Make mSampleData an array of floats instead of an array of arrays

This helps reduce the amount of unused memory in between the mixing channels'
storage, for better cache utilization.

core/device.h

@@ -231,8 +231,7 @@ struct DeviceBase {
     /* Temp storage used for mixer processing. */
     static constexpr size_t MixerLineSize{BufferLineSize + DecoderBase::sMaxPadding};
     static constexpr size_t MixerChannelsMax{16};
-    using MixerBufferLine = std::array<float,MixerLineSize>;
-    alignas(16) std::array<MixerBufferLine,MixerChannelsMax> mSampleData{};
+    alignas(16) std::array<float,MixerLineSize*MixerChannelsMax> mSampleData{};
     alignas(16) std::array<float,MixerLineSize+MaxResamplerPadding> mResampleData{};
 
     alignas(16) std::array<float,BufferLineSize> FilteredData{};

core/voice.cpp

@@ -50,8 +50,7 @@ struct NEONTag;
 #endif
 
 
-static_assert(!(sizeof(DeviceBase::MixerBufferLine)&15),
-    "DeviceBase::MixerBufferLine must be a multiple of 16 bytes");
+static_assert(!(DeviceBase::MixerLineSize&3), "MixerLineSize must be a multiple of 4");
 static_assert(!(MaxResamplerEdge&3), "MaxResamplerEdge is not a multiple of 4");
 
 static_assert((BufferLineSize-1)/MaxPitch > 0, "MaxPitch is too large for BufferLineSize!");
@@ -780,10 +779,16 @@ void Voice::mix(const State vstate, ContextBase *Context, const nanoseconds devi
      */
     std::array<float*,DeviceBase::MixerChannelsMax> SamplePointers;
     const al::span<float*> MixingSamples{SamplePointers.data(), mChans.size()};
-    auto get_bufferline = [](DeviceBase::MixerBufferLine &bufline) noexcept -> float*
-    { return bufline.data(); };
-    std::transform(Device->mSampleData.end() - mChans.size(), Device->mSampleData.end(),
-        MixingSamples.begin(), get_bufferline);
+    {
+        const uint channelStep{(samplesToLoad+3u)&~3u};
+        auto base = Device->mSampleData.end() - mChans.size()*channelStep;
+        std::generate(MixingSamples.begin(), MixingSamples.end(), [&base,channelStep]
+        {
+            const auto ret = base;
+            base += channelStep;
+            return al::to_address(ret);
+        });
+    }
 
     /* UHJ2 and SuperStereo only have 2 buffer channels, but 3 mixing channels
      * (3rd channel is generated from decoding).
@@ -1180,11 +1185,11 @@ void Voice::prepare(DeviceBase *device)
     uint num_channels{(mFmtChannels == FmtUHJ2 || mFmtChannels == FmtSuperStereo) ? 3 :
         (mFmtChannels == FmtMonoDup) ? 2 :
         ChannelsFromFmt(mFmtChannels, std::min(mAmbiOrder, device->mAmbiOrder))};
-    if(num_channels > device->mSampleData.size()) UNLIKELY
+    if(num_channels > device->MixerChannelsMax) UNLIKELY
     {
         ERR("Unexpected channel count: %u (limit: %zu, %d:%d)\n", num_channels,
-            device->mSampleData.size(), mFmtChannels, mAmbiOrder);
-        num_channels = static_cast<uint>(device->mSampleData.size());
+            device->MixerChannelsMax, mFmtChannels, mAmbiOrder);
+        num_channels = device->MixerChannelsMax;
     }
     if(mChans.capacity() > 2 && num_channels < mChans.capacity())
     {