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 dafa6e4dd0991c81a3310d407d4c61daceca954a..574c899d50438b5517dddb2b4437b217a43d5730 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
@@ -91,8 +91,8 @@ public:
        }\r
        \r
        audio_buffer mix()\r
-       {\r
-               auto result = audio_buffer(format_desc_.audio_samples_per_frame, 0);\r
+       {       \r
+               auto intermediate = std::vector<float, tbb::cache_aligned_allocator<float>>(format_desc_.audio_samples_per_frame+128, 0.0f);\r
 \r
                std::map<const void*, core::frame_transform> next_frame_transforms;\r
 \r
@@ -114,6 +114,7 @@ public:
                                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
@@ -121,30 +122,53 @@ public:
                        \r
                        tbb::parallel_for\r
                        (\r
-                               tbb::blocked_range<size_t>(0, format_desc_.audio_samples_per_frame/2),\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 float alpha               = n * delta;\r
-                                               const float volume              = prev_volume * (1.0f - alpha) + next_volume * alpha;\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_loadl_epi64(reinterpret_cast<__m128i*>(&item.audio_data[n*2]));\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_set1_ps(volume));   \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_loadl_epi64(reinterpret_cast<__m128i*>(&result[n*2]));\r
-                                               auto res_sample_ps              = _mm_cvtepi32_ps(res_sample_epi32);    \r
+                                               auto res_sample_ps      = _mm_load_ps(&intermediate[n*4]);\r
+                                               res_sample_ps           = _mm_add_ps(sample_ps, res_sample_ps); \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_storel_epi64(reinterpret_cast<__m128i*>(&result[n*2]), res_sample_epi32);\r
+                                               _mm_store_ps(&intermediate[n*4], res_sample_ps);\r
                                        }\r
                                }\r
                        );\r
                }\r
+               \r
+               auto result = audio_buffer(format_desc_.audio_samples_per_frame+128, 0);        \r
+\r
+               auto intermediate_128 = reinterpret_cast<__m128i*>(intermediate.data());\r
+               auto result_128           = reinterpret_cast<__m128i*>(result.data());\r
+               for(size_t n = 0; n < format_desc_.audio_samples_per_frame/32; ++n)\r
+               {\r
+                       auto xmm0 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm1 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm2 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm3 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm4 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm5 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm6 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       auto xmm7 = _mm_load_ps(reinterpret_cast<float*>(intermediate_128++));\r
+                       \r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm0));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm1));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm2));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm3));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm4));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm5));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm6));\r
+                       _mm_stream_si128(result_128++, _mm_cvtps_epi32(xmm7));\r
+               }\r
 \r
                items.clear();\r
                prev_frame_transforms_ = std::move(next_frame_transforms);      \r