Some general optimization in NS.

No big effort in introducing new style.
Speed improved ~2%.
Bit exact.
Will introduce mulpty-and-accumulate and sqrt_floor next, which increase speed another 2% or so.

Note: In function WebRtcNsx_DataAnalysis, did the block separation because I found one "if" case is more frequent than "else" within a for loop; rest is kind of code re-aligning.
Review URL: http://webrtc-codereview.appspot.com/181002

git-svn-id: http://webrtc.googlecode.com/svn/trunk@692 4adac7df-926f-26a2-2b94-8c16560cd09d

src/modules/audio_processing/ns/main/source/nsx_core.c

@@ -1429,25 +1429,26 @@ void WebRtcNsx_SpeechNoiseProb(NsxInst_t *inst, WebRtc_UWord16 *nonSpeechProbFin
                                                                          tmp16, 14); // Q14
 
     //final speech probability: combine prior model with LR factor:
-    for (i = 0; i < inst->magnLen; i++)
+
-    {
+    memset(nonSpeechProbFinal, 0, sizeof(WebRtc_UWord16) * inst->magnLen);
+
+    if (inst->priorNonSpeechProb > 0) {
+      for (i = 0; i < inst->magnLen; i++) {
         // FLOAT code
         // invLrt = exp(inst->logLrtTimeAvg[i]);
         // invLrt = inst->priorSpeechProb * invLrt;
         // nonSpeechProbFinal[i] = (1.0 - inst->priorSpeechProb) / (1.0 - inst->priorSpeechProb + invLrt);
         // invLrt = (1.0 - inst->priorNonSpeechProb) * invLrt;
         // nonSpeechProbFinal[i] = inst->priorNonSpeechProb / (inst->priorNonSpeechProb + invLrt);
-        nonSpeechProbFinal[i] = 0; // Q8
+        if (inst->logLrtTimeAvgW32[i] < 65300) {
-        if ((inst->logLrtTimeAvgW32[i] < 65300) && (inst->priorNonSpeechProb > 0))
-        {
           tmp32no1 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(inst->logLrtTimeAvgW32[i], 23637),
                                            14); // Q12
           intPart = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32no1, 12);
-            if (intPart < -8)
+          if (intPart < -8) {
-            {
             intPart = -8;
           }
           frac = (WebRtc_Word16)(tmp32no1 & 0x00000fff); // Q12
+
           // Quadratic approximation of 2^frac
           tmp32no2 = WEBRTC_SPL_RSHIFT_W32(frac * frac * 44, 19); // Q12
           tmp32no2 += WEBRTC_SPL_MUL_16_16_RSFT(frac, 84, 7); // Q12
@@ -1456,24 +1457,24 @@ void WebRtcNsx_SpeechNoiseProb(NsxInst_t *inst, WebRtc_UWord16 *nonSpeechProbFin
 
           normTmp = WebRtcSpl_NormW32(invLrtFX);
           normTmp2 = WebRtcSpl_NormW16((16384 - inst->priorNonSpeechProb));
-            if (normTmp + normTmp2 < 15)
+          if (normTmp + normTmp2 >= 7) {
-            {
+            if (normTmp + normTmp2 < 15) {
-                invLrtFX = WEBRTC_SPL_RSHIFT_W32(invLrtFX, 15 - normTmp2 - normTmp); // Q(normTmp+normTmp2-7)
+              invLrtFX = WEBRTC_SPL_RSHIFT_W32(invLrtFX, 15 - normTmp2 - normTmp);
-                tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX, (16384 - inst->priorNonSpeechProb)); // Q(normTmp+normTmp2+7)
+                                                                   // Q(normTmp+normTmp2-7)
+              tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX, (16384 - inst->priorNonSpeechProb));
+                                                                   // Q(normTmp+normTmp2+7)
               invLrtFX = WEBRTC_SPL_SHIFT_W32(tmp32no1, 7 - normTmp - normTmp2); // Q14
-            } else
+            } else {
-            {
               tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX, (16384 - inst->priorNonSpeechProb)); // Q22
               invLrtFX = WEBRTC_SPL_RSHIFT_W32(tmp32no1, 8); // Q14
             }
 
             tmp32no1 = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)inst->priorNonSpeechProb, 8); // Q22
+
             nonSpeechProbFinal[i] = (WebRtc_UWord16)WEBRTC_SPL_DIV(tmp32no1,
                     (WebRtc_Word32)inst->priorNonSpeechProb
                     + invLrtFX); // Q8
-            if (7 - normTmp - normTmp2 > 0)
+          }
-            {
-                nonSpeechProbFinal[i] = 0; // Q8
         }
       }
     }
@@ -1570,9 +1571,24 @@ void WebRtcNsx_DataAnalysis(NsxInst_t *inst, short *speechFrame, WebRtc_UWord16
     inst->sumMagn = (WebRtc_UWord32)magnU16[0]; // Q(normData-stages)
     inst->sumMagn += (WebRtc_UWord32)magnU16[inst->anaLen2];
 
+    if (inst->blockIndex >= END_STARTUP_SHORT) {
+      for (i = 1, j = 2; i < inst->anaLen2; i += 1, j += 2) {
+        inst->real[i] = realImag[j];
+        inst->imag[i] = -realImag[j + 1];
+        // magnitude spectrum
+        // energy in Q(2*(normData-stages))
+        tmpU32no1 = (WebRtc_UWord32)WEBRTC_SPL_MUL_16_16(realImag[j], realImag[j]);
+        tmpU32no1 += (WebRtc_UWord32)WEBRTC_SPL_MUL_16_16(realImag[j + 1], realImag[j + 1]);
+        inst->magnEnergy += tmpU32no1; // Q(2*(normData-stages))
+
+        magnU16[i] = (WebRtc_UWord16)WebRtcSpl_Sqrt(tmpU32no1); // Q(normData-stages)
+        inst->sumMagn += (WebRtc_UWord32)magnU16[i]; // Q(normData-stages)
+      }
+    } else {
+      //
       // Gather information during startup for noise parameter estimation
-    if (inst->blockIndex < END_STARTUP_SHORT)
+      //
-    {
+
       // Switch initMagnEst to Q(minNorm-stages)
       inst->initMagnEst[0] = WEBRTC_SPL_RSHIFT_U32(inst->initMagnEst[0],
                                                    right_shifts_in_initMagnEst);
@@ -1605,11 +1621,9 @@ void WebRtcNsx_DataAnalysis(NsxInst_t *inst, short *speechFrame, WebRtc_UWord16
       sum_log_magn = (WebRtc_Word32)log2; // Q8
       // sum_log_i_log_magn in Q17
       sum_log_i_log_magn = (WEBRTC_SPL_MUL_16_16(kLogIndex[inst->anaLen2], log2) >> 3);
-    }
 
-    for (i = 1; i < inst->anaLen2; i++)
+      for (i = 1, j = 2; i < inst->anaLen2; i += 1, j += 2)
       {
-        j = WEBRTC_SPL_LSHIFT_W16(i, 1);
         inst->real[i] = realImag[j];
         inst->imag[i] = -realImag[j + 1];
         // magnitude spectrum
@@ -1620,8 +1634,7 @@ void WebRtcNsx_DataAnalysis(NsxInst_t *inst, short *speechFrame, WebRtc_UWord16
 
         magnU16[i] = (WebRtc_UWord16)WebRtcSpl_Sqrt(tmpU32no1); // Q(normData-stages)
         inst->sumMagn += (WebRtc_UWord32)magnU16[i]; // Q(normData-stages)
-        if (inst->blockIndex < END_STARTUP_SHORT)
+
-        {
         // Switch initMagnEst to Q(minNorm-stages)
         inst->initMagnEst[i] = WEBRTC_SPL_RSHIFT_U32(inst->initMagnEst[i],
                                                      right_shifts_in_initMagnEst);
@@ -1651,12 +1664,13 @@ void WebRtcNsx_DataAnalysis(NsxInst_t *inst, short *speechFrame, WebRtc_UWord16
           sum_log_i_log_magn += (WEBRTC_SPL_MUL_16_16(kLogIndex[i], log2) >> 3);
         }
       }
-    }
 
+      //
       //compute simplified noise model during startup
-    if (inst->blockIndex < END_STARTUP_SHORT)
+      //
-    {
+
       // Estimate White noise
+
       // Switch whiteNoiseLevel to Q(minNorm-stages)
       inst->whiteNoiseLevel = WEBRTC_SPL_RSHIFT_U32(inst->whiteNoiseLevel,
                                                     right_shifts_in_initMagnEst);