Introduced ARM version of WebRtcSpl_SqrtFloor(). Function cycles reduced by ~ 30% in a real time VOE test in an android device (Nexus-S, ARMv7a).

// Fritz, I added you as a reviewer for the assembly files, just as a warm-up for future storms. :-) The assembly code was from public domain and there's little to touch.
Review URL: https://webrtc-codereview.appspot.com/369017

git-svn-id: http://webrtc.googlecode.com/svn/trunk@1627 4adac7df-926f-26a2-2b94-8c16560cd09d

src/LICENSE_THIRD_PARTY

@@ -4,6 +4,8 @@ licenses than the one provided in the LICENSE file in the root of the source
 tree.
 
 Files governed by third party licenses:
+common_audio/signal_processing/spl_sqrt_floor.c
+common_audio/signal_processing/spl_sqrt_floor.s
 modules/audio_coding/codecs/G711/main/source/g711.h
 modules/audio_coding/codecs/G711/main/source/g711.c
 modules/audio_coding/codecs/G722/main/source/g722_decode.h

src/common_audio/signal_processing/Android.mk

@@ -42,7 +42,6 @@ LOCAL_SRC_FILES := \
     resample_by_2_internal.c \
     resample_fractional.c \
     spl_sqrt.c \
-    spl_sqrt_floor.c \
     spl_version.c \
     splitting_filter.c \
     sqrt_of_one_minus_x_squared.c \
@@ -75,6 +74,14 @@ LOCAL_SRC_FILES += \
     filter_ar_fast_q12.c
 endif
 
+ifeq ($(TARGET_ARCH),arm)
+LOCAL_SRC_FILES += \
+    spl_sqrt_floor.s
+else
+LOCAL_SRC_FILES += \
+    spl_sqrt_floor.c
+endif
+
 LOCAL_SHARED_LIBRARIES := libstlport
 
 ifeq ($(TARGET_OS)-$(TARGET_SIMULATOR),linux-true)

src/common_audio/signal_processing/spl_sqrt_floor.c

@@ -1,21 +1,26 @@
 /*
- *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
+ * Written by Wilco Dijkstra, 1996. Refer to file LICENSE under
- *
+ * trunk/third_party_mods/sqrt_floor.
- *  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.
- */
-
-/*
- * This file contains the function WebRtcSpl_SqrtFloor().
- * The description header can be found in signal_processing_library.h
  *
+ * Minor modifications in code style for WebRTC, 2012.
  */
 
 #include "signal_processing_library.h"
 
+/*
+ * Algorithm:
+ * Successive approximation of the equation (root + delta) ^ 2 = N
+ * until delta < 1. If delta < 1 we have the integer part of SQRT (N).
+ * Use delta = 2^i for i = 15 .. 0.
+ *
+ * Output precision is 16 bits. Note for large input values (close to
+ * 0x7FFFFFFF), bit 15 (the highest bit of the low 16-bit half word)
+ * contains the MSB information (a non-sign value). Do with caution
+ * if you need to cast the output to int16_t type.
+ *
+ * If the input value is negative, it returns 0.
+ */
+
 #define WEBRTC_SPL_SQRT_ITER(N)                 \
   try1 = root + (1 << (N));                     \
   if (value >= try1 << (N))                     \
@@ -24,30 +29,26 @@
     root |= 2 << (N);                           \
   }
 
-// (out) Square root of input parameter
+int32_t WebRtcSpl_SqrtFloor(int32_t value)
-WebRtc_Word32 WebRtcSpl_SqrtFloor(WebRtc_Word32 value)
 {
-    // new routine for performance, 4 cycles/bit in ARM
+  int32_t root = 0, try1;
-    // output precision is 16 bits
 
-    WebRtc_Word32 root = 0, try1;
+  WEBRTC_SPL_SQRT_ITER (15);
+  WEBRTC_SPL_SQRT_ITER (14);
+  WEBRTC_SPL_SQRT_ITER (13);
+  WEBRTC_SPL_SQRT_ITER (12);
+  WEBRTC_SPL_SQRT_ITER (11);
+  WEBRTC_SPL_SQRT_ITER (10);
+  WEBRTC_SPL_SQRT_ITER ( 9);
+  WEBRTC_SPL_SQRT_ITER ( 8);
+  WEBRTC_SPL_SQRT_ITER ( 7);
+  WEBRTC_SPL_SQRT_ITER ( 6);
+  WEBRTC_SPL_SQRT_ITER ( 5);
+  WEBRTC_SPL_SQRT_ITER ( 4);
+  WEBRTC_SPL_SQRT_ITER ( 3);
+  WEBRTC_SPL_SQRT_ITER ( 2);
+  WEBRTC_SPL_SQRT_ITER ( 1);
+  WEBRTC_SPL_SQRT_ITER ( 0);
 
-    WEBRTC_SPL_SQRT_ITER (15);
+  return root >> 1;
-    WEBRTC_SPL_SQRT_ITER (14);
-    WEBRTC_SPL_SQRT_ITER (13);
-    WEBRTC_SPL_SQRT_ITER (12);
-    WEBRTC_SPL_SQRT_ITER (11);
-    WEBRTC_SPL_SQRT_ITER (10);
-    WEBRTC_SPL_SQRT_ITER ( 9);
-    WEBRTC_SPL_SQRT_ITER ( 8);
-    WEBRTC_SPL_SQRT_ITER ( 7);
-    WEBRTC_SPL_SQRT_ITER ( 6);
-    WEBRTC_SPL_SQRT_ITER ( 5);
-    WEBRTC_SPL_SQRT_ITER ( 4);
-    WEBRTC_SPL_SQRT_ITER ( 3);
-    WEBRTC_SPL_SQRT_ITER ( 2);
-    WEBRTC_SPL_SQRT_ITER ( 1);
-    WEBRTC_SPL_SQRT_ITER ( 0);
-
-    return root >> 1;
 }

src/common_audio/signal_processing/spl_sqrt_floor.s

@@ -0,0 +1,88 @@
+@ Written by Wilco Dijkstra, 1996. Refer to file LICENSE under
+@ trunk/third_party_mods/sqrt_floor.
+@
+@ Minor modifications in code style for WebRTC, 2012.
+@ Output is bit-exact with the reference C code in spl_sqrt_floor.c.
+
+@ Input :             r0 32 bit unsigned integer
+@ Output:             r0 = INT (SQRT (r0)), precision is 16 bits
+@ Registers touched:  r1, r2
+
+.global WebRtcSpl_SqrtFloor
+
+.align  2
+WebRtcSpl_SqrtFloor:
+.fnstart
+  mov    r1, #3 << 30
+  mov    r2, #1 << 30
+
+  @ unroll for i = 0 .. 15
+
+  cmp    r0, r2, ror #2 * 0
+  subhs  r0, r0, r2, ror #2 * 0
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 1
+  subhs  r0, r0, r2, ror #2 * 1
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 2
+  subhs  r0, r0, r2, ror #2 * 2
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 3
+  subhs  r0, r0, r2, ror #2 * 3
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 4
+  subhs  r0, r0, r2, ror #2 * 4
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 5
+  subhs  r0, r0, r2, ror #2 * 5
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 6
+  subhs  r0, r0, r2, ror #2 * 6
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 7
+  subhs  r0, r0, r2, ror #2 * 7
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 8
+  subhs  r0, r0, r2, ror #2 * 8
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 9
+  subhs  r0, r0, r2, ror #2 * 9
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 10
+  subhs  r0, r0, r2, ror #2 * 10
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 11
+  subhs  r0, r0, r2, ror #2 * 11
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 12
+  subhs  r0, r0, r2, ror #2 * 12
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 13
+  subhs  r0, r0, r2, ror #2 * 13
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 14
+  subhs  r0, r0, r2, ror #2 * 14
+  adc    r2, r1, r2, lsl #1
+
+  cmp    r0, r2, ror #2 * 15
+  subhs  r0, r0, r2, ror #2 * 15
+  adc    r2, r1, r2, lsl #1
+
+  bic    r0, r2, #3 << 30  @ for rounding add: cmp r0, r2  adc r2, #1
+  bx lr
+
+.fnend

third_party_mods/sqrt_floor/LICENSE

@@ -0,0 +1,26 @@
+The following email record is related to source files spl_sqrt_floor.c
+and spl_sqrt_floor.s in trunk/src/common_audio/signal_processing/.
+
+
+From: Wilco Dijkstra <Wilco.Dijkstra@ntlworld.com>
+Date: Fri, Jun 24, 2011 at 3:20 AM
+Subject: Re: sqrt routine
+To: Kevin Ma <kma@google.com>
+Hi Kevin,
+Thanks for asking. Those routines are public domain (originally posted to 
+comp.sys.arm a long time ago), so you can use them freely for any purpose.
+Cheers,
+Wilco
+
+----- Original Message -----
+From: "Kevin Ma" <kma@google.com>
+To: <Wilco.Dijkstra@ntlworld.com>
+Sent: Thursday, June 23, 2011 11:44 PM
+Subject: Fwd: sqrt routine
+Hi Wilco,
+I saw your sqrt routine from several web sites, including
+http://www.finesse.demon.co.uk/steven/sqrt.html.
+Just wonder if there's any copyright information with your Successive
+approximation routines, or if I can freely use it for any purpose.
+Thanks.
+Kevin