ref: 9b3a3ad7ff3db7f43640be02faa0f869ecff5b99
parent: 6218cba258ae3e29af08253bebe46cfea3c3d39b
author: Jean-Marc Valin <[email protected]>
date: Tue Jan 21 11:23:24 EST 2014
Unrolled version of the comb filter for ARM (eliminates MOVs)
--- a/celt/celt.c
+++ b/celt/celt.c
@@ -97,7 +97,7 @@
x3 = SHL32(x[-T-1], 1);
x2 = SHL32(x[-T], 1);
x1 = SHL32(x[-T+1], 1);
- for (i=0;i<N;i++)
+ for (i=0;i<N-4;i+=5)
{
opus_val32 t;
x0=SHL32(x[i-T+2],1);
@@ -105,12 +105,42 @@
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,