pcm/PcmFormat: ARM NEON optimizations for float->s16

This is nearly 4 times faster than the "portable" algorithm.

4 files changed, 137 insertions, 1 deletions

diff --git a/Makefile.am b/Makefile.am
index e1a5415d8..700511338 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -452,6 +452,7 @@ libpcm_a_SOURCES = \
 	src/pcm/PcmFormat.cxx src/pcm/PcmFormat.hxx \
 	src/pcm/FloatConvert.hxx \
 	src/pcm/ShiftConvert.hxx \
+	src/pcm/Neon.hxx \
 	src/pcm/FormatConverter.cxx src/pcm/FormatConverter.hxx \
 	src/pcm/ChannelsConverter.cxx src/pcm/ChannelsConverter.hxx \
 	src/pcm/Resampler.hxx \
diff --git a/NEWS b/NEWS
index fa5c49298..79bb27b5f 100644
--- a/NEWS
+++ b/NEWS
@@ -36,6 +36,7 @@ ver 0.19 (not yet released)
   - allow playlist directory without music directory
   - use XDG to auto-detect "music_directory" and "db_file"
 * new resampler option using libsoxr
+* ARM NEON optimizations
 * install systemd unit for socket activation
 * Android port
 
diff --git a/src/pcm/Neon.hxx b/src/pcm/Neon.hxx
new file mode 100644
index 000000000..a6f52db42
--- /dev/null
+++ b/src/pcm/Neon.hxx
@@ -0,0 +1,97 @@
+/*
+ * Copyright (C) 2003-2014 The Music Player Daemon Project
+ * http://www.musicpd.org
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#ifndef MPD_PCM_NEON_HXX
+#define MPD_PCM_NEON_HXX
+
+#include "Traits.hxx"
+
+#include <arm_neon.h>
+
+/**
+ * Call a NEON intrinsic for each element in the vector.
+ *
+ * @param func the NEON intrinsic
+ * @param result the vector variable that gets assigned the result
+ * @param vector the input vector
+ */
+#define neon_x4_u(func, result, vector) do { \
+	result.val[0] = func(vector.val[0]); \
+	result.val[1] = func(vector.val[1]); \
+	result.val[2] = func(vector.val[2]); \
+	result.val[3] = func(vector.val[3]); \
+} while (0)
+
+/**
+ * Call a NEON intrinsic for each element in the vector.
+ *
+ * @param func the NEON intrinsic
+ * @param result the vector variable that gets assigned the result
+ * @param vector the input vector
+ * @param arg an additional argument that gets passed to each call
+ */
+#define neon_x4_b(func, result, vector, arg) do { \
+	result.val[0] = func(vector.val[0], arg); \
+	result.val[1] = func(vector.val[1], arg); \
+	result.val[2] = func(vector.val[2], arg); \
+	result.val[3] = func(vector.val[3], arg); \
+} while (0)
+
+/**
+ * Convert floating point samples to 16 bit signed integer using ARM NEON.
+ */
+struct NeonFloatTo16 {
+	static constexpr SampleFormat src_format = SampleFormat::FLOAT;
+	static constexpr SampleFormat dst_format = SampleFormat::S16;
+	typedef SampleTraits<src_format> SrcTraits;
+	typedef SampleTraits<dst_format> DstTraits;
+
+	typedef typename SrcTraits::value_type SV;
+	typedef typename DstTraits::value_type DV;
+
+	static constexpr size_t BLOCK_SIZE = 16;
+
+	void Convert(int16_t *dst, const float *src, const size_t n) const {
+		const float32x4_t factor =
+			vdupq_n_f32(1 << (DstTraits::BITS - 1));
+
+		for (unsigned i = 0; i < n / BLOCK_SIZE;
+		     ++i, src += BLOCK_SIZE, dst += BLOCK_SIZE) {
+			/* load 16 float samples into 4 quad
+			   registers */
+			float32x4x4_t value = vld4q_f32(src);
+
+			/* apply factor */
+			neon_x4_b(vmulq_f32, value, value, factor);
+
+			/* convert to 32 bit integer */
+			int32x4x4_t ivalue;
+			neon_x4_u(vcvtq_s32_f32, ivalue, value);
+
+			/* convert to 16 bit integer with saturation */
+			int16x4x4_t nvalue;
+			neon_x4_u(vqmovn_s32, nvalue, ivalue);
+
+			/* store result */
+			vst4_s16(dst, nvalue);
+		}
+	}
+};
+
+#endif
diff --git a/src/pcm/PcmFormat.cxx b/src/pcm/PcmFormat.cxx
index f2d71a6fb..4cabc05a0 100644
--- a/src/pcm/PcmFormat.cxx
+++ b/src/pcm/PcmFormat.cxx
@@ -76,9 +76,46 @@ struct Convert32To16 {
 };
 
 template<SampleFormat F, class Traits=SampleTraits<F>>
-struct FloatToInteger
+struct PortableFloatToInteger
 	: PerSampleConvert<FloatToIntegerSampleConvert<F, Traits>> {};
 
+template<SampleFormat F, class Traits=SampleTraits<F>>
+struct FloatToInteger : PortableFloatToInteger<F, Traits> {};
+
+/**
+ * A template class that attempts to use the "optimized" algorithm for
+ * large portions of the buffer, and calls the "portable" algorithm"
+ * for the rest when the last block is not full.
+ */
+template<typename Optimized, typename Portable>
+class GlueOptimizedConvert : Optimized, Portable {
+public:
+	typedef typename Portable::SrcTraits SrcTraits;
+	typedef typename Portable::DstTraits DstTraits;
+
+	void Convert(typename DstTraits::pointer_type out,
+		     typename SrcTraits::const_pointer_type in,
+		     size_t n) const {
+		Optimized::Convert(out, in, n);
+
+		/* use the "portable" algorithm for the trailing
+		   samples */
+		size_t remaining = n % Optimized::BLOCK_SIZE;
+		size_t done = n - remaining;
+		Portable::Convert(out + done, in + done, remaining);
+	}
+};
+
+#ifdef __ARM_NEON__
+#include "Neon.hxx"
+
+template<>
+struct FloatToInteger<SampleFormat::S16, SampleTraits<SampleFormat::S16>>
+	: GlueOptimizedConvert<NeonFloatTo16,
+			       PortableFloatToInteger<SampleFormat::S16>> {};
+
+#endif
+
 template<class C>
 static ConstBuffer<typename C::DstTraits::value_type>
 AllocateConvert(PcmBuffer &buffer, C convert,