zyx 0.17.0

Zyx machine learning library
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

// Copyright (C) 2025 zk4x
// SPDX-License-Identifier: LGPL-3.0-only WITH Classpath-exception-2.0

//! Tenstorrent tilized layouts (host-side only).
//!
//! A tilized tensor stores each 32x32 tile contiguously, with the four
//! 16x16 faces inside each tile in face order (the same convention as the
//! `14_tenstorrent` golden kernel's `lin()` mapping). Outer dims are
//! zero-padded up to multiples of 32. This is a pure host-side byte
//! permutation. Tilize once at load time (weights) or once per forward
//! (inputs); on-device kernels chain tilized layouts with no conversion
//! between layers.

use crate::{Tensor, ZyxError, dtype::DType, scalar::Scalar, shape::Dim};

impl Tensor {
    /// Permutes the last two dims into Tenstorrent tilized face order
    /// (host-side), zero-padding them to multiples of 32; dtype is preserved.
    ///
    /// # Example
    ///
    /// ```rust no_run
    /// use zyx::{Tensor, DType};
    /// let t = Tensor::zeros([32, 48], DType::F32);
    /// let til = t.tilize()?;
    /// # Ok::<(), zyx::ZyxError>(())
    /// ```
    ///
    /// # Errors
    /// Returns a device error if the tensor cannot be read to host, or
    /// [`ZyxError::AllocationError`] if the result cannot be allocated.
    pub fn tilize(&self) -> Result<Tensor, ZyxError> {
        fn ceil32(d: i64) -> i64 {
            debug_assert!(d >= 0, "tilize needs a concrete non-negative dim, got {d}");
            (d + 31) / 32 * 32
        }
        fn permute<T: Scalar>(input: &[T], batches: usize, rows: i64, cols: i64, pr: i64, pc: i64) -> Vec<T> {
            let ntc = pc / 32;
            let mut out = vec![T::zero(); batches * (pr * pc) as usize];
            for b in 0..batches {
                for tr in 0..pr / 32 {
                    for tc in 0..ntc {
                        for f in 0..4 {
                            let fr = f / 2;
                            let fc = f % 2;
                            for l in 0..256 {
                                let lr = l / 16;
                                let lc = l % 16;
                                let r = tr * 32 + fr * 16 + lr;
                                let c = tc * 32 + fc * 16 + lc;
                                let p = ((tr * ntc + tc) * 4 + f) * 256 + l;
                                let v = if r < rows && c < cols {
                                    input[b * (rows * cols) as usize + (r * cols + c) as usize]
                                } else {
                                    T::zero()
                                };
                                out[b * (pr * pc) as usize + p as usize] = v;
                            }
                        }
                    }
                }
            }
            out
        }
        fn run<T: Scalar>(t: &Tensor, shape: &[i64]) -> Result<Tensor, ZyxError> {
            let rank = shape.len();
            debug_assert!(rank >= 2, "tilize needs rank >= 2, got {rank}");
            let (rows, cols) = (shape[rank - 2], shape[rank - 1]);
            let batches: usize = shape[..rank - 2].iter().product::<i64>() as usize;
            let input: Vec<T> = t.to_vec()?;
            let (pr, pc) = (ceil32(rows), ceil32(cols));
            let out = permute(&input, batches, rows, cols, pr, pc);
            let mut new_shape: Vec<i64> = shape[..rank - 2].to_vec();
            new_shape.push(pr);
            new_shape.push(pc);
            Tensor::from_vec(out, new_shape)
        }
        let shape: Vec<i64> = self.shape().iter().map(|d| d.item::<Dim>() as i64).collect();
        match self.dtype() {
            DType::F32 => run::<f32>(self, &shape),
            DType::F16 => run::<crate::scalar::f16>(self, &shape),
            DType::BF16 => run::<crate::scalar::bf16>(self, &shape),
            DType::F8E4M3 => run::<crate::scalar::f8e4m3>(self, &shape),
            DType::F8E5M2 => run::<crate::scalar::f8e5m2>(self, &shape),
            DType::F64 => run::<f64>(self, &shape),
            DType::U8 => run::<u8>(self, &shape),
            DType::I8 => run::<i8>(self, &shape),
            DType::U16 => run::<u16>(self, &shape),
            DType::I16 => run::<i16>(self, &shape),
            DType::U32 => run::<u32>(self, &shape),
            DType::I32 => run::<i32>(self, &shape),
            DType::U64 => run::<u64>(self, &shape),
            DType::I64 => run::<i64>(self, &shape),
            DType::Bool => run::<bool>(self, &shape),
        }
    }

    /// Inverse of [`Tensor::tilize`]: strips the 32-aligned padding from the
    /// last two dims, restoring the true `rows` and `cols` shape.
    ///
    /// # Example
    ///
    /// ```rust no_run
    /// use zyx::{Tensor, DType};
    /// let til = Tensor::zeros([64, 64], DType::F32);
    /// let back = til.untilize(32, 48)?;
    /// # Ok::<(), zyx::ZyxError>(())
    /// ```
    ///
    /// # Errors
    /// Returns a device error if the tensor cannot be read to host, or
    /// [`ZyxError::AllocationError`] if the result cannot be allocated.
    pub fn untilize(&self, rows: i64, cols: i64) -> Result<Tensor, ZyxError> {
        fn permute<T: Scalar>(input: &[T], batches: usize, pr: i64, pc: i64, rows: i64, cols: i64) -> Vec<T> {
            let ntc = pc / 32;
            let mut out = vec![T::zero(); batches * (rows * cols) as usize];
            for b in 0..batches {
                for tr in 0..pr / 32 {
                    for tc in 0..ntc {
                        for f in 0..4 {
                            let fr = f / 2;
                            let fc = f % 2;
                            for l in 0..256 {
                                let lr = l / 16;
                                let lc = l % 16;
                                let r = tr * 32 + fr * 16 + lr;
                                let c = tc * 32 + fc * 16 + lc;
                                let p = ((tr * ntc + tc) * 4 + f) * 256 + l;
                                if r < rows && c < cols {
                                    out[b * (rows * cols) as usize + (r * cols + c) as usize] =
                                        input[b * (pr * pc) as usize + p as usize];
                                }
                            }
                        }
                    }
                }
            }
            out
        }
        fn run<T: Scalar>(t: &Tensor, shape: &[i64], rows: i64, cols: i64) -> Result<Tensor, ZyxError> {
            let rank = shape.len();
            debug_assert!(rank >= 2, "untilize needs rank >= 2, got {rank}");
            debug_assert!(shape[rank - 2] % 32 == 0 && shape[rank - 1] % 32 == 0);
            let (pr, pc) = (shape[rank - 2], shape[rank - 1]);
            debug_assert!(rows <= pr && cols <= pc);
            let batches: usize = shape[..rank - 2].iter().product::<i64>() as usize;
            let input: Vec<T> = t.to_vec()?;
            let out = permute(&input, batches, pr, pc, rows, cols);
            let mut new_shape: Vec<i64> = shape[..rank - 2].to_vec();
            new_shape.push(rows);
            new_shape.push(cols);
            Tensor::from_vec(out, new_shape)
        }
        let shape: Vec<i64> = self.shape().iter().map(|d| d.item::<Dim>() as i64).collect();
        match self.dtype() {
            DType::F32 => run::<f32>(self, &shape, rows, cols),
            DType::F16 => run::<crate::scalar::f16>(self, &shape, rows, cols),
            DType::BF16 => run::<crate::scalar::bf16>(self, &shape, rows, cols),
            DType::F8E4M3 => run::<crate::scalar::f8e4m3>(self, &shape, rows, cols),
            DType::F8E5M2 => run::<crate::scalar::f8e5m2>(self, &shape, rows, cols),
            DType::F64 => run::<f64>(self, &shape, rows, cols),
            DType::U8 => run::<u8>(self, &shape, rows, cols),
            DType::I8 => run::<i8>(self, &shape, rows, cols),
            DType::U16 => run::<u16>(self, &shape, rows, cols),
            DType::I16 => run::<i16>(self, &shape, rows, cols),
            DType::U32 => run::<u32>(self, &shape, rows, cols),
            DType::I32 => run::<i32>(self, &shape, rows, cols),
            DType::U64 => run::<u64>(self, &shape, rows, cols),
            DType::I64 => run::<i64>(self, &shape, rows, cols),
            DType::Bool => run::<bool>(self, &shape, rows, cols),
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn tilize_roundtrip_odd_shape() -> Result<(), ZyxError> {
        // Odd, non-tile-multiple shape exercises padding + face order.
        let rows = 37i64;
        let cols = 45i64;
        let data: Vec<f32> = (0..rows * cols).map(|i| i as f32 * 0.5 - 100.0).collect();
        let t = Tensor::from_vec(data.clone(), [rows, cols])?;
        let til = Tensor::tilize(&t)?;
        let shape: Vec<i64> = til.shape().iter().map(|d| d.item::<Dim>() as i64).collect();
        assert_eq!(shape, vec![64, 64]);
        // First tile, first face must hold rows 0..16, cols 0..16 in order.
        let flat: Vec<f32> = til.to_vec()?;
        for lr in 0..16 {
            for lc in 0..16 {
                assert_eq!(flat[(lr * 16 + lc) as usize], (lr * cols + lc) as f32 * 0.5 - 100.0);
            }
        }
        // Padded region reads back as zero.
        assert_eq!(flat[((63 * 64 + 63) % (64 * 64)) as usize], 0.0);
        let back = Tensor::untilize(&til, rows, cols)?;
        let shape: Vec<i64> = back.shape().iter().map(|d| d.item::<Dim>() as i64).collect();
        assert_eq!(shape, vec![rows, cols]);
        let rt: Vec<f32> = back.to_vec()?;
        assert_eq!(rt, data);
        Ok(())
    }

    #[test]
    fn tilize_matches_golden_face_order() -> Result<(), ZyxError> {
        // Conformance with the `14_tenstorrent` golden kernel's `lin()`
        // mapping: face slot -> linear index within a tile.
        let lin = |s: usize| {
            let (face, local) = (s / 256, s % 256);
            let (fr0, fc0) = (face / 2, face % 2);
            fr0 * 16 * 32 + fc0 * 16 + (local / 16) * 32 + local % 16
        };
        let data: Vec<f32> = (0..1024).map(|i| i as f32).collect();
        let t = Tensor::from_vec(data, [32, 32])?;
        let til = Tensor::tilize(&t)?;
        let flat: Vec<f32> = til.to_vec()?;
        for p in 0..1024 {
            assert_eq!(flat[p], lin(p) as f32, "face slot {p}");
        }
        Ok(())
    }

    #[test]
    fn tilize_batched_f16() -> Result<(), ZyxError> {
        let data: Vec<crate::scalar::f16> = (0..2 * 6 * 40).map(|i| crate::scalar::f16::from_f32(i as f32)).collect();
        let t = Tensor::from_vec(data.clone(), [2, 6, 40])?;
        let til = Tensor::tilize(&t)?;
        let shape: Vec<i64> = til.shape().iter().map(|d| d.item::<Dim>() as i64).collect();
        assert_eq!(shape, vec![2, 32, 64]);
        let back = Tensor::untilize(&til, 6, 40)?;
        let rt: Vec<crate::scalar::f16> = back.to_vec()?;
        assert_eq!(rt, data);
        Ok(())
    }
}