From 580b307fd791c0997a8831bc800bba87797bfb7e Mon Sep 17 00:00:00 2001
From: Nils Wallménius <nils@rockbox.org>
Date: Mon, 20 May 2013 22:25:57 +0200
Subject: Sync opus codec to upstream git

Sync opus codec to upstream commit
02fed471a4568852d6618e041c4f2af0d7730ee2 (August 30 2013)

This brings in a lot of optimizations but also makes the diff
between our codec and the upstream much smaller as most of our
optimizations have been upstreamed or supeceded.

Speedups across the board for CELT mode files:

        64kbps      128kbps
H300    9.82MHz     15.48MHz
c200	4.86MHz     9.63MHz
fuze v1 10.32MHz    15.92MHz

For the silk mode test file (16kbps) arm targets get a speedup
of about 2MHz while the H300 is 7.8MHz slower, likely because it's
now using the pseudostack more rather than the real stack which
is in iram. Patches to get around that are upcomming.

Change-Id: Ifecf963e461c51ac42e09dac1e91bc4bc3b12fa3
---
 lib/rbcodec/codecs/libopus/celt/celt_lpc.c | 196 +++++++++++++++++++++++------
 1 file changed, 158 insertions(+), 38 deletions(-)

(limited to 'lib/rbcodec/codecs/libopus/celt/celt_lpc.c')

diff --git a/lib/rbcodec/codecs/libopus/celt/celt_lpc.c b/lib/rbcodec/codecs/libopus/celt/celt_lpc.c
index 66aed1de09..7ffe90a357 100644
--- a/lib/rbcodec/codecs/libopus/celt/celt_lpc.c
+++ b/lib/rbcodec/codecs/libopus/celt/celt_lpc.c
@@ -26,12 +26,13 @@
 */
 
 #ifdef HAVE_CONFIG_H
-#include "opus_config.h"
+#include "config.h"
 #endif
 
 #include "celt_lpc.h"
 #include "stack_alloc.h"
 #include "mathops.h"
+#include "pitch.h"
 
 void _celt_lpc(
       opus_val16       *_lpc, /* out: [0...p-1] LPC coefficients      */
@@ -87,42 +88,71 @@ int          p
 #endif
 }
 
-void celt_fir(const opus_val16 *x,
+void celt_fir(const opus_val16 *_x,
          const opus_val16 *num,
-         opus_val16 *y,
+         opus_val16 *_y,
          int N,
          int ord,
          opus_val16 *mem)
 {
    int i,j;
+   VARDECL(opus_val16, rnum);
+   VARDECL(opus_val16, x);
+   SAVE_STACK;
 
+   ALLOC(rnum, ord, opus_val16);
+   ALLOC(x, N+ord, opus_val16);
+   for(i=0;i<ord;i++)
+      rnum[i] = num[ord-i-1];
+   for(i=0;i<ord;i++)
+      x[i] = mem[ord-i-1];
+   for (i=0;i<N;i++)
+      x[i+ord]=_x[i];
+   for(i=0;i<ord;i++)
+      mem[i] = _x[N-i-1];
+#ifdef SMALL_FOOTPRINT
    for (i=0;i<N;i++)
    {
-      opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
+      opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
       for (j=0;j<ord;j++)
       {
-         sum += MULT16_16(num[j],mem[j]);
-      }
-      for (j=ord-1;j>=1;j--)
-      {
-         mem[j]=mem[j-1];
+         sum = MAC16_16(sum,rnum[j],x[i+j]);
       }
-      mem[0] = x[i];
-      y[i] = ROUND16(sum, SIG_SHIFT);
+      _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
    }
+#else
+   for (i=0;i<N-3;i+=4)
+   {
+      opus_val32 sum[4]={0,0,0,0};
+      xcorr_kernel(rnum, x+i, sum, ord);
+      _y[i  ] = SATURATE16(ADD32(EXTEND32(_x[i  ]), PSHR32(sum[0], SIG_SHIFT)));
+      _y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT)));
+      _y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT)));
+      _y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT)));
+   }
+   for (;i<N;i++)
+   {
+      opus_val32 sum = 0;
+      for (j=0;j<ord;j++)
+         sum = MAC16_16(sum,rnum[j],x[i+j]);
+      _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
+   }
+#endif
+   RESTORE_STACK;
 }
 
-void celt_iir(const opus_val32 *x,
+void celt_iir(const opus_val32 *_x,
          const opus_val16 *den,
-         opus_val32 *y,
+         opus_val32 *_y,
          int N,
          int ord,
          opus_val16 *mem)
 {
+#ifdef SMALL_FOOTPRINT
    int i,j;
    for (i=0;i<N;i++)
    {
-      opus_val32 sum = x[i];
+      opus_val32 sum = _x[i];
       for (j=0;j<ord;j++)
       {
          sum -= MULT16_16(den[j],mem[j]);
@@ -132,11 +162,65 @@ void celt_iir(const opus_val32 *x,
          mem[j]=mem[j-1];
       }
       mem[0] = ROUND16(sum,SIG_SHIFT);
-      y[i] = sum;
+      _y[i] = sum;
    }
+#else
+   int i,j;
+   VARDECL(opus_val16, rden);
+   VARDECL(opus_val16, y);
+   SAVE_STACK;
+
+   celt_assert((ord&3)==0);
+   ALLOC(rden, ord, opus_val16);
+   ALLOC(y, N+ord, opus_val16);
+   for(i=0;i<ord;i++)
+      rden[i] = den[ord-i-1];
+   for(i=0;i<ord;i++)
+      y[i] = -mem[ord-i-1];
+   for(;i<N+ord;i++)
+      y[i]=0;
+   for (i=0;i<N-3;i+=4)
+   {
+      /* Unroll by 4 as if it were an FIR filter */
+      opus_val32 sum[4];
+      sum[0]=_x[i];
+      sum[1]=_x[i+1];
+      sum[2]=_x[i+2];
+      sum[3]=_x[i+3];
+      xcorr_kernel(rden, y+i, sum, ord);
+
+      /* Patch up the result to compensate for the fact that this is an IIR */
+      y[i+ord  ] = -ROUND16(sum[0],SIG_SHIFT);
+      _y[i  ] = sum[0];
+      sum[1] = MAC16_16(sum[1], y[i+ord  ], den[0]);
+      y[i+ord+1] = -ROUND16(sum[1],SIG_SHIFT);
+      _y[i+1] = sum[1];
+      sum[2] = MAC16_16(sum[2], y[i+ord+1], den[0]);
+      sum[2] = MAC16_16(sum[2], y[i+ord  ], den[1]);
+      y[i+ord+2] = -ROUND16(sum[2],SIG_SHIFT);
+      _y[i+2] = sum[2];
+
+      sum[3] = MAC16_16(sum[3], y[i+ord+2], den[0]);
+      sum[3] = MAC16_16(sum[3], y[i+ord+1], den[1]);
+      sum[3] = MAC16_16(sum[3], y[i+ord  ], den[2]);
+      y[i+ord+3] = -ROUND16(sum[3],SIG_SHIFT);
+      _y[i+3] = sum[3];
+   }
+   for (;i<N;i++)
+   {
+      opus_val32 sum = _x[i];
+      for (j=0;j<ord;j++)
+         sum -= MULT16_16(rden[j],y[i+j]);
+      y[i+ord] = ROUND16(sum,SIG_SHIFT);
+      _y[i] = sum;
+   }
+   for(i=0;i<ord;i++)
+      mem[i] = _y[N-i-1];
+   RESTORE_STACK;
+#endif
 }
 
-void _celt_autocorr(
+int _celt_autocorr(
                    const opus_val16 *x,   /*  in: [0...n-1] samples x   */
                    opus_val32       *ac,  /* out: [0...lag-1] ac values */
                    const opus_val16       *window,
@@ -146,43 +230,79 @@ void _celt_autocorr(
                   )
 {
    opus_val32 d;
-   int i;
+   int i, k;
+   int fastN=n-lag;
+   int shift;
+   const opus_val16 *xptr;
    VARDECL(opus_val16, xx);
    SAVE_STACK;
    ALLOC(xx, n, opus_val16);
    celt_assert(n>0);
    celt_assert(overlap>=0);
-   for (i=0;i<n;i++)
-      xx[i] = x[i];
-   for (i=0;i<overlap;i++)
+   if (overlap == 0)
    {
-      xx[i] = MULT16_16_Q15(x[i],window[i]);
-      xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]);
+      xptr = x;
+   } else {
+      for (i=0;i<n;i++)
+         xx[i] = x[i];
+      for (i=0;i<overlap;i++)
+      {
+         xx[i] = MULT16_16_Q15(x[i],window[i]);
+         xx[n-i-1] = MULT16_16_Q15(x[n-i-1],window[i]);
+      }
+      xptr = xx;
    }
+   shift=0;
 #ifdef FIXED_POINT
    {
-      opus_val32 ac0=0;
-      int shift;
-      for(i=0;i<n;i++)
-         ac0 += SHR32(MULT16_16(xx[i],xx[i]),9);
-      ac0 += 1+n;
+      opus_val32 ac0;
+      ac0 = 1+(n<<7);
+      if (n&1) ac0 += SHR32(MULT16_16(xptr[0],xptr[0]),9);
+      for(i=(n&1);i<n;i+=2)
+      {
+         ac0 += SHR32(MULT16_16(xptr[i],xptr[i]),9);
+         ac0 += SHR32(MULT16_16(xptr[i+1],xptr[i+1]),9);
+      }
 
       shift = celt_ilog2(ac0)-30+10;
-      shift = (shift+1)/2;
-      for(i=0;i<n;i++)
-         xx[i] = VSHR32(xx[i], shift);
+      shift = (shift)/2;
+      if (shift>0)
+      {
+         for(i=0;i<n;i++)
+            xx[i] = PSHR32(xptr[i], shift);
+         xptr = xx;
+      } else
+         shift = 0;
    }
 #endif
-   while (lag>=0)
+   celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1);
+   for (k=0;k<=lag;k++)
    {
-      for (i = lag, d = 0; i < n; i++)
-         d += xx[i] * xx[i-lag];
-      ac[lag] = d;
-      /*printf ("%f ", ac[lag]);*/
-      lag--;
+      for (i = k+fastN, d = 0; i < n; i++)
+         d = MAC16_16(d, xptr[i], xptr[i-k]);
+      ac[k] += d;
    }
-   /*printf ("\n");*/
-   ac[0] += 10;
+#ifdef FIXED_POINT
+   shift = 2*shift;
+   if (shift<=0)
+      ac[0] += SHL32((opus_int32)1, -shift);
+   if (ac[0] < 268435456)
+   {
+      int shift2 = 29 - EC_ILOG(ac[0]);
+      for (i=0;i<=lag;i++)
+         ac[i] = SHL32(ac[i], shift2);
+      shift -= shift2;
+   } else if (ac[0] >= 536870912)
+   {
+      int shift2=1;
+      if (ac[0] >= 1073741824)
+         shift2++;
+      for (i=0;i<=lag;i++)
+         ac[i] = SHR32(ac[i], shift2);
+      shift += shift2;
+   }
+#endif
 
    RESTORE_STACK;
+   return shift;
 }
-- 
cgit v1.2.3