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/parallel_for.h>
  45 #include <tbb/parallel_for_each.h>
  46 #include <tbb/concurrent_queue.h>
  47
  48 #include <boost/range/algorithm_ext/erase.hpp>
  49 #include <boost/thread/future.hpp>
  50
  51 #include <algorithm>
  52 #include <cstdint>
  53 #include <vector>
  54 #include <set>
  55 #include <array>
  56
  57 #if defined(_MSC_VER)
  58 #pragma warning (push)
  59 #pragma warning (disable : 4244)
  60 #endif
  61 extern "C"
  62 {
  63         #include <libswscale/swscale.h>
  64         #include <libavcodec/avcodec.h>
  65         #include <libavformat/avformat.h>
  66 }
  67 #if defined(_MSC_VER)
  68 #pragma warning (pop)
  69 #endif
  70
  71 namespace caspar { namespace accelerator { namespace cpu {
  72
  73 struct item
  74 {
  75         core::pixel_format_desc                 pix_desc        = core::pixel_format::invalid;
  76         std::array<const uint8_t*, 4>   data;
  77         core::image_transform                   transform;
  78
  79         item()
  80         {
  81                 data.fill(0);
  82         }
  83 };
  84
  85 bool operator==(const item& lhs, const item& rhs)
  86 {
  87         return lhs.data == rhs.data && lhs.transform == rhs.transform;
  88 }
  89
  90 bool operator!=(const item& lhs, const item& rhs)
  91 {
  92         return !(lhs == rhs);
  93 }
  94
  95 // 100% accurate blending with correct rounding.
  96 inline xmm::s8_x blend(xmm::s8_x d, xmm::s8_x s)
  97 {
  98         using namespace xmm;
  99
 100         // C(S, D) = S + D - (((T >> 8) + T) >> 8);
 101         // T(S, D) = S * D[A] + 0x80
 102
 103         auto aaaa   = s8_x::shuffle(d, s8_x(15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3));
 104         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).
 105
 106         auto xaxa       = s16_x(aaaa) >> 8;
 107
 108         auto t1         = s16_x::multiply_low(s16_x(s) & 0x00FF, xaxa) + 0x80;
 109         auto t2         = s16_x::multiply_low(s16_x(s) >> 8    , xaxa) + 0x80;
 110
 111         auto xyxy       = s8_x(((t1 >> 8) + t1) >> 8);
 112         auto yxyx       = s8_x((t2 >> 8) + t2);
 113         auto argb   = s8_x::blend(xyxy, yxyx, s8_x(-1, 0, -1, 0));
 114
 115         return s8_x(s) + (d - argb);
 116 }
 117
 118 template<typename temporal, typename alignment>
 119 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
 120 {
 121         using namespace xmm;
 122
 123         for(auto n = 0; n < count; n += 32)
 124         {
 125                 auto s0 = s8_x::load<temporal_tag, alignment>(dest+n+0);
 126                 auto s1 = s8_x::load<temporal_tag, alignment>(dest+n+16);
 127
 128                 auto d0 = s8_x::load<temporal_tag, alignment>(source+n+0);
 129                 auto d1 = s8_x::load<temporal_tag, alignment>(source+n+16);
 130
 131                 auto argb0 = blend(d0, s0);
 132                 auto argb1 = blend(d1, s1);
 133
 134                 s8_x::store<temporal, alignment>(argb0, dest+n+0 );
 135                 s8_x::store<temporal, alignment>(argb1, dest+n+16);
 136         }
 137 }
 138
 139 template<typename temporal>
 140 static void kernel(uint8_t* dest, const uint8_t* source, size_t count)
 141 {
 142         using namespace xmm;
 143
 144         if(reinterpret_cast<std::uint64_t>(dest) % 16 != 0 || reinterpret_cast<std::uint64_t>(source) % 16 != 0)
 145                 kernel<temporal_tag, unaligned_tag>(dest, source, count);
 146         else
 147                 kernel<temporal_tag, aligned_tag>(dest, source, count);
 148 }
 149
 150 class image_renderer
 151 {
 152         tbb::concurrent_unordered_map<int64_t, tbb::concurrent_bounded_queue<std::shared_ptr<SwsContext>>>      sws_devices_;
 153         tbb::concurrent_bounded_queue<spl::shared_ptr<buffer>>                                                                                          temp_buffers_;
 154         core::video_format_desc                                                                                                                                                         format_desc_;
 155 public:
 156         std::future<array<const std::uint8_t>> operator()(std::vector<item> items, const core::video_format_desc& format_desc)
 157         {
 158                 if (format_desc != format_desc_)
 159                 {
 160                         format_desc_ = format_desc;
 161                         sws_devices_.clear();
 162                 }
 163
 164                 convert(items, format_desc.width, format_desc.height);
 165
 166                 // Remove first field stills.
 167                 boost::range::remove_erase_if(items, [&](const item& item)
 168                 {
 169                         return item.transform.is_still && item.transform.field_mode == format_desc.field_mode; // only us last field for stills.
 170                 });
 171
 172                 // Stills are progressive
 173                 for (auto& item : items)
 174                 {
 175                         if(item.transform.is_still)
 176                                 item.transform.field_mode = core::field_mode::progressive;
 177                 }
 178
 179                 auto result = spl::make_shared<buffer>(format_desc.size, 0);
 180                 if(format_desc.field_mode != core::field_mode::progressive)
 181                 {
 182                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::upper);
 183                         draw(items, result->data(), format_desc.width, format_desc.height, core::field_mode::lower);
 184                 }
 185                 else
 186                 {
 187                         draw(items, result->data(), format_desc.width, format_desc.height,  core::field_mode::progressive);
 188                 }
 189
 190                 temp_buffers_.clear();
 191
 192                 return make_ready_future(array<const std::uint8_t>(result->data(), format_desc.size, true, result));
 193         }
 194
 195 private:
 196
 197         void draw(std::vector<item> items, uint8_t* dest, std::size_t width, std::size_t height, core::field_mode field_mode)
 198         {
 199                 for (auto& item : items)
 200                         item.transform.field_mode &= field_mode;
 201
 202                 // Remove empty items.
 203                 boost::range::remove_erase_if(items, [&](const item& item)
 204                 {
 205                         return item.transform.field_mode == core::field_mode::empty;
 206                 });
 207
 208                 if(items.empty())
 209                         return;
 210
 211                 auto start = field_mode == core::field_mode::lower ? 1 : 0;
 212                 auto step  = field_mode == core::field_mode::progressive ? 1 : 2;
 213
 214                 // TODO: Add support for fill translations.
 215                 // TODO: Add support for mask rect.
 216                 // TODO: Add support for opacity.
 217                 // TODO: Add support for mix transition.
 218                 // TODO: Add support for push transition.
 219                 // TODO: Add support for wipe transition.
 220                 // TODO: Add support for slide transition.
 221                 tbb::parallel_for(tbb::blocked_range<std::size_t>(0, height/step), [&](const tbb::blocked_range<std::size_t>& r)
 222                 {
 223                         for(auto i = r.begin(); i != r.end(); ++i)
 224                         {
 225                                 auto y = i*step+start;
 226
 227                                 for(std::size_t n = 0; n < items.size()-1; ++n)
 228                                         kernel<xmm::temporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
 229
 230                                 std::size_t n = items.size()-1;
 231                                 kernel<xmm::nontemporal_tag>(dest + y*width*4, items[n].data.at(0) + y*width*4, width*4);
 232                         }
 233
 234                         _mm_mfence();
 235                 });
 236         }
 237
 238         void convert(std::vector<item>& source_items, int width, int height)
 239         {
 240                 std::set<std::array<const uint8_t*, 4>> buffers;
 241
 242                 for (auto& item : source_items)
 243                         buffers.insert(item.data);
 244
 245                 auto dest_items = source_items;
 246
 247                 tbb::parallel_for_each(buffers.begin(), buffers.end(), [&](const std::array<const uint8_t*, 4>& data)
 248                 {
 249                         auto pix_desc = std::find_if(source_items.begin(), source_items.end(), [&](const item& item){return item.data == data;})->pix_desc;
 250
 251                         if(pix_desc.format == core::pixel_format::bgra &&
 252                                 pix_desc.planes.at(0).width == width &&
 253                                 pix_desc.planes.at(0).height == height)
 254                                 return;
 255
 256                         std::array<uint8_t*, 4> data2 = {};
 257                         for(std::size_t n = 0; n < data.size(); ++n)
 258                                 data2.at(n) = const_cast<uint8_t*>(data[n]);
 259
 260                         auto input_av_frame = ffmpeg::make_av_frame(data2, pix_desc);
 261
 262
 263                         int64_t key = ((static_cast<int64_t>(input_av_frame->width)      << 32) & 0xFFFF00000000) |
 264                                                   ((static_cast<int64_t>(input_av_frame->height) << 16) & 0xFFFF0000) |
 265                                                   ((static_cast<int64_t>(input_av_frame->format) <<  8) & 0xFF00);
 266
 267                         auto& pool = sws_devices_[key];
 268
 269                         std::shared_ptr<SwsContext> sws_device;
 270                         if(!pool.try_pop(sws_device))
 271                         {
 272                                 double param;
 273                                 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);
 274                         }
 275
 276                         if(!sws_device)
 277                                 CASPAR_THROW_EXCEPTION(operation_failed() << msg_info("Could not create software scaling device.") << boost::errinfo_api_function("sws_getContext"));
 278
 279                         auto dest_frame = spl::make_shared<buffer>(width*height*4);
 280                         temp_buffers_.push(dest_frame);
 281
 282                         {
 283                                 auto dest_av_frame = ffmpeg::create_frame();
 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.fill(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   = { 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(int channel_id)
 314                 : transform_stack_(1)
 315         {
 316                 CASPAR_LOG(info) << L"Initialized Streaming SIMD Extensions Accelerated CPU Image Mixer for channel " << channel_id;
 317         }
 318
 319         void push(const core::frame_transform& transform)
 320         {
 321                 transform_stack_.push_back(transform_stack_.back()*transform.image_transform);
 322         }
 323
 324         void visit(const core::const_frame& frame)
 325         {
 326                 if(frame.pixel_format_desc().format == core::pixel_format::invalid)
 327                         return;
 328
 329                 if(frame.pixel_format_desc().planes.empty())
 330                         return;
 331
 332                 if(frame.pixel_format_desc().planes.at(0).size < 16)
 333                         return;
 334
 335                 if(transform_stack_.back().field_mode == core::field_mode::empty)
 336                         return;
 337
 338                 item item;
 339                 item.pix_desc   = frame.pixel_format_desc();
 340                 item.transform  = transform_stack_.back();
 341                 for(int n = 0; n < item.pix_desc.planes.size(); ++n)
 342                         item.data.at(n) = frame.image_data(n).begin();
 343
 344                 items_.push_back(item);
 345         }
 346
 347         void pop()
 348         {
 349                 transform_stack_.pop_back();
 350         }
 351
 352         std::future<array<const std::uint8_t>> render(const core::video_format_desc& format_desc)
 353         {
 354                 return renderer_(std::move(items_), format_desc);
 355         }
 356
 357         core::mutable_frame create_frame(const void* tag, const core::pixel_format_desc& desc, const core::audio_channel_layout& channel_layout)
 358         {
 359                 std::vector<array<std::uint8_t>> buffers;
 360                 for (auto& plane : desc.planes)
 361                 {
 362                         auto buf = spl::make_shared<buffer>(plane.size);
 363                         buffers.push_back(array<std::uint8_t>(buf->data(), plane.size, true, buf));
 364                 }
 365                 return core::mutable_frame(std::move(buffers), core::mutable_audio_buffer(), tag, desc, channel_layout);
 366         }
 367 };
 368
 369 image_mixer::image_mixer(int channel_id) : impl_(new impl(channel_id)){}
 370 image_mixer::~image_mixer(){}
 371 void image_mixer::push(const core::frame_transform& transform){impl_->push(transform);}
 372 void image_mixer::visit(const core::const_frame& frame){impl_->visit(frame);}
 373 void image_mixer::pop(){impl_->pop();}
 374 int image_mixer::get_max_frame_size() { return std::numeric_limits<int>::max(); }
 375 std::future<array<const std::uint8_t>> image_mixer::operator()(const core::video_format_desc& format_desc, bool /* straighten_alpha */){return impl_->render(format_desc);}
 376 core::mutable_frame image_mixer::create_frame(const void* tag, const core::pixel_format_desc& desc, const core::audio_channel_layout& channel_layout) {return impl_->create_frame(tag, desc, channel_layout);}
 377
 378 }}}