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
|
/* Copyright (c) 2018 Mozilla
2008-2011 Octasic Inc.
2012-2017 Jean-Marc Valin */
/*
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <stdlib.h>
#include <math.h>
#include "opus_types.h"
#include "arch.h"
#include "nnet.h"
#include "dred_rdovae_constants.h"
#include "plc_data.h"
#include "fargan.h"
#include "os_support.h"
#include "vec.h"
#ifdef ENABLE_OSCE
#include "osce_config.h"
#endif
#ifdef NO_OPTIMIZATIONS
#if defined(_MSC_VER)
#pragma message ("Compiling without any vectorization. This code will be very slow")
#else
#warning Compiling without any vectorization. This code will be very slow
#endif
#endif
#define SOFTMAX_HACK
void compute_generic_dense(const LinearLayer *layer, float *output, const float *input, int activation, int arch)
{
compute_linear(layer, output, input, arch);
compute_activation(output, output, layer->nb_outputs, activation, arch);
}
#ifdef ENABLE_OSCE
#define MAX_RNN_NEURONS_ALL IMAX(IMAX(IMAX(FARGAN_MAX_RNN_NEURONS, PLC_MAX_RNN_NEURONS), DRED_MAX_RNN_NEURONS), OSCE_MAX_RNN_NEURONS)
#else
#define MAX_RNN_NEURONS_ALL IMAX(IMAX(FARGAN_MAX_RNN_NEURONS, PLC_MAX_RNN_NEURONS), DRED_MAX_RNN_NEURONS)
#endif
void compute_generic_gru(const LinearLayer *input_weights, const LinearLayer *recurrent_weights, float *state, const float *in, int arch)
{
int i;
int N;
float zrh[3*MAX_RNN_NEURONS_ALL];
float recur[3*MAX_RNN_NEURONS_ALL];
float *z;
float *r;
float *h;
celt_assert(3*recurrent_weights->nb_inputs == recurrent_weights->nb_outputs);
celt_assert(input_weights->nb_outputs == recurrent_weights->nb_outputs);
N = recurrent_weights->nb_inputs;
z = zrh;
r = &zrh[N];
h = &zrh[2*N];
celt_assert(recurrent_weights->nb_outputs <= 3*MAX_RNN_NEURONS_ALL);
celt_assert(in != state);
compute_linear(input_weights, zrh, in, arch);
compute_linear(recurrent_weights, recur, state, arch);
for (i=0;i<2*N;i++)
zrh[i] += recur[i];
compute_activation(zrh, zrh, 2*N, ACTIVATION_SIGMOID, arch);
for (i=0;i<N;i++)
h[i] += recur[2*N+i]*r[i];
compute_activation(h, h, N, ACTIVATION_TANH, arch);
for (i=0;i<N;i++)
h[i] = z[i]*state[i] + (1-z[i])*h[i];
for (i=0;i<N;i++)
state[i] = h[i];
}
void compute_glu(const LinearLayer *layer, float *output, const float *input, int arch)
{
int i;
float act2[MAX_INPUTS];
celt_assert(layer->nb_inputs == layer->nb_outputs);
compute_linear(layer, act2, input, arch);
compute_activation(act2, act2, layer->nb_outputs, ACTIVATION_SIGMOID, arch);
if (input == output) {
/* Give a vectorization hint to the compiler for the in-place case. */
for (i=0;i<layer->nb_outputs;i++) output[i] = output[i]*act2[i];
} else {
for (i=0;i<layer->nb_outputs;i++) output[i] = input[i]*act2[i];
}
}
void _lpcnet_compute_dense(const DenseLayer *layer, float *output, const float *input, int arch)
{
LinearLayer matrix;
celt_assert(input != output);
matrix.bias = layer->bias;
matrix.subias = NULL;
matrix.float_weights = layer->input_weights;
matrix.weights = NULL;
matrix.weights_idx = NULL;
matrix.diag = NULL;
matrix.nb_inputs = layer->nb_inputs;
matrix.nb_outputs = layer->nb_neurons;
matrix.scale = NULL;
compute_linear(&matrix, output, input, arch);
compute_activation(output, output, layer->nb_neurons, layer->activation, arch);
}
#ifdef USE_SU_BIAS
#define bias_type subias
#else
#define bias_type bias
#endif
#define MAX_IDX_SIZE 8192
void compute_gruB(const GRULayer *gru, const float* gru_b_condition, float *state, const float *input, int arch)
{
LinearLayer in_matrix, rec_matrix;
int i, M, N;
float bias[3*MAX_RNN_NEURONS_ALL];
float scale[3*MAX_RNN_NEURONS_ALL];
M = gru->nb_inputs;
N = gru->nb_neurons;
in_matrix.bias = bias;
in_matrix.diag = NULL;
in_matrix.nb_inputs = M;
in_matrix.nb_outputs = 3*N;
in_matrix.subias = bias;
#ifdef DISABLE_DOT_PROD
for (i=0;i<3*N;i++) bias[i] = gru->bias[i] + gru_b_condition[i];
in_matrix.scale = NULL;
in_matrix.float_weights = gru->input_weights;
in_matrix.weights = NULL;
#else
for (i=0;i<3*N;i++) bias[i] = gru->bias_type[i] + gru_b_condition[i];
for (i=0;i<3*N;i++) scale[i] = SCALE_1;
in_matrix.scale = scale;
in_matrix.weights = gru->input_weights;
in_matrix.float_weights = NULL;
#endif
in_matrix.weights_idx = gru->input_weights_idx;
rec_matrix.bias = &gru->bias[3*N];
rec_matrix.diag = NULL;
rec_matrix.nb_inputs = N;
rec_matrix.nb_outputs = 3*N;
rec_matrix.scale = scale;
rec_matrix.subias = &gru->subias[3*N];
#ifdef DISABLE_DOT_PROD
rec_matrix.scale = NULL;
rec_matrix.float_weights = gru->recurrent_weights;
rec_matrix.weights = NULL;
#else
rec_matrix.scale = scale;
rec_matrix.weights = gru->recurrent_weights;
rec_matrix.float_weights = NULL;
#endif
rec_matrix.weights_idx = NULL;
compute_generic_gru(&in_matrix, &rec_matrix, state, input, arch);
}
#define MAX_CONV_INPUTS_ALL DRED_MAX_CONV_INPUTS
void compute_generic_conv1d(const LinearLayer *layer, float *output, float *mem, const float *input, int input_size, int activation, int arch)
{
float tmp[MAX_CONV_INPUTS_ALL];
celt_assert(input != output);
celt_assert(layer->nb_inputs <= MAX_CONV_INPUTS_ALL);
OPUS_COPY(tmp, mem, layer->nb_inputs-input_size);
OPUS_COPY(&tmp[layer->nb_inputs-input_size], input, input_size);
compute_linear(layer, output, tmp, arch);
compute_activation(output, output, layer->nb_outputs, activation, arch);
OPUS_COPY(mem, &tmp[input_size], layer->nb_inputs-input_size);
}
void compute_generic_conv1d_dilation(const LinearLayer *layer, float *output, float *mem, const float *input, int input_size, int dilation, int activation, int arch)
{
float tmp[MAX_CONV_INPUTS_ALL];
int ksize = layer->nb_inputs/input_size;
int i;
celt_assert(input != output);
celt_assert(layer->nb_inputs <= MAX_CONV_INPUTS_ALL);
if (dilation==1) OPUS_COPY(tmp, mem, layer->nb_inputs-input_size);
else for (i=0;i<ksize-1;i++) OPUS_COPY(&tmp[i*input_size], &mem[i*input_size*dilation], input_size);
OPUS_COPY(&tmp[layer->nb_inputs-input_size], input, input_size);
compute_linear(layer, output, tmp, arch);
compute_activation(output, output, layer->nb_outputs, activation, arch);
if (dilation==1) OPUS_COPY(mem, &tmp[input_size], layer->nb_inputs-input_size);
else {
OPUS_COPY(mem, &mem[input_size], input_size*dilation*(ksize-1)-input_size);
OPUS_COPY(&mem[input_size*dilation*(ksize-1)-input_size], input, input_size);
}
}
|
