• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
gemm_accum(float * out,const opus_int8 * weights,int rows,int cols,int col_stride,const float * x)72 static void gemm_accum(float *out, const opus_int8 *weights, int rows, int cols, int col_stride, const float *x)
73 {
74    int i, j;
75    for (i=0;i<rows;i++)
76    {
77       for (j=0;j<cols;j++)
78          out[i] += weights[j*col_stride + i]*x[j];
79    }
80 }
81 
compute_dense(const DenseLayer * layer,float * output,const float * input)82 void compute_dense(const DenseLayer *layer, float *output, const float *input)
83 {
84    int i;
85    int N, M;
86    int stride;
87    M = layer->nb_inputs;
88    N = layer->nb_neurons;
89    stride = N;
90    for (i=0;i<N;i++)
91       output[i] = layer->bias[i];
92    gemm_accum(output, layer->input_weights, N, M, stride, input);
93    for (i=0;i<N;i++)
94       output[i] *= WEIGHTS_SCALE;
95    if (layer->sigmoid) {
96       for (i=0;i<N;i++)
97          output[i] = sigmoid_approx(output[i]);
98    } else {
99       for (i=0;i<N;i++)
100          output[i] = tansig_approx(output[i]);
101    }
102 }
103 
compute_gru(const GRULayer * gru,float * state,const float * input)104 void compute_gru(const GRULayer *gru, float *state, const float *input)
105 {
106    int i;
107    int N, M;
108    int stride;
109    float tmp[MAX_NEURONS];
110    float z[MAX_NEURONS];
111    float r[MAX_NEURONS];
112    float h[MAX_NEURONS];
113    M = gru->nb_inputs;
114    N = gru->nb_neurons;
115    stride = 3*N;
116    /* Compute update gate. */
117    for (i=0;i<N;i++)
118       z[i] = gru->bias[i];
119    gemm_accum(z, gru->input_weights, N, M, stride, input);
120    gemm_accum(z, gru->recurrent_weights, N, N, stride, state);
121    for (i=0;i<N;i++)
122       z[i] = sigmoid_approx(WEIGHTS_SCALE*z[i]);
123 
124    /* Compute reset gate. */
125    for (i=0;i<N;i++)
126       r[i] = gru->bias[N + i];
127    gemm_accum(r, &gru->input_weights[N], N, M, stride, input);
128    gemm_accum(r, &gru->recurrent_weights[N], N, N, stride, state);
129    for (i=0;i<N;i++)
130       r[i] = sigmoid_approx(WEIGHTS_SCALE*r[i]);
131 
132    /* Compute output. */
133    for (i=0;i<N;i++)
134       h[i] = gru->bias[2*N + i];
135    for (i=0;i<N;i++)
136       tmp[i] = state[i] * r[i];
137    gemm_accum(h, &gru->input_weights[2*N], N, M, stride, input);
138    gemm_accum(h, &gru->recurrent_weights[2*N], N, N, stride, tmp);
139    for (i=0;i<N;i++)
140       h[i] = z[i]*state[i] + (1-z[i])*tansig_approx(WEIGHTS_SCALE*h[i]);
141    for (i=0;i<N;i++)
142       state[i] = h[i];
143 }
144 
145