git.sesse.net Git - casparcg/blob - accelerator/cpu/image/image_mixer.cpp

   1 /*\r
   2 * Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>\r
   3 *\r
   4 * This file is part of CasparCG (www.casparcg.com).\r
   5 *\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
  10 *\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
  15 *\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
  18 *\r
  19 * Author: Robert Nagy, ronag89@gmail.com\r
  20 */\r
  21 \r
  22 #include "../../stdafx.h"\r
  23 \r
  24 #include "image_mixer.h"\r
  25 \r
  26 #include "../util/write_frame.h"\r
  27 #include "../util/blend.h"\r
  28 \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
  33 \r
  34 #include <core/frame/write_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
  38 \r
  39 #include <modules/ffmpeg/producer/util/util.h>\r
  40 \r
  41 #include <asmlib.h>\r
  42 \r
  43 #include <gl/glew.h>\r
  44 \r
  45 #include <tbb/cache_aligned_allocator.h>\r
  46 #include <tbb/parallel_for_each.h>\r
  47 \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
  53 \r
  54 #include <algorithm>\r
  55 #include <vector>\r
  56 \r
  57 #if defined(_MSC_VER)\r
  58 #pragma warning (push)\r
  59 #pragma warning (disable : 4244)\r
  60 #endif\r
  61 extern "C" \r
  62 {\r
  63         #include <libswscale/swscale.h>\r
  64         #include <libavcodec/avcodec.h>\r
  65         #include <libavformat/avformat.h>\r
  66 }\r
  67 #if defined(_MSC_VER)\r
  68 #pragma warning (pop)\r
  69 #endif\r
  70 \r
  71 namespace caspar { namespace accelerator { namespace cpu {\r
  72                 \r
  73 struct item\r
  74 {\r
  75         core::pixel_format_desc                                         pix_desc;\r
  76         std::vector<spl::shared_ptr<host_buffer>>       buffers;\r
  77         core::frame_transform                                           transform;\r
  78 \r
  79         item()\r
  80                 : pix_desc(core::pixel_format::invalid)\r
  81         {\r
  82         }\r
  83 };\r
  84 \r
  85 bool operator==(const item& lhs, const item& rhs)\r
  86 {\r
  87         return lhs.buffers == rhs.buffers && lhs.transform == rhs.transform;\r
  88 }\r
  89 \r
  90 bool operator!=(const item& lhs, const item& rhs)\r
  91 {\r
  92         return !(lhs == rhs);\r
  93 }\r
  94 \r
  95 class image_renderer\r
  96 {\r
  97         std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>>               last_image_;\r
  98         tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>  sws_contexts_;\r
  99 public: \r
 100         boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)\r
 101         {       \r
 102                 if(last_image_.first == items && last_image_.second.has_value())\r
 103                         return last_image_.second;\r
 104 \r
 105                 auto image      = render(items, format_desc);\r
 106                 last_image_ = std::make_pair(std::move(items), image);\r
 107                 return image;\r
 108         }\r
 109 \r
 110 private:\r
 111         boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)\r
 112         {\r
 113                 convert(items, format_desc.width, format_desc.height);          \r
 114                 \r
 115                 auto result = spl::make_shared<host_buffer>(format_desc.size, 0);\r
 116                 if(format_desc.field_mode != core::field_mode::progressive)\r
 117                 {\r
 118                         auto upper = items;\r
 119                         auto lower = items;\r
 120 \r
 121                         BOOST_FOREACH(auto& item, upper)\r
 122                                 item.transform.field_mode &= core::field_mode::upper;\r
 123                                         \r
 124                         BOOST_FOREACH(auto& item, lower)\r
 125                                 item.transform.field_mode &= core::field_mode::lower;\r
 126                         \r
 127                         draw(upper, result->data(), format_desc.width, format_desc.height);\r
 128                         draw(lower, result->data(), format_desc.width, format_desc.height);\r
 129                 }\r
 130                 else\r
 131                 {\r
 132                         draw(items, result->data(), format_desc.width, format_desc.height);\r
 133                 }\r
 134                 \r
 135                 return async(launch_policy::deferred, [=]\r
 136                 {\r
 137                         return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);\r
 138                 });             \r
 139         }\r
 140 \r
 141         void draw(std::vector<item>& items, uint8_t* dest, int width, int height)\r
 142         {\r
 143                 BOOST_FOREACH(auto& item, items)\r
 144                 {\r
 145                         auto field_mode = item.transform.field_mode;            \r
 146 \r
 147                         if(field_mode == core::field_mode::empty)\r
 148                                 continue;\r
 149 \r
 150                         auto start = field_mode == core::field_mode::lower ? 1 : 0;\r
 151                         auto step  = field_mode == core::field_mode::progressive ? 1 : 2;\r
 152 \r
 153                         auto source = item.buffers.at(0)->data();\r
 154 \r
 155                         // TODO: Blend using divide and conquer instead of accumulation.\r
 156                         // TODO: Add support for fill translations.\r
 157                         // TODO: Add support for mask translations.\r
 158                         // TODO: Add support for opacity.\r
 159                         // TODO: Add support for mix transition.\r
 160                         // TODO: Add support for push transition.\r
 161                         // TODO: Add support for wipe transition.\r
 162                         // TODO: Add support for slide transition.\r
 163                         tbb::parallel_for(start, height, step, [&](int y)\r
 164                         {\r
 165                                 cpu::blend(dest + y*width*4, source + y*width*4, width*4);\r
 166                         });\r
 167                 }\r
 168         }\r
 169         \r
 170         void convert(std::vector<item>& items, int width, int height)\r
 171         {\r
 172                 // TODO: Don't convert buffers multiple times just because they are in different items due to e.g. interlacing.\r
 173                 tbb::parallel_for_each(items.begin(), items.end(), [&](item& item)\r
 174                 {\r
 175                         if(item.pix_desc.format == core::pixel_format::bgra && \r
 176                            item.pix_desc.planes.at(0).width == width &&\r
 177                            item.pix_desc.planes.at(0).height == height)\r
 178                                 return;\r
 179 \r
 180                         auto input_av_frame = ffmpeg::make_av_frame(item.buffers, item.pix_desc);\r
 181                                                                 \r
 182                         int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);\r
 183                                                 \r
 184                         auto& pool = sws_contexts_[key];\r
 185 \r
 186                         std::shared_ptr<SwsContext> sws_context;\r
 187                         if(!pool.try_pop(sws_context))\r
 188                         {\r
 189                                 double param;\r
 190                                 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);\r
 191                         }\r
 192                         \r
 193                         if(!sws_context)                                \r
 194                                 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));                           \r
 195                 \r
 196                         auto dest = spl::make_shared<host_buffer>(width*height*4);\r
 197 \r
 198                         spl::shared_ptr<AVFrame> av_frame(avcodec_alloc_frame(), av_free);      \r
 199                         avcodec_get_frame_defaults(av_frame.get());                     \r
 200                         avpicture_fill(reinterpret_cast<AVPicture*>(av_frame.get()), dest->data(), PIX_FMT_BGRA, width, height);\r
 201                                 \r
 202                         sws_scale(sws_context.get(), input_av_frame->data, input_av_frame->linesize, 0, input_av_frame->height, av_frame->data, av_frame->linesize);    \r
 203                         \r
 204                         clamp_alpha_overflow(av_frame->data[0], av_frame->data[0], width*height*4);\r
 205 \r
 206                         item.buffers.clear();\r
 207                         item.buffers.push_back(dest);\r
 208                         item.pix_desc = core::pixel_format_desc(core::pixel_format::bgra);\r
 209                         item.pix_desc.planes.clear();\r
 210                         item.pix_desc.planes.push_back(core::pixel_format_desc::plane(width, height, 4));\r
 211 \r
 212                         pool.push(sws_context);\r
 213                 });\r
 214         }\r
 215 \r
 216         void clamp_alpha_overflow(uint8_t* dest, const uint8_t* source, size_t count)\r
 217         {               \r
 218                 CASPAR_VERIFY(count % 64 == 0);\r
 219 \r
 220                 auto alpha_shuffle      = xmm_epi8(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3);\r
 221 \r
 222                 for(auto n = 0; n < count; n += 64)    \r
 223                 {\r
 224                         auto x0                 = xmm_epi8::load(source+n+0);\r
 225                         auto x1                 = xmm_epi8::load(source+n+16);\r
 226                         auto x2                 = xmm_epi8::load(source+n+32);\r
 227                         auto x3                 = xmm_epi8::load(source+n+48);\r
 228 \r
 229                         auto aaaa0              = xmm_epi8::shuffle(x0, alpha_shuffle);\r
 230                         auto aaaa1              = xmm_epi8::shuffle(x1, alpha_shuffle);\r
 231                         auto aaaa2              = xmm_epi8::shuffle(x2, alpha_shuffle);\r
 232                         auto aaaa3              = xmm_epi8::shuffle(x3, alpha_shuffle);\r
 233 \r
 234                         x0                              = xmm_epi8::umin(x0, aaaa0);\r
 235                         x1                              = xmm_epi8::umin(x1, aaaa1);\r
 236                         x2                              = xmm_epi8::umin(x2, aaaa2);\r
 237                         x3                              = xmm_epi8::umin(x3, aaaa3);\r
 238                         \r
 239                         xmm_epi8::stream(x0, dest+n+0);\r
 240                         xmm_epi8::stream(x1, dest+n+16);\r
 241                         xmm_epi8::stream(x2, dest+n+32);\r
 242                         xmm_epi8::stream(x3, dest+n+48);\r
 243                 } \r
 244         }\r
 245 };\r
 246                 \r
 247 struct image_mixer::impl : boost::noncopyable\r
 248 {       \r
 249         image_renderer                                          renderer_;\r
 250         std::vector<core::frame_transform>      transform_stack_;\r
 251         std::vector<item>                                       items_; // layer/stream/items\r
 252 public:\r
 253         impl() \r
 254                 : transform_stack_(1)   \r
 255         {\r
 256                 CASPAR_LOG(info) << L"Initialized CPU Accelerated Image Mixer";\r
 257         }\r
 258 \r
 259         void begin_layer(core::blend_mode blend_mode)\r
 260         {\r
 261         }\r
 262                 \r
 263         void push(core::frame_transform& transform)\r
 264         {\r
 265                 transform_stack_.push_back(transform_stack_.back()*transform);\r
 266         }\r
 267                 \r
 268         void visit(core::data_frame& frame2)\r
 269         {                       \r
 270                 write_frame* frame = dynamic_cast<write_frame*>(&frame2);\r
 271                 if(frame == nullptr)\r
 272                         return;\r
 273 \r
 274                 if(frame->get_pixel_format_desc().format == core::pixel_format::invalid)\r
 275                         return;\r
 276 \r
 277                 if(frame->get_buffers().empty())\r
 278                         return;\r
 279 \r
 280                 if(transform_stack_.back().field_mode == core::field_mode::empty)\r
 281                         return;\r
 282 \r
 283                 item item;\r
 284                 item.pix_desc                   = frame->get_pixel_format_desc();\r
 285                 item.buffers                    = frame->get_buffers();                         \r
 286                 item.transform                  = transform_stack_.back();\r
 287                 item.transform.volume   = core::frame_transform().volume; // Set volume to default since we don't care about it here.\r
 288 \r
 289                 items_.push_back(item);\r
 290         }\r
 291 \r
 292         void pop()\r
 293         {\r
 294                 transform_stack_.pop_back();\r
 295         }\r
 296 \r
 297         void end_layer()\r
 298         {               \r
 299         }\r
 300         \r
 301         boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)\r
 302         {\r
 303                 return renderer_(std::move(items_), format_desc);\r
 304         }\r
 305         \r
 306         virtual spl::shared_ptr<cpu::write_frame> create_frame(const void* tag, const core::pixel_format_desc& desc)\r
 307         {\r
 308                 return spl::make_shared<cpu::write_frame>(tag, desc);\r
 309         }\r
 310 };\r
 311 \r
 312 image_mixer::image_mixer() : impl_(new impl()){}\r
 313 void image_mixer::push(core::frame_transform& transform){impl_->push(transform);}\r
 314 void image_mixer::visit(core::data_frame& frame){impl_->visit(frame);}\r
 315 void image_mixer::pop(){impl_->pop();}\r
 316 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
 317 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}\r
 318 void image_mixer::end_layer(){impl_->end_layer();}\r
 319 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);}\r
 320 \r
 321 }}}