Sync to upstream libopus

Sync to commit bb4b6885a139644cf3ac14e7deda9f633ec2d93c

This brings in a bunch of optimizations to decode speed
and memory usage. Allocations are switched from using
the pseudostack to using the real stack. Enabled hacks
to reduce stack usage.

This should fix crashes on sansa clip, although some
files will not play due to failing allocations in the
codec buffer.

Speeds up decoding of the following test files:

                 H300 (cf)   C200 (arm7tdmi)  ipod classic (arm9e)
16 kbps (silk)   14.28 MHz   4.00 MHz         2.61 MHz
64 kbps (celt)   4.09 MHz    8.08 MHz         6.24 MHz
128 kbps (celt)  1.93 MHz    8.83 MHz         6.53 MHz

Change-Id: I851733a8a5824b61feb363a173091bc7e6629b58

1 files changed, 73 insertions, 6 deletions

diff --git a/lib/rbcodec/codecs/libopus/celt/celt.c b/lib/rbcodec/codecs/libopus/celt/celt.c
index 3e0ce6e6a5..c0a1e0dab9 100644
--- a/lib/rbcodec/codecs/libopus/celt/celt.c
+++ b/lib/rbcodec/codecs/libopus/celt/celt.c
@@ -54,6 +54,10 @@
 #define PACKAGE_VERSION "unknown"
 #endif
 
+#if defined(MIPSr1_ASM)
+#include "mips/celt_mipsr1.h"
+#endif
+
 
 int resampling_factor(opus_int32 rate)
 {
@@ -86,6 +90,63 @@ int resampling_factor(opus_int32 rate)
 }
 
 #ifndef OVERRIDE_COMB_FILTER_CONST
+/* This version should be faster on ARM */
+#ifdef OPUS_ARM_ASM
+static void comb_filter_const(opus_val32 *y, opus_val32 *x, int T, int N,
+      opus_val16 g10, opus_val16 g11, opus_val16 g12)
+{
+   opus_val32 x0, x1, x2, x3, x4;
+   int i;
+   x4 = SHL32(x[-T-2], 1);
+   x3 = SHL32(x[-T-1], 1);
+   x2 = SHL32(x[-T], 1);
+   x1 = SHL32(x[-T+1], 1);
+   for (i=0;i<N-4;i+=5)
+   {
+      opus_val32 t;
+      x0=SHL32(x[i-T+2],1);
+      t = MAC16_32_Q16(x[i], g10, x2);
+      t = MAC16_32_Q16(t, g11, ADD32(x1,x3));
+      t = MAC16_32_Q16(t, g12, ADD32(x0,x4));
+      y[i] = t;
+      x4=SHL32(x[i-T+3],1);
+      t = MAC16_32_Q16(x[i+1], g10, x1);
+      t = MAC16_32_Q16(t, g11, ADD32(x0,x2));
+      t = MAC16_32_Q16(t, g12, ADD32(x4,x3));
+      y[i+1] = t;
+      x3=SHL32(x[i-T+4],1);
+      t = MAC16_32_Q16(x[i+2], g10, x0);
+      t = MAC16_32_Q16(t, g11, ADD32(x4,x1));
+      t = MAC16_32_Q16(t, g12, ADD32(x3,x2));
+      y[i+2] = t;
+      x2=SHL32(x[i-T+5],1);
+      t = MAC16_32_Q16(x[i+3], g10, x4);
+      t = MAC16_32_Q16(t, g11, ADD32(x3,x0));
+      t = MAC16_32_Q16(t, g12, ADD32(x2,x1));
+      y[i+3] = t;
+      x1=SHL32(x[i-T+6],1);
+      t = MAC16_32_Q16(x[i+4], g10, x3);
+      t = MAC16_32_Q16(t, g11, ADD32(x2,x4));
+      t = MAC16_32_Q16(t, g12, ADD32(x1,x0));
+      y[i+4] = t;
+   }
+#ifdef CUSTOM_MODES
+   for (;i<N;i++)
+   {
+      opus_val32 t;
+      x0=SHL32(x[i-T+2],1);
+      t = MAC16_32_Q16(x[i], g10, x2);
+      t = MAC16_32_Q16(t, g11, ADD32(x1,x3));
+      t = MAC16_32_Q16(t, g12, ADD32(x0,x4));
+      y[i] = t;
+      x4=x3;
+      x3=x2;
+      x2=x1;
+      x1=x0;
+   }
+#endif
+}
+#else
 static void comb_filter_const(opus_val32 *y, opus_val32 *x, int T, int N,
       opus_val16 g10, opus_val16 g11, opus_val16 g12)
 {
@@ -110,7 +171,9 @@ static void comb_filter_const(opus_val32 *y, opus_val32 *x, int T, int N,
 
 }
 #endif
+#endif
 
+#ifndef OVERRIDE_comb_filter
 void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
       opus_val16 g0, opus_val16 g1, int tapset0, int tapset1,
       const opus_val16 *window, int overlap)
@@ -131,16 +194,19 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
          OPUS_MOVE(y, x, N);
       return;
    }
-   g00 = MULT16_16_Q15(g0, gains[tapset0][0]);
-   g01 = MULT16_16_Q15(g0, gains[tapset0][1]);
-   g02 = MULT16_16_Q15(g0, gains[tapset0][2]);
-   g10 = MULT16_16_Q15(g1, gains[tapset1][0]);
-   g11 = MULT16_16_Q15(g1, gains[tapset1][1]);
-   g12 = MULT16_16_Q15(g1, gains[tapset1][2]);
+   g00 = MULT16_16_P15(g0, gains[tapset0][0]);
+   g01 = MULT16_16_P15(g0, gains[tapset0][1]);
+   g02 = MULT16_16_P15(g0, gains[tapset0][2]);
+   g10 = MULT16_16_P15(g1, gains[tapset1][0]);
+   g11 = MULT16_16_P15(g1, gains[tapset1][1]);
+   g12 = MULT16_16_P15(g1, gains[tapset1][2]);
    x1 = x[-T1+1];
    x2 = x[-T1  ];
    x3 = x[-T1-1];
    x4 = x[-T1-2];
+   /* If the filter didn't change, we don't need the overlap */
+   if (g0==g1 && T0==T1 && tapset0==tapset1)
+      overlap=0;
    for (i=0;i<overlap;i++)
    {
       opus_val16 f;
@@ -170,6 +236,7 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
    /* Compute the part with the constant filter. */
    comb_filter_const(y+i, x+i, T1, N-i, g10, g11, g12);
 }
+#endif /* OVERRIDE_comb_filter */
 
 const signed char tf_select_table[4][8] = {
       {0, -1, 0, -1,    0,-1, 0,-1},