echo-cancel: orc-ify some bits for optimisation

This uses Orc to optimise an inner loop in the core NLMS function of the
Adrian echo canceller.

build/orc.mak

@@ -0,0 +1,71 @@
+#
+# This is a makefile.am fragment to build Orc code.
+#
+# Define ORC_SOURCE and then include this file, such as:
+#
+#  ORC_SOURCE=gstadderorc
+#  include $(top_srcdir)/common/orc.mak
+#
+# This fragment will create tmp-orc.c and gstadderorc.h from
+# gstadderorc.orc.
+#
+# When 'make dist' is run at the top level, or 'make orc-update'
+# in a directory including this fragment, the generated source
+# files will be copied to $(ORC_SOURCE)-dist.[ch].  These files
+# should be checked in to git, since they are used if Orc is
+# disabled.
+#
+# Note that this file defines BUILT_SOURCES, so any later usage
+# of BUILT_SOURCES in the Makefile.am that includes this file
+# must use '+='.
+#
+
+
+EXTRA_DIST += $(ORC_SOURCE).orc
+
+ORC_NODIST_SOURCES = tmp-orc.c $(ORC_SOURCE).h
+BUILT_SOURCES += tmp-orc.c $(ORC_SOURCE).h
+
+
+orc-update: tmp-orc.c $(ORC_SOURCE).h
+	cp tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c
+	cp $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h
+
+orcc_v_gen = $(orcc_v_gen_$(V))
+orcc_v_gen_ = $(orcc_v_gen_$(AM_DEFAULT_VERBOSITY))
+orcc_v_gen_0 = @echo "  ORCC   $@";
+
+cp_v_gen = $(cp_v_gen_$(V))
+cp_v_gen_ = $(cp_v_gen_$(AM_DEFAULT_VERBOSITY))
+cp_v_gen_0 = @echo "  CP     $@";
+
+if HAVE_ORC
+tmp-orc.c: $(srcdir)/$(ORC_SOURCE).orc
+	$(orcc_v_gen)$(ORCC) --implementation -o $(builddir)/tmp-orc.c $(srcdir)/$(ORC_SOURCE).orc
+
+$(ORC_SOURCE).h: $(srcdir)/$(ORC_SOURCE).orc
+	mkdir -p $$(dirname $(builddir)/$(ORC_SOURCE).h)
+	$(orcc_v_gen)$(ORCC) --header -o $(builddir)/$(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE).orc
+else
+tmp-orc.c: $(srcdir)/$(ORC_SOURCE).orc
+	$(cp_v_gen)cp $(srcdir)/$(ORC_SOURCE)-dist.c tmp-orc.c
+
+$(ORC_SOURCE).h: $(srcdir)/$(ORC_SOURCE).orc
+	$(cp_v_gen)cp $(srcdir)/$(ORC_SOURCE)-dist.h $(ORC_SOURCE).h
+endif
+
+clean-local: clean-orc
+.PHONY: clean-orc
+clean-orc:
+	rm -f tmp-orc.c $(ORC_SOURCE).h
+
+dist-hook: dist-hook-orc
+.PHONY: dist-hook-orc
+dist-hook-orc: tmp-orc.c $(ORC_SOURCE).h
+	rm -f tmp-orc.c~
+	cmp -s tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c || \
+	  cp tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c
+	cmp -s $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h || \
+	  cp $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h
+	cp -p $(srcdir)/$(ORC_SOURCE)-dist.c $(distdir)/
+	cp -p $(srcdir)/$(ORC_SOURCE)-dist.h $(distdir)/

configure.ac

@@ -1321,6 +1321,9 @@ if test "x${with_fftw}" != "xno"; then
 fi
 AM_CONDITIONAL([HAVE_FFTW], [test "x$HAVE_FFTW" = "x1"])
 
+### ORC (optional) ###
+ORC_CHECK([0.4.9])
+
 ### Build and Install man pages ###
 AC_ARG_ENABLE(manpages,
         AS_HELP_STRING([--disable-manpages],[Disable building and installation of man pages]),
@@ -1573,6 +1576,11 @@ if test "x${HAVE_FFTW}" = "x1" ; then
    ENABLE_FFTW=yes
 fi
 
+ENABLE_ORC=no
+if test "x${HAVE_ORC}" = "xyes" ; then
+   ENABLE_ORC=yes
+fi
+
 ENABLE_OPENSSL=no
 if test "x${HAVE_OPENSSL}" = "x1" ; then
    ENABLE_OPENSSL=yes
@@ -1626,6 +1634,7 @@ echo "
     Enable gdbm:                   ${ENABLE_GDBM}
     Enable simple database:        ${ENABLE_SIMPLEDB}
     Enable fftw:                   ${ENABLE_FFTW}
+    Enable orc:                    ${ENABLE_ORC}
 
     System User:                   ${PA_SYSTEM_USER}
     System Group:                  ${PA_SYSTEM_GROUP}

m4/orc.m4

@@ -0,0 +1,56 @@
+dnl pkg-config-based checks for Orc
+
+dnl specific:
+dnl ORC_CHECK([REQUIRED_VERSION])
+
+AC_DEFUN([ORC_CHECK],
+[
+  ORC_REQ=ifelse([$1], , "0.4.5", [$1])
+
+  AC_ARG_ENABLE(orc,
+  AC_HELP_STRING([--enable-orc],[use Orc if installed]),
+  [case "${enableval}" in
+    auto) enable_orc=auto ;;
+    yes) enable_orc=yes ;;
+    no)  enable_orc=no ;;
+    *) AC_MSG_ERROR(bad value ${enableval} for --enable-orc) ;;
+  esac
+  ],
+  [enable_orc=auto]) dnl Default value
+
+  if test "x$enable_orc" != "xno" ; then
+    PKG_CHECK_MODULES(ORC, orc-0.4 >= $ORC_REQ, [
+      AC_DEFINE(HAVE_ORC, 1, [Use Orc])
+      ORCC=`$PKG_CONFIG --variable=orcc orc-0.4`
+      AC_SUBST(ORCC)
+      HAVE_ORC=yes
+    ], [
+      if test "x$enable_orc" = "xyes" ; then
+        AC_MSG_ERROR([--enable-orc specified, but Orc >= $ORC_REQ not found])
+      fi
+      AC_DEFINE(DISABLE_ORC, 1, [Disable Orc])
+      HAVE_ORC=no
+    ])
+  else
+    AC_DEFINE(DISABLE_ORC, 1, [Disable Orc])
+    HAVE_ORC=no
+  fi
+  AM_CONDITIONAL(HAVE_ORC, test "x$HAVE_ORC" = "xyes")
+
+]))
+
+AC_DEFUN([ORC_OUTPUT],
+[
+  if test "$HAVE_ORC" = yes ; then
+    printf "configure: *** Orc acceleration enabled.\n"
+  else
+    if test "x$enable_orc" = "xno" ; then
+      printf "configure: *** Orc acceleration disabled by --disable-orc.  Slower code paths\n"
+      printf "               will be used.\n"
+    else
+      printf "configure: *** Orc acceleration disabled.  Requires Orc >= $ORC_REQ, which was\n"
+      printf "               not found.  Slower code paths will be used.\n"
+    fi
+  fi
+  printf "\n"
+])

src/Makefile.am

@@ -1708,13 +1708,17 @@ module_suspend_on_idle_la_LIBADD = $(AM_LIBADD) libpulsecore-@PA_MAJORMINORMICRO
 module_suspend_on_idle_la_CFLAGS = $(AM_CFLAGS)
 
 # echo-cancel module
+ORC_SOURCE=modules/echo-cancel/adrian-aec-orc
+include $(top_srcdir)/build/orc.mak
 module_echo_cancel_la_SOURCES = modules/echo-cancel/module-echo-cancel.c modules/echo-cancel/echo-cancel.h \
 				modules/echo-cancel/speex.c \
 				modules/echo-cancel/adrian-aec.c modules/echo-cancel/adrian-aec.h \
-				modules/echo-cancel/adrian.c modules/echo-cancel/adrian.h
+				modules/echo-cancel/adrian.c modules/echo-cancel/adrian.h \
+				$(ORC_SOURCE).orc
+nodist_module_echo_cancel_la_SOURCES = $(ORC_NODIST_SOURCES)
 module_echo_cancel_la_LDFLAGS = $(MODULE_LDFLAGS)
-module_echo_cancel_la_LIBADD = $(AM_LIBADD) libpulsecore-@PA_MAJORMINORMICRO@.la libpulsecommon-@PA_MAJORMINORMICRO@.la libpulse.la $(LIBSPEEX_LIBS)
+module_echo_cancel_la_LIBADD = $(AM_LIBADD) libpulsecore-@PA_MAJORMINORMICRO@.la libpulsecommon-@PA_MAJORMINORMICRO@.la libpulse.la $(LIBSPEEX_LIBS) $(ORC_LIBS)
-module_echo_cancel_la_CFLAGS = $(AM_CFLAGS) $(LIBSPEEX_CFLAGS)
+module_echo_cancel_la_CFLAGS = $(AM_CFLAGS) $(LIBSPEEX_CFLAGS) $(ORC_CFLAGS)
 
 # RTP modules
 module_rtp_send_la_SOURCES = modules/rtp/module-rtp-send.c

src/modules/echo-cancel/adrian-aec-orc-dist.c

@@ -0,0 +1,250 @@
+
+/* autogenerated from adrian-aec-orc.orc */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#ifndef DISABLE_ORC
+#include <orc/orc.h>
+#endif
+
+#ifndef _ORC_INTEGER_TYPEDEFS_
+#define _ORC_INTEGER_TYPEDEFS_
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+#include <stdint.h>
+typedef int8_t orc_int8;
+typedef int16_t orc_int16;
+typedef int32_t orc_int32;
+typedef int64_t orc_int64;
+typedef uint8_t orc_uint8;
+typedef uint16_t orc_uint16;
+typedef uint32_t orc_uint32;
+typedef uint64_t orc_uint64;
+#elif defined(_MSC_VER)
+typedef signed __int8 orc_int8;
+typedef signed __int16 orc_int16;
+typedef signed __int32 orc_int32;
+typedef signed __int64 orc_int64;
+typedef unsigned __int8 orc_uint8;
+typedef unsigned __int16 orc_uint16;
+typedef unsigned __int32 orc_uint32;
+typedef unsigned __int64 orc_uint64;
+#else
+#include <limits.h>
+typedef signed char orc_int8;
+typedef short orc_int16;
+typedef int orc_int32;
+typedef unsigned char orc_uint8;
+typedef unsigned short orc_uint16;
+typedef unsigned int orc_uint32;
+#if INT_MAX == LONG_MAX
+typedef long long orc_int64;
+typedef unsigned long long orc_uint64;
+#else
+typedef long orc_int64;
+typedef unsigned long orc_uint64;
+#endif
+#endif
+typedef union { orc_int16 i; orc_int8 x2[2]; } orc_union16;
+typedef union { orc_int32 i; float f; orc_int16 x2[2]; orc_int8 x4[4]; } orc_union32;
+typedef union { orc_int64 i; double f; orc_int32 x2[2]; orc_int16 x4[4]; } orc_union64;
+#endif
+
+void update_tap_weights (float * d1, const float * s1, float p1, int n);
+
+
+/* begin Orc C target preamble */
+#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x)))
+#define ORC_ABS(a) ((a)<0 ? -(a) : (a))
+#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b))
+#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b))
+#define ORC_SB_MAX 127
+#define ORC_SB_MIN (-1-ORC_SB_MAX)
+#define ORC_UB_MAX 255
+#define ORC_UB_MIN 0
+#define ORC_SW_MAX 32767
+#define ORC_SW_MIN (-1-ORC_SW_MAX)
+#define ORC_UW_MAX 65535
+#define ORC_UW_MIN 0
+#define ORC_SL_MAX 2147483647
+#define ORC_SL_MIN (-1-ORC_SL_MAX)
+#define ORC_UL_MAX 4294967295U
+#define ORC_UL_MIN 0
+#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX)
+#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX)
+#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX)
+#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX)
+#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX)
+#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
+#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
+#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
+#define ORC_SWAP_Q(x) ((((x)&0xffULL)<<56) | (((x)&0xff00ULL)<<40) | (((x)&0xff0000ULL)<<24) | (((x)&0xff000000ULL)<<8) | (((x)&0xff00000000ULL)>>8) | (((x)&0xff0000000000ULL)>>24) | (((x)&0xff000000000000ULL)>>40) | (((x)&0xff00000000000000ULL)>>56))
+#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
+#define ORC_DENORMAL(x) ((x) & ((((x)&0x7f800000) == 0) ? 0xff800000 : 0xffffffff))
+#define ORC_ISNAN(x) ((((x)&0x7f800000) == 0x7f800000) && (((x)&0x007fffff) != 0))
+#define ORC_DENORMAL_DOUBLE(x) ((x) & ((((x)&0x7ff0000000000000ULL) == 0) ? 0xfff0000000000000ULL : 0xffffffffffffffffULL))
+#define ORC_ISNAN_DOUBLE(x) ((((x)&0x7ff0000000000000ULL) == 0x7ff0000000000000ULL) && (((x)&0x000fffffffffffffULL) != 0))
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+#define ORC_RESTRICT restrict
+#elif defined(__GNUC__) && __GNUC__ >= 4
+#define ORC_RESTRICT __restrict__
+#else
+#define ORC_RESTRICT
+#endif
+/* end Orc C target preamble */
+
+
+
+/* update_tap_weights */
+#ifdef DISABLE_ORC
+void
+update_tap_weights (float * d1, const float * s1, float p1, int n){
+  int i;
+  orc_union32 * ORC_RESTRICT ptr0;
+  const orc_union32 * ORC_RESTRICT ptr4;
+  orc_union32 var33;
+  orc_union32 var34;
+  orc_union32 var35;
+  orc_union32 var36;
+  orc_union32 var37;
+
+  ptr0 = (orc_union32 *)d1;
+  ptr4 = (orc_union32 *)s1;
+
+    /* 0: loadpl */
+    var33.i = p1;
+
+  for (i = 0; i < n; i++) {
+    /* 1: loadl */
+    var34 = ptr4[i];
+    /* 2: mulf */
+    {
+       orc_union32 _src1;
+       orc_union32 _src2;
+       orc_union32 _dest1;
+       _src1.i = ORC_DENORMAL(var33.i);
+       _src2.i = ORC_DENORMAL(var34.i);
+       _dest1.f = _src1.f * _src2.f;
+       var37.i = ORC_DENORMAL(_dest1.i);
+    }
+    /* 3: loadl */
+    var35 = ptr0[i];
+    /* 4: addf */
+    {
+       orc_union32 _src1;
+       orc_union32 _src2;
+       orc_union32 _dest1;
+       _src1.i = ORC_DENORMAL(var35.i);
+       _src2.i = ORC_DENORMAL(var37.i);
+       _dest1.f = _src1.f + _src2.f;
+       var36.i = ORC_DENORMAL(_dest1.i);
+    }
+    /* 5: storel */
+    ptr0[i] = var36;
+  }
+
+}
+
+#else
+static void
+_backup_update_tap_weights (OrcExecutor * ORC_RESTRICT ex)
+{
+  int i;
+  int n = ex->n;
+  orc_union32 * ORC_RESTRICT ptr0;
+  const orc_union32 * ORC_RESTRICT ptr4;
+  orc_union32 var33;
+  orc_union32 var34;
+  orc_union32 var35;
+  orc_union32 var36;
+  orc_union32 var37;
+
+  ptr0 = (orc_union32 *)ex->arrays[0];
+  ptr4 = (orc_union32 *)ex->arrays[4];
+
+    /* 0: loadpl */
+    var33.i = ex->params[24];
+
+  for (i = 0; i < n; i++) {
+    /* 1: loadl */
+    var34 = ptr4[i];
+    /* 2: mulf */
+    {
+       orc_union32 _src1;
+       orc_union32 _src2;
+       orc_union32 _dest1;
+       _src1.i = ORC_DENORMAL(var33.i);
+       _src2.i = ORC_DENORMAL(var34.i);
+       _dest1.f = _src1.f * _src2.f;
+       var37.i = ORC_DENORMAL(_dest1.i);
+    }
+    /* 3: loadl */
+    var35 = ptr0[i];
+    /* 4: addf */
+    {
+       orc_union32 _src1;
+       orc_union32 _src2;
+       orc_union32 _dest1;
+       _src1.i = ORC_DENORMAL(var35.i);
+       _src2.i = ORC_DENORMAL(var37.i);
+       _dest1.f = _src1.f + _src2.f;
+       var36.i = ORC_DENORMAL(_dest1.i);
+    }
+    /* 5: storel */
+    ptr0[i] = var36;
+  }
+
+}
+
+void
+update_tap_weights (float * d1, const float * s1, float p1, int n)
+{
+  OrcExecutor _ex, *ex = &_ex;
+  static int p_inited = 0;
+  static OrcProgram *p = 0;
+  void (*func) (OrcExecutor *);
+
+  if (!p_inited) {
+    orc_once_mutex_lock ();
+    if (!p_inited) {
+      OrcCompileResult result;
+
+      p = orc_program_new ();
+      orc_program_set_name (p, "update_tap_weights");
+      orc_program_set_backup_function (p, _backup_update_tap_weights);
+      orc_program_add_destination (p, 4, "d1");
+      orc_program_add_source (p, 4, "s1");
+      orc_program_add_constant (p, 0, 0x00000000, "c1");
+      orc_program_add_constant (p, 0, 0x00000000, "c2");
+      orc_program_add_constant (p, 0, 0x00000000, "c3");
+      orc_program_add_constant (p, 0, 0x00000000, "c4");
+      orc_program_add_constant (p, 0, 0x00000000, "c5");
+      orc_program_add_constant (p, 0, 0x00000000, "c6");
+      orc_program_add_constant (p, 0, 0x00000000, "c7");
+      orc_program_add_constant (p, 0, 0x00000000, "c8");
+      orc_program_add_parameter_float (p, 4, "p1");
+      orc_program_add_temporary (p, 4, "t1");
+
+      orc_program_append_2 (p, "mulf", 0, ORC_VAR_T1, ORC_VAR_P1, ORC_VAR_S1, ORC_VAR_D1);
+      orc_program_append_2 (p, "addf", 0, ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1);
+
+      result = orc_program_compile (p);
+    }
+    p_inited = TRUE;
+    orc_once_mutex_unlock ();
+  }
+  ex->program = p;
+
+  ex->n = n;
+  ex->arrays[ORC_VAR_D1] = d1;
+  ex->arrays[ORC_VAR_S1] = (void *)s1;
+  {
+    orc_union32 tmp;
+    tmp.f = p1;
+    ex->params[ORC_VAR_P1] = tmp.i;
+  }
+
+  func = p->code_exec;
+  func (ex);
+}
+#endif

src/modules/echo-cancel/adrian-aec-orc-dist.h

@@ -0,0 +1,61 @@
+
+/* autogenerated from adrian-aec-orc.orc */
+
+#ifndef ___MODULES_ECHO_CANCEL_ADRIAN_AEC_ORC_H_
+#define ___MODULES_ECHO_CANCEL_ADRIAN_AEC_ORC_H_
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#ifndef _ORC_INTEGER_TYPEDEFS_
+#define _ORC_INTEGER_TYPEDEFS_
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+#include <stdint.h>
+typedef int8_t orc_int8;
+typedef int16_t orc_int16;
+typedef int32_t orc_int32;
+typedef int64_t orc_int64;
+typedef uint8_t orc_uint8;
+typedef uint16_t orc_uint16;
+typedef uint32_t orc_uint32;
+typedef uint64_t orc_uint64;
+#elif defined(_MSC_VER)
+typedef signed __int8 orc_int8;
+typedef signed __int16 orc_int16;
+typedef signed __int32 orc_int32;
+typedef signed __int64 orc_int64;
+typedef unsigned __int8 orc_uint8;
+typedef unsigned __int16 orc_uint16;
+typedef unsigned __int32 orc_uint32;
+typedef unsigned __int64 orc_uint64;
+#else
+#include <limits.h>
+typedef signed char orc_int8;
+typedef short orc_int16;
+typedef int orc_int32;
+typedef unsigned char orc_uint8;
+typedef unsigned short orc_uint16;
+typedef unsigned int orc_uint32;
+#if INT_MAX == LONG_MAX
+typedef long long orc_int64;
+typedef unsigned long long orc_uint64;
+#else
+typedef long orc_int64;
+typedef unsigned long orc_uint64;
+#endif
+#endif
+typedef union { orc_int16 i; orc_int8 x2[2]; } orc_union16;
+typedef union { orc_int32 i; float f; orc_int16 x2[2]; orc_int8 x4[4]; } orc_union32;
+typedef union { orc_int64 i; double f; orc_int32 x2[2]; orc_int16 x4[4]; } orc_union64;
+#endif
+void update_tap_weights (float * d1, const float * s1, float p1, int n);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

src/modules/echo-cancel/adrian-aec-orc.orc

@@ -0,0 +1,8 @@
+.function update_tap_weights
+.dest 4 w float
+.source 4 xf float
+.floatparam 4 mikro_ef
+.temp 4 tmp float
+
+mulf tmp, mikro_ef, xf
+addf w, w, tmp

src/modules/echo-cancel/adrian-aec.c

@@ -17,6 +17,10 @@
 
 #include "adrian-aec.h"
 
+#ifndef DISABLE_ORC
+#include "adrian-aec-orc.h"
+#endif
+
 #ifdef __SSE__
 #include <xmmintrin.h>
 #endif
@@ -190,6 +194,7 @@ static REAL AEC_nlms_pw(AEC *a, REAL d, REAL x_, float stepsize)
     // calculate variable step size
     REAL mikro_ef = stepsize * ef / a->dotp_xf_xf;
 
+#ifdef DISABLE_ORC
     // update tap weights (filter learning)
     int i;
     for (i = 0; i < NLMS_LEN; i += 2) {
@@ -197,6 +202,9 @@ static REAL AEC_nlms_pw(AEC *a, REAL d, REAL x_, float stepsize)
       a->w[i] += mikro_ef * a->xf[i + a->j];
       a->w[i + 1] += mikro_ef * a->xf[i + a->j + 1];
     }
+#else
+    update_tap_weights(a->w, &a->xf[a->j], mikro_ef, NLMS_LEN);
+#endif
   }
 
   if (--(a->j) < 0) {

src/modules/echo-cancel/adrian-aec.h

@@ -23,7 +23,7 @@
 #define WIDEB 2
 
 // use double if your CPU does software-emulation of float
-typedef float REAL;
+#define REAL float
 
 /* dB Values */
 #define M0dB 1.0f
@@ -306,7 +306,7 @@ struct AEC {
   // NLMS-pw
   REAL x[NLMS_LEN + NLMS_EXT];  // tap delayed loudspeaker signal
   REAL xf[NLMS_LEN + NLMS_EXT]; // pre-whitening tap delayed signal
-  REAL w[NLMS_LEN];             // tap weights
+  PA_DECLARE_ALIGNED(16, REAL, w[NLMS_LEN]);             // tap weights
   int j;                        // optimize: less memory copies
   double dotp_xf_xf;            // double to avoid loss of precision
   float delta;                  // noise floor to stabilize NLMS