Ducking fix #3: Removed the state as an input to the FilterAdaptation function

webrtc/modules/audio_processing/aec/aec_core_sse2.c

 /*
  *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 11 matching lines @@
 #include "webrtc/modules/audio_processing/aec/aec_rdft.h"
 
 __inline static float MulRe(float aRe, float aIm, float bRe, float 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 FilterFarSSE2(int num_partitions,
-                          int xfBufBlockPos,
-                          float xfBuf[2][kExtendedNumPartitions * PART_LEN1],
-                          float wfBuf[2][kExtendedNumPartitions * PART_LEN1],
-                          float yf[2][PART_LEN1]) {
+static void FilterFarSSE2(
+    int num_partitions,
+    int x_fft_buf_block_pos,
+    const float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
+    const float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
+    float y_fft[2][PART_LEN1]) {
 
   int i;
   const int num_partitions_local = num_partitions;
   for (i = 0; i < num_partitions_local; i++) {
     int j;
-    int xPos = (i + xfBufBlockPos) * PART_LEN1;
+    int xPos = (i + x_fft_buf_block_pos) * PART_LEN1;
     int pos = i * PART_LEN1;
     // Check for wrap
-    if (i + xfBufBlockPos >= num_partitions_local) {
+    if (i + x_fft_buf_block_pos >= num_partitions_local) {
       xPos -= num_partitions_local * (PART_LEN1);
     }
 
     // vectorized code (four at once)
     for (j = 0; j + 3 < PART_LEN1; j += 4) {
-      const __m128 xfBuf_re = _mm_loadu_ps(&xfBuf[0][xPos + j]);
-      const __m128 xfBuf_im = _mm_loadu_ps(&xfBuf[1][xPos + j]);
-      const __m128 wfBuf_re = _mm_loadu_ps(&wfBuf[0][pos + j]);
-      const __m128 wfBuf_im = _mm_loadu_ps(&wfBuf[1][pos + j]);
-      const __m128 yf_re = _mm_loadu_ps(&yf[0][j]);
-      const __m128 yf_im = _mm_loadu_ps(&yf[1][j]);
-      const __m128 a = _mm_mul_ps(xfBuf_re, wfBuf_re);
-      const __m128 b = _mm_mul_ps(xfBuf_im, wfBuf_im);
-      const __m128 c = _mm_mul_ps(xfBuf_re, wfBuf_im);
-      const __m128 d = _mm_mul_ps(xfBuf_im, wfBuf_re);
+      const __m128 x_fft_buf_re = _mm_loadu_ps(&x_fft_buf[0][xPos + j]);
+      const __m128 x_fft_buf_im = _mm_loadu_ps(&x_fft_buf[1][xPos + j]);
+      const __m128 h_fft_buf_re = _mm_loadu_ps(&h_fft_buf[0][pos + j]);
+      const __m128 h_fft_buf_im = _mm_loadu_ps(&h_fft_buf[1][pos + j]);
+      const __m128 y_fft_re = _mm_loadu_ps(&y_fft[0][j]);
+      const __m128 y_fft_im = _mm_loadu_ps(&y_fft[1][j]);
+      const __m128 a = _mm_mul_ps(x_fft_buf_re, h_fft_buf_re);
+      const __m128 b = _mm_mul_ps(x_fft_buf_im, h_fft_buf_im);
+      const __m128 c = _mm_mul_ps(x_fft_buf_re, h_fft_buf_im);
+      const __m128 d = _mm_mul_ps(x_fft_buf_im, h_fft_buf_re);
       const __m128 e = _mm_sub_ps(a, b);
       const __m128 f = _mm_add_ps(c, d);
-      const __m128 g = _mm_add_ps(yf_re, e);
-      const __m128 h = _mm_add_ps(yf_im, f);
-      _mm_storeu_ps(&yf[0][j], g);
-      _mm_storeu_ps(&yf[1][j], h);
+      const __m128 g = _mm_add_ps(y_fft_re, e);
+      const __m128 h = _mm_add_ps(y_fft_im, f);
+      _mm_storeu_ps(&y_fft[0][j], g);
+      _mm_storeu_ps(&y_fft[1][j], h);
     }
     // scalar code for the remaining items.
     for (; j < PART_LEN1; j++) {
-      yf[0][j] += MulRe(xfBuf[0][xPos + j],
-                        xfBuf[1][xPos + j],
-                        wfBuf[0][pos + j],
-                        wfBuf[1][pos + j]);
-      yf[1][j] += MulIm(xfBuf[0][xPos + j],
-                        xfBuf[1][xPos + j],
-                        wfBuf[0][pos + j],
-                        wfBuf[1][pos + j]);
+      y_fft[0][j] += MulRe(x_fft_buf[0][xPos + j],
+                           x_fft_buf[1][xPos + j],
+                           h_fft_buf[0][pos + j],
+                           h_fft_buf[1][pos + j]);
+      y_fft[1][j] += MulIm(x_fft_buf[0][xPos + j],
+                           x_fft_buf[1][xPos + j],
+                           h_fft_buf[0][pos + j],
+                           h_fft_buf[1][pos + j]);
     }
   }
 }
 
 static void ScaleErrorSignalSSE2(int extended_filter_enabled,
                                  float normal_mu,
                                  float normal_error_threshold,
-                                 float *xPow,
+                                 const float xPow[PART_LEN1],

[hlundin-webrtc, 2015/11/24 13:51:44]

x_pow

[peah-webrtc, 2015/11/26 05:55:17]

Done.

                                  float ef[2][PART_LEN1]) {
   const __m128 k1e_10f = _mm_set1_ps(1e-10f);
   const __m128 kMu = extended_filter_enabled ? _mm_set1_ps(kExtendedMu)
       : _mm_set1_ps(normal_mu);
   const __m128 kThresh = extended_filter_enabled
                              ? _mm_set1_ps(kExtendedErrorThreshold)
                              : _mm_set1_ps(normal_error_threshold);
 
   int i;
   // vectorized code (four at once)
@@ skipping 45 matching lines @@
         ef[1][i] *= abs_ef;
       }
 
       // Stepsize factor
       ef[0][i] *= mu;
       ef[1][i] *= mu;
     }
   }
 }
 
-static void FilterAdaptationSSE2(AecCore* aec,
-                                 float* fft,
-                                 float ef[2][PART_LEN1]) {
+static void FilterAdaptationSSE2(
+    int num_partitions,
+    int x_fft_buf_block_pos,
+    const float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1],
+    const float e_fft[2][PART_LEN1],
+    float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1]) {
+  float fft[PART_LEN2];
   int i, j;
-  const int num_partitions = aec->num_partitions;
   for (i = 0; i < num_partitions; i++) {
-    int xPos = (i + aec->xfBufBlockPos) * (PART_LEN1);
+    int xPos = (i + x_fft_buf_block_pos) * (PART_LEN1);
     int pos = i * PART_LEN1;
     // Check for wrap
-    if (i + aec->xfBufBlockPos >= num_partitions) {
+    if (i + x_fft_buf_block_pos >= num_partitions) {
       xPos -= num_partitions * PART_LEN1;
     }
 
     // Process the whole array...
     for (j = 0; j < PART_LEN; j += 4) {
-      // Load xfBuf and ef.
-      const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
-      const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
-      const __m128 ef_re = _mm_loadu_ps(&ef[0][j]);
-      const __m128 ef_im = _mm_loadu_ps(&ef[1][j]);
-      // Calculate the product of conjugate(xfBuf) by ef.
+      // Load x_fft_buf and e_fft.
+      const __m128 x_fft_buf_re = _mm_loadu_ps(&x_fft_buf[0][xPos + j]);
+      const __m128 x_fft_buf_im = _mm_loadu_ps(&x_fft_buf[1][xPos + j]);
+      const __m128 e_fft_re = _mm_loadu_ps(&e_fft[0][j]);
+      const __m128 e_fft_im = _mm_loadu_ps(&e_fft[1][j]);
+      // Calculate the product of conjugate(x_fft_buf) by e_fft.
       //   re(conjugate(a) * b) = aRe * bRe + aIm * bIm
       //   im(conjugate(a) * b)=  aRe * bIm - aIm * bRe
-      const __m128 a = _mm_mul_ps(xfBuf_re, ef_re);
-      const __m128 b = _mm_mul_ps(xfBuf_im, ef_im);
-      const __m128 c = _mm_mul_ps(xfBuf_re, ef_im);
-      const __m128 d = _mm_mul_ps(xfBuf_im, ef_re);
+      const __m128 a = _mm_mul_ps(x_fft_buf_re, e_fft_re);
+      const __m128 b = _mm_mul_ps(x_fft_buf_im, e_fft_im);
+      const __m128 c = _mm_mul_ps(x_fft_buf_re, e_fft_im);
+      const __m128 d = _mm_mul_ps(x_fft_buf_im, e_fft_re);
       const __m128 e = _mm_add_ps(a, b);
       const __m128 f = _mm_sub_ps(c, d);
       // Interleave real and imaginary parts.
       const __m128 g = _mm_unpacklo_ps(e, f);
       const __m128 h = _mm_unpackhi_ps(e, f);
       // Store
       _mm_storeu_ps(&fft[2 * j + 0], g);
       _mm_storeu_ps(&fft[2 * j + 4], h);
     }
     // ... and fixup the first imaginary entry.
-    fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
-                   -aec->xfBuf[1][xPos + PART_LEN],
-                   ef[0][PART_LEN],
-                   ef[1][PART_LEN]);
+    fft[1] = MulRe(x_fft_buf[0][xPos + PART_LEN],
+                   -x_fft_buf[1][xPos + PART_LEN],
+                   e_fft[0][PART_LEN],
+                   e_fft[1][PART_LEN]);
 
     aec_rdft_inverse_128(fft);
     memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
 
     // fft scaling
     {
       float scale = 2.0f / PART_LEN2;
       const __m128 scale_ps = _mm_load_ps1(&scale);
       for (j = 0; j < PART_LEN; j += 4) {
         const __m128 fft_ps = _mm_loadu_ps(&fft[j]);
         const __m128 fft_scale = _mm_mul_ps(fft_ps, scale_ps);
         _mm_storeu_ps(&fft[j], fft_scale);
       }
     }
     aec_rdft_forward_128(fft);
 
     {
-      float wt1 = aec->wfBuf[1][pos];
-      aec->wfBuf[0][pos + PART_LEN] += fft[1];
+      float wt1 = h_fft_buf[1][pos];
+      h_fft_buf[0][pos + PART_LEN] += fft[1];
       for (j = 0; j < PART_LEN; j += 4) {
-        __m128 wtBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
-        __m128 wtBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
+        __m128 wtBuf_re = _mm_loadu_ps(&h_fft_buf[0][pos + j]);
+        __m128 wtBuf_im = _mm_loadu_ps(&h_fft_buf[1][pos + j]);
         const __m128 fft0 = _mm_loadu_ps(&fft[2 * j + 0]);
         const __m128 fft4 = _mm_loadu_ps(&fft[2 * j + 4]);
         const __m128 fft_re =
             _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2, 0));
         const __m128 fft_im =
             _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3, 1));
         wtBuf_re = _mm_add_ps(wtBuf_re, fft_re);
         wtBuf_im = _mm_add_ps(wtBuf_im, fft_im);
-        _mm_storeu_ps(&aec->wfBuf[0][pos + j], wtBuf_re);
-        _mm_storeu_ps(&aec->wfBuf[1][pos + j], wtBuf_im);
+        _mm_storeu_ps(&h_fft_buf[0][pos + j], wtBuf_re);
+        _mm_storeu_ps(&h_fft_buf[1][pos + j], wtBuf_im);
       }
-      aec->wfBuf[1][pos] = wt1;
+      h_fft_buf[1][pos] = wt1;
     }
   }
 }
 
 static __m128 mm_pow_ps(__m128 a, __m128 b) {
   // a^b = exp2(b * log2(a))
   //   exp2(x) and log2(x) are calculated using polynomial approximations.
   __m128 log2_a, b_log2_a, a_exp_b;
 
   // Calculate log2(x), x = a.
@@ skipping 495 matching lines @@
   }
 }
 
 void WebRtcAec_InitAec_SSE2(void) {
   WebRtcAec_FilterFar = FilterFarSSE2;
   WebRtcAec_ScaleErrorSignal = ScaleErrorSignalSSE2;
   WebRtcAec_FilterAdaptation = FilterAdaptationSSE2;
   WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2;
   WebRtcAec_SubbandCoherence = SubbandCoherenceSSE2;
 }