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/xmm.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 <intrin.h>\r
  55 #include <stdint.h>\r
  56 #include <algorithm>\r
  57 #include <vector>\r
  58 \r
  59 #if defined(_MSC_VER)\r
  60 #pragma warning (push)\r
  61 #pragma warning (disable : 4244)\r
  62 #endif\r
  63 extern "C" \r
  64 {\r
  65         #include <libswscale/swscale.h>\r
  66         #include <libavcodec/avcodec.h>\r
  67         #include <libavformat/avformat.h>\r
  68 }\r
  69 #if defined(_MSC_VER)\r
  70 #pragma warning (pop)\r
  71 #endif\r
  72 \r
  73 namespace caspar { namespace accelerator { namespace cpu {\r
  74                 \r
  75 struct item\r
  76 {\r
  77         core::pixel_format_desc                                         pix_desc;\r
  78         std::vector<spl::shared_ptr<host_buffer>>       buffers;\r
  79         core::frame_transform                                           transform;\r
  80 \r
  81         item()\r
  82                 : pix_desc(core::pixel_format::invalid)\r
  83         {\r
  84         }\r
  85 };\r
  86 \r
  87 bool operator==(const item& lhs, const item& rhs)\r
  88 {\r
  89         return lhs.buffers == rhs.buffers && lhs.transform == rhs.transform;\r
  90 }\r
  91 \r
  92 bool operator!=(const item& lhs, const item& rhs)\r
  93 {\r
  94         return !(lhs == rhs);\r
  95 }\r
  96         \r
  97 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)\r
  98 {       \r
  99         using namespace xmm;\r
 100                 \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
 104 \r
 105         auto xaxa       = s16_x(aaaa) >> 8;             \r
 106                               \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
 109                 \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
 114 \r
 115         return s8_x(s) + (d - argb);\r
 116 }\r
 117         \r
 118 template<typename write_tag>\r
 119 static void kernel(uint8_t* dest, const uint8_t* source, size_t count, const core::frame_transform& transform)\r
 120 {                       \r
 121         using namespace xmm;\r
 122 \r
 123         for(auto n = 0; n < count; n += 32)    \r
 124         {\r
 125                 auto s0 = s8_x::load(dest+n+0);\r
 126                 auto s1 = s8_x::load(dest+n+16);\r
 127 \r
 128                 auto d0 = s8_x::load(source+n+0);\r
 129                 auto d1 = s8_x::load(source+n+16);\r
 130                 \r
 131                 auto argb0 = blend(d0, s0);\r
 132                 auto argb1 = blend(d1, s1);\r
 133 \r
 134                 s8_x::write(argb0, dest+n+0 , write_tag());\r
 135                 s8_x::write(argb1, dest+n+16, write_tag());\r
 136         } \r
 137 }\r
 138 \r
 139 class image_renderer\r
 140 {\r
 141         std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>>               last_image_;\r
 142         tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>  sws_contexts_;\r
 143 public: \r
 144         boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)\r
 145         {       \r
 146                 if(last_image_.first == items && last_image_.second.has_value())\r
 147                         return last_image_.second;\r
 148 \r
 149                 auto image      = render(items, format_desc);\r
 150                 last_image_ = std::make_pair(std::move(items), image);\r
 151                 return image;\r
 152         }\r
 153 \r
 154 private:\r
 155         boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)\r
 156         {\r
 157                 convert(items, format_desc.width, format_desc.height);          \r
 158                 \r
 159                 auto result = spl::make_shared<host_buffer>(format_desc.size, 0);\r
 160                 if(format_desc.field_mode != core::field_mode::progressive)\r
 161                 {                       \r
 162                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);\r
 163                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);\r
 164                 }\r
 165                 else\r
 166                 {\r
 167                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);\r
 168                 }\r
 169                 \r
 170                 return async(launch_policy::deferred, [=]\r
 171                 {\r
 172                         return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);\r
 173                 });             \r
 174         }\r
 175 \r
 176         void draw(std::vector<item> items, uint8_t* dest, int width, int height, core::field_mode field_mode)\r
 177         {               \r
 178                 BOOST_FOREACH(auto& item, items)\r
 179                         item.transform.field_mode &= field_mode;\r
 180 \r
 181                 boost::remove_erase_if(items, [](item& item){return item.transform.field_mode == core::field_mode::empty;});\r
 182 \r
 183                 if(items.empty())\r
 184                         return;\r
 185                 \r
 186                 auto start = field_mode == core::field_mode::lower ? 1 : 0;\r
 187                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;\r
 188                 \r
 189                 // TODO: Add support for fill translations.\r
 190                 // TODO: Add support for mask rect.\r
 191                 // TODO: Add support for opacity.\r
 192                 // TODO: Add support for mix transition.\r
 193                 // TODO: Add support for push transition.\r
 194                 // TODO: Add support for wipe transition.\r
 195                 // TODO: Add support for slide transition.\r
 196                 tbb::parallel_for(tbb::blocked_range<int>(0, height/step), [&](const tbb::blocked_range<int>& r)\r
 197                 {\r
 198                         for(auto n = r.begin(); n != r.end(); ++n)\r
 199                         {\r
 200                                 auto y = n*step+start;\r
 201 \r
 202                                 auto it = items.begin();\r
 203                                 for(; it != items.end()-1; ++it)                        \r
 204                                         kernel<xmm::store_tag>(dest + y*width*4, it->buffers.at(0)->data() + y*width*4, width*4, it->transform);\r
 205 \r
 206                                 kernel<xmm::stream_tag>(dest + y*width*4, it->buffers.at(0)->data() + y*width*4, width*4, it->transform);\r
 207                         }\r
 208                 });\r
 209         }\r
 210         \r
 211         void convert(std::vector<item>& items, int width, int height)\r
 212         {\r
 213                 typedef std::map<std::vector<spl::shared_ptr<host_buffer>>, std::shared_ptr<host_buffer>>       buffer_map_t;\r
 214                 typedef std::map<std::vector<spl::shared_ptr<host_buffer>>, core::pixel_format_desc>            pix_desc_map_t;\r
 215 \r
 216                 buffer_map_t     buffer_map;\r
 217                 pix_desc_map_t   pix_desc_map;\r
 218                 BOOST_FOREACH(auto& item, items)\r
 219                 {\r
 220                         buffer_map[item.buffers] = nullptr;\r
 221                         pix_desc_map[item.buffers] = item.pix_desc;\r
 222                 }\r
 223 \r
 224                 // TODO: Don't convert buffers multiple times just because they are in different items due to e.g. interlacing.\r
 225                 tbb::parallel_for_each(buffer_map.begin(), buffer_map.end(), [&](buffer_map_t::value_type& pair)\r
 226                 {\r
 227                         auto buffers   = pair.first;\r
 228                         auto pix_desc  = pix_desc_map[buffers];\r
 229 \r
 230                         if(pix_desc.format == core::pixel_format::bgra && \r
 231                            pix_desc.planes.at(0).width == width &&\r
 232                            pix_desc.planes.at(0).height == height)\r
 233                                 return;\r
 234 \r
 235                         auto input_av_frame = ffmpeg::make_av_frame(buffers, pix_desc);\r
 236                                                                 \r
 237                         int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);\r
 238                                                 \r
 239                         auto& pool = sws_contexts_[key];\r
 240 \r
 241                         std::shared_ptr<SwsContext> sws_context;\r
 242                         if(!pool.try_pop(sws_context))\r
 243                         {\r
 244                                 double param;\r
 245                                 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
 246                         }\r
 247                         \r
 248                         if(!sws_context)                                \r
 249                                 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));                           \r
 250                 \r
 251                         auto dest = spl::make_shared<host_buffer>(width*height*4);\r
 252 \r
 253                         {\r
 254                                 spl::shared_ptr<AVFrame> av_frame(avcodec_alloc_frame(), av_free);      \r
 255                                 avcodec_get_frame_defaults(av_frame.get());                     \r
 256                                 avpicture_fill(reinterpret_cast<AVPicture*>(av_frame.get()), dest->data(), PIX_FMT_BGRA, width, height);\r
 257                                 \r
 258                                 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
 259                                 pool.push(sws_context);\r
 260                         }\r
 261 \r
 262                         pair.second = dest;\r
 263                 });\r
 264                 \r
 265                 BOOST_FOREACH(auto& item, items)\r
 266                 {               \r
 267                         item.buffers                    = boost::assign::list_of(spl::make_shared_ptr(buffer_map[item.buffers]));\r
 268                         item.pix_desc                   = core::pixel_format_desc(core::pixel_format::bgra);\r
 269                         item.pix_desc.planes    = boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));\r
 270                 }\r
 271         }\r
 272 };\r
 273                 \r
 274 struct image_mixer::impl : boost::noncopyable\r
 275 {       \r
 276         image_renderer                                          renderer_;\r
 277         std::vector<core::frame_transform>      transform_stack_;\r
 278         std::vector<item>                                       items_; // layer/stream/items\r
 279 public:\r
 280         impl() \r
 281                 : transform_stack_(1)   \r
 282         {\r
 283                 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";\r
 284         }\r
 285 \r
 286         void begin_layer(core::blend_mode blend_mode)\r
 287         {\r
 288         }\r
 289                 \r
 290         void push(core::frame_transform& transform)\r
 291         {\r
 292                 transform_stack_.push_back(transform_stack_.back()*transform);\r
 293         }\r
 294                 \r
 295         void visit(core::data_frame& frame2)\r
 296         {                       \r
 297                 write_frame* frame = dynamic_cast<write_frame*>(&frame2);\r
 298                 if(frame == nullptr)\r
 299                         return;\r
 300 \r
 301                 if(frame->get_pixel_format_desc().format == core::pixel_format::invalid)\r
 302                         return;\r
 303 \r
 304                 if(frame->get_buffers().empty())\r
 305                         return;\r
 306 \r
 307                 if(transform_stack_.back().field_mode == core::field_mode::empty)\r
 308                         return;\r
 309 \r
 310                 item item;\r
 311                 item.pix_desc                   = frame->get_pixel_format_desc();\r
 312                 item.buffers                    = frame->get_buffers();                         \r
 313                 item.transform                  = transform_stack_.back();\r
 314                 item.transform.volume   = core::frame_transform().volume; // Set volume to default since we don't care about it here.\r
 315 \r
 316                 items_.push_back(item);\r
 317         }\r
 318 \r
 319         void pop()\r
 320         {\r
 321                 transform_stack_.pop_back();\r
 322         }\r
 323 \r
 324         void end_layer()\r
 325         {               \r
 326         }\r
 327         \r
 328         boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)\r
 329         {\r
 330                 return renderer_(std::move(items_), format_desc);\r
 331         }\r
 332         \r
 333         virtual spl::shared_ptr<cpu::write_frame> create_frame(const void* tag, const core::pixel_format_desc& desc)\r
 334         {\r
 335                 return spl::make_shared<cpu::write_frame>(tag, desc);\r
 336         }\r
 337 };\r
 338 \r
 339 image_mixer::image_mixer() : impl_(new impl()){}\r
 340 void image_mixer::push(core::frame_transform& transform){impl_->push(transform);}\r
 341 void image_mixer::visit(core::data_frame& frame){impl_->visit(frame);}\r
 342 void image_mixer::pop(){impl_->pop();}\r
 343 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
 344 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}\r
 345 void image_mixer::end_layer(){impl_->end_layer();}\r
 346 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
 347 \r
 348 }}}