armv7: celt_pitch_xcorr: Introduce ARM NEON intrinsics
authorViswanath Puttagunta <viswanath.puttagunta@linaro.org>
Fri, 19 Dec 2014 23:25:55 +0000 (17:25 -0600)
committerTimothy B. Terriberry <tterribe@xiph.org>
Thu, 25 Dec 2014 22:36:46 +0000 (14:36 -0800)
Optimize celt_pitch_xcorr function (for floating point)
using ARM NEON intrinsics for SoCs that have NEON VFP unit.

To enable this optimization, use --enable-intrinsics
configure option.

Compile time and runtime checks are also supported to make sure
this optimization is only enabled when the compiler supports
NEON intrinsics.

Signed-off-by: Timothy B. Terriberry <tterribe@xiph.org>
Makefile.am
celt/arm/arm_celt_map.c
celt/arm/celt_neon_intr.c [new file with mode: 0644]
celt/arm/pitch_arm.h
celt/cpu_support.h
celt/pitch.h
celt/tests/test_unit_mathops.c
celt/tests/test_unit_rotation.c
celt_sources.mk
configure.ac

index e20f7b4..95323ca 100644 (file)
@@ -41,6 +41,12 @@ endif
 if CPU_ARM
 CELT_SOURCES += $(CELT_SOURCES_ARM)
 SILK_SOURCES += $(SILK_SOURCES_ARM)
+
+if OPUS_ARM_NEON_INTR
+CELT_SOURCES += $(CELT_SOURCES_ARM_NEON_INTR)
+OPUS_ARM_NEON_INTR_CPPFLAGS = -mfpu=neon
+endif
+
 if OPUS_ARM_EXTERNAL_ASM
 nodist_libopus_la_SOURCES = $(CELT_SOURCES_ARM_ASM:.s=-gnu.S)
 BUILT_SOURCES = $(CELT_SOURCES_ARM_ASM:.s=-gnu.S) \
@@ -260,3 +266,9 @@ if HAVE_SSE2
 $(SSE_OBJ): CFLAGS += -msse2
 endif
 endif
+
+if OPUS_ARM_NEON_INTR
+CELT_ARM_NEON_INTR_OBJ = $(CELT_SOURCES_ARM_NEON_INTR:.c=.lo) \
+                       %test_unit_rotation.o %test_unit_mathops.o
+$(CELT_ARM_NEON_INTR_OBJ): CFLAGS += $(OPUS_ARM_NEON_INTR_CPPFLAGS)
+endif
index 547a84d..68c224d 100644 (file)
@@ -41,9 +41,16 @@ opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
   MAY_HAVE_MEDIA(celt_pitch_xcorr), /* Media */
   MAY_HAVE_NEON(celt_pitch_xcorr)   /* NEON */
 };
-# else
-#  error "Floating-point implementation is not supported by ARM asm yet." \
- "Reconfigure with --disable-rtcd or send patches."
+# else /* !FIXED_POINT */
+#  if defined(OPUS_ARM_NEON_INTR)
+void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *,
+    const opus_val16 *, opus_val32 *, int, int) = {
+  celt_pitch_xcorr_c,              /* ARMv4 */
+  celt_pitch_xcorr_c,              /* EDSP */
+  celt_pitch_xcorr_c,              /* Media */
+  celt_pitch_xcorr_float_neon      /* Neon */
+};
+#  endif
 # endif
 
 #endif
diff --git a/celt/arm/celt_neon_intr.c b/celt/arm/celt_neon_intr.c
new file mode 100644 (file)
index 0000000..4a67413
--- /dev/null
@@ -0,0 +1,245 @@
+/* Copyright (c) 2014-2015 Xiph.Org Foundation
+   Written by Viswanath Puttagunta */
+/**
+   @file celt_neon_intr.c
+   @brief ARM Neon Intrinsic optimizations for celt
+ */
+
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+#include <arm_neon.h>
+#include "../pitch.h"
+
+/*
+ * Function: xcorr_kernel_neon_float
+ * ---------------------------------
+ * Computes 4 correlation values and stores them in sum[4]
+ */
+static void xcorr_kernel_neon_float(const float32_t *x, const float32_t *y,
+      float32_t sum[4], int len) {
+   float32x4_t YY[3];
+   float32x4_t YEXT[3];
+   float32x4_t XX[2];
+   float32x2_t XX_2;
+   float32x4_t SUMM;
+   const float32_t *xi = x;
+   const float32_t *yi = y;
+
+   celt_assert(len>0);
+
+   YY[0] = vld1q_f32(yi);
+   SUMM = vdupq_n_f32(0);
+
+   /* Consume 8 elements in x vector and 12 elements in y
+    * vector. However, the 12'th element never really gets
+    * touched in this loop. So, if len == 8, then we only
+    * must access y[0] to y[10]. y[11] must not be accessed
+    * hence make sure len > 8 and not len >= 8
+    */
+   while (len > 8) {
+      yi += 4;
+      YY[1] = vld1q_f32(yi);
+      yi += 4;
+      YY[2] = vld1q_f32(yi);
+
+      XX[0] = vld1q_f32(xi);
+      xi += 4;
+      XX[1] = vld1q_f32(xi);
+      xi += 4;
+
+      SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0);
+      YEXT[0] = vextq_f32(YY[0], YY[1], 1);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1);
+      YEXT[1] = vextq_f32(YY[0], YY[1], 2);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0);
+      YEXT[2] = vextq_f32(YY[0], YY[1], 3);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1);
+
+      SUMM = vmlaq_lane_f32(SUMM, YY[1], vget_low_f32(XX[1]), 0);
+      YEXT[0] = vextq_f32(YY[1], YY[2], 1);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[1]), 1);
+      YEXT[1] = vextq_f32(YY[1], YY[2], 2);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[1]), 0);
+      YEXT[2] = vextq_f32(YY[1], YY[2], 3);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[1]), 1);
+
+      YY[0] = YY[2];
+      len -= 8;
+   }
+
+   /* Consume 4 elements in x vector and 8 elements in y
+    * vector. However, the 8'th element in y never really gets
+    * touched in this loop. So, if len == 4, then we only
+    * must access y[0] to y[6]. y[7] must not be accessed
+    * hence make sure len>4 and not len>=4
+    */
+   if (len > 4) {
+      yi += 4;
+      YY[1] = vld1q_f32(yi);
+
+      XX[0] = vld1q_f32(xi);
+      xi += 4;
+
+      SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0);
+      YEXT[0] = vextq_f32(YY[0], YY[1], 1);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1);
+      YEXT[1] = vextq_f32(YY[0], YY[1], 2);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0);
+      YEXT[2] = vextq_f32(YY[0], YY[1], 3);
+      SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1);
+
+      YY[0] = YY[1];
+      len -= 4;
+   }
+
+   while (--len > 0) {
+      XX_2 = vld1_dup_f32(xi++);
+      SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0);
+      YY[0]= vld1q_f32(++yi);
+   }
+
+   XX_2 = vld1_dup_f32(xi);
+   SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0);
+
+   vst1q_f32(sum, SUMM);
+}
+
+/*
+ * Function: xcorr_kernel_neon_float_process1
+ * ---------------------------------
+ * Computes single correlation values and stores in *sum
+ */
+static void xcorr_kernel_neon_float_process1(const float32_t *x,
+      const float32_t *y, float32_t *sum, int len) {
+   float32x4_t XX[4];
+   float32x4_t YY[4];
+   float32x2_t XX_2;
+   float32x2_t YY_2;
+   float32x4_t SUMM;
+   float32x2_t SUMM_2[2];
+   const float32_t *xi = x;
+   const float32_t *yi = y;
+
+   SUMM = vdupq_n_f32(0);
+
+   /* Work on 16 values per iteration */
+   while (len >= 16) {
+      XX[0] = vld1q_f32(xi);
+      xi += 4;
+      XX[1] = vld1q_f32(xi);
+      xi += 4;
+      XX[2] = vld1q_f32(xi);
+      xi += 4;
+      XX[3] = vld1q_f32(xi);
+      xi += 4;
+
+      YY[0] = vld1q_f32(yi);
+      yi += 4;
+      YY[1] = vld1q_f32(yi);
+      yi += 4;
+      YY[2] = vld1q_f32(yi);
+      yi += 4;
+      YY[3] = vld1q_f32(yi);
+      yi += 4;
+
+      SUMM = vmlaq_f32(SUMM, YY[0], XX[0]);
+      SUMM = vmlaq_f32(SUMM, YY[1], XX[1]);
+      SUMM = vmlaq_f32(SUMM, YY[2], XX[2]);
+      SUMM = vmlaq_f32(SUMM, YY[3], XX[3]);
+      len -= 16;
+   }
+
+   /* Work on 8 values */
+   if (len >= 8) {
+      XX[0] = vld1q_f32(xi);
+      xi += 4;
+      XX[1] = vld1q_f32(xi);
+      xi += 4;
+
+      YY[0] = vld1q_f32(yi);
+      yi += 4;
+      YY[1] = vld1q_f32(yi);
+      yi += 4;
+
+      SUMM = vmlaq_f32(SUMM, YY[0], XX[0]);
+      SUMM = vmlaq_f32(SUMM, YY[1], XX[1]);
+      len -= 8;
+   }
+
+   /* Work on 4 values */
+   if (len >= 4) {
+      XX[0] = vld1q_f32(xi);
+      xi += 4;
+      YY[0] = vld1q_f32(yi);
+      yi += 4;
+      SUMM = vmlaq_f32(SUMM, YY[0], XX[0]);
+      len -= 4;
+   }
+
+   /* Start accumulating results */
+   SUMM_2[0] = vget_low_f32(SUMM);
+   if (len >= 2) {
+      /* While at it, consume 2 more values if available */
+      XX_2 = vld1_f32(xi);
+      xi += 2;
+      YY_2 = vld1_f32(yi);
+      yi += 2;
+      SUMM_2[0] = vmla_f32(SUMM_2[0], YY_2, XX_2);
+      len -= 2;
+   }
+   SUMM_2[1] = vget_high_f32(SUMM);
+   SUMM_2[0] = vadd_f32(SUMM_2[0], SUMM_2[1]);
+   SUMM_2[0] = vpadd_f32(SUMM_2[0], SUMM_2[0]);
+   /* Ok, now we have result accumulated in SUMM_2[0].0 */
+
+   if (len > 0) {
+      /* Case when you have one value left */
+      XX_2 = vld1_dup_f32(xi);
+      YY_2 = vld1_dup_f32(yi);
+      SUMM_2[0] = vmla_f32(SUMM_2[0], XX_2, YY_2);
+   }
+
+   vst1_lane_f32(sum, SUMM_2[0], 0);
+}
+
+void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y,
+                        opus_val32 *xcorr, int len, int max_pitch) {
+   int i;
+   celt_assert(max_pitch > 0);
+   celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
+
+   for (i = 0; i < (max_pitch-3); i += 4) {
+      xcorr_kernel_neon_float((const float32_t *)_x, (const float32_t *)_y+i,
+            (float32_t *)xcorr+i, len);
+   }
+
+   /* In case max_pitch isn't multiple of 4
+    * compute single correlation value per iteration
+    */
+   for (; i < max_pitch; i++) {
+      xcorr_kernel_neon_float_process1((const float32_t *)_x,
+            (const float32_t *)_y+i, (float32_t *)xcorr+i, len);
+   }
+}
index a07f8ac..125d1bc 100644 (file)
@@ -52,6 +52,17 @@ opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y,
   ((void)(arch),PRESUME_NEON(celt_pitch_xcorr)(_x, _y, xcorr, len, max_pitch))
 #  endif
 
-# endif
+#else /* Start !FIXED_POINT */
+/* Float case */
+#if defined(OPUS_ARM_NEON_INTR)
+void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y,
+                                 opus_val32 *xcorr, int len, int max_pitch);
+#if !defined(OPUS_HAVE_RTCD)
+#define OVERRIDE_PITCH_XCORR (1)
+#   define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \
+   ((void)(arch),celt_pitch_xcorr_float_neon(_x, _y, xcorr, len, max_pitch))
+#endif
+#endif
 
+#endif /* end !FIXED_POINT */
 #endif
index 71efff1..1d62e2f 100644 (file)
@@ -31,7 +31,8 @@
 #include "opus_types.h"
 #include "opus_defines.h"
 
-#if defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_ASM)
+#if defined(OPUS_HAVE_RTCD) && \
+  (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_NEON_INTR))
 #include "arm/armcpu.h"
 
 /* We currently support 4 ARM variants:
index 5c6e551..4368cc5 100644 (file)
@@ -46,7 +46,8 @@
 #include "mips/pitch_mipsr1.h"
 #endif
 
-#if defined(OPUS_ARM_ASM) && defined(FIXED_POINT)
+#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \
+  || defined(OPUS_ARM_NEON_INTR))
 # include "arm/pitch_arm.h"
 #endif
 
@@ -178,7 +179,8 @@ celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
 
 #if !defined(OVERRIDE_PITCH_XCORR)
 /*Is run-time CPU detection enabled on this platform?*/
-# if defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_ASM)
+# if defined(OPUS_HAVE_RTCD) && \
+  (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_NEON_INTR))
 extern
 #  if defined(FIXED_POINT)
 opus_val32
index 3076bbf..b9b1bcf 100644 (file)
 #include "x86/celt_lpc_sse.c"
 #endif
 #include "x86/x86_celt_map.c"
-#elif defined(OPUS_ARM_ASM) && defined(FIXED_POINT)
+#elif ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \
+       || defined(OPUS_ARM_NEON_INTR))
+#if defined(OPUS_ARM_NEON_INTR)
+#include "arm/celt_neon_intr.c"
+#endif
 #include "arm/arm_celt_map.c"
 #endif
 
index 37ba74e..5507884 100644 (file)
 #include "x86/celt_lpc_sse.c"
 #endif
 #include "x86/x86_celt_map.c"
-#elif defined(OPUS_ARM_ASM) && defined(FIXED_POINT)
+#elif ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \
+       || defined(OPUS_ARM_NEON_INTR))
+#if defined(OPUS_ARM_NEON_INTR)
+#include "arm/celt_neon_intr.c"
+#endif
 #include "arm/arm_celt_map.c"
 #endif
 
index 20b1b1b..29ec937 100644 (file)
@@ -32,3 +32,6 @@ celt/arm/celt_pitch_xcorr_arm.s
 
 CELT_AM_SOURCES_ARM_ASM = \
 celt/arm/armopts.s.in
+
+CELT_SOURCES_ARM_NEON_INTR = \
+celt/arm/celt_neon_intr.c
index 95f20b2..87cece9 100644 (file)
@@ -190,14 +190,14 @@ AC_ARG_ENABLE([rtcd],
     [enable_rtcd=yes])
 
 AC_ARG_ENABLE([intrinsics],
-    [AS_HELP_STRING([--enable-intrinsics], [Enable intrinsics optimizations (only for fixed point x86)])],,
+    [AS_HELP_STRING([--enable-intrinsics], [Enable intrinsics optimizations for ARM(float) X86(fixed)])],,
     [enable_intrinsics=no])
 
 rtcd_support=no
 cpu_arm=no
 
 AS_IF([test x"${enable_asm}" = x"yes"],[
-    inline_optimization="No ASM for your platform, please send patches"
+    inline_optimization="No inline ASM for your platform, please send patches"
     case $host_cpu in
       arm*)
         dnl Currently we only have asm for fixed-point
@@ -343,7 +343,6 @@ AS_IF([test x"${enable_asm}" = x"yes"],[
    asm_optimization="disabled"
 ])
 
-AM_CONDITIONAL([CPU_ARM], [test "$cpu_arm" = "yes"])
 AM_CONDITIONAL([OPUS_ARM_INLINE_ASM],
     [test x"${inline_optimization%% *}" = x"ARM"])
 AM_CONDITIONAL([OPUS_ARM_EXTERNAL_ASM],
@@ -351,9 +350,59 @@ AM_CONDITIONAL([OPUS_ARM_EXTERNAL_ASM],
 
 AM_CONDITIONAL([HAVE_SSE4_1], [false])
 AM_CONDITIONAL([HAVE_SSE2], [false])
+
 AS_IF([test x"$enable_intrinsics" = x"yes"],[
-AS_IF([test x"$enable_float" = x"no"],
-[AS_IF([test x"$host_cpu" = x"i386" -o x"$host_cpu" = x"i686" -o x"$host_cpu" = x"x86_64"],[
+   case $host_cpu in
+   arm*)
+      cpu_arm=yes
+      AC_MSG_CHECKING(if compiler supports ARM NEON intrinsics)
+      save_CFLAGS="$CFLAGS"; CFLAGS="-mfpu=neon $CFLAGS"
+      AC_LINK_IFELSE(
+         [
+            AC_LANG_PROGRAM(
+               [[#include <arm_neon.h>
+               ]],
+               [[
+                  static float32x4_t A[2], SUMM;
+                  SUMM = vmlaq_f32(SUMM, A[0], A[1]);
+               ]]
+            )
+         ],[
+            OPUS_ARM_NEON_INTR=1
+            AC_MSG_RESULT([yes])
+         ],[
+            OPUS_ARM_NEON_INTR=0
+            AC_MSG_RESULT([no])
+         ]
+      )
+      CFLAGS="$save_CFLAGS"
+      #Now we know if compiler supports ARM neon intrinsics or not
+
+      #Currently we only have intrinsic optimization for floating point
+      AS_IF([test x"$enable_float" = x"yes"],
+      [
+         AS_IF([test x"$OPUS_ARM_NEON_INTR" = x"1"],
+         [
+            AC_DEFINE([OPUS_ARM_NEON_INTR], 1, [Compiler supports ARMv7 Neon Intrinsics])
+            AS_IF([test x"enable_rtcd" != x""],
+               [rtcd_support="ARM (ARMv7_Neon_Intrinsics)"],[])
+            enable_intrinsics="$enable_intrinsics ARMv7_Neon_Intrinsics"
+            dnl Don't see why defining these is necessary to check features at runtime
+            AC_DEFINE([OPUS_ARM_MAY_HAVE_EDSP], 1, [Define if compiler support EDSP Instructions])
+            AC_DEFINE([OPUS_ARM_MAY_HAVE_MEDIA], 1, [Define if compiler support MEDIA Instructions])
+            AC_DEFINE([OPUS_ARM_MAY_HAVE_NEON], 1, [Define if compiler support NEON instructions])
+         ],
+         [
+            AC_MSG_WARN([Compiler does not support ARM intrinsics])
+            enable_intrinsics=no
+         ])
+      ], [
+            AC_MSG_WARN([Currently on have ARM intrinsics for float])
+            enable_intrinsics=no
+      ])
+   ;;
+   "i386" | "i686" | "x86_64")
+    AS_IF([test x"$enable_float" = x"no"],[
     AS_IF([test x"$enable_rtcd" = x"yes"],[
             get_cpuid_by_asm="no"
             AC_MSG_CHECKING([Get CPU Info])
@@ -423,7 +472,7 @@ AS_IF([test x"$enable_float" = x"no"],
             AM_CONDITIONAL([HAVE_SSE2], [true])
             AS_IF([test x"$get_cpuid_by_asm" = x"yes"],[AC_DEFINE([CPU_INFO_BY_ASM], [1], [Get CPU Info by asm method])],
             [AC_DEFINE([CPU_INFO_BY_C], [1], [Get CPU Info by C method])])
-             ],[
+             ],[ ##### Else case for AS_IF([test x"$?" = x"0"])
                gcc -Q --help=target | grep "\-msse2 "
                AC_MSG_CHECKING([sse2])
                AS_IF([test x"$?" = x"0"],[
@@ -446,13 +495,28 @@ AS_IF([test x"$enable_float" = x"no"],
                   AM_CONDITIONAL([HAVE_SSE2], [true])
                   AS_IF([test x"$get_cpuid_by_asm" = x"yes"],[AC_DEFINE([CPU_INFO_BY_ASM], [1], [Get CPU Info by asm method])],
                   [AC_DEFINE([CPU_INFO_BY_C], [1], [Get CPU Info by c method])])
-            ],[enable_intrinsics="no"])
+            ],[enable_intrinsics="no"]) #End of AS_IF([test x"$?" = x"0"]
         ])
-    ], [enable_intrinsics="no"])
-])
-], [enable_intrinsics="no"])
+    ], [
+        enable_intrinsics="no"
+    ]) ## End of AS_IF([test x"$enable_rtcd" = x"yes"]
+],
+[  ## Else case for AS_IF([test x"$enable_float" = x"no"]
+   AC_MSG_WARN([Disabling intrinsics .. x86 intrinsics only avail for fixed point])
+   enable_intrinsics="no"
+]) ## End of AS_IF([test x"$enable_float" = x"no"]
+   ;;
+   *)
+      AC_MSG_WARN([No intrinsics support for your architecture])
+      enable_intrinsics="no"
+   ;;
+   esac
 ])
 
+AM_CONDITIONAL([CPU_ARM], [test "$cpu_arm" = "yes"])
+AM_CONDITIONAL([OPUS_ARM_NEON_INTR],
+    [test x"$OPUS_ARM_NEON_INTR" = x"1"])
+
 AS_IF([test x"$enable_rtcd" = x"yes"],[
     AS_IF([test x"$rtcd_support" != x"no"],[
         AC_DEFINE([OPUS_HAVE_RTCD], [1],