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 <tbb/parallel_for.h>\r
  28 \r
  29 #include <safeint.h>\r
  30 \r
  31 #include <stack>\r
  32 #include <deque>\r
  33 \r
  34 namespace caspar { namespace core {\r
  35 \r
  36 struct audio_item\r
  37 {\r
  38         const void*                     tag;\r
  39         frame_transform         transform;\r
  40         audio_buffer            audio_data;\r
  41 };\r
  42         \r
  43 struct audio_mixer::implementation\r
  44 {\r
  45         std::stack<core::frame_transform>                               transform_stack_;\r
  46         std::map<const void*, core::frame_transform>    prev_frame_transforms_;\r
  47         const core::video_format_desc                                   format_desc_;\r
  48         std::vector<audio_item>                                                 items;\r
  49 \r
  50 public:\r
  51         implementation(const core::video_format_desc& format_desc)\r
  52                 : format_desc_(format_desc)\r
  53         {\r
  54                 transform_stack_.push(core::frame_transform());\r
  55         }\r
  56         \r
  57         void begin(core::basic_frame& frame)\r
  58         {\r
  59                 transform_stack_.push(transform_stack_.top()*frame.get_frame_transform());\r
  60         }\r
  61 \r
  62         void visit(core::write_frame& frame)\r
  63         {\r
  64                 // We only care about the last field.\r
  65                 if(format_desc_.field_mode == field_mode::upper && transform_stack_.top().field_mode == field_mode::upper)\r
  66                         return;\r
  67 \r
  68                 if(format_desc_.field_mode == field_mode::lower && transform_stack_.top().field_mode == field_mode::lower)\r
  69                         return;\r
  70 \r
  71                 // Skip empty audio.\r
  72                 if(transform_stack_.top().volume < 0.002 || frame.audio_data().empty())\r
  73                         return;\r
  74 \r
  75                 audio_item item;\r
  76                 item.tag                = frame.tag();\r
  77                 item.transform  = transform_stack_.top();\r
  78                 item.audio_data = std::move(frame.audio_data());\r
  79 \r
  80                 items.push_back(item);          \r
  81         }\r
  82 \r
  83         void begin(const core::frame_transform& transform)\r
  84         {\r
  85                 transform_stack_.push(transform_stack_.top()*transform);\r
  86         }\r
  87                 \r
  88         void end()\r
  89         {\r
  90                 transform_stack_.pop();\r
  91         }\r
  92         \r
  93         audio_buffer mix()\r
  94         {       \r
  95                 // NOTE: auto data should be larger than format_desc_.audio_samples_per_frame to allow sse to read/write beyond size.\r
  96 \r
  97                 auto intermediate = std::vector<float, tbb::cache_aligned_allocator<float>>(format_desc_.audio_samples_per_frame+128, 0.0f);\r
  98 \r
  99                 std::map<const void*, core::frame_transform> next_frame_transforms;\r
 100                 \r
 101                 tbb::affinity_partitioner ap;\r
 102 \r
 103                 BOOST_FOREACH(auto& item, items)\r
 104                 {                       \r
 105                         const auto next = item.transform;\r
 106                         auto prev = next;\r
 107 \r
 108                         const auto it = prev_frame_transforms_.find(item.tag);\r
 109                         if(it != prev_frame_transforms_.end())\r
 110                                 prev = it->second;\r
 111                                 \r
 112                         next_frame_transforms[item.tag] = next; // Store all active tags, inactive tags will be removed at the end.\r
 113                                 \r
 114                         if(next.volume < 0.001 && prev.volume < 0.001)\r
 115                                 continue;\r
 116                                                                         \r
 117                         if(static_cast<size_t>(item.audio_data.size()) != format_desc_.audio_samples_per_frame)\r
 118                                 continue;\r
 119 \r
 120                         CASPAR_ASSERT(format_desc_.audio_channels == 2);\r
 121                         CASPAR_ASSERT(format_desc_.audio_samples_per_frame % 4 == 0);\r
 122                                                 \r
 123                         const float prev_volume = static_cast<float>(prev.volume);\r
 124                         const float next_volume = static_cast<float>(next.volume);\r
 125                         const float delta               = 1.0f/static_cast<float>(format_desc_.audio_samples_per_frame/format_desc_.audio_channels);\r
 126                         \r
 127                         tbb::parallel_for\r
 128                         (\r
 129                                 tbb::blocked_range<size_t>(0, format_desc_.audio_samples_per_frame/4),\r
 130                                 [&](const tbb::blocked_range<size_t>& r)\r
 131                                 {                                       \r
 132                                         auto alpha_ps   = _mm_setr_ps(delta, delta, 0.0f, 0.0f);\r
 133                                         auto delta2_ps  = _mm_set_ps1(delta*2.0f);\r
 134                                         auto prev_ps    = _mm_set_ps1(prev_volume);\r
 135                                         auto next_ps    = _mm_set_ps1(next_volume);     \r
 136 \r
 137                                         for(size_t n = r.begin(); n < r.end(); ++n)\r
 138                                         {               \r
 139                                                 auto next2_ps           = _mm_mul_ps(next_ps, alpha_ps);\r
 140                                                 auto prev2_ps           = _mm_sub_ps(prev_ps, _mm_mul_ps(prev_ps, alpha_ps));\r
 141                                                 auto volume_ps          = _mm_add_ps(next2_ps, prev2_ps);\r
 142 \r
 143                                                 auto sample_ps          = _mm_cvtepi32_ps(_mm_load_si128(reinterpret_cast<__m128i*>(&item.audio_data[n*4])));\r
 144                                                 auto res_sample_ps      = _mm_load_ps(&intermediate[n*4]);                                                                                      \r
 145                                                 sample_ps                       = _mm_mul_ps(sample_ps, volume_ps);     \r
 146                                                 res_sample_ps           = _mm_add_ps(sample_ps, res_sample_ps); \r
 147 \r
 148                                                 alpha_ps                        = _mm_add_ps(alpha_ps, delta2_ps);\r
 149 \r
 150                                                 _mm_store_ps(&intermediate[n*4], res_sample_ps);\r
 151                                         }\r
 152                                 }\r
 153                         , ap);\r
 154                 }\r
 155                 \r
 156                 auto result = audio_buffer(format_desc_.audio_samples_per_frame+128, 0);        \r
 157                 \r
 158                 tbb::parallel_for\r
 159                 (\r
 160                         tbb::blocked_range<size_t>(0, format_desc_.audio_samples_per_frame/32),\r
 161                         [&](const tbb::blocked_range<size_t>& r)\r
 162                         {               \r
 163                                 auto intermediate_128 = reinterpret_cast<__m128i*>(intermediate.data()+r.begin()*32);\r
 164                                 auto result_128           = reinterpret_cast<__m128i*>(result.data()+r.begin()*32);\r
 165                                 \r
 166                                 for(size_t n = r.begin(); n < r.end(); ++n)\r
 167                                 {       \r
 168                                         auto xmm0 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 169                                         auto xmm1 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 170                                         auto xmm2 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 171                                         auto xmm3 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 172                                         auto xmm4 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 173                                         auto xmm5 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 174                                         auto xmm6 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 175                                         auto xmm7 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
 176                         \r
 177                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm0));\r
 178                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm1));\r
 179                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm2));\r
 180                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm3));\r
 181                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm4));\r
 182                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm5));\r
 183                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm6));\r
 184                                         _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm7));\r
 185                                 }\r
 186                         }\r
 187                 , ap);\r
 188 \r
 189                 items.clear();\r
 190                 prev_frame_transforms_ = std::move(next_frame_transforms);      \r
 191 \r
 192                 result.resize(format_desc_.audio_samples_per_frame);\r
 193                 return std::move(result);\r
 194         }\r
 195 };\r
 196 \r
 197 audio_mixer::audio_mixer(const core::video_format_desc& format_desc) : impl_(new implementation(format_desc)){}\r
 198 void audio_mixer::begin(core::basic_frame& frame){impl_->begin(frame);}\r
 199 void audio_mixer::visit(core::write_frame& frame){impl_->visit(frame);}\r
 200 void audio_mixer::end(){impl_->end();}\r
 201 audio_buffer audio_mixer::mix(){return impl_->mix();}\r
 202 audio_mixer& audio_mixer::operator=(audio_mixer&& other)\r
 203 {\r
 204         impl_ = std::move(other.impl_);\r
 205         return *this;\r
 206 }\r
 207 \r
 208 }}