teeny_kernels/nn/pool/
avgpool3d.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 teeny_core::dtype::Num;
20use teeny_macros::kernel;
21use teeny_triton::triton::{
22    types::{AddOffsets, Comparison, Tensor},
23    *,
24};
25
26/// 3-D average-pooling forward pass.
27///
28/// Grid: `pid = (((b * C + c) * OD + od) * OH + oh) * num_ow_tiles + ow_tile`
29///
30/// **Constraints**: no padding;
31/// `OD = (D - KD) / STRIDE_D + 1`, `OH = (H - KH) / STRIDE_H + 1`,
32/// `OW = (W - KW) / STRIDE_W + 1`.
33#[kernel]
34pub fn avgpool3d_forward<
35    T: Triton,
36    D: Num,
37    const KD: i32,
38    const KH: i32,
39    const KW: i32,
40    const STRIDE_D: i32,
41    const STRIDE_H: i32,
42    const STRIDE_W: i32,
43    const BLOCK_OW: i32,
44>(
45    input_ptr: T::Pointer<D>,
46    output_ptr: T::Pointer<D>,
47    _B: i32,
48    C: i32,
49    Dv: i32,
50    H: i32,
51    W: i32,
52    OD: i32,
53    OH: i32,
54    OW: i32,
55) where
56    T::I32Tensor: Tensor<i32, 1>,
57    T::I32Tensor: Comparison<i32, BoolTensor = 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 in_bc_base = (b * C + c) * Dv * H * W;
77    let out_base = ((b * C + c) * OD * OH * OW) + od * OH * OW + oh * OW;
78
79    let mut acc = T::zeros::<D>(&[BLOCK_OW]);
80
81    let loop_bound = KD * KH * KW;
82    for idx in 0..loop_bound {
83        let kw = idx % KW;
84        let tmp = idx / KW;
85        let kh = tmp % KH;
86        let kd = tmp / KH;
87
88        let id = od * STRIDE_D + kd;
89        let ih = oh * STRIDE_H + kh;
90        let iw_range = ow_range * STRIDE_W + kw;
91        let in_offsets = iw_range + (in_bc_base + id * H * W + ih * W);
92        let tile = T::load(
93            input_ptr.add_offsets(in_offsets),
94            Some(ow_mask),
95            Some(T::zeros::<D>(&[BLOCK_OW])),
96            &[],
97            None,
98            None,
99            None,
100            false,
101        );
102        acc = acc + tile;
103    }
104
105    let ksize_1 = T::full::<i32>(&[1], KD * KH * KW);
106    let ksize_f_1 = T::cast::<i32, D>(ksize_1, None, false);
107    let ksize = T::broadcast_to(ksize_f_1, &[BLOCK_OW]);
108    let result = acc / ksize;
109
110    let out_offsets = ow_range + out_base;
111    T::store(
112        output_ptr.add_offsets(out_offsets),
113        result,
114        Some(ow_mask),
115        &[],
116        None,
117        None,
118    );
119}
120
121/// 3-D average-pooling backward pass.
122///
123/// Grid: `pid = (((b * C + c) * OD + od) * OH + oh) * num_ow_tiles + ow_tile`
124///
125/// `dx` must be zero-initialised before launch.
126#[kernel]
127pub fn avgpool3d_backward<
128    T: Triton,
129    D: Num,
130    const KD: i32,
131    const KH: i32,
132    const KW: i32,
133    const STRIDE_D: i32,
134    const STRIDE_H: i32,
135    const STRIDE_W: i32,
136    const BLOCK_OW: i32,
137>(
138    dy_ptr: T::Pointer<D>,
139    dx_ptr: T::Pointer<D>,
140    _B: i32,
141    C: i32,
142    Dv: i32,
143    H: i32,
144    W: i32,
145    OD: i32,
146    OH: i32,
147    OW: i32,
148) where
149    T::I32Tensor: Tensor<i32, 1>,
150    T::I32Tensor: Comparison<i32, BoolTensor = T::BoolTensor>,
151    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
152{
153    let pid = T::program_id(Axis::X);
154    let num_ow_tiles = T::cdiv(OW, BLOCK_OW);
155
156    let ow_tile = pid % num_ow_tiles;
157    let rest = pid / num_ow_tiles;
158    let oh = rest % OH;
159    let rest2 = rest / OH;
160    let od = rest2 % OD;
161    let bco = rest2 / OD;
162    let c = bco % C;
163    let b = bco / C;
164
165    let ow_start = ow_tile * BLOCK_OW;
166    let ow_range = T::arange(0, BLOCK_OW) + ow_start;
167    let ow_mask = ow_range.lt(OW);
168
169    let dy_base = ((b * C + c) * OD * OH * OW) + od * OH * OW + oh * OW;
170    let dx_bc_base = (b * C + c) * Dv * H * W;
171
172    let dy_offsets = ow_range + dy_base;
173    let dy_tile = T::load(
174        dy_ptr.add_offsets(dy_offsets),
175        Some(ow_mask),
176        Some(T::zeros::<D>(&[BLOCK_OW])),
177        &[],
178        None,
179        None,
180        None,
181        false,
182    );
183    let ksize_1 = T::full::<i32>(&[1], KD * KH * KW);
184    let ksize_f_1 = T::cast::<i32, D>(ksize_1, None, false);
185    let ksize = T::broadcast_to(ksize_f_1, &[BLOCK_OW]);
186    let grad = dy_tile / ksize;
187
188    let loop_bound = KD * KH * KW;
189    for idx in 0..loop_bound {
190        let kw = idx % KW;
191        let tmp = idx / KW;
192        let kh = tmp % KH;
193        let kd = tmp / KH;
194
195        let id = od * STRIDE_D + kd;
196        let ih = oh * STRIDE_H + kh;
197        let iw_range = ow_range * STRIDE_W + kw;
198        let dx_offsets = iw_range + (dx_bc_base + id * H * W + ih * W);
199        T::atomic_add(
200            dx_ptr.add_offsets(dx_offsets),
201            grad,
202            Some(ow_mask),
203            None,
204            None,
205        );
206    }
207}
208
209pub struct Avgpool3dOp<'a, T: Num> {
210    pub forward: Avgpool3dForward<T>,
211    pub backward: Avgpool3dBackward<T>,
212    _marker: core::marker::PhantomData<&'a ()>,
213}
teeny_kernels/nn/pool/avgpool3d.rs

teeny_kernels/nn/pool/
avgpool3d.rs
