1 /* Copyright (c) 2008-2011 Octasic Inc.
2 2012-2017 Jean-Marc Valin */
3 /*
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions
6 are met:
7
8 - Redistributions of source code must retain the above copyright
9 notice, this list of conditions and the following disclaimer.
10
11 - Redistributions in binary form must reproduce the above copyright
12 notice, this list of conditions and the following disclaimer in the
13 documentation and/or other materials provided with the distribution.
14
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
18 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
19 CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
20 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
21 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
22 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
23 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
24 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 */
27
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31
32 #include <math.h>
33 #include "opus_types.h"
34 #include "opus_defines.h"
35 #include "arch.h"
36 #include "tansig_table.h"
37 #include "mlp.h"
38
tansig_approx(float x)39 static OPUS_INLINE float tansig_approx(float x)
40 {
41 int i;
42 float y, dy;
43 float sign=1;
44 /* Tests are reversed to catch NaNs */
45 if (!(x<8))
46 return 1;
47 if (!(x>-8))
48 return -1;
49 #ifndef FIXED_POINT
50 /* Another check in case of -ffast-math */
51 if (celt_isnan(x))
52 return 0;
53 #endif
54 if (x<0)
55 {
56 x=-x;
57 sign=-1;
58 }
59 i = (int)floor(.5f+25*x);
60 x -= .04f*i;
61 y = tansig_table[i];
62 dy = 1-y*y;
63 y = y + x*dy*(1 - y*x);
64 return sign*y;
65 }
66
sigmoid_approx(float x)67 static OPUS_INLINE float sigmoid_approx(float x)
68 {
69 return .5f + .5f*tansig_approx(.5f*x);
70 }
71
compute_dense(const DenseLayer * layer,float * output,const float * input)72 void compute_dense(const DenseLayer *layer, float *output, const float *input)
73 {
74 int i, j;
75 int N, M;
76 int stride;
77 M = layer->nb_inputs;
78 N = layer->nb_neurons;
79 stride = N;
80 for (i=0;i<N;i++)
81 {
82 /* Compute update gate. */
83 float sum = layer->bias[i];
84 for (j=0;j<M;j++)
85 sum += layer->input_weights[j*stride + i]*input[j];
86 output[i] = WEIGHTS_SCALE*sum;
87 }
88 if (layer->sigmoid) {
89 for (i=0;i<N;i++)
90 output[i] = sigmoid_approx(output[i]);
91 } else {
92 for (i=0;i<N;i++)
93 output[i] = tansig_approx(output[i]);
94 }
95 }
96
compute_gru(const GRULayer * gru,float * state,const float * input)97 void compute_gru(const GRULayer *gru, float *state, const float *input)
98 {
99 int i, j;
100 int N, M;
101 int stride;
102 float z[MAX_NEURONS];
103 float r[MAX_NEURONS];
104 float h[MAX_NEURONS];
105 M = gru->nb_inputs;
106 N = gru->nb_neurons;
107 stride = 3*N;
108 for (i=0;i<N;i++)
109 {
110 /* Compute update gate. */
111 float sum = gru->bias[i];
112 for (j=0;j<M;j++)
113 sum += gru->input_weights[j*stride + i]*input[j];
114 for (j=0;j<N;j++)
115 sum += gru->recurrent_weights[j*stride + i]*state[j];
116 z[i] = sigmoid_approx(WEIGHTS_SCALE*sum);
117 }
118 for (i=0;i<N;i++)
119 {
120 /* Compute reset gate. */
121 float sum = gru->bias[N + i];
122 for (j=0;j<M;j++)
123 sum += gru->input_weights[N + j*stride + i]*input[j];
124 for (j=0;j<N;j++)
125 sum += gru->recurrent_weights[N + j*stride + i]*state[j];
126 r[i] = sigmoid_approx(WEIGHTS_SCALE*sum);
127 }
128 for (i=0;i<N;i++)
129 {
130 /* Compute output. */
131 float sum = gru->bias[2*N + i];
132 for (j=0;j<M;j++)
133 sum += gru->input_weights[2*N + j*stride + i]*input[j];
134 for (j=0;j<N;j++)
135 sum += gru->recurrent_weights[2*N + j*stride + i]*state[j]*r[j];
136 h[i] = z[i]*state[i] + (1-z[i])*tansig_approx(WEIGHTS_SCALE*sum);
137 }
138 for (i=0;i<N;i++)
139 state[i] = h[i];
140 }
141
142