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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
use crate::scratch::Scratch;
use crate::tensor::TensorView;
pub struct Conv5dBackwardArgs<'t> {
pub input: &'t TensorView<'t>,
pub kernel: &'t TensorView<'t>,
pub doutput: &'t TensorView<'t>,
pub dinput: &'t mut TensorView<'t>,
pub dkernel: &'t mut TensorView<'t>,
pub dbias: Option<&'t mut [f32]>,
pub pad: (usize, usize, usize),
pub stride: (usize, usize, usize),
pub scratch: &'t mut Scratch<'t>,
}
pub fn conv5d_forward(
input: &TensorView<'_>,
kernel: &TensorView<'_>,
bias: Option<&[f32]>,
output: &mut TensorView<'_>,
pad: (usize, usize, usize),
stride: (usize, usize, usize),
) {
if stride.0 == 0 || stride.1 == 0 || stride.2 == 0 {
return;
}
if !input.is_valid_layout() || !kernel.is_valid_layout() || !output.is_valid_layout() {
return;
}
let n = input.shape[0];
let c = input.shape[1];
let d = input.shape[2];
let h = input.shape[3];
let w = input.shape[4];
let cout = kernel.shape[0];
if kernel.shape[1] != c {
return;
}
if output.shape[0] != n || output.shape[1] != cout {
return;
}
if let Some(b) = bias {
if b.len() < cout {
return;
}
}
let pad_d = pad.0 as isize;
let pad_h = pad.1 as isize;
let pad_w = pad.2 as isize;
let d_isize = d as isize;
let h_isize = h as isize;
let w_isize = w as isize;
for n in 0..n {
for co in 0..cout {
for od in 0..output.shape[2] {
for oh in 0..output.shape[3] {
for ow in 0..output.shape[4] {
let mut acc = 0f32;
for ci in 0..c {
for kd in 0..kernel.shape[2] {
for kh in 0..kernel.shape[3] {
for kw in 0..kernel.shape[4] {
let id = (od * stride.0 + kd) as isize - pad_d;
let ih = (oh * stride.1 + kh) as isize - pad_h;
let iw = (ow * stride.2 + kw) as isize - pad_w;
if id >= 0
&& ih >= 0
&& iw >= 0
&& id < d_isize
&& ih < h_isize
&& iw < w_isize
{
let idu = id as usize;
let ihu = ih as usize;
let iwu = iw as usize;
let i_idx = match input.idx_linear(n, ci, idu, ihu, iwu)
{
Some(v) => v,
None => continue,
};
let k_idx = match kernel.idx_linear(co, ci, kd, kh, kw)
{
Some(v) => v,
None => continue,
};
acc += input.data[i_idx] * kernel.data[k_idx];
}
}
}
}
}
if let Some(out_i) = output.idx_linear(n, co, od, oh, ow) {
output.data[out_i] = acc + bias.map(|b| b[co]).unwrap_or(0.0);
}
}
}
}
}
}
}
pub fn conv5d_backward<'a>(args: Conv5dBackwardArgs<'a>) {
let Conv5dBackwardArgs {
input,
kernel,
doutput,
dinput,
dkernel,
mut dbias,
pad,
stride,
scratch,
} = args;
if stride.0 == 0 || stride.1 == 0 || stride.2 == 0 {
return;
}
if !input.is_valid_layout()
|| !kernel.is_valid_layout()
|| !doutput.is_valid_layout()
|| !dinput.is_valid_layout()
|| !dkernel.is_valid_layout()
{
return;
}
for v in dinput.data.iter_mut() {
*v = 0.0;
}
for v in dkernel.data.iter_mut() {
*v = 0.0;
}
if let Some(b) = dbias.as_mut() {
for x in b.iter_mut() {
*x = 0.0;
}
}
let n = input.shape[0];
let c = input.shape[1];
let d = input.shape[2];
let h = input.shape[3];
let w = input.shape[4];
let cout = kernel.shape[0];
if kernel.shape[1] != c {
return;
}
if doutput.shape[0] != n || doutput.shape[1] != cout {
return;
}
if dinput.shape != input.shape {
return;
}
if dkernel.shape != kernel.shape {
return;
}
if let Some(b) = dbias.as_ref() {
if b.len() < cout {
return;
}
}
let pad_d = pad.0 as isize;
let pad_h = pad.1 as isize;
let pad_w = pad.2 as isize;
let d_isize = d as isize;
let h_isize = h as isize;
let w_isize = w as isize;
if let Some(b) = dbias.as_mut() {
for co in 0..cout {
let mut sum = 0f32;
for n in 0..n {
for od in 0..doutput.shape[2] {
for oh in 0..doutput.shape[3] {
for ow in 0..doutput.shape[4] {
let idx = match doutput.idx_linear(n, co, od, oh, ow) {
Some(v) => v,
None => continue,
};
sum += doutput.data[idx];
}
}
}
}
b[co] = sum;
}
}
scratch.base_ptr();
for co in 0..cout {
for ci in 0..c {
for kd in 0..kernel.shape[2] {
for kh in 0..kernel.shape[3] {
for kw in 0..kernel.shape[4] {
let mut acc = 0f32;
for n in 0..n {
for od in 0..doutput.shape[2] {
for oh in 0..doutput.shape[3] {
for ow in 0..doutput.shape[4] {
let id = (od * stride.0 + kd) as isize - pad_d;
let ih = (oh * stride.1 + kh) as isize - pad_h;
let iw = (ow * stride.2 + kw) as isize - pad_w;
if id >= 0
&& ih >= 0
&& iw >= 0
&& id < d_isize
&& ih < h_isize
&& iw < w_isize
{
let idu = id as usize;
let ihu = ih as usize;
let iwu = iw as usize;
let in_idx =
match input.idx_linear(n, ci, idu, ihu, iwu) {
Some(v) => v,
None => continue,
};
let dout_idx =
match doutput.idx_linear(n, co, od, oh, ow) {
Some(v) => v,
None => continue,
};
acc += input.data[in_idx] * doutput.data[dout_idx];
}
}
}
}
}
if let Some(k_idx) = dkernel.idx_linear(co, ci, kd, kh, kw) {
dkernel.data[k_idx] = acc;
}
}
}
}
}
}
for n in 0..n {
for ci in 0..c {
for id in 0..d {
for ih in 0..h {
for iw in 0..w {
let mut acc = 0f32;
for co in 0..cout {
for kd in 0..kernel.shape[2] {
for kh in 0..kernel.shape[3] {
for kw in 0..kernel.shape[4] {
let od_num = (id + pad.0) as isize - kd as isize;
let oh_num = (ih + pad.1) as isize - kh as isize;
let ow_num = (iw + pad.2) as isize - kw as isize;
if od_num < 0 || oh_num < 0 || ow_num < 0 {
continue;
}
let od = od_num as usize;
let oh = oh_num as usize;
let ow = ow_num as usize;
if !od.is_multiple_of(stride.0)
|| !oh.is_multiple_of(stride.1)
|| !ow.is_multiple_of(stride.2)
{
continue;
}
let od = od / stride.0;
let oh = oh / stride.1;
let ow = ow / stride.2;
if od < doutput.shape[2]
&& oh < doutput.shape[3]
&& ow < doutput.shape[4]
{
let dout_idx =
match doutput.idx_linear(n, co, od, oh, ow) {
Some(v) => v,
None => continue,
};
let k_idx = match kernel.idx_linear(co, ci, kd, kh, kw)
{
Some(v) => v,
None => continue,
};
acc += doutput.data[dout_idx] * kernel.data[k_idx];
}
}
}
}
}
if let Some(i_idx) = dinput.idx_linear(n, ci, id, ih, iw) {
dinput.data[i_idx] = acc;
}
}
}
}
}
}
}