mirrors/pulseaudio
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pulseaudio/pulseaudio.git synced 2025-11-03 09:01:50 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

echo-cancel: Add SSE optimisation to the adrian module

Browse source 
Optimises the core inner-product function, which takes the most CPU. The
SSE-optimised bits of the adrian echo canceller only if the CPU that PA
is running on actually supports SSE.
This commit is contained in:
Arun Raghavan 2010-09-21 20:42:32 +05:30
parent ab4223e9cf
commit 963250abb9
7 changed files with 61 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

37
src/modules/echo-cancel/adrian-aec.c
View file

@ -17,6 +17,10 @@
#include "adrian-aec.h"
#ifdef __SSE__
#include <xmmintrin.h>
#endif
/* Vector Dot Product */
static REAL dotp(REAL a[], REAL b[])
{
@ -31,8 +35,32 @@ static REAL dotp(REAL a[], REAL b[])
return sum0 + sum1;
}
static REAL dotp_sse(REAL a[], REAL b[]) __attribute__((noinline));
static REAL dotp_sse(REAL a[], REAL b[])
{
#ifdef __SSE__
/* This is taken from speex's inner product implementation */
int j;
REAL sum;
__m128 acc = _mm_setzero_ps();
AEC* AEC_init(int RATE)
for (j=0;j<NLMS_LEN;j+=8)
{
acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j), _mm_loadu_ps(b+j)));
acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j+4), _mm_loadu_ps(b+j+4)));
}
acc = _mm_add_ps(acc, _mm_movehl_ps(acc, acc));
acc = _mm_add_ss(acc, _mm_shuffle_ps(acc, acc, 0x55));
_mm_store_ss(&sum, acc);
return sum;
#else
return dotp(a, b);
#endif
}
AEC* AEC_init(int RATE, int have_vector)
{
AEC *a = pa_xnew(AEC, 1);
a->hangover = 0;
@ -57,6 +85,11 @@ AEC* AEC_init(int RATE)
a->dumpcnt = 0;
memset(a->ws, 0, sizeof(a->ws));
if (have_vector)
a->dotp = dotp_sse;
else
a->dotp = dotp;
return a;
}
@ -146,7 +179,7 @@ static REAL AEC_nlms_pw(AEC *a, REAL d, REAL x_, float stepsize)
// (mic signal - estimated mic signal from spk signal)
e = d;
if (a->hangover > 0) {
e -= dotp(a->w, a->x + a->j);
e -= a->dotp(a->w, a->x + a->j);
}
ef = IIR1_highpass(a->Fe, e); // pre-whitening of e