From c014a34385671110102c9b00778cf5e86d773505 Mon Sep 17 00:00:00 2001
From: Paul Ferrand <paulfd@outlook.fr>
Date: Wed, 31 Jul 2019 01:45:58 +0200
Subject: [PATCH] WIP on vector operations

---
 benchmarks/AudioBuffers.cpp |  70 +++++++--
 sources/AudioBuffer.h       | 106 ++++++++++++-
 sources/Buffer.h            |   4 +-
 sources/Globals.h           |   4 +-
 tests/AudioBuffer.cpp       | 297 ++++++++++++++++++++++++++++++++++++
 5 files changed, 460 insertions(+), 21 deletions(-)

diff --git a/benchmarks/AudioBuffers.cpp b/benchmarks/AudioBuffers.cpp
index 1c2c9fba..bd3a766b 100644
--- a/benchmarks/AudioBuffers.cpp
+++ b/benchmarks/AudioBuffers.cpp
@@ -1,9 +1,10 @@
 #include <benchmark/benchmark.h>
 #include "../sources/AudioBuffer.h"
 #include <algorithm>
+#include <numeric>
 
-static void AB_Joint_Fill_float(benchmark::State& state) {
-  AudioBuffer<float> buffer (65535);
+static void Joint_Fill_float(benchmark::State& state) {
+  AudioBuffer<float> buffer (100001);
   float fillValue = 0.0f;
   for (auto _ : state) {
     buffer.fill(fillValue);
@@ -11,20 +12,43 @@ static void AB_Joint_Fill_float(benchmark::State& state) {
   }
 }
 // Register the function as a benchmark
-BENCHMARK(AB_Joint_Fill_float);
+BENCHMARK(Joint_Fill_float);
 
-static void AB_Split_Fill_float(benchmark::State& state) {
-  SplitAudioBuffer<float> buffer (65535);
+static void Split_Fill_float(benchmark::State& state) {
+  SplitAudioBuffer<float> buffer (100001);
   float fillValue = 0.0f;
   for (auto _ : state) {
     buffer.fill(fillValue);
     fillValue += 1.0f;
   }
 }
-BENCHMARK(AB_Split_Fill_float);
+BENCHMARK(Split_Fill_float);
 
-static void AB_Joint_Fill_double(benchmark::State& state) {
-  AudioBuffer<double> buffer (65535);
+static void Joint_Fill_float_SSE(benchmark::State& state) {
+  AudioBuffer<float> buffer (100001);
+  float fillValue = 0.0f;
+  for (auto _ : state) {
+    buffer.fill<VectorOperations::sse>(fillValue);
+    fillValue += 1.0f;
+  }
+}
+// Register the function as a benchmark
+BENCHMARK(Joint_Fill_float_SSE);
+
+
+static void Split_Fill_float_SSE(benchmark::State& state) {
+  SplitAudioBuffer<float> buffer (100001);
+  float fillValue = 0.0f;
+  for (auto _ : state) {
+    buffer.fill<VectorOperations::sse>(fillValue);
+    fillValue += 1.0f;
+  }
+}
+
+BENCHMARK(Split_Fill_float_SSE);
+
+static void Joint_Fill_double(benchmark::State& state) {
+  AudioBuffer<double> buffer (100001);
   double fillValue = 0.0;
   for (auto _ : state) {
     buffer.fill(fillValue);
@@ -32,15 +56,37 @@ static void AB_Joint_Fill_double(benchmark::State& state) {
   }
 }
 // Register the function as a benchmark
-BENCHMARK(AB_Joint_Fill_double);
+BENCHMARK(Joint_Fill_double);
 
-static void AB_Split_Fill_double(benchmark::State& state) {
-  SplitAudioBuffer<double> buffer (65535);
+static void Split_Fill_double(benchmark::State& state) {
+  SplitAudioBuffer<double> buffer (100001);
   double fillValue = 0.0;
   for (auto _ : state) {
     buffer.fill(fillValue);
     fillValue += 1.0;
   }
 }
-BENCHMARK(AB_Split_Fill_double);
+
+BENCHMARK(Split_Fill_double);
+
+constexpr int size { 1039247 };
+static void Interleaved_Read(benchmark::State& state) {
+  AudioBuffer<float> buffer (size);
+  std::array<float, 2*size> input;
+  std::iota(input.begin(), input.end(), 1.0f);
+  for (auto _ : state) {
+    buffer.readInterleaved(input.data(), size);
+  }
+}
+BENCHMARK(Interleaved_Read);
+
+static void Interleaved_Read_SSE(benchmark::State& state) {
+  AudioBuffer<float> buffer (size);
+  std::array<float, 2*size> input;
+  std::iota(input.begin(), input.end(), 1.0f);
+  for (auto _ : state) {
+    buffer.readInterleaved<VectorOperations::sse>(input.data(), size);
+  }
+}
+BENCHMARK(Interleaved_Read_SSE);
 BENCHMARK_MAIN();
\ No newline at end of file
diff --git a/sources/AudioBuffer.h b/sources/AudioBuffer.h
index 0c33a175..373d36f9 100644
--- a/sources/AudioBuffer.h
+++ b/sources/AudioBuffer.h
@@ -2,6 +2,8 @@
 #include "Buffer.h"
 #include "Helpers.h"
 #include "Globals.h"
+#include <array>
+#include <type_traits>
 
 template<class Type, unsigned int NumChannels = sfz::config::numChannels, unsigned int Alignment = config::defaultAlignment>
 
@@ -35,14 +37,19 @@ public:
     {
         ASSERT(channelIndex >= 0);
         ASSERT(sampleIndex >= 0);
-        return *(buffer.data() + numFrames * channelIndex + sampleIndex);
+        return *(channels[channelIndex] + sampleIndex);
     }
 
+    template<VectorOperations op = VectorOperations::standard>
     void fill(Type value) noexcept
     {
-        if constexpr (config::vectorOperation == config::VectorOperations::sse)
+        if constexpr (op == VectorOperations::sse && std::is_same<Type, float>::value)
         {
-
+            static_assert(Alignment == 16, "Wrong alignment");
+            const __m128 mmValue = _mm_set_ps1(value);
+            for(auto i = 0; i < NumChannels; ++i)
+                for (auto mm = (__m128*)alignedBegin(i); mm < (__m128*)alignedEnd(i); mm++)
+                    _mm_store_ps((float*)mm, mmValue);
         }
         else
         {
@@ -51,6 +58,65 @@ public:
         }
     }
 
+    template<VectorOperations op = VectorOperations::standard>
+    void readInterleaved(float* input, int numFrames) noexcept
+    {
+        ASSERT(this->numFrames >= numFrames);
+        if constexpr (op == VectorOperations::sse && std::is_same<Type, float>::value && NumChannels == 2)
+        {
+            static_assert(Alignment == 16, "Wrong alignment");
+            const int residualFrames = numFrames & (2 * TypeAlignment - 1);
+            const int lastAligned = numFrames - residualFrames;
+            float* in = input;
+            float* out0 = getChannel(0);
+            float* out1 = getChannel(1);
+            const float* end = input + 2 * lastAligned;
+            while (in < end)
+            {
+                auto input0 = _mm_load_ps(in);
+                in += 4;
+                auto input1 = _mm_load_ps(in);
+                in += 4;
+                auto intermediate0 = _mm_unpacklo_ps(input0, input1);
+                auto intermediate1 = _mm_unpackhi_ps(input0, input1);
+                auto output0 = _mm_unpacklo_ps(intermediate0, intermediate1);
+                auto output1 = _mm_unpackhi_ps(intermediate0, intermediate1);
+                _mm_store_ps(out0, output0);
+                _mm_store_ps(out1, output1);
+                out0 += 4;
+                out1 += 4;
+            }
+
+            for (auto chanIdx = 0; chanIdx < NumChannels; chanIdx++)
+            {
+                auto* _in = input + 2 * lastAligned + chanIdx;
+                auto* _end = input + numFrames * NumChannels;
+                auto* _out = getChannel(chanIdx) + lastAligned;
+                while (_in < _end)
+                {
+                    *_out = *_in;
+                    _in += 2;
+                    _out += 1;
+                }
+            }
+        }
+        else
+        {
+            for (auto chanIdx = 0; chanIdx < NumChannels; chanIdx++)
+            {
+                auto* _in = input + chanIdx;
+                auto* _end = input + numFrames * NumChannels;
+                auto* _out = getChannel(chanIdx);
+                while (_in < _end)
+                {
+                    *_out = *_in;
+                    _in += NumChannels;
+                    _out += 1;
+                }
+            }
+        }
+    }
+
     Type* getChannel(int channelIndex) noexcept
     {
         return channels[channelIndex];
@@ -58,12 +124,22 @@ public:
 
     Type* begin(int channelIndex) noexcept
     {
-        return buffer.data() + numFrames * channelIndex;
+        return channels[channelIndex];
     }
 
     Type* end(int channelIndex) noexcept
     {
-        return buffer.data() + numFrames * (channelIndex + 1);
+        return buffer.data() + numFrames * (channelIndex + 1) + padding * channelIndex;
+    }
+
+    Type* alignedBegin(int channelIndex) noexcept
+    {
+        return begin(channelIndex);
+    }
+
+    Type* alignedEnd(int channelIndex) noexcept
+    {
+        return buffer.data() + (channelIndex + 1) * (numFrames + padding);
     }
 
     Type& operator()(int channelIndex, int sampleIndex) noexcept
@@ -84,6 +160,7 @@ private:
     int padding { 0 };
     std::array<Type*, NumChannels> channels;
     Buffer<Type, Alignment> buffer {};
+    Buffer<Type, Alignment> tempBuffer {2 * Alignment};
 };
 
 template<class Type, unsigned int NumChannels = sfz::config::numChannels, unsigned int Alignment = config::defaultAlignment>
@@ -113,11 +190,16 @@ public:
         return resizedOK;
     }
 
+    template<VectorOperations op = VectorOperations::standard>
     void fill(Type value) noexcept
     {
-        if constexpr (config::vectorOperation == config::VectorOperations::sse)
+        if constexpr (op == VectorOperations::sse)
         {
-
+            static_assert(Alignment == 16, "Wrong alignment");
+            const __m128 mmValue = _mm_set_ps1(value);
+            for(auto i = 0; i < NumChannels; ++i)
+                for (auto mm = (__m128*)alignedBegin(i); mm < (__m128*)alignedEnd(i); mm++)
+                    _mm_store_ps((float*)mm, mmValue);
         }
         else
         {
@@ -153,6 +235,16 @@ public:
         return buffers[channelIndex].end();
     }
 
+    Type* alignedBegin(int channelIndex) noexcept
+    {
+        return begin(channelIndex);
+    }
+
+    Type* alignedEnd(int channelIndex) noexcept
+    {
+        return buffers[channelIndex].alignedEnd();
+    }
+
     int getNumFrames() const noexcept { return numFrames; }
     int getNumChannels() const noexcept { return NumChannels; }
     bool empty() const noexcept { return numFrames == 0; }
diff --git a/sources/Buffer.h b/sources/Buffer.h
index b5ffaa69..7cd0a812 100644
--- a/sources/Buffer.h
+++ b/sources/Buffer.h
@@ -9,6 +9,7 @@ class Buffer
 {
 static_assert(std::is_arithmetic<Type>::value, "Type should be arithmetic");
 static_assert(Alignment == 0 || Alignment == 4 || Alignment == 8 || Alignment == 16, "Bad alignment value");
+static constexpr auto AlignmentMask { Alignment - 1 };
 public:
     Buffer() { }
     Buffer(size_t size)
@@ -23,7 +24,7 @@ public:
             return true;
         }
 
-        auto tempSize = newSize + Alignment;
+        auto tempSize = newSize + 2 * AlignmentMask; // To ensure that we have leeway at the beginning and at the end
         auto* newData = largerData != nullptr ? std::realloc(largerData, tempSize * sizeof(Type)) : std::malloc(tempSize * sizeof(Type));
         if (newData == nullptr)
             return false;
@@ -54,6 +55,7 @@ public:
     
     Type* begin() noexcept { return data(); }
     Type* end() noexcept { return data() + alignedSize; }
+    Type* alignedEnd() noexcept { return data() + alignedSize; }
     // const Type* cbegin() const noexcept { return data(); }
     // const Type* cend() const noexcept { return data() + alignedSize; }
 private:
diff --git a/sources/Globals.h b/sources/Globals.h
index 6ef6b00c..ff8e188e 100644
--- a/sources/Globals.h
+++ b/sources/Globals.h
@@ -26,4 +26,6 @@ namespace config
     inline constexpr unsigned int defaultAlignment { 16 };
     
     inline constexpr VectorOperations vectorOperation { VectorOperations::standard };
-} // namespace config
\ No newline at end of file
+} // namespace config
+
+#include <x86intrin.h>
\ No newline at end of file
diff --git a/tests/AudioBuffer.cpp b/tests/AudioBuffer.cpp
index 774e057c..708e99bb 100644
--- a/tests/AudioBuffer.cpp
+++ b/tests/AudioBuffer.cpp
@@ -215,4 +215,301 @@ TEST_CASE("[AudioBuffer/SplitBuffer] Channel alignments (floats)")
     channelAlignmentTest<float, 2, 8>(65537);
     channelAlignmentTest<float, 2, 8>(65536);
     channelAlignmentTest<float, 2, 8>(65535);
+}
+
+TEST_CASE("[AudioBuffer/SplitBuffer] Channel alignments (doubles)")
+{
+    channelAlignmentTest<double, 1>(4);
+    channelAlignmentTest<double, 1>(5);
+    channelAlignmentTest<double, 1>(8);
+    channelAlignmentTest<double, 1>(256);
+    channelAlignmentTest<double, 1>(257);
+    channelAlignmentTest<double, 1>(1023);
+    channelAlignmentTest<double, 1>(1024);
+    channelAlignmentTest<double, 1>(65537);
+    channelAlignmentTest<double, 1>(65536);
+    channelAlignmentTest<double, 1>(65535);
+
+    channelAlignmentTest<double, 2>(4);
+    channelAlignmentTest<double, 2>(5);
+    channelAlignmentTest<double, 2>(8);
+    channelAlignmentTest<double, 2>(256);
+    channelAlignmentTest<double, 2>(257);
+    channelAlignmentTest<double, 2>(1023);
+    channelAlignmentTest<double, 2>(1024);
+    channelAlignmentTest<double, 2>(65537);
+    channelAlignmentTest<double, 2>(65536);
+    channelAlignmentTest<double, 2>(65535);
+
+    channelAlignmentTest<double, 3>(4);
+    channelAlignmentTest<double, 3>(5);
+    channelAlignmentTest<double, 3>(8);
+    channelAlignmentTest<double, 3>(256);
+    channelAlignmentTest<double, 3>(257);
+    channelAlignmentTest<double, 3>(1023);
+    channelAlignmentTest<double, 3>(1024);
+    channelAlignmentTest<double, 3>(65537);
+    channelAlignmentTest<double, 3>(65536);
+    channelAlignmentTest<double, 3>(65535);
+
+    channelAlignmentTest<double, 4>(4);
+    channelAlignmentTest<double, 4>(5);
+    channelAlignmentTest<double, 4>(8);
+    channelAlignmentTest<double, 4>(256);
+    channelAlignmentTest<double, 4>(257);
+    channelAlignmentTest<double, 4>(1023);
+    channelAlignmentTest<double, 4>(1024);
+    channelAlignmentTest<double, 4>(65537);
+    channelAlignmentTest<double, 4>(65536);
+    channelAlignmentTest<double, 4>(65535);
+
+    channelAlignmentTest<double, 1, 8>(4);
+    channelAlignmentTest<double, 1, 8>(5);
+    channelAlignmentTest<double, 1, 8>(8);
+    channelAlignmentTest<double, 1, 8>(256);
+    channelAlignmentTest<double, 1, 8>(257);
+    channelAlignmentTest<double, 1, 8>(1023);
+    channelAlignmentTest<double, 1, 8>(1024);
+    channelAlignmentTest<double, 1, 8>(65537);
+    channelAlignmentTest<double, 1, 8>(65536);
+    channelAlignmentTest<double, 1, 8>(65535);
+
+    channelAlignmentTest<double, 2, 8>(4);
+    channelAlignmentTest<double, 2, 8>(5);
+    channelAlignmentTest<double, 2, 8>(8);
+    channelAlignmentTest<double, 2, 8>(256);
+    channelAlignmentTest<double, 2, 8>(257);
+    channelAlignmentTest<double, 2, 8>(1023);
+    channelAlignmentTest<double, 2, 8>(1024);
+    channelAlignmentTest<double, 2, 8>(65537);
+    channelAlignmentTest<double, 2, 8>(65536);
+    channelAlignmentTest<double, 2, 8>(65535);
+}
+
+TEST_CASE("[AudioBuffer] fills")
+{
+    SECTION("Floats - 0.0")
+    {
+        AudioBuffer<float> buffer(10);
+        buffer.fill(0.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 1.0")
+    {
+        AudioBuffer<float> buffer(10);
+        buffer.fill(1.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 1.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Doubles - 0.0")
+    {
+        AudioBuffer<double> buffer(10);
+        buffer.fill(0.0);
+        std::array<double, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0);
+        std::array<double, 10> real { 2.0 };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Doubles - 1.0")
+    {
+        AudioBuffer<double> buffer(10);
+        buffer.fill(1.0);
+        std::array<double, 10> expected;
+        std::fill(expected.begin(), expected.end(), 1.0);
+        std::array<double, 10> real { 2.0 };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 0.0 - SSE")
+    {
+        AudioBuffer<float> buffer(10);
+        buffer.fill<VectorOperations::sse>(0.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 1.0 - SSE")
+    {
+        AudioBuffer<float> buffer(10);
+        buffer.fill<VectorOperations::sse>(1.0f);
+        std::array<float, 10> expected { 1.0f };
+        std::fill(expected.begin(), expected.end(), 1.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+}
+
+TEST_CASE("[SplitAudioBuffer] fills")
+{
+    SECTION("Floats - 0.0")
+    {
+        SplitAudioBuffer<float> buffer(10);
+        buffer.fill(0.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 1.0")
+    {
+        SplitAudioBuffer<float> buffer(10);
+        buffer.fill(1.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 1.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Doubles - 0.0")
+    {
+        SplitAudioBuffer<double> buffer(10);
+        buffer.fill(0.0);
+        std::array<double, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0);
+        std::array<double, 10> real { 2.0 };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Doubles - 1.0")
+    {
+        SplitAudioBuffer<double> buffer(10);
+        buffer.fill(1.0);
+        std::array<double, 10> expected;
+        std::fill(expected.begin(), expected.end(), 1.0);
+        std::array<double, 10> real { 2.0 };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 0.0 - SSE")
+    {
+        SplitAudioBuffer<float> buffer(10);
+        buffer.fill<VectorOperations::sse>(0.0f);
+        std::array<float, 10> expected;
+        std::fill(expected.begin(), expected.end(), 0.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+
+    SECTION("Floats - 1.0 - SSE")
+    {
+        SplitAudioBuffer<float> buffer(10);
+        buffer.fill<VectorOperations::sse>(1.0f);
+        std::array<float, 10> expected { 1.0f };
+        std::fill(expected.begin(), expected.end(), 1.0f);
+        std::array<float, 10> real { 2.0f };
+        for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        {
+            for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+                real[frameIdx] = buffer(chanIdx, frameIdx);
+            REQUIRE( real == expected );
+        }
+    }
+}
+
+TEST_CASE("[AudioBuffer] Interleave read")
+{
+    AudioBuffer<float> buffer(8);
+    std::array<float, 16> input = { 0.0f, 10.0f, 1.0f, 11.0f, 2.0f, 12.0f, 3.0f, 13.0f, 4.0f, 14.0f, 5.0f, 15.0f, 6.0f, 16.0f, 7.0f, 17.0f};
+    std::array<float, 16> expected = { 0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f, 16.0f, 17.0f };
+    buffer.readInterleaved(input.data(), 8);
+    std::array<float, 16> real { 0.0f };
+    auto realIdx = 0;
+    for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+            real[realIdx++] = buffer(chanIdx, frameIdx);
+    REQUIRE( real == expected );
+}
+
+TEST_CASE("[AudioBuffer] Interleave read -- SSE")
+{
+    AudioBuffer<float> buffer(8);
+    std::array<float, 16> input = { 0.0f, 10.0f, 1.0f, 11.0f, 2.0f, 12.0f, 3.0f, 13.0f, 4.0f, 14.0f, 5.0f, 15.0f, 6.0f, 16.0f, 7.0f, 17.0f};
+    std::array<float, 16> expected = { 0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f, 16.0f, 17.0f };
+    buffer.readInterleaved<VectorOperations::sse>(input.data(), 8);
+    std::array<float, 16> real { 0.0f };
+    auto realIdx = 0;
+    for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+            real[realIdx++] = buffer(chanIdx, frameIdx);
+    REQUIRE( real == expected );
+}
+
+TEST_CASE("[AudioBuffer] Interleave read unaligned end -- SSE")
+{
+    AudioBuffer<float> buffer(10);
+    std::array<float, 20> input = { 0.0f, 10.0f, 1.0f, 11.0f, 2.0f, 12.0f, 3.0f, 13.0f, 4.0f, 14.0f, 5.0f, 15.0f, 6.0f, 16.0f, 7.0f, 17.0f, 8.0f, 18.0f, 9.0f, 19.0f};
+    std::array<float, 20> expected = { 0.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f, 16.0f, 17.0f, 18.0f, 19.0f};
+    buffer.readInterleaved<VectorOperations::sse>(input.data(), 10);
+    std::array<float, 20> real { 0.0f };
+    auto realIdx = 0;
+    for (auto chanIdx = 0; chanIdx < buffer.getNumChannels(); ++chanIdx)
+        for (auto frameIdx = 0; frameIdx < buffer.getNumFrames(); ++frameIdx)
+            real[realIdx++] = buffer(chanIdx, frameIdx);
+    REQUIRE( real == expected );
 }
\ No newline at end of file