generic-library/webrtc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Neon version of FilterFar()

Browse Source 
The performance gain on a Nexus 7 reported by audioproc is ~3.5%.

The output is bit exact.

BUG=3131
TESTED=verified performance manually, passed trybots
R=bjornv@webrtc.org, cd@webrtc.org, kwiberg@webrtc.org

Review URL: https://webrtc-codereview.appspot.com/12799005

Patch from Scott LaVarnway <slavarnw@gmail.com>.

git-svn-id: http://webrtc.googlecode.com/svn/trunk@6535 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
bjornv@webrtc.org 2014-06-25 12:15:46 +00:00
parent 1ed1af9b31
commit d5075bdbb5
1 changed files with 48 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
webrtc/modules/audio_processing/aec/aec_core_neon.c
View File

@ -30,6 +30,53 @@ __inline static float MulRe(float aRe, float aIm, float bRe, float bIm) {
return aRe * bRe - aIm * bIm; return aRe * bRe - aIm * bIm;
} }
__inline static float MulIm(float aRe, float aIm, float bRe, float bIm) {
return aRe * bIm + aIm * bRe;
}
static void FilterFarNEON(AecCore* aec, float yf[2][PART_LEN1]) {
int i;
const int num_partitions = aec->num_partitions;
for (i = 0; i < num_partitions; i++) {
int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1;
// Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) {
xPos -= num_partitions * PART_LEN1;
}
// vectorized code (four at once)
for (j = 0; j + 3 < PART_LEN1; j += 4) {
const float32x4_t xfBuf_re = vld1q_f32(&aec->xfBuf[0][xPos + j]);
const float32x4_t xfBuf_im = vld1q_f32(&aec->xfBuf[1][xPos + j]);
const float32x4_t wfBuf_re = vld1q_f32(&aec->wfBuf[0][pos + j]);
const float32x4_t wfBuf_im = vld1q_f32(&aec->wfBuf[1][pos + j]);
const float32x4_t yf_re = vld1q_f32(&yf[0][j]);
const float32x4_t yf_im = vld1q_f32(&yf[1][j]);
const float32x4_t a = vmulq_f32(xfBuf_re, wfBuf_re);
const float32x4_t e = vmlsq_f32(a, xfBuf_im, wfBuf_im);
const float32x4_t c = vmulq_f32(xfBuf_re, wfBuf_im);
const float32x4_t f = vmlaq_f32(c, xfBuf_im, wfBuf_re);
const float32x4_t g = vaddq_f32(yf_re, e);
const float32x4_t h = vaddq_f32(yf_im, f);
vst1q_f32(&yf[0][j], g);
vst1q_f32(&yf[1][j], h);
}
// scalar code for the remaining items.
for (; j < PART_LEN1; j++) {
yf[0][j] += MulRe(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
yf[1][j] += MulIm(aec->xfBuf[0][xPos + j],
aec->xfBuf[1][xPos + j],
aec->wfBuf[0][pos + j],
aec->wfBuf[1][pos + j]);
}
}
}
static float32x4_t vdivq_f32(float32x4_t a, float32x4_t b) { static float32x4_t vdivq_f32(float32x4_t a, float32x4_t b) {
int i; int i;
float32x4_t x = vrecpeq_f32(b); float32x4_t x = vrecpeq_f32(b);
@ -396,6 +443,7 @@ static void OverdriveAndSuppressNEON(AecCore* aec,
} }
void WebRtcAec_InitAec_neon(void) { void WebRtcAec_InitAec_neon(void) {
WebRtcAec_FilterFar = FilterFarNEON;
WebRtcAec_ScaleErrorSignal = ScaleErrorSignalNEON; WebRtcAec_ScaleErrorSignal = ScaleErrorSignalNEON;
WebRtcAec_FilterAdaptation = FilterAdaptationNEON; WebRtcAec_FilterAdaptation = FilterAdaptationNEON;
WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressNEON; WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressNEON;