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/xmm.h"\r
  27 \r
  28 #include <common/assert.h>\r
  29 #include <common/gl/gl_check.h>\r
  30 #include <common/concurrency/async.h>\r
  31 #include <common/memory/memcpy.h>\r
  32 \r
  33 #include <core/frame/frame.h>\r
  34 #include <core/frame/frame_transform.h>\r
  35 #include <core/frame/pixel_format.h>\r
  36 #include <core/video_format.h>\r
  37 \r
  38 #include <modules/ffmpeg/producer/util/util.h>\r
  39 \r
  40 #include <asmlib.h>\r
  41 \r
  42 #include <gl/glew.h>\r
  43 \r
  44 #include <tbb/cache_aligned_allocator.h>\r
  45 #include <tbb/parallel_for_each.h>\r
  46 #include <tbb/concurrent_queue.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::array<const uint8_t*, 4>   data;\r
  78         core::image_transform                   transform;\r
  79 \r
  80         item()\r
  81                 : pix_desc(core::pixel_format::invalid)\r
  82         {\r
  83                 data.fill(0);\r
  84         }\r
  85 };\r
  86 \r
  87 bool operator==(const item& lhs, const item& rhs)\r
  88 {\r
  89         return lhs.data == rhs.data && 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 // 100% accurate blending with correct rounding.\r
  98 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)\r
  99 {       \r
 100         using namespace xmm;\r
 101                 \r
 102         // C(S, D) = S + D - (((T >> 8) + T) >> 8);\r
 103         // T(S, D) = S * D[A] + 0x80\r
 104 \r
 105         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
 106         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).\r
 107 \r
 108         auto xaxa       = s16_x(aaaa) >> 8;             \r
 109                               \r
 110         auto t1         = s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;    \r
 111         auto t2         = s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;\r
 112                 \r
 113         auto xyxy       = s8_x(((t1 >> 8) + t1) >> 8);      \r
 114         auto yxyx       = s8_x((t2 >> 8) + t2);    \r
 115         auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));\r
 116 \r
 117         return s8_x(s) + (d - argb);\r
 118 }\r
 119         \r
 120 template<typename temporal, typename alignment>\r
 121 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)\r
 122 {                       \r
 123         using namespace xmm;\r
 124 \r
 125         for(auto n = 0; n < count; n += 32)    \r
 126         {\r
 127                 auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);\r
 128                 auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);\r
 129 \r
 130                 auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);\r
 131                 auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);\r
 132                 \r
 133                 auto argb0 = blend(d0, s0);\r
 134                 auto argb1 = blend(d1, s1);\r
 135 \r
 136                 s8_x::store<temporal, alignment>(argb0, dest+n+0 );\r
 137                 s8_x::store<temporal, alignment>(argb1, dest+n+16);\r
 138         } \r
 139 }\r
 140 \r
 141 template<typename temporal>\r
 142 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)\r
 143 {                       \r
 144         using namespace xmm;\r
 145 \r
 146         if(reinterpret_cast<int>(dest) % 16 != 0 || reinterpret_cast<int>(source) % 16 != 0)\r
 147                 kernel<temporal_tag, unaligned_tag>(dest, source, count);\r
 148         else\r
 149                 kernel<temporal_tag, aligned_tag>(dest, source, count);\r
 150 }\r
 151 \r
 152 class image_renderer\r
 153 {\r
 154         tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>  sws_devices_;\r
 155         tbb::concurrent_bounded_queue<spl::shared_ptr<buffer>>                                                                                  temp_buffers_;\r
 156 public: \r
 157         boost::unique_future<core::const_array> operator()(std::vector<item> items, const core::video_format_desc& format_desc)\r
 158         {       \r
 159                 convert(items, format_desc.width, format_desc.height);          \r
 160                 \r
 161                 auto result = spl::make_shared<buffer>(format_desc.size, 0);\r
 162                 if(format_desc.field_mode != core::field_mode::progressive)\r
 163                 {                       \r
 164                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);\r
 165                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);\r
 166                 }\r
 167                 else\r
 168                 {\r
 169                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);\r
 170                 }\r
 171 \r
 172                 temp_buffers_.clear();\r
 173                 \r
 174                 return async(launch_policy::deferred, [=]\r
 175                 {\r
 176                         return core::const_array(result->data(), format_desc.size, result);\r
 177                 });     \r
 178         }\r
 179 \r
 180 private:\r
 181 \r
 182         void draw(std::vector<item> items, uint8_t* dest, std::size_t width, std::size_t height, core::field_mode field_mode)\r
 183         {               \r
 184                 BOOST_FOREACH(auto& item, items)\r
 185                         item.transform.field_mode &= field_mode;\r
 186                 \r
 187                 // Remove empty items.\r
 188                 boost::range::remove_erase_if(items, [&](const item& item)\r
 189                 {\r
 190                         return item.transform.field_mode == core::field_mode::empty;\r
 191                 });\r
 192                 \r
 193                 // Remove first field stills.\r
 194                 boost::range::remove_erase_if(items, [&](const item& item)\r
 195                 {\r
 196                         return item.transform.is_still && item.transform.field_mode == field_mode; // only us last field for stills.\r
 197                 });\r
 198 \r
 199                 if(items.empty())\r
 200                         return;\r
 201                 \r
 202                 auto start = field_mode == core::field_mode::lower ? 1 : 0;\r
 203                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;\r
 204                 \r
 205                 // TODO: Add support for fill translations.\r
 206                 // TODO: Add support for mask rect.\r
 207                 // TODO: Add support for opacity.\r
 208                 // TODO: Add support for mix transition.\r
 209                 // TODO: Add support for push transition.\r
 210                 // TODO: Add support for wipe transition.\r
 211                 // TODO: Add support for slide transition.\r
 212                 tbb::parallel_for(tbb::blocked_range<std::size_t>(0, height/step), [&](const tbb::blocked_range<std::size_t>& r)\r
 213                 {\r
 214                         for(auto i = r.begin(); i != r.end(); ++i)\r
 215                         {\r
 216                                 auto y = i*step+start;\r
 217 \r
 218                                 for(std::size_t n = 0; n < items.size()-1; ++n)\r
 219                                         kernel<xmm::temporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);\r
 220                                 \r
 221                                 std::size_t n = items.size()-1;                         \r
 222                                 kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);\r
 223                         }\r
 224 \r
 225                         _mm_mfence();\r
 226                 });\r
 227         }\r
 228                 \r
 229         void convert(std::vector<item>& source_items, int width, int height)\r
 230         {\r
 231                 std::set<std::array<const uint8_t*, 4>> buffers;\r
 232 \r
 233                 BOOST_FOREACH(auto& item, source_items)\r
 234                         buffers.insert(item.data);\r
 235                 \r
 236                 auto dest_items = source_items;\r
 237 \r
 238                 tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::array<const uint8_t*, 4>& data)\r
 239                 {                       \r
 240                         auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.data == data;})->pix_desc;\r
 241 \r
 242                         if(pix_desc.format == core::pixel_format::bgra && \r
 243                                 pix_desc.planes.at(0).width == width &&\r
 244                                 pix_desc.planes.at(0).height == height)\r
 245                                 return;\r
 246 \r
 247                         std::array<uint8_t*, 4> data2 = {};\r
 248                         for(std::size_t n = 0; n < data.size(); ++n)\r
 249                                 data2.at(n) = const_cast<uint8_t*>(data[n]);\r
 250 \r
 251                         auto input_av_frame = ffmpeg::make_av_frame(data2, 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_devices_[key];\r
 256 \r
 257                         std::shared_ptr<SwsContext> sws_device;\r
 258                         if(!pool.try_pop(sws_device))\r
 259                         {\r
 260                                 double param;\r
 261                                 sws_device.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_device)                         \r
 265                                 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling device.") << boost::errinfo_api_function("sws_getContext"));                            \r
 266                 \r
 267                         auto dest_frame = spl::make_shared<buffer>(width*height*4);\r
 268                         temp_buffers_.push(dest_frame);\r
 269 \r
 270                         {\r
 271                                 spl::shared_ptr<AVFrame> dest_av_frame(avcodec_alloc_frame(), av_free); \r
 272                                 avcodec_get_frame_defaults(dest_av_frame.get());                        \r
 273                                 avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), PIX_FMT_BGRA, width, height);\r
 274                                 \r
 275                                 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);                           \r
 276                                 pool.push(sws_device);\r
 277                         }\r
 278                                         \r
 279                         for(std::size_t n = 0; n < source_items.size(); ++n)\r
 280                         {\r
 281                                 if(source_items[n].data == data)\r
 282                                 {\r
 283                                         dest_items[n].data.assign(0);\r
 284                                         dest_items[n].data[0]                   = dest_frame->data();\r
 285                                         dest_items[n].pix_desc                  = core::pixel_format_desc(core::pixel_format::bgra);\r
 286                                         dest_items[n].pix_desc.planes   = boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));\r
 287                                         dest_items[n].transform                 = source_items[n].transform;\r
 288                                 }\r
 289                         }\r
 290                 });     \r
 291 \r
 292                 source_items = std::move(dest_items);\r
 293         }\r
 294 };\r
 295                 \r
 296 struct image_mixer::impl : boost::noncopyable\r
 297 {       \r
 298         image_renderer                                          renderer_;\r
 299         std::vector<core::image_transform>      transform_stack_;\r
 300         std::vector<item>                                       items_; // layer/stream/items\r
 301 public:\r
 302         impl() \r
 303                 : transform_stack_(1)   \r
 304         {\r
 305                 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";\r
 306         }\r
 307 \r
 308         void begin_layer(core::blend_mode blend_mode)\r
 309         {\r
 310         }\r
 311                 \r
 312         void push(const core::frame_transform& transform)\r
 313         {\r
 314                 transform_stack_.push_back(transform_stack_.back()*transform.image_transform);\r
 315         }\r
 316                 \r
 317         void visit(const core::mutable_frame& frame)\r
 318         {                       \r
 319                 if(frame.pixel_format_desc().format == core::pixel_format::invalid)\r
 320                         return;\r
 321 \r
 322                 if(frame.pixel_format_desc().planes.empty())\r
 323                         return;\r
 324                 \r
 325                 if(frame.pixel_format_desc().planes.at(0).size < 16)\r
 326                         return;\r
 327 \r
 328                 if(transform_stack_.back().field_mode == core::field_mode::empty)\r
 329                         return;\r
 330 \r
 331                 item item;\r
 332                 item.pix_desc   = frame.pixel_format_desc();\r
 333                 item.transform  = transform_stack_.back();\r
 334                 for(int n = 0; n < item.pix_desc.planes.size(); ++n)\r
 335                         item.data.at(n) = frame.image_data(n).begin();          \r
 336 \r
 337                 items_.push_back(item);\r
 338         }\r
 339 \r
 340         void pop()\r
 341         {\r
 342                 transform_stack_.pop_back();\r
 343         }\r
 344 \r
 345         void end_layer()\r
 346         {               \r
 347         }\r
 348         \r
 349         boost::unique_future<core::const_array> render(const core::video_format_desc& format_desc)\r
 350         {\r
 351                 return renderer_(std::move(items_), format_desc);\r
 352         }\r
 353         \r
 354         virtual core::mutable_frame create_frame(const void* tag, const core::pixel_format_desc& desc, double frame_rate, core::field_mode field_mode)\r
 355         {\r
 356                 std::vector<core::mutable_array> buffers;\r
 357                 BOOST_FOREACH(auto& plane, desc.planes)\r
 358                 {\r
 359                         auto buf = spl::make_shared<buffer>(plane.size);\r
 360                         buffers.push_back(core::mutable_array(buf->data(), plane.size, buf));\r
 361                 }\r
 362                 return core::mutable_frame(std::move(buffers), core::audio_buffer(), tag, desc, frame_rate, field_mode);\r
 363         }\r
 364 };\r
 365 \r
 366 image_mixer::image_mixer() : impl_(new impl()){}\r
 367 image_mixer::~image_mixer(){}\r
 368 void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}\r
 369 void image_mixer::visit(const core::mutable_frame& frame){impl_->visit(frame);}\r
 370 void image_mixer::pop(){impl_->pop();}\r
 371 boost::unique_future<core::const_array> image_mixer::operator()(const core::video_format_desc& format_desc){return impl_->render(format_desc);}\r
 372 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}\r
 373 void image_mixer::end_layer(){impl_->end_layer();}\r
 374 core::mutable_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
 375 \r
 376 }}}