git-svn-id: https://casparcg.svn.sourceforge.net/svnroot/casparcg/server/branches...

[casparcg] / core / mixer / audio / audio_util.h

diff --git a/core/mixer/audio/audio_util.h b/core/mixer/audio/audio_util.h
index 0cbebc3d1dbfc27158b33c3ce63eb9e00e9f427e..d422bc0861fcf7650c9446c3d55b8d325bee4c55 100644 (file)
--- a/core/mixer/audio/audio_util.h
+++ b/core/mixer/audio/audio_util.h
@@ -6,16 +6,23 @@
 \r
 #include <boost/range/iterator_range.hpp>\r
 \r
+#include <tbb/cache_aligned_allocator.h>\r
+\r
 namespace caspar { namespace core {\r
 \r
-static std::vector<int16_t> audio_32_to_16(const boost::iterator_range<int32_t*>& input)\r
+static std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>> audio_32_to_16_sse(const boost::iterator_range<int32_t*>& input)\r
 {      \r
-       std::vector<int16_t> audio16(input.size());\r
-       auto audio32_ptr = reinterpret_cast<const uint32_t*>(input.begin());\r
-       auto audio16_ptr = reinterpret_cast<uint32_t*>(audio16.data());\r
-       auto size                = input.size()/2;\r
-       for(int n = 0; n < size; ++n)           \r
-               audio16_ptr[n] = (audio32_ptr[n*2+1] & 0xffff0000) | (audio32_ptr[n*2+0] >> 16);        \r
+       std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>> audio16(input.size());\r
+       auto audio32_ptr = reinterpret_cast<const __m128i*>(input.begin());\r
+       auto audio16_ptr = reinterpret_cast<__m128i*>(audio16.data());\r
+       auto size                = input.size();\r
+       for(int n = 0; n < size/8; ++n)         \r
+       {\r
+               auto xmm0 = _mm_srai_epi32(_mm_load_si128(audio32_ptr++), 16);\r
+               auto xmm1 = _mm_srai_epi32(_mm_load_si128(audio32_ptr++), 16);\r
+               auto xmm3 = _mm_packs_epi32(xmm0, xmm1);\r
+               _mm_store_si128(audio16_ptr++, xmm3);\r
+       }\r
        return audio16;\r
 }\r
 \r