2.0. audio_mixer: Optimized.

[casparcg] / core / mixer / audio / audio_mixer.cpp

diff --git a/core/mixer/audio/audio_mixer.cpp b/core/mixer/audio/audio_mixer.cpp
index 1c3a77364a8fe2bb725fb063d49233ea993f1c63..34e77a56a5da93d58f7edcd90b3b8bf957bf8eba 100644 (file)
--- a/core/mixer/audio/audio_mixer.cpp
+++ b/core/mixer/audio/audio_mixer.cpp
@@ -35,9 +35,9 @@ namespace caspar { namespace core {
 \r
 struct audio_item\r
 {\r
-       const void*                             tag;\r
-       frame_transform                 transform;\r
-       std::vector<int32_t>    audio_data;\r
+       const void*                     tag;\r
+       frame_transform         transform;\r
+       audio_buffer            audio_data;\r
 };\r
        \r
 struct audio_mixer::implementation\r
@@ -59,7 +59,7 @@ public:
                transform_stack_.push(transform_stack_.top()*frame.get_frame_transform());\r
        }\r
 \r
-       void visit(const core::write_frame& frame)\r
+       void visit(core::write_frame& frame)\r
        {\r
                // We only care about the last field.\r
                if(format_desc_.field_mode == field_mode::upper && transform_stack_.top().field_mode == field_mode::upper)\r
@@ -75,7 +75,7 @@ public:
                audio_item item;\r
                item.tag                = frame.tag();\r
                item.transform  = transform_stack_.top();\r
-               item.audio_data = std::vector<int32_t>(frame.audio_data().begin(), frame.audio_data().end());\r
+               item.audio_data = std::move(frame.audio_data());\r
 \r
                items.push_back(item);          \r
        }\r
@@ -90,9 +90,9 @@ public:
                transform_stack_.pop();\r
        }\r
        \r
-       std::vector<int32_t> mix()\r
+       audio_buffer mix()\r
        {\r
-               auto result = std::vector<int32_t>(format_desc_.audio_samples_per_frame);\r
+               auto result = audio_buffer(format_desc_.audio_samples_per_frame, 0);\r
 \r
                std::map<const void*, core::frame_transform> next_frame_transforms;\r
 \r
@@ -109,30 +109,40 @@ public:
                                \r
                        if(next.volume < 0.001 && prev.volume < 0.001)\r
                                continue;\r
-               \r
-                       static const int BASE = 1<<31;\r
-\r
-                       const auto next_volume = static_cast<int64_t>(next.volume*BASE);\r
-                       const auto prev_volume = static_cast<int64_t>(prev.volume*BASE);\r
-               \r
-                       const int n_samples = result.size();\r
-               \r
-                       const auto in_size = static_cast<size_t>(item.audio_data.size());\r
-                       CASPAR_VERIFY(in_size == 0 || in_size == result.size());\r
-\r
-                       if(in_size > result.size())\r
+                                                                       \r
+                       if(static_cast<size_t>(item.audio_data.size()) != format_desc_.audio_samples_per_frame)\r
                                continue;\r
 \r
+                       CASPAR_ASSERT(format_desc_.audio_channels == 2);\r
+                       CASPAR_ASSERT(format_desc_.audio_samples_per_frame % 4 == 0);\r
+                                               \r
+                       const float prev_volume = static_cast<float>(prev.volume);\r
+                       const float next_volume = static_cast<float>(next.volume);\r
+                       const float delta               = 1.0f/static_cast<float>(format_desc_.audio_samples_per_frame/2);\r
+                       \r
                        tbb::parallel_for\r
                        (\r
-                               tbb::blocked_range<size_t>(0, item.audio_data.size()),\r
+                               tbb::blocked_range<size_t>(0, format_desc_.audio_samples_per_frame/4),\r
                                [&](const tbb::blocked_range<size_t>& r)\r
                                {\r
                                        for(size_t n = r.begin(); n < r.end(); ++n)\r
                                        {\r
-                                               const auto sample_volume = (prev_volume - (prev_volume * n)/n_samples) + (next_volume * n)/n_samples;\r
-                                               const auto sample = static_cast<int32_t>((static_cast<int64_t>(item.audio_data[n])*sample_volume)/BASE);\r
-                                               result[n] = result[n] + sample;\r
+                                               const float alpha0              = (n*2) * delta;\r
+                                               const float volume0             = prev_volume * (1.0f - alpha0) + next_volume * alpha0;\r
+                                               const float volume1             = prev_volume * (1.0f - alpha0 + delta) + next_volume * (alpha0 + delta);\r
+\r
+                                               auto sample_epi32               = _mm_load_si128(reinterpret_cast<__m128i*>(&item.audio_data[n*4]));\r
+\r
+                                               auto sample_ps                  = _mm_cvtepi32_ps(sample_epi32);                                                                                                \r
+                                               sample_ps                               = _mm_mul_ps(sample_ps, _mm_setr_ps(volume1, volume1, volume0, volume0));       \r
+\r
+                                               auto res_sample_epi32   = _mm_load_si128(reinterpret_cast<__m128i*>(&result[n*4]));\r
+                                               auto res_sample_ps              = _mm_cvtepi32_ps(res_sample_epi32);    \r
+\r
+                                               res_sample_ps                   = _mm_add_ps(sample_ps, res_sample_ps);\r
+                                               res_sample_epi32                = _mm_cvtps_epi32(res_sample_ps);\r
+                                               \r
+                                               _mm_store_si128(reinterpret_cast<__m128i*>(&result[n*4]), res_sample_epi32);\r
                                        }\r
                                }\r
                        );\r
@@ -149,7 +159,7 @@ audio_mixer::audio_mixer(const core::video_format_desc& format_desc) : impl_(new
 void audio_mixer::begin(core::basic_frame& frame){impl_->begin(frame);}\r
 void audio_mixer::visit(core::write_frame& frame){impl_->visit(frame);}\r
 void audio_mixer::end(){impl_->end();}\r
-std::vector<int32_t> audio_mixer::mix(){return impl_->mix();}\r
+audio_buffer audio_mixer::mix(){return impl_->mix();}\r
 audio_mixer& audio_mixer::operator=(audio_mixer&& other)\r
 {\r
        impl_ = std::move(other.impl_);\r