_permute() {
+        let a = [[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]];
+        let mut b = [[[0.0; 1]; 2]; 3];
+        <Cpu as DevicePermute3D<_, _, 2, 1, 0>>::permute(&a, &mut b);
+        assert_eq!(b, [[[1.0], [4.0]], [[2.0], [5.0]], [[3.0], [6.0]]]);
+
+        let mut c = [[[0.0; 3]; 2]; 1];
+        <Cpu as DevicePermute3D<_, _, 2, 1, 0>>::inverse_permute(&mut c, &b);
+        assert_eq!(a, c);
+    }
+
+    #[test]
+    fn test_4d_permute() {
+        let mut rng = thread_rng();
+        let mut a = [[[[0.0; 9]; 7]; 5]; 3];
+        Cpu::fill(&mut a, &mut |v| *v = rng.gen());
+
+        let mut b = [[[[0.0; 3]; 5]; 9]; 7];
+        <Cpu as DevicePermute4D<_, _, 2, 3, 1, 0>>::permute(&a, &mut b);
+        assert_ne!(b, [[[[0.0; 3]; 5]; 9]; 7]);
+
+        let mut c = [[[[0.0; 9]; 7]; 5]; 3];
+        <Cpu as DevicePermute4D<_, _, 2, 3, 1, 0>>::inverse_permute(&mut c, &b);
+
+        assert_eq!(a, c);
+    }
+}

src/tensor_ops/mod.rs

@@ -117,6 +117,7 @@ mod impl_sum;
 mod impl_sum_axis;
 mod map;
 mod matmul;
+mod permute;
 mod reduce;
 mod select;
 pub(crate) mod utils;
@@ -146,6 +147,7 @@ pub use impl_sum::*;
 pub use impl_sum_axis::*;
 pub use map::*;
 pub use matmul::*;
+pub use permute::{Permute2DSugar, Permute3DSugar, Permute4DSugar};
 pub use reduce::*;
 pub use select::*;