From 72f8a6d77bf0d9815ab0657cbf7c419d35bf93db Mon Sep 17 00:00:00 2001
From: "kma@webrtc.org" <kma@webrtc.org@4adac7df-926f-26a2-2b94-8c16560cd09d>
Date: Mon, 9 Jul 2012 23:27:02 +0000
Subject: [PATCH] Optimized PCorr2Q32() in iSAC with intrinsics in ARM Neon
 platform. Review URL: https://webrtc-codereview.appspot.com/634004

git-svn-id: http://webrtc.googlecode.com/svn/trunk@2497 4adac7df-926f-26a2-2b94-8c16560cd09d
---
 .../codecs/iSAC/fix/source/pitch_estimator.c  | 45 ++++++++++++++++---
 .../codecs/iSAC/fix/source/pitch_estimator.h  | 10 ++---
 2 files changed, 43 insertions(+), 12 deletions(-)

diff --git a/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.c b/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.c
index 1702098ed..d95c19b3d 100644
--- a/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.c
+++ b/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.c
@@ -15,7 +15,9 @@
  *
  */
 
-#include <string.h>
+#ifdef WEBRTC_ARCH_ARM_NEON
+#include <arm_neon.h>
+#endif
 
 #include "signal_processing_library.h"
 #include "pitch_estimator.h"
@@ -201,15 +203,44 @@ static void PCorr2Q32(const WebRtc_Word16 *in, WebRtc_Word32 *logcorQ8)
     inptr = &in[k];
     ysum32 -= WEBRTC_SPL_MUL_16_16_RSFT( (WebRtc_Word16) in[k-1],(WebRtc_Word16) in[k-1], scaling);
     ysum32 += WEBRTC_SPL_MUL_16_16_RSFT( (WebRtc_Word16) in[PITCH_CORR_LEN2 + k - 1],(WebRtc_Word16) in[PITCH_CORR_LEN2 + k - 1], scaling);
-    csum32 = 0;
-    prod32 = WEBRTC_SPL_MUL_16_16_RSFT( (WebRtc_Word16) x[0],(WebRtc_Word16) inptr[0], scaling);
 
-    for (n = 1; n < PITCH_CORR_LEN2; n++) {
-      csum32 += prod32;
-      prod32 = WEBRTC_SPL_MUL_16_16_RSFT( (WebRtc_Word16) x[n],(WebRtc_Word16) inptr[n], scaling);
+#ifdef WEBRTC_ARCH_ARM_NEON
+    {
+      int32_t vbuff[4];
+      int32x4_t int_32x4_sum = vmovq_n_s32(0);
+      // Can't shift a Neon register to right with a non-constant shift value.
+      int32x4_t int_32x4_scale = vdupq_n_s32(-scaling);
+      // Assert a codition used in loop unrolling at compile-time.
+      WEBRTC_STATIC_ASSERT(PITCH_CORR_LEN2, PITCH_CORR_LEN2 %4 == 0);
+
+      for (n = 0; n < PITCH_CORR_LEN2; n += 4) {
+        int16x4_t int_16x4_x = vld1_s16(&x[n]);
+        int16x4_t int_16x4_in = vld1_s16(&inptr[n]);
+        int32x4_t int_32x4 = vmull_s16(int_16x4_x, int_16x4_in);
+        int_32x4 = vshlq_s32(int_32x4, int_32x4_scale);
+        int_32x4_sum = vaddq_s32(int_32x4_sum, int_32x4);
+      }
+
+      // Use vector store to avoid long stall from data trasferring
+      // from vector to general register.
+      vst1q_s32(vbuff, int_32x4_sum);
+      csum32 = vbuff[0] + vbuff[1];
+      csum32 += vbuff[2];
+      csum32 += vbuff[3];
     }
+#else
+    csum32 = 0;
+    if(scaling == 0) {
+      for (n = 0; n < PITCH_CORR_LEN2; n++) {
+        csum32 += x[n] * inptr[n];
+      }
+    } else {
+      for (n = 0; n < PITCH_CORR_LEN2; n++) {
+        csum32 += (x[n] * inptr[n]) >> scaling;
+      }
+    }
+#endif
 
-    csum32 += prod32;
     logcorQ8--;
 
     lys=Log2Q8((WebRtc_UWord32)ysum32); // Q8
diff --git a/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.h b/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.h
index afdc9785d..8ba034c3b 100644
--- a/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.h
+++ b/src/modules/audio_coding/codecs/iSAC/fix/source/pitch_estimator.h
@@ -1,5 +1,5 @@
 /*
- *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
+ *  Copyright (c) 2012 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
@@ -20,7 +20,10 @@
 
 #include "structs.h"
 
-
+// TODO(andrew): put this into general WebRTC so other modules can use it.
+// Define a compiler-time assertion.
+#define WEBRTC_STATIC_ASSERT(name, boolean_cond) \
+  static char const static_assert_##name[(boolean_cond) ? 1 : -1] = {'!'}
 
 void WebRtcIsacfix_PitchAnalysis(const WebRtc_Word16 *in,               /* PITCH_FRAME_LEN samples */
                                  WebRtc_Word16 *outQ0,                  /* PITCH_FRAME_LEN+QLOOKAHEAD samples */
@@ -28,7 +31,6 @@ void WebRtcIsacfix_PitchAnalysis(const WebRtc_Word16 *in,               /* PITCH
                                  WebRtc_Word16 *lagsQ7,
                                  WebRtc_Word16 *PitchGains_Q12);
 
-
 void WebRtcIsacfix_InitialPitch(const WebRtc_Word16 *in,
                                 PitchAnalysisStruct *State,
                                 WebRtc_Word16 *qlags);
@@ -45,8 +47,6 @@ void WebRtcIsacfix_PitchFilterGains(const WebRtc_Word16 *indatQ0,
                                     WebRtc_Word16 *lagsQ7,
                                     WebRtc_Word16 *gainsQ12);
 
-
-
 void WebRtcIsacfix_DecimateAllpass32(const WebRtc_Word16 *in,
                                      WebRtc_Word32 *state_in,        /* array of size: 2*ALLPASSSECTIONS+1 */
                                      WebRtc_Word16 N,                   /* number of input samples */