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 <algorithm>\r
  55 #include <stdint.h>\r
  56 #include <vector>\r
  57 \r
  58 #if defined(_MSC_VER)\r
  59 #pragma warning (push)\r
  60 #pragma warning (disable : 4244)\r
  61 #endif\r
  62 extern "C" \r
  63 {\r
  64         #include <libswscale/swscale.h>\r
  65         #include <libavcodec/avcodec.h>\r
  66         #include <libavformat/avformat.h>\r
  67 }\r
  68 #if defined(_MSC_VER)\r
  69 #pragma warning (pop)\r
  70 #endif\r
  71 \r
  72 namespace caspar { namespace accelerator { namespace cpu {\r
  73                 \r
  74 struct item\r
  75 {\r
  76         core::pixel_format_desc                                         pix_desc;\r
  77         std::vector<spl::shared_ptr<host_buffer>>       buffers;\r
  78         core::image_transform                                           transform;\r
  79 \r
  80         item()\r
  81                 : pix_desc(core::pixel_format::invalid)\r
  82         {\r
  83         }\r
  84 };\r
  85 \r
  86 bool operator==(const item& lhs, const item& rhs)\r
  87 {\r
  88         return lhs.buffers == rhs.buffers && lhs.transform == rhs.transform;\r
  89 }\r
  90 \r
  91 bool operator!=(const item& lhs, const item& rhs)\r
  92 {\r
  93         return !(lhs == rhs);\r
  94 }\r
  95         \r
  96 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)\r
  97 {       \r
  98         using namespace xmm;\r
  99                 \r
 100         // T(S, D) = S * D[A] + 0x80\r
 101         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
 102         d                       = s8_x(u8_x::min(u8_x(d), u8_x(aaaa))); // overflow guard\r
 103 \r
 104         auto xaxa       = s16_x(aaaa) >> 8;             \r
 105                               \r
 106         auto t1         = s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;    \r
 107         auto t2         = s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;\r
 108                 \r
 109         // C(S, D) = S + D - (((T >> 8) + T) >> 8);\r
 110         auto xyxy       = s8_x(((t1 >> 8) + t1) >> 8);      \r
 111         auto yxyx       = s8_x((t2 >> 8) + t2);    \r
 112         auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));\r
 113 \r
 114         return s8_x(s) + (d - argb);\r
 115 }\r
 116         \r
 117 template<typename temporal, typename alignment>\r
 118 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)\r
 119 {                       \r
 120         using namespace xmm;\r
 121 \r
 122         for(auto n = 0; n < count; n += 32)    \r
 123         {\r
 124                 auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);\r
 125                 auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);\r
 126 \r
 127                 auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);\r
 128                 auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);\r
 129                 \r
 130                 auto argb0 = blend(d0, s0);\r
 131                 auto argb1 = blend(d1, s1);\r
 132 \r
 133                 s8_x::store<temporal, alignment>(argb0, dest+n+0 );\r
 134                 s8_x::store<temporal, alignment>(argb1, dest+n+16);\r
 135         } \r
 136 }\r
 137 \r
 138 template<typename temporal>\r
 139 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)\r
 140 {                       \r
 141         using namespace xmm;\r
 142 \r
 143         if(reinterpret_cast<int>(dest) % 16 != 0 || reinterpret_cast<int>(source) % 16 != 0)\r
 144                 kernel<temporal_tag, unaligned_tag>(dest, source, count);\r
 145         else\r
 146                 kernel<temporal_tag, aligned_tag>(dest, source, count);\r
 147 }\r
 148 \r
 149 class image_renderer\r
 150 {\r
 151         std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>>               last_image_;\r
 152         tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>  sws_contexts_;\r
 153 public: \r
 154         boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)\r
 155         {       \r
 156                 if(last_image_.first == items && last_image_.second.has_value())\r
 157                         return last_image_.second;\r
 158 \r
 159                 auto image      = render(items, format_desc);\r
 160                 last_image_ = std::make_pair(std::move(items), image);\r
 161                 return image;\r
 162         }\r
 163 \r
 164 private:\r
 165         boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)\r
 166         {\r
 167                 convert(items, format_desc.width, format_desc.height);          \r
 168                 \r
 169                 auto result = spl::make_shared<host_buffer>(format_desc.size, 0);\r
 170                 if(format_desc.field_mode != core::field_mode::progressive)\r
 171                 {                       \r
 172                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);\r
 173                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);\r
 174                 }\r
 175                 else\r
 176                 {\r
 177                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);\r
 178                 }\r
 179                 \r
 180                 return async(launch_policy::deferred, [=]\r
 181                 {\r
 182                         return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);\r
 183                 });             \r
 184         }\r
 185 \r
 186         void draw(std::vector<item> items, uint8_t* dest, int width, int height, core::field_mode field_mode)\r
 187         {               \r
 188                 BOOST_FOREACH(auto& item, items)\r
 189                         item.transform.field_mode &= field_mode;\r
 190                 \r
 191                 // Remove empty items.\r
 192                 boost::range::remove_erase_if(items, [&](const item& item)\r
 193                 {\r
 194                         return item.transform.field_mode == core::field_mode::empty;\r
 195                 });\r
 196                 \r
 197                 // Remove first field stills.\r
 198                 boost::range::remove_erase_if(items, [&](const item& item)\r
 199                 {\r
 200                         return item.transform.is_still && item.transform.field_mode == field_mode; // only us last field for stills.\r
 201                 });\r
 202 \r
 203                 if(items.empty())\r
 204                         return;\r
 205                 \r
 206                 auto start = field_mode == core::field_mode::lower ? 1 : 0;\r
 207                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;\r
 208                 \r
 209                 // TODO: Add support for fill translations.\r
 210                 // TODO: Add support for mask rect.\r
 211                 // TODO: Add support for opacity.\r
 212                 // TODO: Add support for mix transition.\r
 213                 // TODO: Add support for push transition.\r
 214                 // TODO: Add support for wipe transition.\r
 215                 // TODO: Add support for slide transition.\r
 216                 tbb::parallel_for(tbb::blocked_range<int>(0, height/step), [&](const tbb::blocked_range<int>& r)\r
 217                 {\r
 218                         for(auto i = r.begin(); i != r.end(); ++i)\r
 219                         {\r
 220                                 auto y = i*step+start;\r
 221 \r
 222                                 for(std::size_t n = 0; n < items.size()-1; ++n)\r
 223                                         kernel<xmm::temporal_tag>(dest + y*width*4, items[n].buffers.at(0)->data() + y*width*4, width*4);\r
 224                                 \r
 225                                 std::size_t n = items.size()-1;                         \r
 226                                 kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].buffers.at(0)->data() + y*width*4, width*4);\r
 227                         }\r
 228 \r
 229                         _mm_mfence();\r
 230                 });\r
 231         }\r
 232                 \r
 233         void convert(std::vector<item>& source_items, int width, int height)\r
 234         {\r
 235                 std::set<std::vector<spl::shared_ptr<host_buffer>>> buffers;\r
 236 \r
 237                 BOOST_FOREACH(auto& item, source_items)\r
 238                         buffers.insert(item.buffers);\r
 239                 \r
 240                 auto dest_items = source_items;\r
 241 \r
 242                 tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::vector<spl::shared_ptr<host_buffer>>& buffers)\r
 243                 {                       \r
 244                         auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.buffers == buffers;})->pix_desc;\r
 245 \r
 246                         if(pix_desc.format == core::pixel_format::bgra && \r
 247                                 pix_desc.planes.at(0).width == width &&\r
 248                                 pix_desc.planes.at(0).height == height)\r
 249                                 return;\r
 250 \r
 251                         auto input_av_frame = ffmpeg::make_av_frame(buffers, pix_desc);\r
 252                                                                 \r
 253                         int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);\r
 254                                                 \r
 255                         auto& pool = sws_contexts_[key];\r
 256 \r
 257                         std::shared_ptr<SwsContext> sws_context;\r
 258                         if(!pool.try_pop(sws_context))\r
 259                         {\r
 260                                 double param;\r
 261                                 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
 262                         }\r
 263                         \r
 264                         if(!sws_context)                                \r
 265                                 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));                           \r
 266                 \r
 267                         auto dest_frame = spl::make_shared<host_buffer>(width*height*4);\r
 268 \r
 269                         {\r
 270                                 spl::shared_ptr<AVFrame> dest_av_frame(avcodec_alloc_frame(), av_free); \r
 271                                 avcodec_get_frame_defaults(dest_av_frame.get());                        \r
 272                                 avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), PIX_FMT_BGRA, width, height);\r
 273                                 \r
 274                                 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
 275                                 pool.push(sws_context);\r
 276                         }\r
 277                 \r
 278                         for(std::size_t n = 0; n < source_items.size(); ++n)\r
 279                         {\r
 280                                 if(source_items[n].buffers == buffers)\r
 281                                 {\r
 282                                         dest_items[n].buffers                   = boost::assign::list_of(dest_frame);\r
 283                                         dest_items[n].pix_desc                  = core::pixel_format_desc(core::pixel_format::bgra);\r
 284                                         dest_items[n].pix_desc.planes   = boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));\r
 285                                         dest_items[n].transform                 = source_items[n].transform;\r
 286                                 }\r
 287                         }\r
 288                 });     \r
 289 \r
 290                 source_items = std::move(dest_items);\r
 291         }\r
 292 };\r
 293                 \r
 294 struct image_mixer::impl : boost::noncopyable\r
 295 {       \r
 296         image_renderer                                          renderer_;\r
 297         std::vector<core::image_transform>      transform_stack_;\r
 298         std::vector<item>                                       items_; // layer/stream/items\r
 299 public:\r
 300         impl() \r
 301                 : transform_stack_(1)   \r
 302         {\r
 303                 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";\r
 304         }\r
 305 \r
 306         void begin_layer(core::blend_mode blend_mode)\r
 307         {\r
 308         }\r
 309                 \r
 310         void push(const core::frame_transform& transform)\r
 311         {\r
 312                 transform_stack_.push_back(transform_stack_.back()*transform.image_transform);\r
 313         }\r
 314                 \r
 315         void visit(const core::data_frame& frame2)\r
 316         {                       \r
 317                 auto frame = dynamic_cast<const write_frame*>(&frame2);\r
 318                 if(frame == nullptr)\r
 319                         return;\r
 320 \r
 321                 if(frame->get_pixel_format_desc().format == core::pixel_format::invalid)\r
 322                         return;\r
 323 \r
 324                 if(frame->get_buffers().empty())\r
 325                         return;\r
 326 \r
 327                 if(transform_stack_.back().field_mode == core::field_mode::empty)\r
 328                         return;\r
 329 \r
 330                 item item;\r
 331                 item.pix_desc                   = frame->get_pixel_format_desc();\r
 332                 item.buffers                    = frame->get_buffers();                         \r
 333                 item.transform                  = transform_stack_.back();\r
 334 \r
 335                 items_.push_back(item);\r
 336         }\r
 337 \r
 338         void pop()\r
 339         {\r
 340                 transform_stack_.pop_back();\r
 341         }\r
 342 \r
 343         void end_layer()\r
 344         {               \r
 345         }\r
 346         \r
 347         boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)\r
 348         {\r
 349                 return renderer_(std::move(items_), format_desc);\r
 350         }\r
 351         \r
 352         virtual spl::shared_ptr<cpu::write_frame> create_frame(const void* tag, const core::pixel_format_desc& desc)\r
 353         {\r
 354                 return spl::make_shared<cpu::write_frame>(tag, desc);\r
 355         }\r
 356 };\r
 357 \r
 358 image_mixer::image_mixer() : impl_(new impl()){}\r
 359 void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}\r
 360 void image_mixer::visit(const core::data_frame& frame){impl_->visit(frame);}\r
 361 void image_mixer::pop(){impl_->pop();}\r
 362 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
 363 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}\r
 364 void image_mixer::end_layer(){impl_->end_layer();}\r
 365 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
 366 \r
 367 }}}