ref: fb19ba9126988e123dbde1b2384b87bd1165fc55
parent: 9ac942660dd662851f025a97784fdcbf21be87fe
author: Jean-Marc Valin <jmvalin@jmvalin.ca>
date: Mon Aug 15 13:26:20 EDT 2016
Reducing the overhead due to dependency chains in dc_reject() for stereo Gives another ~2% speedup
--- a/src/opus_encoder.c
+++ b/src/opus_encoder.c
@@ -403,24 +403,57 @@
#else
static void dc_reject(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs)
{- int c, i;
- float coef;
-
+ int i;
+ float coef, coef2;
coef = 4.0f*cutoff_Hz/Fs;
- for (c=0;c<channels;c++)
+ coef2 = 1-coef;
+ if (channels==2)
{+ float m0, m1, m2, m3;
+ m0 = hp_mem[0];
+ m1 = hp_mem[1];
+ m2 = hp_mem[2];
+ m3 = hp_mem[3];
for (i=0;i<len;i++)
{+ opus_val32 x0, x1, tmp0, tmp1, y0, y1;
+ x0 = in[2*i+0];
+ x1 = in[2*i+1];
+ /* First stage */
+ tmp0 = x0-m0;
+ tmp1 = x1-m2;
+ m0 = coef*x0 + VERY_SMALL - coef2*m0;
+ m2 = coef*x1 + VERY_SMALL - coef2*m2;
+ /* Second stage */
+ y0 = tmp0 - m1;
+ y1 = tmp1 - m3;
+ m1 = coef*tmp0 + VERY_SMALL + coef2*m1;
+ m3 = coef*tmp1 + VERY_SMALL + coef2*m3;
+ out[2*i+0] = y0;
+ out[2*i+1] = y1;
+ }
+ hp_mem[0] = m0;
+ hp_mem[1] = m1;
+ hp_mem[2] = m2;
+ hp_mem[3] = m3;
+ } else {+ float m0, m1;
+ m0 = hp_mem[0];
+ m1 = hp_mem[1];
+ for (i=0;i<len;i++)
+ {opus_val32 x, tmp, y;
- x = in[channels*i+c];
+ x = in[i];
/* First stage */
- tmp = x-hp_mem[2*c];
- hp_mem[2*c] = coef*x + VERY_SMALL - (1-coef)*hp_mem[2*c];
+ tmp = x-m0;
+ m0 = coef*x + VERY_SMALL - coef2*m0;
/* Second stage */
- y = tmp - hp_mem[2*c+1];
- hp_mem[2*c+1] = coef*tmp + VERY_SMALL + (1-coef)*hp_mem[2*c+1];
- out[channels*i+c] = y;
+ y = tmp - m1;
+ m1 = coef*tmp + VERY_SMALL + coef2*m1;
+ out[i] = y;
}
+ hp_mem[0] = m0;
+ hp_mem[1] = m1;
}
}
#endif