X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=accelerator%2Fcpu%2Fimage%2Fimage_mixer.cpp;h=315d3f9b7b27ab116902e453c6c6e7109624512c;hb=e0e1a2b1fdcab66e8da72e3eaa2db5564f8f9d72;hp=eb08afc728df941e93b163a12866225964948456;hpb=9de14e806e98c49b336d0a11987721ae051690d2;p=casparcg

diff --git a/accelerator/cpu/image/image_mixer.cpp b/accelerator/cpu/image/image_mixer.cpp
index eb08afc72..315d3f9b7 100644
--- a/accelerator/cpu/image/image_mixer.cpp
+++ b/accelerator/cpu/image/image_mixer.cpp
@@ -1,358 +1,378 @@
-/*
-* Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>
-*
-* This file is part of CasparCG (www.casparcg.com).
-*
-* CasparCG 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 3 of the License, or
-* (at your option) any later version.
-*
-* CasparCG 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 CasparCG. If not, see <http://www.gnu.org/licenses/>.
-*
-* Author: Robert Nagy, ronag89@gmail.com
-*/
-
-#include "../../stdafx.h"
-
-#include "image_mixer.h"
-
-#include "../util/write_frame.h"
-#include "../util/xmm.h"
-
-#include <common/assert.h>
-#include <common/gl/gl_check.h>
-#include <common/concurrency/async.h>
-#include <common/memory/memcpy.h>
-
-#include <core/frame/write_frame.h>
-#include <core/frame/frame_transform.h>
-#include <core/frame/pixel_format.h>
-#include <core/video_format.h>
-
-#include <modules/ffmpeg/producer/util/util.h>
-
-#include <asmlib.h>
-
-#include <gl/glew.h>
-
-#include <tbb/cache_aligned_allocator.h>
-#include <tbb/parallel_for_each.h>
-
-#include <boost/assign.hpp>
-#include <boost/foreach.hpp>
-#include <boost/range.hpp>
-#include <boost/range/algorithm_ext/erase.hpp>
-#include <boost/thread/future.hpp>
-
-#include <algorithm>
-#include <stdint.h>
-#include <vector>
-
-#if defined(_MSC_VER)
-#pragma warning (push)
-#pragma warning (disable : 4244)
-#endif
-extern "C" 
-{
-	#include <libswscale/swscale.h>
-	#include <libavcodec/avcodec.h>
-	#include <libavformat/avformat.h>
-}
-#if defined(_MSC_VER)
-#pragma warning (pop)
-#endif
-
-namespace caspar { namespace accelerator { namespace cpu {
-		
-struct item
-{
-	core::pixel_format_desc						pix_desc;
-	std::vector<spl::shared_ptr<host_buffer>>	buffers;
-	core::frame_transform						transform;
-
-	item()
-		: pix_desc(core::pixel_format::invalid)
-	{
-	}
-};
-
-bool operator==(const item& lhs, const item& rhs)
-{
-	return lhs.buffers == rhs.buffers && lhs.transform == rhs.transform;
-}
-
-bool operator!=(const item& lhs, const item& rhs)
-{
-	return !(lhs == rhs);
-}
-	
-inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)
-{	
-	using namespace xmm;
-		
-	// T(S, D) = S * D[A] + 0x80
-	auto aaaa   = s8_x::shuffle(d, s8_x(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));
-	d			= s8_x(u8_x::min(u8_x(d), u8_x(aaaa))); // overflow guard
-
-	auto xaxa	= s16_x(aaaa) >> 8;		
-			      
-	auto t1		= s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;    
-	auto t2		= s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;
-		
-	// C(S, D) = S + D - (((T >> 8) + T) >> 8);
-	auto xyxy	= s8_x(((t1 >> 8) + t1) >> 8);      
-	auto yxyx	= s8_x((t2 >> 8) + t2);    
-	auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));
-
-	return s8_x(s) + (d - argb);
-}
-	
-template<typename temporal, typename alignment>
-static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
-{			
-	using namespace xmm;
-
-	for(auto n = 0; n < count; n += 32)    
-	{
-		auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);
-		auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);
-
-		auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);
-		auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);
-		
-		auto argb0 = blend(d0, s0);
-		auto argb1 = blend(d1, s1);
-
-		s8_x::store<temporal, alignment>(argb0, dest+n+0 );
-		s8_x::store<temporal, alignment>(argb1, dest+n+16);
-	} 
-}
-
-template<typename temporal>
-static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
-{			
-	using namespace xmm;
-
-	if(reinterpret_cast<int>(dest) % 16 != 0 || reinterpret_cast<int>(source) % 16 != 0)
-		kernel<temporal_tag, unaligned_tag>(dest, source, count);
-	else
-		kernel<temporal_tag, aligned_tag>(dest, source, count);
-}
-
-class image_renderer
-{
-	std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>>		last_image_;
-	tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>	sws_contexts_;
-public:	
-	boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)
-	{	
-		if(last_image_.first == items && last_image_.second.has_value())
-			return last_image_.second;
-
-		auto image	= render(items, format_desc);
-		last_image_ = std::make_pair(std::move(items), image);
-		return image;
-	}
-
-private:
-	boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)
-	{
-		convert(items, format_desc.width, format_desc.height);		
-		
-		auto result = spl::make_shared<host_buffer>(format_desc.size, 0);
-		if(format_desc.field_mode != core::field_mode::progressive)
-		{			
-			draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);
-			draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);
-		}
-		else
-		{
-			draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);
-		}
-		
-		return async(launch_policy::deferred, [=]
-		{
-			return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);
-		});		
-	}
-
-	void draw(std::vector<item> items, uint8_t* dest, int width, int height, core::field_mode field_mode)
-	{		
-		BOOST_FOREACH(auto& item, items)
-			item.transform.field_mode &= field_mode;
-
-		boost::remove_erase_if(items, [](item& item){return item.transform.field_mode == core::field_mode::empty;});
-
-		if(items.empty())
-			return;
-		
-		auto start = field_mode == core::field_mode::lower ? 1 : 0;
-		auto step  = field_mode == core::field_mode::progressive ? 1 : 2;
-		
-		// TODO: Add support for fill translations.
-		// TODO: Add support for mask rect.
-		// TODO: Add support for opacity.
-		// TODO: Add support for mix transition.
-		// TODO: Add support for push transition.
-		// TODO: Add support for wipe transition.
-		// TODO: Add support for slide transition.
-		tbb::parallel_for(tbb::blocked_range<int>(0, height/step), [&](const tbb::blocked_range<int>& r)
-		{
-			for(auto i = r.begin(); i != r.end(); ++i)
-			{
-				auto y = i*step+start;
-
-				for(std::size_t n = 0; n < items.size()-1; ++n)
-					kernel<xmm::temporal_tag>(dest + y*width*4, items[n].buffers.at(0)->data() + y*width*4, width*4);
-				
-				std::size_t n = items.size()-1;				
-				kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].buffers.at(0)->data() + y*width*4, width*4);
-			}
-
-			_mm_mfence();
-		});
-	}
-		
-	void convert(std::vector<item>& source_items, int width, int height)
-	{
-		std::set<std::vector<spl::shared_ptr<host_buffer>>> buffers;
-
-		BOOST_FOREACH(auto& item, source_items)
-			buffers.insert(item.buffers);
-		
-		auto dest_items = source_items;
-
-		tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::vector<spl::shared_ptr<host_buffer>>& buffers)
-		{			
-			auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.buffers == buffers;})->pix_desc;
-
-			if(pix_desc.format == core::pixel_format::bgra && 
-				pix_desc.planes.at(0).width == width &&
-				pix_desc.planes.at(0).height == height)
-				return;
-
-			auto input_av_frame = ffmpeg::make_av_frame(buffers, pix_desc);
-								
-			int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);
-						
-			auto& pool = sws_contexts_[key];
-
-			std::shared_ptr<SwsContext> sws_context;
-			if(!pool.try_pop(sws_context))
-			{
-				double param;
-				sws_context.reset(sws_getContext(input_av_frame->width, input_av_frame->height, static_cast<PixelFormat>(input_av_frame->format), width, height, PIX_FMT_BGRA, SWS_BILINEAR, nullptr, nullptr, &param), sws_freeContext);
-			}
-			
-			if(!sws_context)				
-				BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));				
-		
-			auto dest_frame = spl::make_shared<host_buffer>(width*height*4);
-
-			{
-				spl::shared_ptr<AVFrame> dest_av_frame(avcodec_alloc_frame(), av_free);	
-				avcodec_get_frame_defaults(dest_av_frame.get());			
-				avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), PIX_FMT_BGRA, width, height);
-				
-				sws_scale(sws_context.get(), input_av_frame->data, input_av_frame->linesize, 0, input_av_frame->height, dest_av_frame->data, dest_av_frame->linesize);				
-				pool.push(sws_context);
-			}
-		
-			for(std::size_t n = 0; n < source_items.size(); ++n)
-			{
-				if(source_items[n].buffers == buffers)
-				{
-					dest_items[n].buffers			= boost::assign::list_of(dest_frame);
-					dest_items[n].pix_desc			= core::pixel_format_desc(core::pixel_format::bgra);
-					dest_items[n].pix_desc.planes	= boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));
-					dest_items[n].transform			= source_items[n].transform;
-				}
-			}
-		});	
-
-		source_items = std::move(dest_items);
-	}
-};
-		
-struct image_mixer::impl : boost::noncopyable
-{	
-	image_renderer						renderer_;
-	std::vector<core::frame_transform>	transform_stack_;
-	std::vector<item>					items_; // layer/stream/items
-public:
-	impl() 
-		: transform_stack_(1)	
-	{
-		CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";
-	}
-
-	void begin_layer(core::blend_mode blend_mode)
-	{
-	}
-		
-	void push(core::frame_transform& transform)
-	{
-		transform_stack_.push_back(transform_stack_.back()*transform);
-	}
-		
-	void visit(core::data_frame& frame2)
-	{			
-		write_frame* frame = dynamic_cast<write_frame*>(&frame2);
-		if(frame == nullptr)
-			return;
-
-		if(frame->get_pixel_format_desc().format == core::pixel_format::invalid)
-			return;
-
-		if(frame->get_buffers().empty())
-			return;
-
-		if(transform_stack_.back().field_mode == core::field_mode::empty)
-			return;
-
-		item item;
-		item.pix_desc			= frame->get_pixel_format_desc();
-		item.buffers			= frame->get_buffers();				
-		item.transform			= transform_stack_.back();
-		item.transform.volume	= core::frame_transform().volume; // Set volume to default since we don't care about it here.
-
-		items_.push_back(item);
-	}
-
-	void pop()
-	{
-		transform_stack_.pop_back();
-	}
-
-	void end_layer()
-	{		
-	}
-	
-	boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)
-	{
-		return renderer_(std::move(items_), format_desc);
-	}
-	
-	virtual spl::shared_ptr<cpu::write_frame> create_frame(const void* tag, const core::pixel_format_desc& desc)
-	{
-		return spl::make_shared<cpu::write_frame>(tag, desc);
-	}
-};
-
-image_mixer::image_mixer() : impl_(new impl()){}
-void image_mixer::push(core::frame_transform& transform){impl_->push(transform);}
-void image_mixer::visit(core::data_frame& frame){impl_->visit(frame);}
-void image_mixer::pop(){impl_->pop();}
-boost::shared_future<boost::iterator_range<const uint8_t*>> image_mixer::operator()(const core::video_format_desc& format_desc){return impl_->render(format_desc);}
-void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}
-void image_mixer::end_layer(){impl_->end_layer();}
-spl::shared_ptr<core::write_frame> image_mixer::create_frame(const void* tag, const core::pixel_format_desc& desc) {return impl_->create_frame(tag, desc);}
-
-}}}
\ No newline at end of file
+/*
+* Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>
+*
+* This file is part of CasparCG (www.casparcg.com).
+*
+* CasparCG 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 3 of the License, or
+* (at your option) any later version.
+*
+* CasparCG 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 CasparCG. If not, see <http://www.gnu.org/licenses/>.
+*
+* Author: Robert Nagy, ronag89@gmail.com
+*/
+
+#include "../../StdAfx.h"
+
+#include "image_mixer.h"
+
+#include "../util/xmm.h"
+
+#include <common/assert.h>
+#include <common/gl/gl_check.h>
+#include <common/future.h>
+#include <common/array.h>
+
+#include <core/frame/frame.h>
+#include <core/frame/frame_transform.h>
+#include <core/frame/pixel_format.h>
+#include <core/video_format.h>
+
+#include <modules/ffmpeg/producer/util/util.h>
+
+#include <asmlib.h>
+
+#include <GL/glew.h>
+
+#include <tbb/parallel_for.h>
+#include <tbb/parallel_for_each.h>
+#include <tbb/concurrent_queue.h>
+
+#include <boost/range/algorithm_ext/erase.hpp>
+#include <boost/thread/future.hpp>
+
+#include <algorithm>
+#include <cstdint>
+#include <vector>
+#include <set>
+#include <array>
+
+#if defined(_MSC_VER)
+#pragma warning (push)
+#pragma warning (disable : 4244)
+#endif
+extern "C"
+{
+	#include <libswscale/swscale.h>
+	#include <libavcodec/avcodec.h>
+	#include <libavformat/avformat.h>
+}
+#if defined(_MSC_VER)
+#pragma warning (pop)
+#endif
+
+namespace caspar { namespace accelerator { namespace cpu {
+
+struct item
+{
+	core::pixel_format_desc			pix_desc	= core::pixel_format::invalid;
+	std::array<const uint8_t*, 4>	data;
+	core::image_transform			transform;
+
+	item()
+	{
+		data.fill(0);
+	}
+};
+
+bool operator==(const item& lhs, const item& rhs)
+{
+	return lhs.data == rhs.data && lhs.transform == rhs.transform;
+}
+
+bool operator!=(const item& lhs, const item& rhs)
+{
+	return !(lhs == rhs);
+}
+
+// 100% accurate blending with correct rounding.
+inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)
+{
+	using namespace xmm;
+
+	// C(S, D) = S + D - (((T >> 8) + T) >> 8);
+	// T(S, D) = S * D[A] + 0x80
+
+	auto aaaa   = s8_x::shuffle(d, s8_x(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));
+	d			= s8_x(u8_x::min(u8_x(d), u8_x(aaaa))); // Overflow guard. Some source files have color values which incorrectly exceed pre-multiplied alpha values, e.g. red(255) > alpha(254).
+
+	auto xaxa	= s16_x(aaaa) >> 8;
+
+	auto t1		= s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;
+	auto t2		= s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;
+
+	auto xyxy	= s8_x(((t1 >> 8) + t1) >> 8);
+	auto yxyx	= s8_x((t2 >> 8) + t2);
+	auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));
+
+	return s8_x(s) + (d - argb);
+}
+
+template<typename temporal, typename alignment>
+static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
+{
+	using namespace xmm;
+
+	for(auto n = 0; n < count; n += 32)
+	{
+		auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);
+		auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);
+
+		auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);
+		auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);
+
+		auto argb0 = blend(d0, s0);
+		auto argb1 = blend(d1, s1);
+
+		s8_x::store<temporal, alignment>(argb0, dest+n+0 );
+		s8_x::store<temporal, alignment>(argb1, dest+n+16);
+	}
+}
+
+template<typename temporal>
+static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
+{
+	using namespace xmm;
+
+	if(reinterpret_cast<std::uint64_t>(dest) % 16 != 0 || reinterpret_cast<std::uint64_t>(source) % 16 != 0)
+		kernel<temporal_tag, unaligned_tag>(dest, source, count);
+	else
+		kernel<temporal_tag, aligned_tag>(dest, source, count);
+}
+
+class image_renderer
+{
+	tbb::concurrent_unordered_map<int64_t, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>	sws_devices_;
+	tbb::concurrent_bounded_queue<spl::shared_ptr<buffer>>												temp_buffers_;
+	core::video_format_desc																				format_desc_;
+public:
+	std::future<array<const std::uint8_t>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)
+	{
+		if (format_desc != format_desc_)
+		{
+			format_desc_ = format_desc;
+			sws_devices_.clear();
+		}
+
+		convert(items, format_desc.width, format_desc.height);
+
+		// Remove first field stills.
+		boost::range::remove_erase_if(items, [&](const item& item)
+		{
+			return item.transform.is_still && item.transform.field_mode == format_desc.field_mode; // only us last field for stills.
+		});
+
+		// Stills are progressive
+		for (auto& item : items)
+		{
+			if(item.transform.is_still)
+				item.transform.field_mode = core::field_mode::progressive;
+		}
+
+		auto result = spl::make_shared<buffer>(format_desc.size, 0);
+		if(format_desc.field_mode != core::field_mode::progressive)
+		{
+			draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);
+			draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);
+		}
+		else
+		{
+			draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);
+		}
+
+		temp_buffers_.clear();
+
+		return make_ready_future(array<const std::uint8_t>(result->data(), format_desc.size, true, result));
+	}
+
+private:
+
+	void draw(std::vector<item> items, uint8_t* dest, std::size_t width, std::size_t height, core::field_mode field_mode)
+	{
+		for (auto& item : items)
+			item.transform.field_mode &= field_mode;
+
+		// Remove empty items.
+		boost::range::remove_erase_if(items, [&](const item& item)
+		{
+			return item.transform.field_mode == core::field_mode::empty;
+		});
+
+		if(items.empty())
+			return;
+
+		auto start = field_mode == core::field_mode::lower ? 1 : 0;
+		auto step  = field_mode == core::field_mode::progressive ? 1 : 2;
+
+		// TODO: Add support for fill translations.
+		// TODO: Add support for mask rect.
+		// TODO: Add support for opacity.
+		// TODO: Add support for mix transition.
+		// TODO: Add support for push transition.
+		// TODO: Add support for wipe transition.
+		// TODO: Add support for slide transition.
+		tbb::parallel_for(tbb::blocked_range<std::size_t>(0, height/step), [&](const tbb::blocked_range<std::size_t>& r)
+		{
+			for(auto i = r.begin(); i != r.end(); ++i)
+			{
+				auto y = i*step+start;
+
+				for(std::size_t n = 0; n < items.size()-1; ++n)
+					kernel<xmm::temporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
+
+				std::size_t n = items.size()-1;
+				kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
+			}
+
+			_mm_mfence();
+		});
+	}
+
+	void convert(std::vector<item>& source_items, int width, int height)
+	{
+		std::set<std::array<const uint8_t*, 4>> buffers;
+
+		for (auto& item : source_items)
+			buffers.insert(item.data);
+
+		auto dest_items = source_items;
+
+		tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::array<const uint8_t*, 4>& data)
+		{
+			auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.data == data;})->pix_desc;
+
+			if(pix_desc.format == core::pixel_format::bgra &&
+				pix_desc.planes.at(0).width == width &&
+				pix_desc.planes.at(0).height == height)
+				return;
+
+			std::array<uint8_t*, 4> data2 = {};
+			for(std::size_t n = 0; n < data.size(); ++n)
+				data2.at(n) = const_cast<uint8_t*>(data[n]);
+
+			auto input_av_frame = ffmpeg::make_av_frame(data2, pix_desc);
+
+
+			int64_t key = ((static_cast<int64_t>(input_av_frame->width)	 << 32) & 0xFFFF00000000) |
+						  ((static_cast<int64_t>(input_av_frame->height) << 16) & 0xFFFF0000) |
+						  ((static_cast<int64_t>(input_av_frame->format) <<  8) & 0xFF00);
+
+			auto& pool = sws_devices_[key];
+
+			std::shared_ptr<SwsContext> sws_device;
+			if(!pool.try_pop(sws_device))
+			{
+				double param;
+				sws_device.reset(sws_getContext(input_av_frame->width, input_av_frame->height, static_cast<AVPixelFormat>(input_av_frame->format), width, height, AVPixelFormat::AV_PIX_FMT_BGRA, SWS_BILINEAR, nullptr, nullptr, &param), sws_freeContext);
+			}
+
+			if(!sws_device)
+				CASPAR_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling device.") << boost::errinfo_api_function("sws_getContext"));
+
+			auto dest_frame = spl::make_shared<buffer>(width*height*4);
+			temp_buffers_.push(dest_frame);
+
+			{
+				auto dest_av_frame = ffmpeg::create_frame();
+				avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), AVPixelFormat::AV_PIX_FMT_BGRA, width, height);
+
+				sws_scale(sws_device.get(), input_av_frame->data, input_av_frame->linesize, 0, input_av_frame->height, dest_av_frame->data, dest_av_frame->linesize);
+				pool.push(sws_device);
+			}
+
+			for(std::size_t n = 0; n < source_items.size(); ++n)
+			{
+				if(source_items[n].data == data)
+				{
+					dest_items[n].data.fill(0);
+					dest_items[n].data[0]			= dest_frame->data();
+					dest_items[n].pix_desc			= core::pixel_format_desc(core::pixel_format::bgra);
+					dest_items[n].pix_desc.planes	= { core::pixel_format_desc::plane(width, height, 4) };
+					dest_items[n].transform			= source_items[n].transform;
+				}
+			}
+		});
+
+		source_items = std::move(dest_items);
+	}
+};
+
+struct image_mixer::impl : boost::noncopyable
+{
+	image_renderer						renderer_;
+	std::vector<core::image_transform>	transform_stack_;
+	std::vector<item>					items_; // layer/stream/items
+public:
+	impl(int channel_id)
+		: transform_stack_(1)
+	{
+		CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer for channel " << channel_id;
+	}
+
+	void push(const core::frame_transform& transform)
+	{
+		transform_stack_.push_back(transform_stack_.back()*transform.image_transform);
+	}
+
+	void visit(const core::const_frame& frame)
+	{
+		if(frame.pixel_format_desc().format == core::pixel_format::invalid)
+			return;
+
+		if(frame.pixel_format_desc().planes.empty())
+			return;
+
+		if(frame.pixel_format_desc().planes.at(0).size < 16)
+			return;
+
+		if(transform_stack_.back().field_mode == core::field_mode::empty)
+			return;
+
+		item item;
+		item.pix_desc	= frame.pixel_format_desc();
+		item.transform	= transform_stack_.back();
+		for(int n = 0; n < item.pix_desc.planes.size(); ++n)
+			item.data.at(n) = frame.image_data(n).begin();
+
+		items_.push_back(item);
+	}
+
+	void pop()
+	{
+		transform_stack_.pop_back();
+	}
+
+	std::future<array<const std::uint8_t>> render(const core::video_format_desc& format_desc)
+	{
+		return renderer_(std::move(items_), format_desc);
+	}
+
+	core::mutable_frame create_frame(const void* tag, const core::pixel_format_desc& desc, const core::audio_channel_layout& channel_layout)
+	{
+		std::vector<array<std::uint8_t>> buffers;
+		for (auto& plane : desc.planes)
+		{
+			auto buf = spl::make_shared<buffer>(plane.size);
+			buffers.push_back(array<std::uint8_t>(buf->data(), plane.size, true, buf));
+		}
+		return core::mutable_frame(std::move(buffers), core::mutable_audio_buffer(), tag, desc, channel_layout);
+	}
+};
+
+image_mixer::image_mixer(int channel_id) : impl_(new impl(channel_id)){}
+image_mixer::~image_mixer(){}
+void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}
+void image_mixer::visit(const core::const_frame& frame){impl_->visit(frame);}
+void image_mixer::pop(){impl_->pop();}
+int image_mixer::get_max_frame_size() { return std::numeric_limits<int>::max(); }
+std::future<array<const std::uint8_t>> image_mixer::operator()(const core::video_format_desc& format_desc, bool /* straighten_alpha */){return impl_->render(format_desc);}
+core::mutable_frame image_mixer::create_frame(const void* tag, const core::pixel_format_desc& desc, const core::audio_channel_layout& channel_layout) {return impl_->create_frame(tag, desc, channel_layout);}
+
+}}}