-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconvlayer.hpp
371 lines (324 loc) · 15.9 KB
/
convlayer.hpp
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
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
#ifndef __STANN_HLS_CONV_HPP__
#define __STANN_HLS_CONV_HPP__
#include "stann.hpp"
/**
* Namespace for convolutions.
*/
namespace ConvLayer {
/**
* Namespace for direct convolutions.
*/
namespace Direct {
/**
* Anonymous namespace for internal functions.
*/
namespace {
/**
* Internal base function for 2D convolutional layer inference.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam OUTPUT_CHANNEL number of kernels
* @tparam T data type the function is working on
*
* @param[in] input input data
* @param[in] filter filter weights of the conv layer
* @param[out] output output data
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int OUTPUT_CHANNEL, typename T>
void forward_convolution_base(T *input,T *filter, T *output){
const int OUTPUT_HI = INPUT_HI - (FILTER_HI-1);
const int OUTPUT_WI = INPUT_WI - (FILTER_WI-1);
for(int ib = 0; ib < INPUT_BATCH_SIZE; ib ++){
for(int kd = 0; kd < OUTPUT_CHANNEL; kd ++){
for(int y = 0; y <= INPUT_HI - FILTER_HI; y += 1){
for(int x = 0; x <= INPUT_WI - FILTER_WI; x += 1){
int output_idx = x + y * OUTPUT_HI + (kd + ib * OUTPUT_CHANNEL) * OUTPUT_WI * OUTPUT_HI;
output[output_idx] = 0;
for(int c = 0 ; c < INPUT_CHANNEL; c ++){
for(int ky = 0; ky < FILTER_HI; ky ++){
for(int kx = 0; kx < FILTER_WI; kx ++){
#pragma HLS PIPELINE II=3
output[output_idx] +=
input[(ib * INPUT_CHANNEL + c) * INPUT_WI * INPUT_HI + (y + ky) * INPUT_WI + (x + kx)] *
filter[(kd * INPUT_CHANNEL + c) * FILTER_WI * FILTER_HI + ky * FILTER_WI + kx];
}
}
}
}
}
}
}
}
/**
* Internal base function for 2D convolutional layer backpropagation.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam OUTPUT_CHANNEL number of kernels
* @tparam T data type the function is working on
*
* @param[in] input input data
* @param[in] filter filter weights of the conv layer
* @param[out] output output data
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int OUTPUT_CHANNEL, typename T>
void backward_convolution_base(T *input,T *filter, T *output){
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
for (int ibs = 0; ibs < INPUT_BATCH_SIZE; ibs++){
for(int fd = 0; fd < OUTPUT_CHANNEL; fd++){
for(int oh = 0; oh < OUTPUT_HI ; oh++){
for(int ow = 0; ow < OUTPUT_WI ; ow++){
for(int ic = 0; ic < INPUT_CHANNEL; ic++){
for(int h = 0; h < FILTER_HI; h++){
for(int w = 0; w < FILTER_WI; w++){
#pragma HLS PIPELINE II=3
int INPUT_BATCH_SIZEInputOverhead = ibs * INPUT_WI * INPUT_HI * INPUT_CHANNEL ;
int INPUT_BATCH_SIZEOutputOverhead = ibs * OUTPUT_WI * OUTPUT_HI * OUTPUT_CHANNEL;
int currentOutputElement = oh * OUTPUT_WI + ow + fd * OUTPUT_HI * OUTPUT_WI + INPUT_BATCH_SIZEOutputOverhead;
int currentInputElement = ic * INPUT_HI * INPUT_WI + ow + w + (oh + h) * INPUT_WI + INPUT_BATCH_SIZEInputOverhead;
int currentFilterElement = (fd * INPUT_CHANNEL + ic) * FILTER_HI * FILTER_WI + w + h * FILTER_WI;
if (oh == 0 && ow == 0 && ic == 0 && h == 0 && w == 0) {
input[currentInputElement] = 0;
}
input[currentInputElement] += output[currentOutputElement] * filter[currentFilterElement];
}
}
}
}
}
}
}
}
/**
* Internal base function for 2D convolutional layer weight update.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam OUTPUT_CHANNEL number of kernels
* @tparam T data type the function is working on
*
* @param[in] input input data
* @param[in,out] filter filter weights of the conv layer
* @param[in] output output data of this layer
* @param[in] learning_rate learning rate for the update
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int OUTPUT_CHANNEL, typename T>
void update_convolution_base(T *input,T *filter, T *output, T learning_rate){
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
for (int ibs = 0; ibs < INPUT_BATCH_SIZE; ibs++){
for(int fd = 0; fd < OUTPUT_CHANNEL; fd++){
for(int oh = 0; oh < OUTPUT_HI ; oh++){
for(int ow = 0; ow < OUTPUT_WI ; ow++){
for(int ic = 0; ic < INPUT_CHANNEL; ic++){
for(int h = 0; h < FILTER_HI; h++){
for(int w = 0; w < FILTER_WI; w++){
#pragma HLS PIPELINE II=3
int INPUT_BATCH_SIZEInputOverhead = ibs * INPUT_WI * INPUT_HI * INPUT_CHANNEL ;
int INPUT_BATCH_SIZEOutputOverhead = ibs * OUTPUT_WI * OUTPUT_HI * OUTPUT_CHANNEL;
int currentOutputElement = oh * OUTPUT_WI + ow + fd * OUTPUT_HI * OUTPUT_WI + INPUT_BATCH_SIZEOutputOverhead;
int currentInputElement = ic * INPUT_HI * INPUT_WI + ow + w + (oh + h) * INPUT_WI + INPUT_BATCH_SIZEInputOverhead;
int currentFilterElement = (fd * INPUT_CHANNEL + ic) * FILTER_HI * FILTER_WI + w + h * FILTER_WI;
// TODO: divide by batchsize?
filter[currentFilterElement] -= learning_rate*(input[currentInputElement] * output[currentOutputElement]);
}
}
}
}
}
}
}
}
/**
* Internal base function for 2D convolutional layer bias update.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam OUTPUT_CHANNEL number of kernels
* @tparam T data type the function is working on
*
* @param[in,out] biases input data
* @param[in] this_output output of this layer
* @param[in] learning_rate learning rate for the update
*
* @note TODO not working yet
*/
template<int INPUT_WI,int INPUT_HI, int INPUT_BATCH_SIZE,int FILTER_WI, int FILTER_HI, int OUTPUT_CHANNEL, typename T>
void update_biases( T *biases, T *this_output, T learning_rate) {
int t = 0;
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
for (int c = 0; c < INPUT_BATCH_SIZE; c++){
int INPUT_BATCH_SIZEOutputOverhead = c * OUTPUT_HI * OUTPUT_WI * OUTPUT_CHANNEL;
for(int b = 0; b < OUTPUT_CHANNEL; b++){
int outputbatchoversize = b * OUTPUT_WI * OUTPUT_HI;
for(int i = 0; i < OUTPUT_HI; i++){
for(int j = 0; j < OUTPUT_WI; j++){
#pragma HLS PIPELINE II=5
biases[c * OUTPUT_CHANNEL + b] += this_output[i * OUTPUT_WI + j + outputbatchoversize + INPUT_BATCH_SIZEOutputOverhead];
}
}
}
}
}
} // anonymous namespace
namespace Float {
/**
* Inference of direct convolution layer.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam FILTER_DEPTH number of kernels
*
* @param[in] input input data
* @param[in] weights filter weights
* @param[in] bias biases of the layer
* @param[out] output convolution output
* @param[in] act activation constant
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int FILTER_DEPTH>
void forward(float *input, float *weights, float *bias, float *output, activation_t act){
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
forward_convolution_base<INPUT_WI, INPUT_HI, INPUT_CHANNEL, INPUT_BATCH_SIZE, FILTER_WI, FILTER_HI, FILTER_DEPTH,float>(input,weights,output);
for (int ibs = 0; ibs < INPUT_BATCH_SIZE; ibs++){
for(int fd = 0; fd < FILTER_DEPTH; fd++){
for (int ow = 0; ow < OUTPUT_WI * OUTPUT_HI; ow++){
#pragma HLS PIPELINE II=2
output[ow + fd * OUTPUT_WI *OUTPUT_HI ] += bias[fd];//bias[fd + FILTER_DEPTH * ibs];
}
}
}
if (act == LEAKY_RELU){
ActivationLayer::Float::leaky_relu_inplace<OUTPUT_WI*OUTPUT_HI*FILTER_DEPTH, INPUT_BATCH_SIZE>(output);
}
}
template<int INPUT_HEIGHT, int INPUT_WIDTH, int INPUT_CHANNELS, int OUTPUT_CHANNELS, int KERNEL_SIZE, int INPUT_BATCH_SIZE, int PE1, int PE2, int PE3>
void forward_stream(hls::stream<float> &input_stream, float *weights, float *bias, hls::stream<float> &output_stream, activation_t act){
float input[INPUT_WIDTH*INPUT_HEIGHT*INPUT_CHANNELS*INPUT_BATCH_SIZE];
float output[(INPUT_WIDTH-4)*(INPUT_HEIGHT-4)*OUTPUT_CHANNELS*INPUT_BATCH_SIZE];
StreamUtil::toarray<INPUT_WIDTH*INPUT_HEIGHT*INPUT_CHANNELS*INPUT_BATCH_SIZE>(input_stream, input);
forward<INPUT_WIDTH, INPUT_HEIGHT, INPUT_CHANNELS, INPUT_BATCH_SIZE, KERNEL_SIZE, KERNEL_SIZE, OUTPUT_CHANNELS>(input, weights, bias, output, act);
StreamUtil::tostream<(INPUT_WIDTH-4)*(INPUT_HEIGHT-4)*OUTPUT_CHANNELS*INPUT_BATCH_SIZE>(output, output_stream);
}
/**
* Backpropagation of direct convolution layer.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam FILTER_DEPTH number of kernels
*
* @param[in] old_input input data
* @param[in] filter filter weights
* @param[in] bias biases of the layer
* @param[in] delta partial errors of next layer
* @param[out] newdelta partial errors of this layer
* @param[in] act activation constant
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int FILTER_DEPTH>
void backward(float *old_input, float *filter, float *bias,float *delta, float *newdelta, activation_t act = NONE) {
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
backward_convolution_base<INPUT_WI, INPUT_HI, INPUT_CHANNEL , INPUT_BATCH_SIZE, FILTER_WI, FILTER_HI, FILTER_DEPTH, float> (newdelta, filter, delta);
for (int j = 0; j < INPUT_BATCH_SIZE; j++) {
for (int i = 0; i < OUTPUT_WI*OUTPUT_HI; i++) {
#pragma HLS PIPELINE II=8
if (act == LEAKY_RELU) {
newdelta[j * OUTPUT_WI*OUTPUT_HI + i] = newdelta[j * OUTPUT_WI*OUTPUT_HI + i] * Activation::leaky_relu_simple_derivative(old_input[j * OUTPUT_WI*OUTPUT_HI + i]);
} else {
newdelta[j * OUTPUT_WI*OUTPUT_HI + i] = newdelta[j * OUTPUT_WI*OUTPUT_HI + i] * old_input[j * OUTPUT_WI*OUTPUT_HI + i];
}
}
}
}
/**
* Weight update of direct convolution layer.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam FILTER_DEPTH number of kernels
*
* @param[in] delta partial errors for this layer
* @param[in,out] filter filter weights
* @param[in,out] bias biases of the layer
* @param[in] prev_input input to this layer
* @param[in] learning_rate learning rate for the update
*
* @note TODO not working yet
*/
template<int INPUT_WI, int INPUT_HI, int INPUT_CHANNEL, int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int FILTER_DEPTH>
void update(float *delta, float *filter, float *bias, float *prev_input, float learning_rate){
update_convolution_base<INPUT_WI, INPUT_HI, INPUT_CHANNEL, INPUT_BATCH_SIZE, FILTER_WI, FILTER_HI, FILTER_DEPTH, float>(prev_input, filter, delta, learning_rate);
update_biases<INPUT_WI, INPUT_HI, INPUT_BATCH_SIZE, FILTER_WI, FILTER_HI, FILTER_DEPTH, float>(bias, delta, learning_rate);
}
} // namespace Float
namespace Half {
/**
* Inference of direct convolution layer.
*
* @tparam INPUT_WI width of input image
* @tparam INPUT_HI height of input image
* @tparam INPUT_CHANNEL number channels of input image
* @tparam INPUT_BATCH_SIZE number channels of input image
* @tparam FILTER_WI width of kernel
* @tparam FILTER_HI height of kernel
* @tparam FILTER_DEPTH number of kernels
*
* @param[in] input input data
* @param[in] weights filter weights
* @param[in] bias biases of the layer
* @param[out] output convolution output
* @param[in] act activation constant
*/
template<int INPUT_WI,int INPUT_HI,int INPUT_CHANNEL , int INPUT_BATCH_SIZE, int FILTER_WI, int FILTER_HI, int FILTER_DEPTH>
void forward(half *input, half *weights, half *bias, half *output, activation_t act){
const int OUTPUT_WI = INPUT_WI - FILTER_WI + 1;
const int OUTPUT_HI = INPUT_HI - FILTER_HI + 1;
forward_convolution_base<INPUT_WI, INPUT_HI, INPUT_CHANNEL, INPUT_BATCH_SIZE, FILTER_WI, FILTER_HI, FILTER_DEPTH, half>(input,weights,output);
for (int ibs = 0; ibs < INPUT_BATCH_SIZE; ibs++){
for(int fd = 0; fd < FILTER_DEPTH; fd++){
for (int ow = 0; ow < OUTPUT_WI * OUTPUT_HI; ow++){
#pragma HLS PIPELINE II=2
output[ow + fd * OUTPUT_WI *OUTPUT_HI ] += bias[fd + FILTER_DEPTH * ibs];
}
}
}
if (act == LEAKY_RELU){
ActivationLayer::Half::leaky_relu_inplace<OUTPUT_WI*OUTPUT_HI*FILTER_DEPTH, INPUT_BATCH_SIZE>(output);
}
}
} // namespace Half
namespace Fixed {
} // namespace Fixed
} // namespace Direct
} // namespace ConvLayer
#endif