From 7dd8bdd6999be5bf21f6963a7b4ed312806244e4 Mon Sep 17 00:00:00 2001
From: ronag <ronag@362d55ac-95cf-4e76-9f9a-cbaa9c17b72d>
Date: Sun, 4 Sep 2011 00:09:12 +0000
Subject: [PATCH] git-svn-id:
 https://casparcg.svn.sourceforge.net/svnroot/casparcg/server/branches/2.0.0.2@1343
 362d55ac-95cf-4e76-9f9a-cbaa9c17b72d

---
 core/mixer/audio/audio_util.h         | 21 ++++++++++++++-------
 modules/oal/consumer/oal_consumer.cpp | 14 +++++++-------
 2 files changed, 21 insertions(+), 14 deletions(-)

diff --git a/core/mixer/audio/audio_util.h b/core/mixer/audio/audio_util.h
index 0cbebc3d1..d422bc086 100644
--- a/core/mixer/audio/audio_util.h
+++ b/core/mixer/audio/audio_util.h
@@ -6,16 +6,23 @@
 
 #include <boost/range/iterator_range.hpp>
 
+#include <tbb/cache_aligned_allocator.h>
+
 namespace caspar { namespace core {
 
-static std::vector<int16_t> audio_32_to_16(const boost::iterator_range<int32_t*>& input)
+static std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>> audio_32_to_16_sse(const boost::iterator_range<int32_t*>& input)
 {	
-	std::vector<int16_t> audio16(input.size());
-	auto audio32_ptr = reinterpret_cast<const uint32_t*>(input.begin());
-	auto audio16_ptr = reinterpret_cast<uint32_t*>(audio16.data());
-	auto size		 = input.size()/2;
-	for(int n = 0; n < size; ++n)		
-		audio16_ptr[n] = (audio32_ptr[n*2+1] & 0xffff0000) | (audio32_ptr[n*2+0] >> 16);	
+	std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>> audio16(input.size());
+	auto audio32_ptr = reinterpret_cast<const __m128i*>(input.begin());
+	auto audio16_ptr = reinterpret_cast<__m128i*>(audio16.data());
+	auto size		 = input.size();
+	for(int n = 0; n < size/8; ++n)		
+	{
+		auto xmm0 = _mm_srai_epi32(_mm_load_si128(audio32_ptr++), 16);
+		auto xmm1 = _mm_srai_epi32(_mm_load_si128(audio32_ptr++), 16);
+		auto xmm3 = _mm_packs_epi32(xmm0, xmm1);
+		_mm_store_si128(audio16_ptr++, xmm3);
+	}
 	return audio16;
 }
 
diff --git a/modules/oal/consumer/oal_consumer.cpp b/modules/oal/consumer/oal_consumer.cpp
index 7186dc74c..cd0130a5f 100644
--- a/modules/oal/consumer/oal_consumer.cpp
+++ b/modules/oal/consumer/oal_consumer.cpp
@@ -46,8 +46,8 @@ struct oal_consumer : public core::frame_consumer,  public sf::SoundStream
 	safe_ptr<diagnostics::graph>						graph_;
 	boost::timer										perf_timer_;
 
-	tbb::concurrent_bounded_queue<std::shared_ptr<std::vector<short>>>	input_;
-	boost::circular_buffer<std::vector<short>>			container_;
+	tbb::concurrent_bounded_queue<std::shared_ptr<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>>	input_;
+	boost::circular_buffer<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>			container_;
 	tbb::atomic<bool>									is_running_;
 
 	core::video_format_desc								format_desc_;
@@ -70,8 +70,8 @@ public:
 	~oal_consumer()
 	{
 		is_running_ = false;
-		input_.try_push(std::make_shared<std::vector<short>>());
-		input_.try_push(std::make_shared<std::vector<short>>());
+		input_.try_push(std::make_shared<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>());
+		input_.try_push(std::make_shared<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>());
 		Stop();
 		CASPAR_LOG(info) << print() << L" Shutting down.";	
 	}
@@ -87,19 +87,19 @@ public:
 	{			
 		if(preroll_count_ < input_.capacity())
 		{
-			while(input_.try_push(std::make_shared<std::vector<int16_t>>(format_desc_.audio_samples_per_frame, 0)))
+			while(input_.try_push(std::make_shared<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>(format_desc_.audio_samples_per_frame, 0)))
 				++preroll_count_;
 			Play();		
 		}
 
-		input_.push(std::make_shared<std::vector<int16_t>>(core::audio_32_to_16(frame->audio_data())));
+		input_.push(std::make_shared<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>>(core::audio_32_to_16_sse(frame->audio_data())));
 
 		return true;
 	}
 	
 	virtual bool OnGetData(sf::SoundStream::Chunk& data)
 	{		
-		std::shared_ptr<std::vector<short>> audio_data;		
+		std::shared_ptr<std::vector<int16_t, tbb::cache_aligned_allocator<int16_t>>> audio_data;		
 		input_.pop(audio_data);
 				
 		container_.push_back(std::move(*audio_data));
-- 
2.39.2