ref: a0bd7d3b8420be9c96f4e65e6113e11a05a23717
parent: 184a4767aad0e813cffa4aa178e0c161a4d17253
author: Jean-Marc Valin <[email protected]>
date: Fri Jul 22 05:13:00 EDT 2016
fixes overflows in celt_iir() by doing proper saturation
--- a/celt/arch.h
+++ b/celt/arch.h
@@ -206,6 +206,7 @@
#define SATURATE16(x) (x)
#define ROUND16(a,shift) (a)
+#define SROUND16(a,shift) (a)
#define HALF16(x) (.5f*(x))
#define HALF32(x) (.5f*(x))
--- a/celt/celt_lpc.c
+++ b/celt/celt_lpc.c
@@ -166,7 +166,7 @@
{mem[j]=mem[j-1];
}
- mem[0] = ROUND16(sum,SIG_SHIFT);
+ mem[0] = SROUND16(sum, SIG_SHIFT);
_y[i] = sum;
}
#else
@@ -195,20 +195,20 @@
xcorr_kernel(rden, y+i, sum, ord, arch);
/* Patch up the result to compensate for the fact that this is an IIR */
- y[i+ord ] = -ROUND16(sum[0],SIG_SHIFT);
+ y[i+ord ] = -SROUND16(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+ord+1] = -SROUND16(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+ord+2] = -SROUND16(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+ord+3] = -SROUND16(sum[3],SIG_SHIFT);
_y[i+3] = sum[3];
}
for (;i<N;i++)
@@ -216,7 +216,7 @@
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+ord] = SROUND16(sum,SIG_SHIFT);
_y[i] = sum;
}
for(i=0;i<ord;i++)
--- a/celt/fixed_debug.h
+++ b/celt/fixed_debug.h
@@ -233,6 +233,8 @@
#define VSHR32(a, shift) (((shift)>0) ? SHR32(a, shift) : SHL32(a, -(shift)))
#define ROUND16(x,a) (celt_mips--,EXTRACT16(PSHR32((x),(a))))
+#define SROUND16(x,a) (celt_mips--,EXTRACT16(SATURATE(PSHR32(x,a), 32767)));
+
#define HALF16(x) (SHR16(x,1))
#define HALF32(x) (SHR32(x,1))
--- a/celt/fixed_generic.h
+++ b/celt/fixed_generic.h
@@ -104,6 +104,9 @@
/** Shift by a and round-to-neareast 32-bit value. Result is a 16-bit value */
#define ROUND16(x,a) (EXTRACT16(PSHR32((x),(a))))
+/** Shift by a and round-to-neareast 32-bit value. Result is a saturated 16-bit value */
+#define SROUND16(x,a) EXTRACT16(SATURATE(PSHR32(x,a), 32767));
+
/** Divide by two */
#define HALF16(x) (SHR16(x,1))
#define HALF32(x) (SHR32(x,1))