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 namespace caspar { namespace core {\r
  31 \r
  32 struct audio_item\r
  33 {\r
  34         const void*                     tag;\r
  35         frame_transform         transform;\r
  36         audio_buffer            audio_data;\r
  37 };\r
  38         \r
  39 struct audio_mixer::implementation\r
  40 {\r
  41         std::stack<core::frame_transform>                               transform_stack_;\r
  42         std::map<const void*, core::frame_transform>    prev_frame_transforms_;\r
  43         const core::video_format_desc                                   format_desc_;\r
  44         std::vector<audio_item>                                                 items_;\r
  45 \r
  46 public:\r
  47         implementation(const core::video_format_desc& format_desc)\r
  48                 : format_desc_(format_desc)\r
  49         {\r
  50                 transform_stack_.push(core::frame_transform());\r
  51         }\r
  52         \r
  53         void begin(core::basic_frame& frame)\r
  54         {\r
  55                 transform_stack_.push(transform_stack_.top()*frame.get_frame_transform());\r
  56         }\r
  57 \r
  58         void visit(core::write_frame& frame)\r
  59         {\r
  60                 // We only care about the last field.\r
  61                 if(format_desc_.field_mode == field_mode::upper && transform_stack_.top().field_mode == field_mode::upper)\r
  62                         return;\r
  63 \r
  64                 if(format_desc_.field_mode == field_mode::lower && transform_stack_.top().field_mode == field_mode::lower)\r
  65                         return;\r
  66 \r
  67                 // Skip empty audio.\r
  68                 if(transform_stack_.top().volume < 0.002 || frame.audio_data().empty())\r
  69                         return;\r
  70 \r
  71                 audio_item item;\r
  72                 item.tag                = frame.tag();\r
  73                 item.transform  = transform_stack_.top();\r
  74                 item.audio_data = std::move(frame.audio_data());\r
  75 \r
  76                 items_.push_back(item);         \r
  77         }\r
  78 \r
  79         void begin(const core::frame_transform& transform)\r
  80         {\r
  81                 transform_stack_.push(transform_stack_.top()*transform);\r
  82         }\r
  83                 \r
  84         void end()\r
  85         {\r
  86                 transform_stack_.pop();\r
  87         }\r
  88         \r
  89         audio_buffer mix()\r
  90         {       \r
  91                 // NOTE: auto data should be larger than format_desc_.audio_samples_per_frame to allow sse to read/write beyond size.\r
  92 \r
  93                 auto intermediate = std::vector<float, tbb::cache_aligned_allocator<float>>(format_desc_.audio_samples_per_frame+128, 0.0f);\r
  94 \r
  95                 std::map<const void*, core::frame_transform> next_frame_transforms;\r
  96                 \r
  97                 BOOST_FOREACH(auto& item, items_)\r
  98                 {                       \r
  99                         const auto next = item.transform;\r
 100                         auto prev = next;\r
 101 \r
 102                         const auto it = prev_frame_transforms_.find(item.tag);\r
 103                         if(it != prev_frame_transforms_.end())\r
 104                                 prev = it->second;\r
 105                                 \r
 106                         next_frame_transforms[item.tag] = next; // Store all active tags, inactive tags will be removed at the end.\r
 107 \r
 108                         if(next.volume < 0.001 && prev.volume < 0.001)\r
 109                                 continue;\r
 110                                                                         \r
 111                         if(static_cast<size_t>(item.audio_data.size()) != format_desc_.audio_samples_per_frame)\r
 112                                 continue;\r
 113 \r
 114                         CASPAR_ASSERT(format_desc_.audio_channels == 2);\r
 115                         CASPAR_ASSERT(format_desc_.audio_samples_per_frame % 4 == 0);\r
 116                                                 \r
 117                         const float prev_volume = static_cast<float>(prev.volume);\r
 118                         const float next_volume = static_cast<float>(next.volume);\r
 119                         const float delta               = 1.0f/static_cast<float>(format_desc_.audio_samples_per_frame/format_desc_.audio_channels);\r
 120                                                 \r
 121                         auto alpha_ps   = _mm_setr_ps(0.0f, 0.0f, delta, delta);\r
 122                         auto delta2_ps  = _mm_set_ps1(delta*2.0f);\r
 123                         auto prev_ps    = _mm_set_ps1(prev_volume);\r
 124                         auto next_ps    = _mm_set_ps1(next_volume);     \r
 125 \r
 126                         for(size_t n = 0; n < format_desc_.audio_samples_per_frame/4; ++n)\r
 127                         {               \r
 128                                 auto next2_ps           = _mm_mul_ps(next_ps, alpha_ps);\r
 129                                 auto prev2_ps           = _mm_sub_ps(prev_ps, _mm_mul_ps(prev_ps, alpha_ps));\r
 130                                 auto volume_ps          = _mm_add_ps(next2_ps, prev2_ps);\r
 131 \r
 132                                 auto sample_ps          = _mm_cvtepi32_ps(_mm_load_si128(reinterpret_cast<__m128i*>(&item.audio_data[n*4])));\r
 133                                 auto res_sample_ps      = _mm_load_ps(&intermediate[n*4]);                                                                                      \r
 134                                 sample_ps                       = _mm_mul_ps(sample_ps, volume_ps);     \r
 135                                 res_sample_ps           = _mm_add_ps(sample_ps, res_sample_ps); \r
 136 \r
 137                                 alpha_ps                        = _mm_add_ps(alpha_ps, delta2_ps);\r
 138 \r
 139                                 _mm_store_ps(&intermediate[n*4], res_sample_ps);\r
 140                         }\r
 141                 }\r
 142                 \r
 143                 auto result = audio_buffer(format_desc_.audio_samples_per_frame+128);   \r
 144                         \r
 145                 auto intermediate_128 = reinterpret_cast<__m128i*>(intermediate.data());\r
 146                 auto result_128           = reinterpret_cast<__m128i*>(result.data());\r
 147                                 \r
 148                 for(size_t n = 0; n < format_desc_.audio_samples_per_frame/32; ++n)\r
 149                 {       \r
 150                         auto xmm0 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 151                         auto xmm1 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 152                         auto xmm2 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 153                         auto xmm3 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 154                         auto xmm4 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 155                         auto xmm5 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 156                         auto xmm6 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 157                         auto xmm7 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 158                         \r
 159                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm0));\r
 160                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm1));\r
 161                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm2));\r
 162                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm3));\r
 163                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm4));\r
 164                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm5));\r
 165                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm6));\r
 166                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm7));\r
 167                 }\r
 168 \r
 169                 items_.clear();\r
 170                 prev_frame_transforms_ = std::move(next_frame_transforms);      \r
 171 \r
 172                 result.resize(format_desc_.audio_samples_per_frame);\r
 173                 return std::move(result);\r
 174         }\r
 175 };\r
 176 \r
 177 audio_mixer::audio_mixer(const core::video_format_desc& format_desc) : impl_(new implementation(format_desc)){}\r
 178 void audio_mixer::begin(core::basic_frame& frame){impl_->begin(frame);}\r
 179 void audio_mixer::visit(core::write_frame& frame){impl_->visit(frame);}\r
 180 void audio_mixer::end(){impl_->end();}\r
 181 audio_buffer audio_mixer::mix(){return impl_->mix();}\r
 182 audio_mixer& audio_mixer::operator=(audio_mixer&& other)\r
 183 {\r
 184         impl_ = std::move(other.impl_);\r
 185         return *this;\r
 186 }\r
 187 \r
 188 }}