diff options
| author | Max Kellermann <max@duempel.org> | 2014-03-14 23:45:27 +0100 |
|---|---|---|
| committer | Max Kellermann <max@duempel.org> | 2014-03-15 11:41:01 +0100 |
| commit | 5696f91a1e5e0cd1022c60737465c29b429ea599 (patch) | |
| tree | 3b81b96802113065c28faf97af12503260c29551 /src/pcm/PcmFormat.cxx | |
| parent | 3932e62fc7d8bc05e4f77c1b3cd69fd314cee3d1 (diff) | |
| download | mpd-5696f91a1e5e0cd1022c60737465c29b429ea599.tar.gz mpd-5696f91a1e5e0cd1022c60737465c29b429ea599.tar.xz mpd-5696f91a1e5e0cd1022c60737465c29b429ea599.zip | |
pcm/PcmFormat: ARM NEON optimizations for float->s16
This is nearly 4 times faster than the "portable" algorithm.
Diffstat (limited to 'src/pcm/PcmFormat.cxx')
| -rw-r--r-- | src/pcm/PcmFormat.cxx | 39 |
1 files changed, 38 insertions, 1 deletions
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, |