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

Bit-exact with non-Neon version.

Browse Source 
Review URL: http://webrtc-codereview.appspot.com/180002

git-svn-id: http://webrtc.googlecode.com/svn/trunk@660 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
kma@google.com 2011-09-28 16:03:38 +00:00
parent 87d49798ca
commit c611b1a950
1 changed files with 39 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
src/modules/audio_processing/ns/main/source/nsx_core_neon.c
View File

@ -11,7 +11,9 @@
#if defined(WEBRTC_ARCH_ARM_NEON) && defined(WEBRTC_ANDROID)
#include "nsx_core.h"
#include <arm_neon.h>
#include <assert.h>
void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWord32 *noise,
WebRtc_Word16 *qNoise)
@ -28,6 +30,8 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
numerator = FACTOR_Q16;
tabind = inst->stages - inst->normData;
assert(tabind < 9);
assert(tabind > -9);
if (tabind < 0)
{
logval = -WebRtcNsx_kLogTable[-tabind];
@ -36,6 +40,8 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
logval = WebRtcNsx_kLogTable[tabind];
}
int16x8_t logval_16x8 = vdupq_n_s16(logval);
// lmagn(i)=log(magn(i))=log(2)*log2(magn(i))
// magn is in Q(-stages), and the real lmagn values are:
// real_lmagn(i)=log(magn(i)*2^stages)=log(magn(i))+log(2^stages)
@ -46,6 +52,7 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
{
zeros = WebRtcSpl_NormU32((WebRtc_UWord32)magn[i]);
frac = (WebRtc_Word16)((((WebRtc_UWord32)magn[i] << zeros) & 0x7FFFFFFF) >> 23);
assert(frac < 256);
// log2(magn(i))
log2 = (WebRtc_Word16)(((31 - zeros) << 8) + WebRtcNsx_kLogTableFrac[frac]);
// log2(magn(i))*log(2)
@ -62,6 +69,10 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
int16x8_t WIDTHQ8_16x8 = vdupq_n_s16(WIDTH_Q8);
int16x8_t WIDTHFACTOR_16x8 = vdupq_n_s16(widthFactor);
WebRtc_Word16 factor = FACTOR_Q7;
if (inst->blockIndex < END_STARTUP_LONG)
factor = FACTOR_Q7_STARTUP;
// Loop over simultaneous estimates
for (s = 0; s < SIMULT; s++)
{
@ -69,11 +80,12 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
// Get counter values from state
counter = inst->noiseEstCounter[s];
assert(counter < 201);
countDiv = WebRtcNsx_kCounterDiv[counter];
countProd = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16(counter, countDiv);
// quant_est(...)
WebRtc_Word16 delta_[8];
WebRtc_Word16 deltaBuff[8];
int16x4_t tmp16x4_0;
int16x4_t tmp16x4_1;
int16x4_t countDiv_16x4 = vdup_n_s16(countDiv);
@ -88,22 +100,25 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
int32x4_t tmp32x4;
for (i = 0; i < inst->magnLen - 7; i += 8) {
// compute delta
tmp16x8_0 = vdupq_n_s16(FACTOR_Q7);
vst1q_s16(delta_, tmp16x8_0);
// Compute delta.
// Smaller step size during startup. This prevents from using
// unrealistic values causing overflow.
tmp16x8_0 = vdupq_n_s16(factor);
vst1q_s16(deltaBuff, tmp16x8_0);
int j;
for (j = 0; j < 8; j++) {
if (inst->noiseEstDensity[offset + i + j] > 512)
delta_[j] = WebRtcSpl_DivW32W16ResW16(numerator,
deltaBuff[j] = WebRtcSpl_DivW32W16ResW16(numerator,
inst->noiseEstDensity[offset + i + j]);
}
// Update log quantile estimate
// tmp16 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(delta, countDiv, 14);
tmp32x4 = vmull_s16(vld1_s16(&delta_[0]), countDiv_16x4);
tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[0]), countDiv_16x4);
tmp16x4_1 = vshrn_n_s32(tmp32x4, 14);
tmp32x4 = vmull_s16(vld1_s16(&delta_[4]), countDiv_16x4);
tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[4]), countDiv_16x4);
tmp16x4_0 = vshrn_n_s32(tmp32x4, 14);
tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // Keep for several lines.
@ -133,6 +148,11 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // keep
tmp16x8_0 = vsubq_s16(tmp16x8_2, tmp16x8_0);
// logval is the smallest fixed point representation we can have. Values below
// that will correspond to values in the interval [0, 1], which can't possibly
// occur.
tmp16x8_0 = vmaxq_s16(tmp16x8_0, logval_16x8);
// Do the if-else branches:
tmp16x8_3 = vld1q_s16(&lmagn[i]); // keep for several lines
tmp16x8_5 = vsubq_s16(tmp16x8_3, tmp16x8_2);
@ -165,6 +185,11 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
} else
{
delta = FACTOR_Q7;
if (inst->blockIndex < END_STARTUP_LONG) {
// Smaller step size during startup. This prevents from using
// unrealistic values causing overflow.
delta = FACTOR_Q7_STARTUP;
}
}
// update log quantile estimate
@ -172,7 +197,7 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
if (lmagn[i] > inst->noiseEstLogQuantile[offset + i])
{
// +=QUANTILE*delta/(inst->counter[s]+1) QUANTILE=0.25, =1 in Q2
// CounterDiv=1/inst->counter[s] in Q15
// CounterDiv=1/(inst->counter[s]+1) in Q15
tmp16 += 2;
tmp16no1 = WEBRTC_SPL_RSHIFT_W16(tmp16, 2);
inst->noiseEstLogQuantile[offset + i] += tmp16no1;
@ -183,6 +208,12 @@ void WebRtcNsx_NoiseEstimation(NsxInst_t *inst, WebRtc_UWord16 *magn, WebRtc_UWo
// *(1-QUANTILE), in Q2 QUANTILE=0.25, 1-0.25=0.75=3 in Q2
tmp16no2 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, 3, 1);
inst->noiseEstLogQuantile[offset + i] -= tmp16no2;
if (inst->noiseEstLogQuantile[offset + i] < logval) {
// logval is the smallest fixed point representation we can have.
// Values below that will correspond to values in the interval
// [0, 1], which can't possibly occur.
inst->noiseEstLogQuantile[offset + i] = logval;
}
}
// update density estimate