git.sesse.net Git - casparcg/blob - core/mixer/audio/audio_mixer.cpp

   1 /*\r
   2 * copyright (c) 2010 Sveriges Television AB <info@casparcg.com>\r
   3 *\r
   4 *  This file is part of CasparCG.\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 */\r
  20 #include "../../stdafx.h"\r
  21 \r
  22 #include "audio_mixer.h"\r
  23 \r
  24 #include <core/mixer/write_frame.h>\r
  25 #include <core/producer/frame/frame_transform.h>\r
  26 \r
  27 #include <stack>\r
  28 #include <vector>\r
  29 \r
  30 #include <tbb/parallel_for.h>\r
  31 \r
  32 namespace caspar { namespace core {\r
  33 \r
  34 struct audio_item\r
  35 {\r
  36         const void*                     tag;\r
  37         frame_transform         transform;\r
  38         audio_buffer            audio_data;\r
  39 };\r
  40         \r
  41 struct audio_mixer::implementation\r
  42 {\r
  43         std::stack<core::frame_transform>                               transform_stack_;\r
  44         std::map<const void*, core::frame_transform>    prev_frame_transforms_;\r
  45         const core::video_format_desc                                   format_desc_;\r
  46         std::vector<audio_item>                                                 items_;\r
  47 \r
  48 public:\r
  49         implementation(const core::video_format_desc& format_desc)\r
  50                 : format_desc_(format_desc)\r
  51         {\r
  52                 transform_stack_.push(core::frame_transform());\r
  53         }\r
  54         \r
  55         void begin(core::basic_frame& frame)\r
  56         {\r
  57                 transform_stack_.push(transform_stack_.top()*frame.get_frame_transform());\r
  58         }\r
  59 \r
  60         void visit(core::write_frame& frame)\r
  61         {\r
  62                 // We only care about the last field.\r
  63                 if(format_desc_.field_mode == field_mode::upper && transform_stack_.top().field_mode == field_mode::upper)\r
  64                         return;\r
  65 \r
  66                 if(format_desc_.field_mode == field_mode::lower && transform_stack_.top().field_mode == field_mode::lower)\r
  67                         return;\r
  68 \r
  69                 // Skip empty audio.\r
  70                 if(transform_stack_.top().volume < 0.002 || frame.audio_data().empty())\r
  71                         return;\r
  72 \r
  73                 audio_item item;\r
  74                 item.tag                = frame.tag();\r
  75                 item.transform  = transform_stack_.top();\r
  76                 item.audio_data = std::move(frame.audio_data());\r
  77 \r
  78                 items_.push_back(item);         \r
  79         }\r
  80 \r
  81         void begin(const core::frame_transform& transform)\r
  82         {\r
  83                 transform_stack_.push(transform_stack_.top()*transform);\r
  84         }\r
  85                 \r
  86         void end()\r
  87         {\r
  88                 transform_stack_.pop();\r
  89         }\r
  90         \r
  91         audio_buffer mix()\r
  92         {       \r
  93                 // NOTE: auto data should be larger than format_desc_.audio_samples_per_frame to allow sse to read/write beyond size.\r
  94 \r
  95                 auto intermediate = std::vector<float, tbb::cache_aligned_allocator<float>>(format_desc_.audio_samples_per_frame+128, 0.0f);\r
  96                 auto result               = audio_buffer(format_desc_.audio_samples_per_frame+128);     \r
  97 \r
  98                 std::map<const void*, core::frame_transform> next_frame_transforms;\r
  99                 \r
 100                 BOOST_FOREACH(auto& item, items_)\r
 101                 {                       \r
 102                         const auto next = item.transform;\r
 103                         auto prev = next;\r
 104 \r
 105                         const auto it = prev_frame_transforms_.find(item.tag);\r
 106                         if(it != prev_frame_transforms_.end())\r
 107                                 prev = it->second;\r
 108                                 \r
 109                         next_frame_transforms[item.tag] = next; // Store all active tags, inactive tags will be removed at the end.\r
 110 \r
 111                         if(next.volume < 0.001 && prev.volume < 0.001)\r
 112                                 continue;\r
 113                                                                         \r
 114                         if(static_cast<size_t>(item.audio_data.size()) != format_desc_.audio_samples_per_frame)\r
 115                                 continue;\r
 116 \r
 117                         CASPAR_ASSERT(format_desc_.audio_channels == 2);\r
 118                         CASPAR_ASSERT(format_desc_.audio_samples_per_frame % 4 == 0);\r
 119                                                 \r
 120                         const float prev_volume = static_cast<float>(prev.volume);\r
 121                         const float next_volume = static_cast<float>(next.volume);\r
 122                                                                         \r
 123                         auto alpha              = (next_volume-prev_volume)/static_cast<float>(format_desc_.audio_samples_per_frame/format_desc_.audio_channels);\r
 124                         auto alpha_ps   = _mm_set_ps1(alpha*2.0f);\r
 125                         auto volume_ps  = _mm_setr_ps(prev_volume, prev_volume, prev_volume+alpha, prev_volume+alpha);\r
 126 \r
 127                         for(size_t n = 0; n < format_desc_.audio_samples_per_frame/4; ++n)\r
 128                         {               \r
 129                                 auto sample_ps          = _mm_cvtepi32_ps(_mm_load_si128(reinterpret_cast<__m128i*>(&item.audio_data[n*4])));\r
 130                                 auto res_sample_ps      = _mm_load_ps(&intermediate[n*4]);                                                                                      \r
 131                                 sample_ps                       = _mm_mul_ps(sample_ps, volume_ps);     \r
 132                                 res_sample_ps           = _mm_add_ps(sample_ps, res_sample_ps); \r
 133 \r
 134                                 volume_ps                       = _mm_add_ps(volume_ps, alpha_ps);\r
 135 \r
 136                                 _mm_store_ps(&intermediate[n*4], res_sample_ps);\r
 137                         }\r
 138                 }               \r
 139                         \r
 140                 auto intermediate_128 = reinterpret_cast<__m128i*>(intermediate.data());\r
 141                 auto result_128           = reinterpret_cast<__m128i*>(result.data());\r
 142                                 \r
 143                 for(size_t n = 0; n < format_desc_.audio_samples_per_frame/32; ++n)\r
 144                 {       \r
 145                         auto xmm0 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 146                         auto xmm1 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 147                         auto xmm2 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 148                         auto xmm3 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 149                         auto xmm4 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 150                         auto xmm5 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 151                         auto xmm6 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 152                         auto xmm7 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 153                         \r
 154                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm0));\r
 155                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm1));\r
 156                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm2));\r
 157                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm3));\r
 158                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm4));\r
 159                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm5));\r
 160                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm6));\r
 161                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm7));\r
 162                 }\r
 163 \r
 164                 items_.clear();\r
 165                 prev_frame_transforms_ = std::move(next_frame_transforms);      \r
 166 \r
 167                 result.resize(format_desc_.audio_samples_per_frame);\r
 168                 return std::move(result);\r
 169         }\r
 170 };\r
 171 \r
 172 audio_mixer::audio_mixer(const core::video_format_desc& format_desc) : impl_(new implementation(format_desc)){}\r
 173 void audio_mixer::begin(core::basic_frame& frame){impl_->begin(frame);}\r
 174 void audio_mixer::visit(core::write_frame& frame){impl_->visit(frame);}\r
 175 void audio_mixer::end(){impl_->end();}\r
 176 audio_buffer audio_mixer::mix(){return impl_->mix();}\r
 177 audio_mixer& audio_mixer::operator=(audio_mixer&& other)\r
 178 {\r
 179         impl_ = std::move(other.impl_);\r
 180         return *this;\r
 181 }\r
 182 \r
 183 }}