Revert "Update opus to 1.3.1 and opusfile to 0.11"

This reverts commit e00426c512a7905f5f925d382c443bab7a0ca693.

The way we handle platform-specific intrinsics is not good, so the
current state will not compile on armv8. This commit also requires
SSE4.1 support, which is likely not a good idea for portable binaries.

We'll have to redo this with more caution after 3.2 is released, or
we might simply drop opus as we're only using it as dependency for
theora right now.

Fixes #33606.

1 files changed, 78 insertions, 77 deletions

diff --git a/thirdparty/opus/mlp.c b/thirdparty/opus/mlp.c
index 964c6a98f6..ff9e50df47 100644
--- a/thirdparty/opus/mlp.c
+++ b/thirdparty/opus/mlp.c
@@ -1,5 +1,5 @@
 /* Copyright (c) 2008-2011 Octasic Inc.
-                 2012-2017 Jean-Marc Valin */
+   Written by Jean-Marc Valin */
 /*
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions
@@ -29,13 +29,42 @@
 #include "config.h"
 #endif
 
-#include <math.h>
 #include "opus_types.h"
 #include "opus_defines.h"
+
+#include <math.h>
+#include "mlp.h"
 #include "arch.h"
 #include "tansig_table.h"
-#include "mlp.h"
+#define MAX_NEURONS 100
 
+#if 0
+static OPUS_INLINE opus_val16 tansig_approx(opus_val32 _x) /* Q19 */
+{
+    int i;
+    opus_val16 xx; /* Q11 */
+    /*double x, y;*/
+    opus_val16 dy, yy; /* Q14 */
+    /*x = 1.9073e-06*_x;*/
+    if (_x>=QCONST32(8,19))
+        return QCONST32(1.,14);
+    if (_x<=-QCONST32(8,19))
+        return -QCONST32(1.,14);
+    xx = EXTRACT16(SHR32(_x, 8));
+    /*i = lrint(25*x);*/
+    i = SHR32(ADD32(1024,MULT16_16(25, xx)),11);
+    /*x -= .04*i;*/
+    xx -= EXTRACT16(SHR32(MULT16_16(20972,i),8));
+    /*x = xx*(1./2048);*/
+    /*y = tansig_table[250+i];*/
+    yy = tansig_table[250+i];
+    /*y = yy*(1./16384);*/
+    dy = 16384-MULT16_16_Q14(yy,yy);
+    yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx,dy),(16384 - MULT16_16_Q11(yy,xx)));
+    return yy;
+}
+#else
+/*extern const float tansig_table[501];*/
 static OPUS_INLINE float tansig_approx(float x)
 {
     int i;
@@ -63,82 +92,54 @@ static OPUS_INLINE float tansig_approx(float x)
     y = y + x*dy*(1 - y*x);
     return sign*y;
 }
+#endif
 
-static OPUS_INLINE float sigmoid_approx(float x)
-{
-   return .5f + .5f*tansig_approx(.5f*x);
-}
-
-static void gemm_accum(float *out, const opus_int8 *weights, int rows, int cols, int col_stride, const float *x)
-{
-   int i, j;
-   for (i=0;i<rows;i++)
-   {
-      for (j=0;j<cols;j++)
-         out[i] += weights[j*col_stride + i]*x[j];
-   }
-}
-
-void compute_dense(const DenseLayer *layer, float *output, const float *input)
+#if 0
+void mlp_process(const MLP *m, const opus_val16 *in, opus_val16 *out)
 {
-   int i;
-   int N, M;
-   int stride;
-   M = layer->nb_inputs;
-   N = layer->nb_neurons;
-   stride = N;
-   for (i=0;i<N;i++)
-      output[i] = layer->bias[i];
-   gemm_accum(output, layer->input_weights, N, M, stride, input);
-   for (i=0;i<N;i++)
-      output[i] *= WEIGHTS_SCALE;
-   if (layer->sigmoid) {
-      for (i=0;i<N;i++)
-         output[i] = sigmoid_approx(output[i]);
-   } else {
-      for (i=0;i<N;i++)
-         output[i] = tansig_approx(output[i]);
-   }
+    int j;
+    opus_val16 hidden[MAX_NEURONS];
+    const opus_val16 *W = m->weights;
+    /* Copy to tmp_in */
+    for (j=0;j<m->topo[1];j++)
+    {
+        int k;
+        opus_val32 sum = SHL32(EXTEND32(*W++),8);
+        for (k=0;k<m->topo[0];k++)
+            sum = MAC16_16(sum, in[k],*W++);
+        hidden[j] = tansig_approx(sum);
+    }
+    for (j=0;j<m->topo[2];j++)
+    {
+        int k;
+        opus_val32 sum = SHL32(EXTEND32(*W++),14);
+        for (k=0;k<m->topo[1];k++)
+            sum = MAC16_16(sum, hidden[k], *W++);
+        out[j] = tansig_approx(EXTRACT16(PSHR32(sum,17)));
+    }
 }
-
-void compute_gru(const GRULayer *gru, float *state, const float *input)
+#else
+void mlp_process(const MLP *m, const float *in, float *out)
 {
-   int i;
-   int N, M;
-   int stride;
-   float tmp[MAX_NEURONS];
-   float z[MAX_NEURONS];
-   float r[MAX_NEURONS];
-   float h[MAX_NEURONS];
-   M = gru->nb_inputs;
-   N = gru->nb_neurons;
-   stride = 3*N;
-   /* Compute update gate. */
-   for (i=0;i<N;i++)
-      z[i] = gru->bias[i];
-   gemm_accum(z, gru->input_weights, N, M, stride, input);
-   gemm_accum(z, gru->recurrent_weights, N, N, stride, state);
-   for (i=0;i<N;i++)
-      z[i] = sigmoid_approx(WEIGHTS_SCALE*z[i]);
-
-   /* Compute reset gate. */
-   for (i=0;i<N;i++)
-      r[i] = gru->bias[N + i];
-   gemm_accum(r, &gru->input_weights[N], N, M, stride, input);
-   gemm_accum(r, &gru->recurrent_weights[N], N, N, stride, state);
-   for (i=0;i<N;i++)
-      r[i] = sigmoid_approx(WEIGHTS_SCALE*r[i]);
-
-   /* Compute output. */
-   for (i=0;i<N;i++)
-      h[i] = gru->bias[2*N + i];
-   for (i=0;i<N;i++)
-      tmp[i] = state[i] * r[i];
-   gemm_accum(h, &gru->input_weights[2*N], N, M, stride, input);
-   gemm_accum(h, &gru->recurrent_weights[2*N], N, N, stride, tmp);
-   for (i=0;i<N;i++)
-      h[i] = z[i]*state[i] + (1-z[i])*tansig_approx(WEIGHTS_SCALE*h[i]);
-   for (i=0;i<N;i++)
-      state[i] = h[i];
+    int j;
+    float hidden[MAX_NEURONS];
+    const float *W = m->weights;
+    /* Copy to tmp_in */
+    for (j=0;j<m->topo[1];j++)
+    {
+        int k;
+        float sum = *W++;
+        for (k=0;k<m->topo[0];k++)
+            sum = sum + in[k]**W++;
+        hidden[j] = tansig_approx(sum);
+    }
+    for (j=0;j<m->topo[2];j++)
+    {
+        int k;
+        float sum = *W++;
+        for (k=0;k<m->topo[1];k++)
+            sum = sum + hidden[k]**W++;
+        out[j] = tansig_approx(sum);
+    }
 }
-
+#endif