core: Add ARM NEON optimized sample conversion code

final:
* includes some minor style fixes and build-time changes to allow
  building a single binary for neon and non-neon systems
v4:
* fix for sample length < 4
v3:
* convert from intrinsics to inline assembly
v2:
* load and store data with vld1/vld1q and vst1/vst1q, resp., to work
  around alignment issues of compiler-generated vldmia instruction
* remove redundant check for NEON flags

Ubuntu/Linaro gcc 4.6.3
arm-linux-gnueabi-gcc -O2 -mcpu=cortex-a8 -mfloat-abi=softfp -mfpu=neon

runtime on beagle-xm:

D: [pulseaudio] sconv_neon.c: checking NEON sconv_s16le_from_float
I: [pulseaudio] sconv_neon.c: NEON: 3754 usec.
I: [pulseaudio] sconv_neon.c: ref: 58594 usec.
D: [pulseaudio] sconv_neon.c: checking NEON sconv_s16le_to_float
I: [pulseaudio] sconv_neon.c: NEON: 1831 usec.
I: [pulseaudio] sconv_neon.c: ref: 10528 usec.
I: [pulseaudio] sconv_neon.c: Initialising ARM NEON optimized conversions.

conversion may be off by one for some samples due to rounding issues

configure.ac

@@ -317,6 +317,34 @@ case $host in
   ;;
 esac
 
+#### NEON optimisations ####
+AC_ARG_ENABLE([neon-opt],
+    AS_HELP_STRING([--enable-neon-opt], [Enable NEON optimisations on ARM CPUs that support it]))
+
+AS_IF([test "x$enable_neon_opt" != "xno"],
+    [save_CFLAGS="$CFLAGS"; CFLAGS="$CFLAGS -mfpu=neon"
+     AC_COMPILE_IFELSE(
+        AC_LANG_PROGRAM([], []),
+        [
+         HAVE_NEON=1
+         NEON_CFLAGS="-mfpu=neon"
+        ],
+        [
+         HAVE_NEON=0
+         NEON_CFLAGS=
+        ])
+     CFLAGS="$save_CFLAGS"
+    ],
+    [HAVE_NEON=0])
+
+AS_IF([test "x$enable_neon_opt" = "xyes" && test "x$HAVE_NEON" = "x0"],
+      [AC_MSG_ERROR([*** Compiler does not support -mfpu=neon])])
+
+AC_SUBST(HAVE_NEON)
+AC_SUBST(NEON_CFLAGS)
+AM_CONDITIONAL([HAVE_NEON], [test "x$HAVE_NEON" = x1])
+AS_IF([test "x$HAVE_NEON" = "x1"], AC_DEFINE([HAVE_NEON], 1, [Have NEON support?]))
+
 
 #### libtool stuff ####
 

src/Makefile.am

@@ -813,6 +813,7 @@ libpulsedsp_la_LDFLAGS = $(AM_LDFLAGS) -avoid-version -disable-static
 ###################################
 
 lib_LTLIBRARIES += libpulsecore-@PA_MAJORMINOR@.la
+noinst_LTLIBRARIES =
 
 # Pure core stuff
 libpulsecore_@PA_MAJORMINOR@_la_SOURCES = \
@@ -870,6 +871,13 @@ libpulsecore_@PA_MAJORMINOR@_la_CFLAGS = $(AM_CFLAGS) $(SERVER_CFLAGS) $(LIBSAMP
 libpulsecore_@PA_MAJORMINOR@_la_LDFLAGS = $(AM_LDFLAGS) -avoid-version
 libpulsecore_@PA_MAJORMINOR@_la_LIBADD = $(AM_LIBADD) $(LIBLTDL) $(LIBSAMPLERATE_LIBS) $(LIBSPEEX_LIBS) $(LIBSNDFILE_LIBS) $(WINSOCK_LIBS) $(LTLIBICONV) libpulsecommon-@PA_MAJORMINOR@.la libpulse.la libpulsecore-foreign.la
 
+if HAVE_NEON
+noinst_LTLIBRARIES += libpulsecore_sconv_neon.la
+libpulsecore_sconv_neon_la_SOURCES = pulsecore/sconv_neon.c
+libpulsecore_sconv_neon_la_CFLAGS = $(AM_CFLAGS) $(NEON_CFLAGS)
+libpulsecore_@PA_MAJORMINOR@_la_LIBADD += libpulsecore_sconv_neon.la
+endif
+
 if HAVE_ORC
 ORC_SOURCE += pulsecore/svolume
 libpulsecore_@PA_MAJORMINOR@_la_SOURCES += pulsecore/svolume_orc.c
@@ -909,7 +917,7 @@ libpulsecore_@PA_MAJORMINOR@_la_SOURCES += pulsecore/database-simple.c
 endif
 
 # We split the foreign code off to not be annoyed by warnings we don't care about
-noinst_LTLIBRARIES = libpulsecore-foreign.la
+noinst_LTLIBRARIES += libpulsecore-foreign.la
 
 libpulsecore_foreign_la_SOURCES = \
 		pulsecore/ffmpeg/resample2.c pulsecore/ffmpeg/avcodec.h pulsecore/ffmpeg/dsputil.h

src/pulsecore/cpu-arm.c

@@ -142,6 +142,10 @@ pa_bool_t pa_cpu_init_arm(pa_cpu_arm_flag_t *flags) {
 
     if (*flags & PA_CPU_ARM_V6)
         pa_volume_func_init_arm(*flags);
+#ifdef HAVE_NEON
+    if (*flags & PA_CPU_ARM_NEON)
+        pa_convert_func_init_neon(*flags);
+#endif
 
     return TRUE;
 

src/pulsecore/cpu-arm.h

@@ -26,6 +26,10 @@
 #include <stdint.h>
 #include <pulsecore/macro.h>
 
+#ifndef PACKAGE
+#error "Please include config.h before including this file!"
+#endif
+
 typedef enum pa_cpu_arm_flag {
     PA_CPU_ARM_V6       = (1 << 0),
     PA_CPU_ARM_V7       = (1 << 1),
@@ -41,4 +45,8 @@ pa_bool_t pa_cpu_init_arm(pa_cpu_arm_flag_t *flags);
 /* some optimized functions */
 void pa_volume_func_init_arm(pa_cpu_arm_flag_t flags);
 
+#ifdef HAVE_NEON
+void pa_convert_func_init_neon(pa_cpu_arm_flag_t flags);
+#endif
+
 #endif /* foocpuarmhfoo */

src/pulsecore/sconv_neon.c

@@ -0,0 +1,110 @@
+/***
+  This file is part of PulseAudio.
+
+  Copyright 2012 Peter Meerwald <p.meerwald@bct-electronic.com>
+
+  PulseAudio is free software; you can redistribute it and/or modify
+  it under the terms of the GNU Lesser General Public License as published
+  by the Free Software Foundation; either version 2.1 of the License,
+  or (at your option) any later version.
+
+  PulseAudio is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  General Public License for more details.
+***/
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <pulse/rtclock.h>
+
+#include <pulsecore/macro.h>
+#include <pulsecore/endianmacros.h>
+
+#include "cpu-arm.h"
+#include "sconv.h"
+
+#include <math.h>
+#include <arm_neon.h>
+
+static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t *dst) {
+    unsigned i = n & 3;
+
+    __asm__ __volatile__ (
+        "movs       %[n], %[n], lsr #2      \n\t"
+        "beq        2f                      \n\t"
+
+        "vdup.f32   q2, %[plusone]          \n\t"
+        "vneg.f32   q3, q2                  \n\t"
+        "vdup.f32   q4, %[scale]            \n\t"
+        "vdup.u32   q5, %[mask]             \n\t"
+        "vdup.f32   q6, %[half]             \n\t"
+
+        "1:                                 \n\t"
+        "vld1.32    {q0}, [%[src]]!         \n\t"
+        "vmin.f32   q0, q0, q2              \n\t" /* clamp */
+        "vmax.f32   q0, q0, q3              \n\t"
+        "vmul.f32   q0, q0, q4              \n\t" /* scale */
+        "vand.u32   q1, q0, q5              \n\t"
+        "vorr.u32   q1, q1, q6              \n\t" /* round */
+        "vadd.f32   q0, q0, q1              \n\t"
+        "vcvt.s32.f32 q0, q0                \n\t" /* narrow */
+        "vmovn.i32  d0, q0                  \n\t"
+        "subs       %[n], %[n], #1          \n\t"
+        "vst1.16    {d0}, [%[dst]]!         \n\t"
+        "bgt        1b                      \n\t"
+
+        "2:                                 \n\t"
+
+        : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */
+        : [plusone] "r" (1.0f), [scale] "r" (32767.0f), [half] "r" (0.5f), [mask] "r" (0x80000000) /* input operands */
+        : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q6" /* clobber list */
+    );
+
+    /* leftovers */
+    while (i--) {
+        *dst++ = (int16_t) lrintf(PA_CLAMP_UNLIKELY(*src, -1.0f, 1.0f) * 0x7FFF);
+        src++;
+    }
+}
+
+static void pa_sconv_s16le_to_f32ne_neon(unsigned n, const int16_t *src, float *dst) {
+    unsigned i = n & 3;
+
+    const float invscale = 1.0f / 0x7FFF;
+
+    __asm__ __volatile__ (
+        "movs        %[n], %[n], lsr #2     \n\t"
+        "beq        2f                      \n\t"
+
+        "vdup.f32   q1, %[invscale]         \n\t"
+
+        "1:                                 \n\t"
+        "vld1.16    {d0}, [%[src]]!         \n\t"
+        "vmovl.s16  q0, d0                  \n\t"
+        "vcvt.f32.s32 q0, q0                \n\t"
+        "vmul.f32   q0, q0, q1              \n\t"
+        "subs       %[n], %[n], #1          \n\t"
+        "vst1.32    {q0}, [%[dst]]!         \n\t"
+        "bgt        1b                      \n\t"
+
+        "2:                                 \n\t"
+
+        : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */
+        : [invscale] "r" (invscale) /* input operands */
+        : "memory", "cc", "q0", "q1" /* clobber list */
+    );
+
+    /* leftovers */
+    while (i--) {
+        *dst++ = *src++ * invscale;
+    }
+}
+
+void pa_convert_func_init_neon(pa_cpu_arm_flag_t flags) {
+    pa_log_info("Initialising ARM NEON optimized conversions.");
+    pa_set_convert_from_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_from_f32ne_neon);
+    pa_set_convert_to_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_to_f32ne_neon);
+}

src/tests/cpu-test.c

@@ -282,6 +282,35 @@ START_TEST (sconv_sse_test) {
 END_TEST
 #endif /* defined (__i386__) || defined (__amd64__) */
 
+#if defined (__arm__) && defined (__linux__)
+START_TEST (sconv_neon_test) {
+    pa_cpu_arm_flag_t flags = 0;
+    pa_convert_func_t orig_func, neon_func;
+
+    pa_cpu_get_arm_flags(&flags);
+
+    if (!(flags & PA_CPU_ARM_NEON)) {
+        pa_log_info("NEON not supported. Skipping");
+        return;
+    }
+
+    orig_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE);
+    pa_convert_func_init_neon(flags);
+    neon_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE);
+
+    pa_log_debug("Checking NEON sconv (s16 -> float)");
+    run_conv_test_float_to_s16(neon_func, orig_func, 0, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 1, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 2, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 3, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 4, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 5, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 6, TRUE, FALSE);
+    run_conv_test_float_to_s16(neon_func, orig_func, 7, TRUE, TRUE);
+}
+END_TEST
+#endif /* defined (__arm__) && defined (__linux__) */
+
 #undef SAMPLES
 #undef TIMES
 /* End conversion tests */
@@ -313,6 +342,11 @@ int main(int argc, char *argv[]) {
     tc = tcase_create("sconv");
 #if defined (__i386__) || defined (__amd64__)
     tcase_add_test(tc, sconv_sse_test);
+#endif
+#if defined (__arm__) && defined (__linux__)
+#if HAVE_NEON
+    tcase_add_test(tc, sconv_neon_test);
+#endif
 #endif
     suite_add_tcase(s, tc);