2.0. audio_mixer: Optimized.

git-svn-id: https://casparcg.svn.sourceforge.net/svnroot/casparcg/server/branches/2.0.0.2@1334 362d55ac-95cf-4e76-9f9a-cbaa9c17b72d

diff --git a/core/mixer/audio/audio_mixer.cpp b/core/mixer/audio/audio_mixer.cpp
index 34e77a56a5da93d58f7edcd90b3b8bf957bf8eba..574c899d50438b5517dddb2b4437b217a43d5730 100644 (file)
--- a/core/mixer/audio/audio_mixer.cpp
+++ b/core/mixer/audio/audio_mixer.cpp
@@ -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
@@ -131,26 +131,49 @@ public:
                                                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
+                                               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
+                                               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
+                                               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_store_si128(reinterpret_cast<__m128i*>(&result[n*4]), 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
 \r
+               result.resize(format_desc_.audio_samples_per_frame);\r
                return std::move(result);\r
        }\r
 };\r