teeny_kernels/nn/pad/
reflection_pad3d.rs

1/*
2 * Copyright (c) 2026 Teenygrad.
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *   http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#![allow(non_snake_case)]
18
19use core::ops::{BitAnd, BitOr};
20
21use teeny_core::dtype::Num;
22use teeny_macros::kernel;
23use teeny_triton::triton::{
24    types::{AddOffsets, Comparison, Tensor},
25    *,
26};
27
28/// 3-D reflection padding forward pass.
29///
30/// Grid: `pid = (((b*C+c)*OD+od)*OH+oh)*num_ow_tiles + ow_tile`
31#[kernel]
32pub fn reflection_pad3d_forward<
33    T: Triton,
34    D: Num,
35    const PD1: i32,
36    const PD2: i32,
37    const PH1: i32,
38    const PH2: i32,
39    const PW1: i32,
40    const PW2: i32,
41    const BLOCK_OW: i32,
42>(
43    input_ptr: T::Pointer<D>,
44    output_ptr: T::Pointer<D>,
45    _B: i32,
46    C: i32,
47    Dv: i32,
48    H: i32,
49    W: i32,
50    OD: i32,
51    OH: i32,
52    OW: i32,
53) where
54    T::I32Tensor: Tensor<i32, 1>,
55    T::I32Tensor: Comparison<i32, BoolTensor = T::BoolTensor>,
56    T::BoolTensor: BitAnd<Output = T::BoolTensor>,
57    T::BoolTensor: BitOr<Output = T::BoolTensor>,
58    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
59{
60    let pid = T::program_id(Axis::X);
61    let num_ow_tiles = T::cdiv(OW, BLOCK_OW);
62
63    let ow_tile = pid % num_ow_tiles;
64    let rest = pid / num_ow_tiles;
65    let oh = rest % OH;
66    let rest2 = rest / OH;
67    let od = rest2 % OD;
68    let bco = rest2 / OD;
69    let c = bco % C;
70    let b = bco / C;
71
72    let ow_start = ow_tile * BLOCK_OW;
73    let ow_range = T::arange(0, BLOCK_OW) + ow_start;
74    let ow_mask = ow_range.lt(OW);
75
76    let id_raw = od - PD1;
77    let id = if id_raw < 0 {
78        -id_raw
79    } else if id_raw >= Dv {
80        2 * (Dv - 1) - id_raw
81    } else {
82        id_raw
83    };
84
85    let ih_raw = oh - PH1;
86    let ih = if ih_raw < 0 {
87        -ih_raw
88    } else if ih_raw >= H {
89        2 * (H - 1) - ih_raw
90    } else {
91        ih_raw
92    };
93
94    let in_bc_base = ((b * C + c) * Dv + id) * H * W + ih * W;
95    let out_bc_base = (((b * C + c) * OD + od) * OH + oh) * OW;
96
97    let iw_raw = ow_range - PW1;
98    let left_cond = iw_raw.lt(0);
99    let right_cond = iw_raw.ge(W);
100    let in_bounds = iw_raw.ge(0) & iw_raw.lt(W);
101
102    let iw_left = iw_raw * (-1);
103    let iw_right = iw_raw * (-1) + (2 * (W - 1));
104
105    let zeros = T::zeros::<D>(&[BLOCK_OW]);
106    let val_center = T::load(
107        input_ptr.add_offsets(iw_raw + in_bc_base),
108        Some(ow_mask & in_bounds),
109        Some(zeros),
110        &[],
111        None,
112        None,
113        None,
114        false,
115    );
116    let val_left = T::load(
117        input_ptr.add_offsets(iw_left + in_bc_base),
118        Some(ow_mask & left_cond),
119        Some(zeros),
120        &[],
121        None,
122        None,
123        None,
124        false,
125    );
126    let val_right = T::load(
127        input_ptr.add_offsets(iw_right + in_bc_base),
128        Some(ow_mask & right_cond),
129        Some(zeros),
130        &[],
131        None,
132        None,
133        None,
134        false,
135    );
136
137    let result = T::where_(
138        left_cond,
139        val_left,
140        T::where_(right_cond, val_right, val_center),
141    );
142
143    let out_offsets = ow_range + out_bc_base;
144    T::store(
145        output_ptr.add_offsets(out_offsets),
146        result,
147        Some(ow_mask),
148        &[],
149        None,
150        None,
151    );
152}
153
154/// 3-D reflection padding backward pass.
155#[kernel]
156pub fn reflection_pad3d_backward<
157    T: Triton,
158    D: Num,
159    const PD1: i32,
160    const PD2: i32,
161    const PH1: i32,
162    const PH2: i32,
163    const PW1: i32,
164    const PW2: i32,
165    const BLOCK_OW: i32,
166>(
167    dy_ptr: T::Pointer<D>,
168    dx_ptr: T::Pointer<D>,
169    _B: i32,
170    C: i32,
171    Dv: i32,
172    H: i32,
173    W: i32,
174    OD: i32,
175    OH: i32,
176    OW: i32,
177) where
178    T::I32Tensor: Tensor<i32, 1>,
179    T::I32Tensor: Comparison<i32, BoolTensor = T::BoolTensor>,
180    T::BoolTensor: BitAnd<Output = T::BoolTensor>,
181    T::BoolTensor: BitOr<Output = T::BoolTensor>,
182    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
183{
184    let pid = T::program_id(Axis::X);
185    let num_ow_tiles = T::cdiv(OW, BLOCK_OW);
186
187    let ow_tile = pid % num_ow_tiles;
188    let rest = pid / num_ow_tiles;
189    let oh = rest % OH;
190    let rest2 = rest / OH;
191    let od = rest2 % OD;
192    let bco = rest2 / OD;
193    let c = bco % C;
194    let b = bco / C;
195
196    let ow_start = ow_tile * BLOCK_OW;
197    let ow_range = T::arange(0, BLOCK_OW) + ow_start;
198    let ow_mask = ow_range.lt(OW);
199
200    let id_raw = od - PD1;
201    let id = if id_raw < 0 {
202        -id_raw
203    } else if id_raw >= Dv {
204        2 * (Dv - 1) - id_raw
205    } else {
206        id_raw
207    };
208
209    let ih_raw = oh - PH1;
210    let ih = if ih_raw < 0 {
211        -ih_raw
212    } else if ih_raw >= H {
213        2 * (H - 1) - ih_raw
214    } else {
215        ih_raw
216    };
217
218    let dy_bc_base = (((b * C + c) * OD + od) * OH + oh) * OW;
219    let dx_bc_base = ((b * C + c) * Dv + id) * H * W + ih * W;
220
221    let dy_offsets = ow_range + dy_bc_base;
222    let dy_tile = T::load(
223        dy_ptr.add_offsets(dy_offsets),
224        Some(ow_mask),
225        Some(T::zeros::<D>(&[BLOCK_OW])),
226        &[],
227        None,
228        None,
229        None,
230        false,
231    );
232
233    let iw_raw = ow_range - PW1;
234    let left_cond = iw_raw.lt(0);
235    let right_cond = iw_raw.ge(W);
236    let in_bounds = iw_raw.ge(0) & iw_raw.lt(W);
237
238    let iw_left = iw_raw * (-1);
239    let iw_right = iw_raw * (-1) + (2 * (W - 1));
240
241    T::atomic_add(
242        dx_ptr.add_offsets(iw_raw + dx_bc_base),
243        dy_tile,
244        Some(ow_mask & in_bounds),
245        None,
246        None,
247    );
248    T::atomic_add(
249        dx_ptr.add_offsets(iw_left + dx_bc_base),
250        dy_tile,
251        Some(ow_mask & left_cond),
252        None,
253        None,
254    );
255    T::atomic_add(
256        dx_ptr.add_offsets(iw_right + dx_bc_base),
257        dy_tile,
258        Some(ow_mask & right_cond),
259        None,
260        None,
261    );
262}
263
264pub struct ReflectionPad3dOp<'a, T: Num> {
265    pub forward: ReflectionPad3dForward<T>,
266    pub backward: ReflectionPad3dBackward<T>,
267    _marker: core::marker::PhantomData<&'a ()>,
268}
teeny_kernels/nn/pad/reflection_pad3d.rs

teeny_kernels/nn/pad/
reflection_pad3d.rs
