2 * Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>
\r
4 * This file is part of CasparCG (www.casparcg.com).
\r
6 * CasparCG is free software: you can redistribute it and/or modify
\r
7 * it under the terms of the GNU General Public License as published by
\r
8 * the Free Software Foundation, either version 3 of the License, or
\r
9 * (at your option) any later version.
\r
11 * CasparCG is distributed in the hope that it will be useful,
\r
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
\r
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
\r
14 * GNU General Public License for more details.
\r
16 * You should have received a copy of the GNU General Public License
\r
17 * along with CasparCG. If not, see <http://www.gnu.org/licenses/>.
\r
19 * Author: Robert Nagy, ronag89@gmail.com
\r
22 #include "../../stdafx.h"
\r
24 #include "image_mixer.h"
\r
26 #include "../util/data_frame.h"
\r
27 #include "../util/xmm.h"
\r
29 #include <common/assert.h>
\r
30 #include <common/gl/gl_check.h>
\r
31 #include <common/concurrency/async.h>
\r
32 #include <common/memory/memcpy.h>
\r
34 #include <core/frame/data_frame.h>
\r
35 #include <core/frame/frame_transform.h>
\r
36 #include <core/frame/pixel_format.h>
\r
37 #include <core/video_format.h>
\r
39 #include <modules/ffmpeg/producer/util/util.h>
\r
43 #include <gl/glew.h>
\r
45 #include <tbb/cache_aligned_allocator.h>
\r
46 #include <tbb/parallel_for_each.h>
\r
48 #include <boost/assign.hpp>
\r
49 #include <boost/foreach.hpp>
\r
50 #include <boost/range.hpp>
\r
51 #include <boost/range/algorithm_ext/erase.hpp>
\r
52 #include <boost/thread/future.hpp>
\r
54 #include <algorithm>
\r
58 #if defined(_MSC_VER)
\r
59 #pragma warning (push)
\r
60 #pragma warning (disable : 4244)
\r
64 #include <libswscale/swscale.h>
\r
65 #include <libavcodec/avcodec.h>
\r
66 #include <libavformat/avformat.h>
\r
68 #if defined(_MSC_VER)
\r
69 #pragma warning (pop)
\r
72 namespace caspar { namespace accelerator { namespace cpu {
\r
76 core::pixel_format_desc pix_desc;
\r
77 std::array<const uint8_t*, 4> data;
\r
78 core::image_transform transform;
\r
81 : pix_desc(core::pixel_format::invalid)
\r
87 bool operator==(const item& lhs, const item& rhs)
\r
89 return lhs.data == rhs.data && lhs.transform == rhs.transform;
\r
92 bool operator!=(const item& lhs, const item& rhs)
\r
94 return !(lhs == rhs);
\r
97 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)
\r
99 using namespace xmm;
\r
101 // T(S, D) = S * D[A] + 0x80
\r
102 auto aaaa = s8_x::shuffle(d, s8_x(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));
\r
103 d = s8_x(u8_x::min(u8_x(d), u8_x(aaaa))); // overflow guard
\r
105 auto xaxa = s16_x(aaaa) >> 8;
\r
107 auto t1 = s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;
\r
108 auto t2 = s16_x::multiply_low(s16_x(s) >> 8 , xaxa) + 0x80;
\r
110 // C(S, D) = S + D - (((T >> 8) + T) >> 8);
\r
111 auto xyxy = s8_x(((t1 >> 8) + t1) >> 8);
\r
112 auto yxyx = s8_x((t2 >> 8) + t2);
\r
113 auto argb = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));
\r
115 return s8_x(s) + (d - argb);
\r
118 template<typename temporal, typename alignment>
\r
119 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
\r
121 using namespace xmm;
\r
123 for(auto n = 0; n < count; n += 32)
\r
125 auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);
\r
126 auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);
\r
128 auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);
\r
129 auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);
\r
131 auto argb0 = blend(d0, s0);
\r
132 auto argb1 = blend(d1, s1);
\r
134 s8_x::store<temporal, alignment>(argb0, dest+n+0 );
\r
135 s8_x::store<temporal, alignment>(argb1, dest+n+16);
\r
139 template<typename temporal>
\r
140 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
\r
142 using namespace xmm;
\r
144 if(reinterpret_cast<int>(dest) % 16 != 0 || reinterpret_cast<int>(source) % 16 != 0)
\r
145 kernel<temporal_tag, unaligned_tag>(dest, source, count);
\r
147 kernel<temporal_tag, aligned_tag>(dest, source, count);
\r
150 class image_renderer
\r
152 std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>> last_image_;
\r
153 tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>> sws_contexts_;
\r
154 std::vector<spl::shared_ptr<host_buffer>> temp_buffers_;
\r
156 boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)
\r
158 if(last_image_.first == items && last_image_.second.has_value())
\r
159 return last_image_.second;
\r
161 auto image = render(items, format_desc);
\r
162 last_image_ = std::make_pair(std::move(items), image);
\r
167 boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)
\r
169 convert(items, format_desc.width, format_desc.height);
\r
171 auto result = spl::make_shared<host_buffer>(format_desc.size, 0);
\r
172 if(format_desc.field_mode != core::field_mode::progressive)
\r
174 draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);
\r
175 draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);
\r
179 draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::progressive);
\r
182 return async(launch_policy::deferred, [=]
\r
184 return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);
\r
188 void draw(std::vector<item> items, uint8_t* dest, int width, int height, core::field_mode field_mode)
\r
190 BOOST_FOREACH(auto& item, items)
\r
191 item.transform.field_mode &= field_mode;
\r
193 // Remove empty items.
\r
194 boost::range::remove_erase_if(items, [&](const item& item)
\r
196 return item.transform.field_mode == core::field_mode::empty;
\r
199 // Remove first field stills.
\r
200 boost::range::remove_erase_if(items, [&](const item& item)
\r
202 return item.transform.is_still && item.transform.field_mode == field_mode; // only us last field for stills.
\r
208 auto start = field_mode == core::field_mode::lower ? 1 : 0;
\r
209 auto step = field_mode == core::field_mode::progressive ? 1 : 2;
\r
211 // TODO: Add support for fill translations.
\r
212 // TODO: Add support for mask rect.
\r
213 // TODO: Add support for opacity.
\r
214 // TODO: Add support for mix transition.
\r
215 // TODO: Add support for push transition.
\r
216 // TODO: Add support for wipe transition.
\r
217 // TODO: Add support for slide transition.
\r
218 tbb::parallel_for(tbb::blocked_range<int>(0, height/step), [&](const tbb::blocked_range<int>& r)
\r
220 for(auto i = r.begin(); i != r.end(); ++i)
\r
222 auto y = i*step+start;
\r
224 for(std::size_t n = 0; n < items.size()-1; ++n)
\r
225 kernel<xmm::temporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
\r
227 std::size_t n = items.size()-1;
\r
228 kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
\r
234 temp_buffers_.clear();
\r
237 void convert(std::vector<item>& source_items, int width, int height)
\r
239 std::set<std::array<const uint8_t*, 4>> buffers;
\r
241 BOOST_FOREACH(auto& item, source_items)
\r
242 buffers.insert(item.data);
\r
244 auto dest_items = source_items;
\r
246 tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::array<const uint8_t*, 4>& data)
\r
248 auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.data == data;})->pix_desc;
\r
250 if(pix_desc.format == core::pixel_format::bgra &&
\r
251 pix_desc.planes.at(0).width == width &&
\r
252 pix_desc.planes.at(0).height == height)
\r
255 std::array<uint8_t*, 4> data2 = {};
\r
256 for(std::size_t n = 0; n < data.size(); ++n)
\r
257 data2.at(n) = const_cast<uint8_t*>(data[n]);
\r
259 auto input_av_frame = ffmpeg::make_av_frame(data2, pix_desc);
\r
261 int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);
\r
263 auto& pool = sws_contexts_[key];
\r
265 std::shared_ptr<SwsContext> sws_context;
\r
266 if(!pool.try_pop(sws_context))
\r
269 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, ¶m), sws_freeContext);
\r
273 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));
\r
275 auto dest_frame = spl::make_shared<host_buffer>(width*height*4);
\r
278 spl::shared_ptr<AVFrame> dest_av_frame(avcodec_alloc_frame(), av_free);
\r
279 avcodec_get_frame_defaults(dest_av_frame.get());
\r
280 avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), PIX_FMT_BGRA, width, height);
\r
282 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);
\r
283 pool.push(sws_context);
\r
286 temp_buffers_.push_back(dest_frame);
\r
288 for(std::size_t n = 0; n < source_items.size(); ++n)
\r
290 if(source_items[n].data == data)
\r
292 dest_items[n].data.assign(0);
\r
293 dest_items[n].data[0] = dest_frame->data();
\r
294 dest_items[n].pix_desc = core::pixel_format_desc(core::pixel_format::bgra);
\r
295 dest_items[n].pix_desc.planes = boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));
\r
296 dest_items[n].transform = source_items[n].transform;
\r
301 source_items = std::move(dest_items);
\r
305 struct image_mixer::impl : boost::noncopyable
\r
307 image_renderer renderer_;
\r
308 std::vector<core::image_transform> transform_stack_;
\r
309 std::vector<item> items_; // layer/stream/items
\r
312 : transform_stack_(1)
\r
314 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";
\r
317 void begin_layer(core::blend_mode blend_mode)
\r
321 void push(const core::frame_transform& transform)
\r
323 transform_stack_.push_back(transform_stack_.back()*transform.image_transform);
\r
326 void visit(const core::data_frame& frame)
\r
328 if(frame.pixel_format_desc().format == core::pixel_format::invalid)
\r
331 if(frame.pixel_format_desc().planes.empty())
\r
334 if(transform_stack_.back().field_mode == core::field_mode::empty)
\r
338 item.pix_desc = frame.pixel_format_desc();
\r
339 item.transform = transform_stack_.back();
\r
340 for(int n = 0; n < item.pix_desc.planes.size(); ++n)
\r
341 item.data.at(n) = frame.image_data(n).begin();
\r
343 items_.push_back(item);
\r
348 transform_stack_.pop_back();
\r
355 boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)
\r
357 return renderer_(std::move(items_), format_desc);
\r
360 virtual spl::unique_ptr<core::data_frame> create_frame(const void* tag, const core::pixel_format_desc& desc, double frame_rate, core::field_mode field_mode)
\r
362 return spl::make_unique<cpu::data_frame>(tag, desc, frame_rate, field_mode);
\r
366 image_mixer::image_mixer() : impl_(new impl()){}
\r
367 void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}
\r
368 void image_mixer::visit(const core::data_frame& frame){impl_->visit(frame);}
\r
369 void image_mixer::pop(){impl_->pop();}
\r
370 boost::shared_future<boost::iterator_range<const uint8_t*>> image_mixer::operator()(const core::video_format_desc& format_desc){return impl_->render(format_desc);}
\r
371 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}
\r
372 void image_mixer::end_layer(){impl_->end_layer();}
\r
373 spl::unique_ptr<core::data_frame> image_mixer::create_frame(const void* tag, const core::pixel_format_desc& desc, double frame_rate, core::field_mode field_mode) {return impl_->create_frame(tag, desc, frame_rate, field_mode);}
\r