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, 31 insertions, 3 deletions

diff --git a/lib/rbcodec/codecs/libopus/celt/arch.h b/lib/rbcodec/codecs/libopus/celt/arch.h
index b2d26c4ee4..035b92ff29 100644
--- a/lib/rbcodec/codecs/libopus/celt/arch.h
+++ b/lib/rbcodec/codecs/libopus/celt/arch.h
@@ -69,11 +69,9 @@ static OPUS_INLINE void _celt_fatal(const char *str, const char *file, int line)
 
 #define IMUL32(a,b) ((a)*(b))
 
-#define ABS(x) ((x) < 0 ? (-(x)) : (x))      /**< Absolute integer value. */
-#define ABS16(x) ((x) < 0 ? (-(x)) : (x))    /**< Absolute 16-bit value.  */
+#define ABS(x) ((x) < 0 ? (-(x)) : (x))
 #define MIN16(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum 16-bit value.   */
 #define MAX16(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum 16-bit value.   */
-#define ABS32(x) ((x) < 0 ? (-(x)) : (x))    /**< Absolute 32-bit value.  */
 #define MIN32(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum 32-bit value.   */
 #define MAX32(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum 32-bit value.   */
 #define IMIN(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum int value.   */
@@ -108,6 +106,13 @@ typedef opus_val32 celt_ener;
 #define SCALEIN(a)      (a)
 #define SCALEOUT(a)     (a)
 
+#define ABS16(x) ((x) < 0 ? (-(x)) : (x))
+#define ABS32(x) ((x) < 0 ? (-(x)) : (x))
+
+static OPUS_INLINE opus_int16 SAT16(opus_int32 x) {
+   return x > 32767 ? 32767 : x < -32768 ? -32768 : (opus_int16)x;
+}
+
 #ifdef FIXED_DEBUG
 #include "fixed_debug.h"
 #else
@@ -139,6 +144,22 @@ typedef float celt_sig;
 typedef float celt_norm;
 typedef float celt_ener;
 
+#ifdef FLOAT_APPROX
+/* This code should reliably detect NaN/inf even when -ffast-math is used.
+   Assumes IEEE 754 format. */
+static OPUS_INLINE int celt_isnan(float x)
+{
+   union {float f; opus_uint32 i;} in;
+   in.f = x;
+   return ((in.i>>23)&0xFF)==0xFF && (in.i&0x007FFFFF)!=0;
+}
+#else
+#ifdef __FAST_MATH__
+#error Cannot build libopus with -ffast-math unless FLOAT_APPROX is defined. This could result in crashes on extreme (e.g. NaN) input
+#endif
+#define celt_isnan(x) ((x)!=(x))
+#endif
+
 #define Q15ONE 1.0f
 
 #define NORM_SCALING 1.f
@@ -148,6 +169,10 @@ typedef float celt_ener;
 #define VERY_LARGE16 1e15f
 #define Q15_ONE ((opus_val16)1.f)
 
+/* This appears to be the same speed as C99's fabsf() but it's more portable. */
+#define ABS16(x) ((float)fabs(x))
+#define ABS32(x) ((float)fabs(x))
+
 #define QCONST16(x,bits) (x)
 #define QCONST32(x,bits) (x)
 
@@ -186,6 +211,7 @@ typedef float celt_ener;
 #define MULT32_32_Q31(a,b)     ((a)*(b))
 
 #define MAC16_32_Q15(c,a,b)     ((c)+(a)*(b))
+#define MAC16_32_Q16(c,a,b)     ((c)+(a)*(b))
 
 #define MULT16_16_Q11_32(a,b)     ((a)*(b))
 #define MULT16_16_Q11(a,b)     ((a)*(b))
@@ -203,6 +229,8 @@ typedef float celt_ener;
 #define SCALEIN(a)      ((a)*CELT_SIG_SCALE)
 #define SCALEOUT(a)     ((a)*(1/CELT_SIG_SCALE))
 
+#define SIG2WORD16(x) (x)
+
 #endif /* !FIXED_POINT */
 
 #ifndef GLOBAL_STACK_SIZE