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

   1 /*
   2 * Copyright (c) 2011 Sveriges Television AB <info@casparcg.com>
   3 *
   4 * This file is part of CasparCG (www.casparcg.com).
   5 *
   6 * CasparCG is free software: you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License as published by
   8 * the Free Software Foundation, either version 3 of the License, or
   9 * (at your option) any later version.
  10 *
  11 * CasparCG is distributed in the hope that it will be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with CasparCG. If not, see <http://www.gnu.org/licenses/>.
  18 *
  19 * Author: Robert Nagy, ronag89@gmail.com
  20 */
  21
  22 #include "../../stdafx.h"
  23
  24 #include "image_mixer.h"
  25
  26 #include "../util/xmm.h"
  27
  28 #include <common/assert.h>
  29 #include <common/gl/gl_check.h>
  30 #include <common/future.h>
  31 #include <common/array.h>
  32
  33 #include <core/frame/frame.h>
  34 #include <core/frame/frame_transform.h>
  35 #include <core/frame/pixel_format.h>
  36 #include <core/video_format.h>
  37
  38 #include <modules/ffmpeg/producer/util/util.h>
  39
  40 #include <asmlib.h>
  41
  42 #include <gl/glew.h>
  43
  44 #include <tbb/cache_aligned_allocator.h>
  45 #include <tbb/parallel_for.h>
  46 #include <tbb/parallel_for_each.h>
  47 #include <tbb/concurrent_queue.h>
  48
  49 #include <boost/assign.hpp>
  50 #include <boost/foreach.hpp>
  51 #include <boost/range.hpp>
  52 #include <boost/range/algorithm_ext/erase.hpp>
  53 #include <boost/thread/future.hpp>
  54
  55 #include <algorithm>
  56 #include <cstdint>
  57 #include <vector>
  58
  59 #if defined(_MSC_VER)
  60 #pragma warning (push)
  61 #pragma warning (disable : 4244)
  62 #endif
  63 extern "C"
  64 {
  65         #include <libswscale/swscale.h>
  66         #include <libavcodec/avcodec.h>
  67         #include <libavformat/avformat.h>
  68 }
  69 #if defined(_MSC_VER)
  70 #pragma warning (pop)
  71 #endif
  72
  73 namespace caspar { namespace accelerator { namespace cpu {
  74
  75 struct item
  76 {
  77         core::pixel_format_desc                 pix_desc;
  78         std::array<const uint8_t*, 4>   data;
  79         core::image_transform                   transform;
  80
  81         item()
  82                 : pix_desc(core::pixel_format::invalid)
  83         {
  84                 data.fill(0);
  85         }
  86 };
  87
  88 bool operator==(const item& lhs, const item& rhs)
  89 {
  90         return lhs.data == rhs.data && lhs.transform == rhs.transform;
  91 }
  92
  93 bool operator!=(const item& lhs, const item& rhs)
  94 {
  95         return !(lhs == rhs);
  96 }
  97
  98 // 100% accurate blending with correct rounding.
  99 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)
 100 {
 101         using namespace xmm;
 102
 103         // C(S, D) = S + D - (((T >> 8) + T) >> 8);
 104         // T(S, D) = S * D[A] + 0x80
 105
 106         auto aaaa   = s8_x::shuffle(d, s8_x(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));
 107         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).
 108
 109         auto xaxa       = s16_x(aaaa) >> 8;
 110
 111         auto t1         = s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;
 112         auto t2         = s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;
 113
 114         auto xyxy       = s8_x(((t1 >> 8) + t1) >> 8);
 115         auto yxyx       = s8_x((t2 >> 8) + t2);
 116         auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));
 117
 118         return s8_x(s) + (d - argb);
 119 }
 120
 121 template<typename temporal, typename alignment>
 122 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
 123 {
 124         using namespace xmm;
 125
 126         for(auto n = 0; n < count; n += 32)
 127         {
 128                 auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);
 129                 auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);
 130
 131                 auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);
 132                 auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);
 133
 134                 auto argb0 = blend(d0, s0);
 135                 auto argb1 = blend(d1, s1);
 136
 137                 s8_x::store<temporal, alignment>(argb0, dest+n+0 );
 138                 s8_x::store<temporal, alignment>(argb1, dest+n+16);
 139         }
 140 }
 141
 142 template<typename temporal>
 143 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
 144 {
 145         using namespace xmm;
 146
 147         if(reinterpret_cast<int>(dest) % 16 != 0 || reinterpret_cast<int>(source) % 16 != 0)
 148                 kernel<temporal_tag, unaligned_tag>(dest, source, count);
 149         else
 150                 kernel<temporal_tag, aligned_tag>(dest, source, count);
 151 }
 152
 153 class image_renderer
 154 {
 155         tbb::concurrent_unordered_map<int64_t, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>      sws_devices_;
 156         tbb::concurrent_bounded_queue<spl::shared_ptr<buffer>>                                                                                          temp_buffers_;
 157 public:
 158         boost::unique_future<array<const std::uint8_t>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)
 159         {
 160                 convert(items, format_desc.width, format_desc.height);
 161
 162                 // Remove first field stills.
 163                 boost::range::remove_erase_if(items, [&](const item& item)
 164                 {
 165                         return item.transform.is_still && item.transform.field_mode == format_desc.field_mode; // only us last field for stills.
 166                 });
 167
 168                 // Stills are progressive
 169                 BOOST_FOREACH(auto item, items)
 170                 {
 171                         if(item.transform.is_still)
 172                                 item.transform.field_mode = core::field_mode::progressive;
 173                 }
 174
 175                 auto result = spl::make_shared<buffer>(format_desc.size, 0);
 176                 if(format_desc.field_mode != core::field_mode::progressive)
 177                 {
 178                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);
 179                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);
 180                 }
 181                 else
 182                 {
 183                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);
 184                 }
 185
 186                 temp_buffers_.clear();
 187
 188                 return async(launch::deferred, [=]
 189                 {
 190                         return array<const std::uint8_t>(result->data(), format_desc.size, true, result);
 191                 });
 192         }
 193
 194 private:
 195
 196         void draw(std::vector<item> items, uint8_t* dest, std::size_t width, std::size_t height, core::field_mode field_mode)
 197         {
 198                 BOOST_FOREACH(auto& item, items)
 199                         item.transform.field_mode &= field_mode;
 200
 201                 // Remove empty items.
 202                 boost::range::remove_erase_if(items, [&](const item& item)
 203                 {
 204                         return item.transform.field_mode == core::field_mode::empty;
 205                 });
 206
 207                 if(items.empty())
 208                         return;
 209
 210                 auto start = field_mode == core::field_mode::lower ? 1 : 0;
 211                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;
 212
 213                 // TODO: Add support for fill translations.
 214                 // TODO: Add support for mask rect.
 215                 // TODO: Add support for opacity.
 216                 // TODO: Add support for mix transition.
 217                 // TODO: Add support for push transition.
 218                 // TODO: Add support for wipe transition.
 219                 // TODO: Add support for slide transition.
 220                 tbb::parallel_for(tbb::blocked_range<std::size_t>(0, height/step), [&](const tbb::blocked_range<std::size_t>& r)
 221                 {
 222                         for(auto i = r.begin(); i != r.end(); ++i)
 223                         {
 224                                 auto y = i*step+start;
 225
 226                                 for(std::size_t n = 0; n < items.size()-1; ++n)
 227                                         kernel<xmm::temporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
 228
 229                                 std::size_t n = items.size()-1;
 230                                 kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
 231                         }
 232
 233                         _mm_mfence();
 234                 });
 235         }
 236
 237         void convert(std::vector<item>& source_items, int width, int height)
 238         {
 239                 std::set<std::array<const uint8_t*, 4>> buffers;
 240
 241                 BOOST_FOREACH(auto& item, source_items)
 242                         buffers.insert(item.data);
 243
 244                 auto dest_items = source_items;
 245
 246                 tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::array<const uint8_t*, 4>& data)
 247                 {
 248                         auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.data == data;})->pix_desc;
 249
 250                         if(pix_desc.format == core::pixel_format::bgra &&
 251                                 pix_desc.planes.at(0).width == width &&
 252                                 pix_desc.planes.at(0).height == height)
 253                                 return;
 254
 255                         std::array<uint8_t*, 4> data2 = {};
 256                         for(std::size_t n = 0; n < data.size(); ++n)
 257                                 data2.at(n) = const_cast<uint8_t*>(data[n]);
 258
 259                         auto input_av_frame = ffmpeg::make_av_frame(data2, pix_desc);
 260
 261
 262                         int64_t key = ((static_cast<int64_t>(input_av_frame->width)      << 32) & 0xFFFF00000000) |
 263                                                   ((static_cast<int64_t>(input_av_frame->height) << 16) & 0xFFFF0000) |
 264                                                   ((static_cast<int64_t>(input_av_frame->format) <<  8) & 0xFF00);
 265
 266                         auto& pool = sws_devices_[key];
 267
 268                         std::shared_ptr<SwsContext> sws_device;
 269                         if(!pool.try_pop(sws_device))
 270                         {
 271                                 double param;
 272                                 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);
 273                         }
 274
 275                         if(!sws_device)
 276                                 CASPAR_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling device.") << boost::errinfo_api_function("sws_getContext"));
 277
 278                         auto dest_frame = spl::make_shared<buffer>(width*height*4);
 279                         temp_buffers_.push(dest_frame);
 280
 281                         {
 282                                 spl::shared_ptr<AVFrame> dest_av_frame(avcodec_alloc_frame(), av_free);
 283                                 avcodec_get_frame_defaults(dest_av_frame.get());
 284                                 avpicture_fill(reinterpret_cast<AVPicture*>(dest_av_frame.get()), dest_frame->data(), PIX_FMT_BGRA, width, height);
 285
 286                                 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);
 287                                 pool.push(sws_device);
 288                         }
 289
 290                         for(std::size_t n = 0; n < source_items.size(); ++n)
 291                         {
 292                                 if(source_items[n].data == data)
 293                                 {
 294                                         dest_items[n].data.assign(0);
 295                                         dest_items[n].data[0]                   = dest_frame->data();
 296                                         dest_items[n].pix_desc                  = core::pixel_format_desc(core::pixel_format::bgra);
 297                                         dest_items[n].pix_desc.planes   = boost::assign::list_of(core::pixel_format_desc::plane(width, height, 4));
 298                                         dest_items[n].transform                 = source_items[n].transform;
 299                                 }
 300                         }
 301                 });
 302
 303                 source_items = std::move(dest_items);
 304         }
 305 };
 306
 307 struct image_mixer::impl : boost::noncopyable
 308 {
 309         image_renderer                                          renderer_;
 310         std::vector<core::image_transform>      transform_stack_;
 311         std::vector<item>                                       items_; // layer/stream/items
 312 public:
 313         impl()
 314                 : transform_stack_(1)
 315         {
 316                 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer";
 317         }
 318
 319         void begin_layer(core::blend_mode blend_mode)
 320         {
 321         }
 322
 323         void push(const core::frame_transform& transform)
 324         {
 325                 transform_stack_.push_back(transform_stack_.back()*transform.image_transform);
 326         }
 327
 328         void visit(const core::const_frame& frame)
 329         {
 330                 if(frame.pixel_format_desc().format == core::pixel_format::invalid)
 331                         return;
 332
 333                 if(frame.pixel_format_desc().planes.empty())
 334                         return;
 335
 336                 if(frame.pixel_format_desc().planes.at(0).size < 16)
 337                         return;
 338
 339                 if(transform_stack_.back().field_mode == core::field_mode::empty)
 340                         return;
 341
 342                 item item;
 343                 item.pix_desc   = frame.pixel_format_desc();
 344                 item.transform  = transform_stack_.back();
 345                 for(int n = 0; n < item.pix_desc.planes.size(); ++n)
 346                         item.data.at(n) = frame.image_data(n).begin();
 347
 348                 items_.push_back(item);
 349         }
 350
 351         void pop()
 352         {
 353                 transform_stack_.pop_back();
 354         }
 355
 356         void end_layer()
 357         {
 358         }
 359
 360         boost::unique_future<array<const std::uint8_t>> render(const core::video_format_desc& format_desc)
 361         {
 362                 return renderer_(std::move(items_), format_desc);
 363         }
 364
 365         virtual core::mutable_frame create_frame(const void* tag, const core::pixel_format_desc& desc)
 366         {
 367                 std::vector<array<std::uint8_t>> buffers;
 368                 BOOST_FOREACH(auto& plane, desc.planes)
 369                 {
 370                         auto buf = spl::make_shared<buffer>(plane.size);
 371                         buffers.push_back(array<std::uint8_t>(buf->data(), plane.size, true, buf));
 372                 }
 373                 return core::mutable_frame(std::move(buffers), core::audio_buffer(), tag, desc);
 374         }
 375 };
 376
 377 image_mixer::image_mixer() : impl_(new impl()){}
 378 image_mixer::~image_mixer(){}
 379 void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}
 380 void image_mixer::visit(const core::const_frame& frame){impl_->visit(frame);}
 381 void image_mixer::pop(){impl_->pop();}
 382 boost::unique_future<array<const std::uint8_t>> image_mixer::operator()(const core::video_format_desc& format_desc){return impl_->render(format_desc);}
 383 void image_mixer::begin_layer(core::blend_mode blend_mode){impl_->begin_layer(blend_mode);}
 384 void image_mixer::end_layer(){impl_->end_layer();}
 385 core::mutable_frame image_mixer::create_frame(const void* tag, const core::pixel_format_desc& desc) {return impl_->create_frame(tag, desc);}
 386
 387 }}}