tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

mlp.c (4181B)


      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 "mlp.h"
     37 
     38 #define fmadd(a, b, c) ((a)*(b)+(c))
     39 static OPUS_INLINE float tansig_approx(float x)
     40 {
     41    const float N0 = 952.52801514f;
     42    const float N1 = 96.39235687f;
     43    const float N2 = 0.60863042f;
     44    const float D0 = 952.72399902f;
     45    const float D1 = 413.36801147f;
     46    const float D2 = 11.88600922f;
     47    float X2, num, den;
     48    X2 = x*x;
     49    num = fmadd(fmadd(N2, X2, N1), X2, N0);
     50    den = fmadd(fmadd(D2, X2, D1), X2, D0);
     51    num = num*x/den;
     52    return MAX32(-1.f, MIN32(1.f, num));
     53 }
     54 
     55 static OPUS_INLINE float sigmoid_approx(float x)
     56 {
     57   return .5f + .5f*tansig_approx(.5f*x);
     58 }
     59 
     60 static void gemm_accum(float *out, const opus_int8 *weights, int rows, int cols, int col_stride, const float *x)
     61 {
     62   int i, j;
     63   for (i=0;i<rows;i++)
     64   {
     65      for (j=0;j<cols;j++)
     66         out[i] += weights[j*col_stride + i]*x[j];
     67   }
     68 }
     69 
     70 void analysis_compute_dense(const AnalysisDenseLayer *layer, float *output, const float *input)
     71 {
     72   int i;
     73   int N, M;
     74   int stride;
     75   M = layer->nb_inputs;
     76   N = layer->nb_neurons;
     77   stride = N;
     78   for (i=0;i<N;i++)
     79      output[i] = layer->bias[i];
     80   gemm_accum(output, layer->input_weights, N, M, stride, input);
     81   for (i=0;i<N;i++)
     82      output[i] *= WEIGHTS_SCALE;
     83   if (layer->sigmoid) {
     84      for (i=0;i<N;i++)
     85         output[i] = sigmoid_approx(output[i]);
     86   } else {
     87      for (i=0;i<N;i++)
     88         output[i] = tansig_approx(output[i]);
     89   }
     90 }
     91 
     92 void analysis_compute_gru(const AnalysisGRULayer *gru, float *state, const float *input)
     93 {
     94   int i;
     95   int N, M;
     96   int stride;
     97   float tmp[MAX_NEURONS];
     98   float z[MAX_NEURONS];
     99   float r[MAX_NEURONS];
    100   float h[MAX_NEURONS];
    101   M = gru->nb_inputs;
    102   N = gru->nb_neurons;
    103   stride = 3*N;
    104   /* Compute update gate. */
    105   for (i=0;i<N;i++)
    106      z[i] = gru->bias[i];
    107   gemm_accum(z, gru->input_weights, N, M, stride, input);
    108   gemm_accum(z, gru->recurrent_weights, N, N, stride, state);
    109   for (i=0;i<N;i++)
    110      z[i] = sigmoid_approx(WEIGHTS_SCALE*z[i]);
    111 
    112   /* Compute reset gate. */
    113   for (i=0;i<N;i++)
    114      r[i] = gru->bias[N + i];
    115   gemm_accum(r, &gru->input_weights[N], N, M, stride, input);
    116   gemm_accum(r, &gru->recurrent_weights[N], N, N, stride, state);
    117   for (i=0;i<N;i++)
    118      r[i] = sigmoid_approx(WEIGHTS_SCALE*r[i]);
    119 
    120   /* Compute output. */
    121   for (i=0;i<N;i++)
    122      h[i] = gru->bias[2*N + i];
    123   for (i=0;i<N;i++)
    124      tmp[i] = state[i] * r[i];
    125   gemm_accum(h, &gru->input_weights[2*N], N, M, stride, input);
    126   gemm_accum(h, &gru->recurrent_weights[2*N], N, N, stride, tmp);
    127   for (i=0;i<N;i++)
    128      h[i] = z[i]*state[i] + (1-z[i])*tansig_approx(WEIGHTS_SCALE*h[i]);
    129   for (i=0;i<N;i++)
    130      state[i] = h[i];
    131 }