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/simd.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_epi8 blend(xmm_epi8 dest, xmm_epi8 source)\r
  98 {       \r
  99         auto s = xmm_cast<xmm_epi16>(source);\r
 100         auto d = dest;\r
 101 \r
 102         const xmm_epi16 round   = 128;\r
 103         const xmm_epi16 lomask  = 0x00FF;\r
 104 \r
 105         // T(S, D) = S * D[A] + 0x80\r
 106         auto aaaa   = xmm_epi8::shuffle(d, xmm_epi8(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));\r
 107         d                       = xmm_epi8::umin(d, aaaa); // overflow guard\r
 108 \r
 109         auto xaxa       = xmm_cast<xmm_epi16>(aaaa) & lomask;           \r
 110                               \r
 111         auto xrxb       = s & lomask;\r
 112         auto t1         = xmm_epi16::multiply_low(xrxb, xaxa) + round;    \r
 113                         \r
 114         auto xaxg       = s >> 8;\r
 115         auto t2         = xmm_epi16::multiply_low(xaxg, xaxa) + round;\r
 116                 \r
 117         // C(S, D) = S + D - (((T >> 8) + T) >> 8);\r
 118         auto rxbx       = xmm_cast<xmm_epi8>(((t1 >> 8) + t1) >> 8);      \r
 119         auto axgx       = xmm_cast<xmm_epi8>((t2 >> 8) + t2);    \r
 120         auto argb   = xmm_epi8::blend(rxbx, axgx, xmm_epi8(-1, 0, -1, 0));\r
 121 \r
 122         return xmm_cast<xmm_epi8>(s) + (d - argb);\r
 123 }\r
 124         \r
 125 template<typename write_op>\r
 126 static void kernel(uint8_t* dest, const uint8_t* source, size_t count, const core::frame_transform& transform)\r
 127 {                               \r
 128         for(auto n = 0; n < count; n += 32)    \r
 129         {\r
 130                 auto s0 = xmm_epi8::load(dest+n+0);\r
 131                 auto s1 = xmm_epi8::load(dest+n+16);\r
 132 \r
 133                 auto d0 = xmm_epi8::load(source+n+0);\r
 134                 auto d1 = xmm_epi8::load(source+n+16);\r
 135                 \r
 136                 auto argb0 = blend(d0, s0);\r
 137                 auto argb1 = blend(d1, s1);\r
 138 \r
 139                 xmm_epi8::write<write_op>(argb0, dest+n+0);\r
 140                 xmm_epi8::write<write_op>(argb1, dest+n+16);\r
 141         } \r
 142 }\r
 143 \r
 144 class image_renderer\r
 145 {\r
 146         std::pair<std::vector<item>, boost::shared_future<boost::iterator_range<const uint8_t*>>>               last_image_;\r
 147         tbb::concurrent_unordered_map<int, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>  sws_contexts_;\r
 148 public: \r
 149         boost::shared_future<boost::iterator_range<const uint8_t*>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)\r
 150         {       \r
 151                 if(last_image_.first == items && last_image_.second.has_value())\r
 152                         return last_image_.second;\r
 153 \r
 154                 auto image      = render(items, format_desc);\r
 155                 last_image_ = std::make_pair(std::move(items), image);\r
 156                 return image;\r
 157         }\r
 158 \r
 159 private:\r
 160         boost::shared_future<boost::iterator_range<const uint8_t*>> render(std::vector<item> items, const core::video_format_desc& format_desc)\r
 161         {\r
 162                 convert(items, format_desc.width, format_desc.height);          \r
 163                 \r
 164                 auto result = spl::make_shared<host_buffer>(format_desc.size, 0);\r
 165                 if(format_desc.field_mode != core::field_mode::progressive)\r
 166                 {                       \r
 167                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);\r
 168                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);\r
 169                 }\r
 170                 else\r
 171                 {\r
 172                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);\r
 173                 }\r
 174                 \r
 175                 return async(launch_policy::deferred, [=]\r
 176                 {\r
 177                         return boost::iterator_range<const uint8_t*>(result->data(), result->data() + format_desc.size);\r
 178                 });             \r
 179         }\r
 180 \r
 181         void draw(std::vector<item> items, uint8_t* dest, int width, int height, core::field_mode field_mode)\r
 182         {               \r
 183                 BOOST_FOREACH(auto& item, items)\r
 184                         item.transform.field_mode &= field_mode;\r
 185 \r
 186                 boost::remove_erase_if(items, [](item& item){return item.transform.field_mode == core::field_mode::empty;});\r
 187 \r
 188                 if(items.empty())\r
 189                         return;\r
 190 \r
 191                 static const int CACHE_SIZE = 8192;\r
 192 \r
 193                 auto start = field_mode == core::field_mode::lower ? 1 : 0;\r
 194                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;\r
 195                 \r
 196                 // TODO: Add support for fill translations.\r
 197                 // TODO: Add support for mask rect.\r
 198                 // TODO: Add support for opacity.\r
 199                 // TODO: Add support for mix transition.\r
 200                 // TODO: Add support for push transition.\r
 201                 // TODO: Add support for wipe transition.\r
 202                 // TODO: Add support for slide transition.\r
 203                 tbb::parallel_for(tbb::blocked_range<int>(0, height/step, CACHE_SIZE/(width*4)), [&](const tbb::blocked_range<int>& r)\r
 204                 {\r
 205                         for(auto n = r.begin(); n != r.end(); ++n)\r
 206                         {\r
 207                                 auto y = n*step+start;\r
 208 \r
 209                                 auto it = items.begin();\r
 210                                 for(; it != items.end()-1; ++it)                        \r
 211                                         kernel<store_write>(dest + y*width*4, it->buffers.at(0)->data() + y*width*4, width*4, it->transform);\r
 212 \r
 213                                 kernel<stream_write>(dest + y*width*4, it->buffers.at(0)->data() + y*width*4, width*4, it->transform);\r
 214                         }\r
 215                 });\r
 216         }\r
 217         \r
 218         void convert(std::vector<item>& items, int width, int height)\r
 219         {\r
 220                 // TODO: Don't convert buffers multiple times just because they are in different items due to e.g. interlacing.\r
 221                 tbb::parallel_for_each(items.begin(), items.end(), [&](item& item)\r
 222                 {\r
 223                         if(item.pix_desc.format == core::pixel_format::bgra && \r
 224                            item.pix_desc.planes.at(0).width == width &&\r
 225                            item.pix_desc.planes.at(0).height == height)\r
 226                                 return;\r
 227 \r
 228                         auto input_av_frame = ffmpeg::make_av_frame(item.buffers, item.pix_desc);\r
 229                                                                 \r
 230                         int key = ((input_av_frame->width << 22) & 0xFFC00000) | ((input_av_frame->height << 6) & 0x003FC000) | ((input_av_frame->format << 7) & 0x00007F00);\r
 231                                                 \r
 232                         auto& pool = sws_contexts_[key];\r
 233 \r
 234                         std::shared_ptr<SwsContext> sws_context;\r
 235                         if(!pool.try_pop(sws_context))\r
 236                         {\r
 237                                 double param;\r
 238                                 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
 239                         }\r
 240                         \r
 241                         if(!sws_context)                                \r
 242                                 BOOST_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling context.") << boost::errinfo_api_function("sws_getContext"));                           \r
 243                 \r
 244                         auto dest = spl::make_shared<host_buffer>(width*height*4);\r
 245 \r
 246                         spl::shared_ptr<AVFrame> av_frame(avcodec_alloc_frame(), av_free);      \r
 247                         avcodec_get_frame_defaults(av_frame.get());                     \r
 248                         avpicture_fill(reinterpret_cast<AVPicture*>(av_frame.get()), dest->data(), PIX_FMT_BGRA, width, height);\r
 249                                 \r
 250                         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
 251                         \r
 252                         item.buffers.clear();\r
 253                         item.buffers.push_back(dest);\r
 254                         item.pix_desc = core::pixel_format_desc(core::pixel_format::bgra);\r
 255                         item.pix_desc.planes.clear();\r
 256                         item.pix_desc.planes.push_back(core::pixel_format_desc::plane(width, height, 4));\r
 257 \r
 258                         pool.push(sws_context);\r
 259                 });\r
 260         }\r
 261 };\r
 262                 \r
 263 struct image_mixer::impl : boost::noncopyable\r
 264 {       \r
 265         image_renderer                                          renderer_;\r
 266         std::vector<core::frame_transform>      transform_stack_;\r
 267         std::vector<item>                                       items_; // layer/stream/items\r
 268 public:\r
 269         impl() \r
 270                 : transform_stack_(1)   \r
 271         {\r
 272                 CASPAR_LOG(info) << L"Initialized CPU Accelerated Image Mixer";\r
 273         }\r
 274 \r
 275         void begin_layer(core::blend_mode blend_mode)\r
 276         {\r
 277         }\r
 278                 \r
 279         void push(core::frame_transform& transform)\r
 280         {\r
 281                 transform_stack_.push_back(transform_stack_.back()*transform);\r
 282         }\r
 283                 \r
 284         void visit(core::data_frame& frame2)\r
 285         {                       \r
 286                 write_frame* frame = dynamic_cast<write_frame*>(&frame2);\r
 287                 if(frame == nullptr)\r
 288                         return;\r
 289 \r
 290                 if(frame->get_pixel_format_desc().format == core::pixel_format::invalid)\r
 291                         return;\r
 292 \r
 293                 if(frame->get_buffers().empty())\r
 294                         return;\r
 295 \r
 296                 if(transform_stack_.back().field_mode == core::field_mode::empty)\r
 297                         return;\r
 298 \r
 299                 item item;\r
 300                 item.pix_desc                   = frame->get_pixel_format_desc();\r
 301                 item.buffers                    = frame->get_buffers();                         \r
 302                 item.transform                  = transform_stack_.back();\r
 303                 item.transform.volume   = core::frame_transform().volume; // Set volume to default since we don't care about it here.\r
 304 \r
 305                 items_.push_back(item);\r
 306         }\r
 307 \r
 308         void pop()\r
 309         {\r
 310                 transform_stack_.pop_back();\r
 311         }\r
 312 \r
 313         void end_layer()\r
 314         {               \r
 315         }\r
 316         \r
 317         boost::shared_future<boost::iterator_range<const uint8_t*>> render(const core::video_format_desc& format_desc)\r
 318         {\r
 319                 return renderer_(std::move(items_), format_desc);\r
 320         }\r
 321         \r
 322         virtual spl::shared_ptr<cpu::write_frame> create_frame(const void* tag, const core::pixel_format_desc& desc)\r
 323         {\r
 324                 return spl::make_shared<cpu::write_frame>(tag, desc);\r
 325         }\r
 326 };\r
 327 \r
 328 image_mixer::image_mixer() : impl_(new impl()){}\r
 329 void image_mixer::push(core::frame_transform& transform){impl_->push(transform);}\r
 330 void image_mixer::visit(core::data_frame& frame){impl_->visit(frame);}\r
 331 void image_mixer::pop(){impl_->pop();}\r
 332 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
 333 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}\r
 334 void image_mixer::end_layer(){impl_->end_layer();}\r
 335 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
 336 \r
 337 }}}