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
// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
// SPDX-License-Identifier: MIT OR Apache-2.0
//! Naive NCDHW 3-D conv / max-pool backward (CPU reference + host fallback).
#[allow(clippy::too_many_arguments)]
pub fn conv3d_backward_input_ncdhw(
dy: &[f32],
weight: &[f32],
dx: &mut [f32],
n: usize,
c_in: usize,
c_out: usize,
d: usize,
h: usize,
w: usize,
d_out: usize,
h_out: usize,
w_out: usize,
kd: usize,
kh: usize,
kw: usize,
sd: usize,
sh: usize,
sw: usize,
pd: usize,
ph: usize,
pw: usize,
dd: usize,
dh: usize,
dw: usize,
groups: usize,
) {
let c_in_per_g = c_in / groups.max(1);
let c_out_per_g = c_out / groups.max(1);
dx.fill(0.0);
for ni in 0..n {
for g in 0..groups {
for co_off in 0..c_out_per_g {
let co = g * c_out_per_g + co_off;
for do_ in 0..d_out {
for ho in 0..h_out {
for wo in 0..w_out {
let dyv =
dy[(((ni * c_out + co) * d_out + do_) * h_out + ho) * w_out + wo];
for ci_off in 0..c_in_per_g {
let ci = g * c_in_per_g + ci_off;
for kz in 0..kd {
let id = do_ as isize * sd as isize + kz as isize * dd as isize
- pd as isize;
if id < 0 || id >= d as isize {
continue;
}
for ki in 0..kh {
let ih = ho as isize * sh as isize
+ ki as isize * dh as isize
- ph as isize;
if ih < 0 || ih >= h as isize {
continue;
}
for kj in 0..kw {
let iw = wo as isize * sw as isize
+ kj as isize * dw as isize
- pw as isize;
if iw < 0 || iw >= w as isize {
continue;
}
let wv = weight[(((co * c_in_per_g + ci_off) * kd
+ kz)
* kh
+ ki)
* kw
+ kj];
let dx_i = (((ni * c_in + ci) * d + id as usize) * h
+ ih as usize)
* w
+ iw as usize;
dx[dx_i] += dyv * wv;
}
}
}
}
}
}
}
}
}
}
}
#[allow(clippy::too_many_arguments)]
pub fn conv3d_backward_weight_ncdhw(
x: &[f32],
dy: &[f32],
dw: &mut [f32],
n: usize,
c_in: usize,
c_out: usize,
d: usize,
h: usize,
w: usize,
d_out: usize,
h_out: usize,
w_out: usize,
kd: usize,
kh: usize,
kw: usize,
sd: usize,
sh: usize,
sw: usize,
pd: usize,
ph: usize,
pw: usize,
dd: usize,
dh: usize,
dw_dil: usize,
groups: usize,
) {
let c_in_per_g = c_in / groups.max(1);
let c_out_per_g = c_out / groups.max(1);
dw.fill(0.0);
for ni in 0..n {
for g in 0..groups {
for co_off in 0..c_out_per_g {
let co = g * c_out_per_g + co_off;
for do_ in 0..d_out {
for ho in 0..h_out {
for wo in 0..w_out {
let dyv =
dy[(((ni * c_out + co) * d_out + do_) * h_out + ho) * w_out + wo];
for ci_off in 0..c_in_per_g {
let ci = g * c_in_per_g + ci_off;
for kz in 0..kd {
let id = do_ as isize * sd as isize + kz as isize * dd as isize
- pd as isize;
if id < 0 || id >= d as isize {
continue;
}
for ki in 0..kh {
let ih = ho as isize * sh as isize
+ ki as isize * dh as isize
- ph as isize;
if ih < 0 || ih >= h as isize {
continue;
}
for kj in 0..kw {
let iw = wo as isize * sw as isize
+ kj as isize * dw_dil as isize
- pw as isize;
if iw < 0 || iw >= w as isize {
continue;
}
let xv = x[(((ni * c_in + ci) * d + id as usize) * h
+ ih as usize)
* w
+ iw as usize];
let dw_i =
(((co * c_in_per_g + ci_off) * kd + kz) * kh + ki)
* kw
+ kj;
dw[dw_i] += dyv * xv;
}
}
}
}
}
}
}
}
}
}
}
#[allow(clippy::too_many_arguments)]
pub fn maxpool3d_backward_ncdhw(
x: &[f32],
dy: &[f32],
dx: &mut [f32],
n: usize,
c: usize,
d: usize,
h: usize,
w: usize,
d_out: usize,
h_out: usize,
w_out: usize,
kd: usize,
kh: usize,
kw: usize,
sd: usize,
sh: usize,
sw: usize,
pd: usize,
ph: usize,
pw: usize,
) {
dx.fill(0.0);
for ni in 0..n {
for cc in 0..c {
let base = (ni * c + cc) * d * h * w;
for do_ in 0..d_out {
for ho in 0..h_out {
for wo in 0..w_out {
let dstart = do_ as isize * sd as isize - pd as isize;
let hstart = ho as isize * sh as isize - ph as isize;
let wstart = wo as isize * sw as isize - pw as isize;
let mut best = f32::NEG_INFINITY;
let mut best_idx = 0usize;
let mut found = false;
for kz in 0..kd {
let id = dstart + kz as isize;
if id < 0 || id >= d as isize {
continue;
}
for ki in 0..kh {
let ih = hstart + ki as isize;
if ih < 0 || ih >= h as isize {
continue;
}
for kj in 0..kw {
let iw = wstart + kj as isize;
if iw < 0 || iw >= w as isize {
continue;
}
let idx =
base + ((id as usize * h + ih as usize) * w + iw as usize);
let v = x[idx];
if !found || v > best {
best = v;
best_idx = idx;
found = true;
}
}
}
}
if found {
let dyv = dy[(((ni * c + cc) * d_out + do_) * h_out + ho) * w_out + wo];
dx[best_idx] += dyv;
}
}
}
}
}
}
}