35 files changed, 754 insertions, 154 deletions

diff --git a/src/audio_core/CMakeLists.txt b/src/audio_core/CMakeLists.txt
index eba0a5697..a72a907ef 100644
--- a/src/audio_core/CMakeLists.txt
+++ b/src/audio_core/CMakeLists.txt
@@ -3,6 +3,7 @@ set(SRCS
             codec.cpp
             hle/dsp.cpp
             hle/filter.cpp
+            hle/mixers.cpp
             hle/pipe.cpp
             hle/source.cpp
             interpolate.cpp
@@ -16,6 +17,7 @@ set(HEADERS
             hle/common.h
             hle/dsp.h
             hle/filter.h
+            hle/mixers.h
             hle/pipe.h
             hle/source.h
             interpolate.h
diff --git a/src/audio_core/hle/dsp.cpp b/src/audio_core/hle/dsp.cpp
index 5113ad8ca..0640e1eff 100644
--- a/src/audio_core/hle/dsp.cpp
+++ b/src/audio_core/hle/dsp.cpp
@@ -6,6 +6,7 @@
 #include <memory>
 
 #include "audio_core/hle/dsp.h"
+#include "audio_core/hle/mixers.h"
 #include "audio_core/hle/pipe.h"
 #include "audio_core/hle/source.h"
 #include "audio_core/sink.h"
@@ -14,6 +15,8 @@
 namespace DSP {
 namespace HLE {
 
+// Region management
+
 std::array<SharedMemory, 2> g_regions;
 
 static size_t CurrentRegionIndex() {
@@ -41,16 +44,57 @@ static SharedMemory& WriteRegion() {
     return g_regions[1 - CurrentRegionIndex()];
 }
 
+// Audio processing and mixing
+
 static std::array<Source, num_sources> sources = {
     Source(0), Source(1), Source(2), Source(3), Source(4), Source(5),
     Source(6), Source(7), Source(8), Source(9), Source(10), Source(11),
     Source(12), Source(13), Source(14), Source(15), Source(16), Source(17),
     Source(18), Source(19), Source(20), Source(21), Source(22), Source(23)
 };
+static Mixers mixers;
+
+static StereoFrame16 GenerateCurrentFrame() {
+    SharedMemory& read = ReadRegion();
+    SharedMemory& write = WriteRegion();
+
+    std::array<QuadFrame32, 3> intermediate_mixes = {};
+
+    // Generate intermediate mixes
+    for (size_t i = 0; i < num_sources; i++) {
+        write.source_statuses.status[i] = sources[i].Tick(read.source_configurations.config[i], read.adpcm_coefficients.coeff[i]);
+        for (size_t mix = 0; mix < 3; mix++) {
+            sources[i].MixInto(intermediate_mixes[mix], mix);
+        }
+    }
+
+    // Generate final mix
+    write.dsp_status = mixers.Tick(read.dsp_configuration, read.intermediate_mix_samples, write.intermediate_mix_samples, intermediate_mixes);
+
+    StereoFrame16 output_frame = mixers.GetOutput();
+
+    // Write current output frame to the shared memory region
+    for (size_t samplei = 0; samplei < output_frame.size(); samplei++) {
+        for (size_t channeli = 0; channeli < output_frame[0].size(); channeli++) {
+            write.final_samples.pcm16[samplei][channeli] = s16_le(output_frame[samplei][channeli]);
+        }
+    }
+
+    return output_frame;
+}
+
+// Audio output
 
 static std::unique_ptr<AudioCore::Sink> sink;
 static AudioCore::TimeStretcher time_stretcher;
 
+static void OutputCurrentFrame(const StereoFrame16& frame) {
+    time_stretcher.AddSamples(&frame[0][0], frame.size());
+    sink->EnqueueSamples(time_stretcher.Process(sink->SamplesInQueue()));
+}
+
+// Public Interface
+
 void Init() {
     DSP::HLE::ResetPipes();
 
@@ -58,6 +102,8 @@ void Init() {
         source.Reset();
     }
 
+    mixers.Reset();
+
     time_stretcher.Reset();
     if (sink) {
         time_stretcher.SetOutputSampleRate(sink->GetNativeSampleRate());
@@ -75,17 +121,12 @@ void Shutdown() {
 }
 
 bool Tick() {
-    SharedMemory& read = ReadRegion();
-    SharedMemory& write = WriteRegion();
+    StereoFrame16 current_frame = {};
 
-    std::array<QuadFrame32, 3> intermediate_mixes = {};
+    // TODO: Check dsp::DSP semaphore (which indicates emulated application has finished writing to shared memory region)
+    current_frame = GenerateCurrentFrame();
 
-    for (size_t i = 0; i < num_sources; i++) {
-        write.source_statuses.status[i] = sources[i].Tick(read.source_configurations.config[i], read.adpcm_coefficients.coeff[i]);
-        for (size_t mix = 0; mix < 3; mix++) {
-            sources[i].MixInto(intermediate_mixes[mix], mix);
-        }
-    }
+    OutputCurrentFrame(current_frame);
 
     return true;
 }
diff --git a/src/audio_core/hle/dsp.h b/src/audio_core/hle/dsp.h
index f6e53f68f..9275cd7de 100644
--- a/src/audio_core/hle/dsp.h
+++ b/src/audio_core/hle/dsp.h
@@ -428,7 +428,7 @@ ASSERT_DSP_STRUCT(DspStatus, 32);
 /// Final mixed output in PCM16 stereo format, what you hear out of the speakers.
 /// When the application writes to this region it has no effect.
 struct FinalMixSamples {
-    s16_le pcm16[2 * samples_per_frame];
+    s16_le pcm16[samples_per_frame][2];
 };
 ASSERT_DSP_STRUCT(FinalMixSamples, 640);
 
diff --git a/src/audio_core/hle/mixers.cpp b/src/audio_core/hle/mixers.cpp
new file mode 100644
index 000000000..18335f7f0
--- /dev/null
+++ b/src/audio_core/hle/mixers.cpp
@@ -0,0 +1,201 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <cstddef>
+
+#include "audio_core/hle/common.h"
+#include "audio_core/hle/dsp.h"
+#include "audio_core/hle/mixers.h"
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "common/math_util.h"
+
+namespace DSP {
+namespace HLE {
+
+void Mixers::Reset() {
+    current_frame.fill({});
+    state = {};
+}
+
+DspStatus Mixers::Tick(DspConfiguration& config,
+        const IntermediateMixSamples& read_samples,
+        IntermediateMixSamples& write_samples,
+        const std::array<QuadFrame32, 3>& input)
+{
+    ParseConfig(config);
+
+    AuxReturn(read_samples);
+    AuxSend(write_samples, input);
+
+    MixCurrentFrame();
+
+    return GetCurrentStatus();
+}
+
+void Mixers::ParseConfig(DspConfiguration& config) {
+    if (!config.dirty_raw) {
+        return;
+    }
+
+    if (config.mixer1_enabled_dirty) {
+        config.mixer1_enabled_dirty.Assign(0);
+        state.mixer1_enabled = config.mixer1_enabled != 0;
+        LOG_TRACE(Audio_DSP, "mixers mixer1_enabled = %hu", config.mixer1_enabled);
+    }
+
+    if (config.mixer2_enabled_dirty) {
+        config.mixer2_enabled_dirty.Assign(0);
+        state.mixer2_enabled = config.mixer2_enabled != 0;
+        LOG_TRACE(Audio_DSP, "mixers mixer2_enabled = %hu", config.mixer2_enabled);
+    }
+
+    if (config.volume_0_dirty) {
+        config.volume_0_dirty.Assign(0);
+        state.intermediate_mixer_volume[0] = config.volume[0];
+        LOG_TRACE(Audio_DSP, "mixers volume[0] = %f", config.volume[0]);
+    }
+
+    if (config.volume_1_dirty) {
+        config.volume_1_dirty.Assign(0);
+        state.intermediate_mixer_volume[1] = config.volume[1];
+        LOG_TRACE(Audio_DSP, "mixers volume[1] = %f", config.volume[1]);
+    }
+
+    if (config.volume_2_dirty) {
+        config.volume_2_dirty.Assign(0);
+        state.intermediate_mixer_volume[2] = config.volume[2];
+        LOG_TRACE(Audio_DSP, "mixers volume[2] = %f", config.volume[2]);
+    }
+
+    if (config.output_format_dirty) {
+        config.output_format_dirty.Assign(0);
+        state.output_format = config.output_format;
+        LOG_TRACE(Audio_DSP, "mixers output_format = %zu", static_cast<size_t>(config.output_format));
+    }
+
+    if (config.headphones_connected_dirty) {
+        config.headphones_connected_dirty.Assign(0);
+        // Do nothing.
+        // (Note: Whether headphones are connected does affect coefficients used for surround sound.)
+        LOG_TRACE(Audio_DSP, "mixers headphones_connected=%hu", config.headphones_connected);
+    }
+
+    if (config.dirty_raw) {
+        LOG_DEBUG(Audio_DSP, "mixers remaining_dirty=%x", config.dirty_raw);
+    }
+
+    config.dirty_raw = 0;
+}
+
+static s16 ClampToS16(s32 value) {
+    return static_cast<s16>(MathUtil::Clamp(value, -32768, 32767));
+}
+
+static std::array<s16, 2> AddAndClampToS16(const std::array<s16, 2>& a, const std::array<s16, 2>& b) {
+    return {
+        ClampToS16(static_cast<s32>(a[0]) + static_cast<s32>(b[0])),
+        ClampToS16(static_cast<s32>(a[1]) + static_cast<s32>(b[1]))
+    };
+}
+
+void Mixers::DownmixAndMixIntoCurrentFrame(float gain, const QuadFrame32& samples) {
+    // TODO(merry): Limiter. (Currently we're performing final mixing assuming a disabled limiter.)
+
+    switch (state.output_format) {
+    case OutputFormat::Mono:
+        std::transform(current_frame.begin(), current_frame.end(), samples.begin(), current_frame.begin(),
+            [gain](const std::array<s16, 2>& accumulator, const std::array<s32, 4>& sample) -> std::array<s16, 2> {
+                // Downmix to mono
+                s16 mono = ClampToS16(static_cast<s32>((gain * sample[0] + gain * sample[1] + gain * sample[2] + gain * sample[3]) / 2));
+                // Mix into current frame
+                return AddAndClampToS16(accumulator, { mono, mono });
+            });
+        return;
+
+    case OutputFormat::Surround:
+        // TODO(merry): Implement surround sound.
+        // fallthrough
+
+    case OutputFormat::Stereo:
+        std::transform(current_frame.begin(), current_frame.end(), samples.begin(), current_frame.begin(),
+            [gain](const std::array<s16, 2>& accumulator, const std::array<s32, 4>& sample) -> std::array<s16, 2> {
+                // Downmix to stereo
+                s16 left = ClampToS16(static_cast<s32>(gain * sample[0] + gain * sample[2]));
+                s16 right = ClampToS16(static_cast<s32>(gain * sample[1] + gain * sample[3]));
+                // Mix into current frame
+                return AddAndClampToS16(accumulator, { left, right });
+            });
+        return;
+    }
+
+    UNREACHABLE_MSG("Invalid output_format %zu", static_cast<size_t>(state.output_format));
+}
+
+void Mixers::AuxReturn(const IntermediateMixSamples& read_samples) {
+    // NOTE: read_samples.mix{1,2}.pcm32 annoyingly have their dimensions in reverse order to QuadFrame32.
+
+    if (state.mixer1_enabled) {
+        for (size_t sample = 0; sample < samples_per_frame; sample++) {
+            for (size_t channel = 0; channel < 4; channel++) {
+                state.intermediate_mix_buffer[1][sample][channel] = read_samples.mix1.pcm32[channel][sample];
+            }
+        }
+    }
+
+    if (state.mixer2_enabled) {
+        for (size_t sample = 0; sample < samples_per_frame; sample++) {
+            for (size_t channel = 0; channel < 4; channel++) {
+                state.intermediate_mix_buffer[2][sample][channel] = read_samples.mix2.pcm32[channel][sample];
+            }
+        }
+    }
+}
+
+void Mixers::AuxSend(IntermediateMixSamples& write_samples, const std::array<QuadFrame32, 3>& input) {
+    // NOTE: read_samples.mix{1,2}.pcm32 annoyingly have their dimensions in reverse order to QuadFrame32.
+
+    state.intermediate_mix_buffer[0] = input[0];
+
+    if (state.mixer1_enabled) {
+        for (size_t sample = 0; sample < samples_per_frame; sample++) {
+            for (size_t channel = 0; channel < 4; channel++) {
+                write_samples.mix1.pcm32[channel][sample] = input[1][sample][channel];
+            }
+        }
+    } else {
+        state.intermediate_mix_buffer[1] = input[1];
+    }
+
+    if (state.mixer2_enabled) {
+        for (size_t sample = 0; sample < samples_per_frame; sample++) {
+            for (size_t channel = 0; channel < 4; channel++) {
+                write_samples.mix2.pcm32[channel][sample] = input[2][sample][channel];
+            }
+        }
+    } else {
+        state.intermediate_mix_buffer[2] = input[2];
+    }
+}
+
+void Mixers::MixCurrentFrame() {
+    current_frame.fill({});
+
+    for (size_t mix = 0; mix < 3; mix++) {
+        DownmixAndMixIntoCurrentFrame(state.intermediate_mixer_volume[mix], state.intermediate_mix_buffer[mix]);
+    }
+
+    // TODO(merry): Compressor. (We currently assume a disabled compressor.)
+}
+
+DspStatus Mixers::GetCurrentStatus() const {
+    DspStatus status;
+    status.unknown = 0;
+    status.dropped_frames = 0;
+    return status;
+}
+
+} // namespace HLE
+} // namespace DSP
diff --git a/src/audio_core/hle/mixers.h b/src/audio_core/hle/mixers.h
new file mode 100644
index 000000000..b52952eb5
--- /dev/null
+++ b/src/audio_core/hle/mixers.h
@@ -0,0 +1,63 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+
+#include "audio_core/hle/common.h"
+#include "audio_core/hle/dsp.h"
+
+namespace DSP {
+namespace HLE {
+
+class Mixers final {
+public:
+    Mixers() {
+        Reset();
+    }
+
+    void Reset();
+
+    DspStatus Tick(DspConfiguration& config,
+                   const IntermediateMixSamples& read_samples,
+                   IntermediateMixSamples& write_samples,
+                   const std::array<QuadFrame32, 3>& input);
+
+    StereoFrame16 GetOutput() const {
+        return current_frame;
+    }
+
+private:
+    StereoFrame16 current_frame = {};
+
+    using OutputFormat = DspConfiguration::OutputFormat;
+
+    struct {
+        std::array<float, 3> intermediate_mixer_volume = {};
+
+        bool mixer1_enabled = false;
+        bool mixer2_enabled = false;
+        std::array<QuadFrame32, 3> intermediate_mix_buffer = {};
+
+        OutputFormat output_format = OutputFormat::Stereo;
+
+    } state;
+
+    /// INTERNAL: Update our internal state based on the current config.
+    void ParseConfig(DspConfiguration& config);
+    /// INTERNAL: Read samples from shared memory that have been modified by the ARM11.
+    void AuxReturn(const IntermediateMixSamples& read_samples);
+    /// INTERNAL: Write samples to shared memory for the ARM11 to modify.
+    void AuxSend(IntermediateMixSamples& write_samples, const std::array<QuadFrame32, 3>& input);
+    /// INTERNAL: Mix current_frame.
+    void MixCurrentFrame();
+    /// INTERNAL: Downmix from quadraphonic to stereo based on status.output_format and accumulate into current_frame.
+    void DownmixAndMixIntoCurrentFrame(float gain, const QuadFrame32& samples);
+    /// INTERNAL: Generate DspStatus based on internal state.
+    DspStatus GetCurrentStatus() const;
+};
+
+} // namespace HLE
+} // namespace DSP
diff --git a/src/citra_qt/CMakeLists.txt b/src/citra_qt/CMakeLists.txt
index 3f0099200..0a5d4624b 100644
--- a/src/citra_qt/CMakeLists.txt
+++ b/src/citra_qt/CMakeLists.txt
@@ -20,6 +20,7 @@ set(SRCS
             util/spinbox.cpp
             util/util.cpp
             bootmanager.cpp
+            configure_audio.cpp
             configure_debug.cpp
             configure_dialog.cpp
             configure_general.cpp
@@ -51,6 +52,7 @@ set(HEADERS
             util/spinbox.h
             util/util.h
             bootmanager.h
+            configure_audio.h
             configure_debug.h
             configure_dialog.h
             configure_general.h
@@ -69,6 +71,7 @@ set(UIS
             debugger/profiler.ui
             debugger/registers.ui
             configure.ui
+            configure_audio.ui
             configure_debug.ui
             configure_general.ui
             hotkeys.ui
diff --git a/src/citra_qt/configure.ui b/src/citra_qt/configure.ui
index 6ae056ff9..e1624bbef 100644
--- a/src/citra_qt/configure.ui
+++ b/src/citra_qt/configure.ui
@@ -29,6 +29,11 @@
        <string>Input</string>
       </attribute>
      </widget>
+      <widget class="ConfigureAudio" name="audioTab">
+        <attribute name="title">
+          <string>Audio</string>
+        </attribute>
+      </widget>
      <widget class="ConfigureDebug" name="debugTab">
       <attribute name="title">
        <string>Debug</string>
@@ -53,6 +58,12 @@
    <container>1</container>
   </customwidget>
   <customwidget>
+   <class>ConfigureAudio</class>
+   <extends>QWidget</extends>
+   <header>configure_audio.h</header>
+   <container>1</container>
+  </customwidget>
+  <customwidget>
    <class>ConfigureDebug</class>
    <extends>QWidget</extends>
    <header>configure_debug.h</header>
diff --git a/src/citra_qt/configure_audio.cpp b/src/citra_qt/configure_audio.cpp
new file mode 100644
index 000000000..cedfa2f2a
--- /dev/null
+++ b/src/citra_qt/configure_audio.cpp
@@ -0,0 +1,44 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "audio_core/sink_details.h"
+
+#include "citra_qt/configure_audio.h"
+#include "ui_configure_audio.h"
+
+#include "core/settings.h"
+
+ConfigureAudio::ConfigureAudio(QWidget* parent) :
+        QWidget(parent),
+        ui(std::make_unique<Ui::ConfigureAudio>())
+{
+    ui->setupUi(this);
+
+    ui->output_sink_combo_box->clear();
+    ui->output_sink_combo_box->addItem("auto");
+    for (const auto& sink_detail : AudioCore::g_sink_details) {
+        ui->output_sink_combo_box->addItem(sink_detail.id);
+    }
+
+    this->setConfiguration();
+}
+
+ConfigureAudio::~ConfigureAudio() {
+}
+
+void ConfigureAudio::setConfiguration() {
+    int new_sink_index = 0;
+    for (int index = 0; index < ui->output_sink_combo_box->count(); index++) {
+        if (ui->output_sink_combo_box->itemText(index).toStdString() == Settings::values.sink_id) {
+            new_sink_index = index;
+            break;
+        }
+    }
+    ui->output_sink_combo_box->setCurrentIndex(new_sink_index);
+}
+
+void ConfigureAudio::applyConfiguration() {
+    Settings::values.sink_id = ui->output_sink_combo_box->itemText(ui->output_sink_combo_box->currentIndex()).toStdString();
+    Settings::Apply();
+}
diff --git a/src/citra_qt/configure_audio.h b/src/citra_qt/configure_audio.h
new file mode 100644
index 000000000..51df2e27b
--- /dev/null
+++ b/src/citra_qt/configure_audio.h
@@ -0,0 +1,27 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <memory>
+#include <QWidget>
+
+namespace Ui {
+class ConfigureAudio;
+}
+
+class ConfigureAudio : public QWidget {
+    Q_OBJECT
+
+public:
+    explicit ConfigureAudio(QWidget* parent = nullptr);
+    ~ConfigureAudio();
+
+    void applyConfiguration();
+
+private:
+    void setConfiguration();
+
+    std::unique_ptr<Ui::ConfigureAudio> ui;
+};
diff --git a/src/citra_qt/configure_audio.ui b/src/citra_qt/configure_audio.ui
new file mode 100644
index 000000000..d7f6946ca
--- /dev/null
+++ b/src/citra_qt/configure_audio.ui
@@ -0,0 +1,48 @@
+<?xml version="1.0" encoding="utf-8"?>
+
+<ui version="4.0">
+ <class>ConfigureAudio</class>
+ <widget class="QWidget" name="ConfigureAudio">
+  <layout class="QVBoxLayout">
+   <item>
+    <widget class="QGroupBox">
+     <property name="title">
+      <string>Audio</string>
+     </property>
+     <layout class="QVBoxLayout">
+      <item>
+       <layout class="QHBoxLayout">
+        <item>
+         <widget class="QLabel">
+          <property name="text">
+           <string>Output Engine:</string>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QComboBox" name="output_sink_combo_box">
+         </widget>
+        </item>
+       </layout>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item>
+    <spacer>
+     <property name="orientation">
+      <enum>Qt::Vertical</enum>
+     </property>
+     <property name="sizeHint" stdset="0">
+      <size>
+       <width>20</width>
+       <height>40</height>
+      </size>
+     </property>
+    </spacer>
+   </item>
+  </layout>
+ </widget>
+ <resources />
+ <connections />
+</ui>
diff --git a/src/citra_qt/configure_dialog.cpp b/src/citra_qt/configure_dialog.cpp
index 87c26c715..2f0317fe0 100644
--- a/src/citra_qt/configure_dialog.cpp
+++ b/src/citra_qt/configure_dialog.cpp
@@ -25,5 +25,6 @@ void ConfigureDialog::setConfiguration() {
 
 void ConfigureDialog::applyConfiguration() {
     ui->generalTab->applyConfiguration();
+    ui->audioTab->applyConfiguration();
     ui->debugTab->applyConfiguration();
 }
diff --git a/src/citra_qt/debugger/profiler.cpp b/src/citra_qt/debugger/profiler.cpp
index 7bb010f77..585ac049a 100644
--- a/src/citra_qt/debugger/profiler.cpp
+++ b/src/citra_qt/debugger/profiler.cpp
@@ -151,6 +151,8 @@ private:
     /// This timer is used to redraw the widget's contents continuously. To save resources, it only
     /// runs while the widget is visible.
     QTimer update_timer;
+    /// Scale the coordinate system appropriately when physical DPI != logical DPI.
+    qreal x_scale, y_scale;
 };
 
 #endif
@@ -220,11 +222,17 @@ MicroProfileWidget::MicroProfileWidget(QWidget* parent) : QWidget(parent) {
     MicroProfileInitUI();
 
     connect(&update_timer, SIGNAL(timeout()), SLOT(update()));
+
+    QPainter painter(this);
+    x_scale = qreal(painter.device()->physicalDpiX()) / qreal(painter.device()->logicalDpiX());
+    y_scale = qreal(painter.device()->physicalDpiY()) / qreal(painter.device()->logicalDpiY());
 }
 
 void MicroProfileWidget::paintEvent(QPaintEvent* ev) {
     QPainter painter(this);
 
+    painter.scale(x_scale, y_scale);
+
     painter.setBackground(Qt::black);
     painter.eraseRect(rect());
 
@@ -248,24 +256,24 @@ void MicroProfileWidget::hideEvent(QHideEvent* ev) {
 }
 
 void MicroProfileWidget::mouseMoveEvent(QMouseEvent* ev) {
-    MicroProfileMousePosition(ev->x(), ev->y(), 0);
+    MicroProfileMousePosition(ev->x() / x_scale, ev->y() / y_scale, 0);
     ev->accept();
 }
 
 void MicroProfileWidget::mousePressEvent(QMouseEvent* ev) {
-    MicroProfileMousePosition(ev->x(), ev->y(), 0);
+    MicroProfileMousePosition(ev->x() / x_scale, ev->y() / y_scale, 0);
     MicroProfileMouseButton(ev->buttons() & Qt::LeftButton, ev->buttons() & Qt::RightButton);
     ev->accept();
 }
 
 void MicroProfileWidget::mouseReleaseEvent(QMouseEvent* ev) {
-    MicroProfileMousePosition(ev->x(), ev->y(), 0);
+    MicroProfileMousePosition(ev->x() / x_scale, ev->y() / y_scale, 0);
     MicroProfileMouseButton(ev->buttons() & Qt::LeftButton, ev->buttons() & Qt::RightButton);
     ev->accept();
 }
 
 void MicroProfileWidget::wheelEvent(QWheelEvent* ev) {
-    MicroProfileMousePosition(ev->x(), ev->y(), ev->delta() / 120);
+    MicroProfileMousePosition(ev->x() / x_scale, ev->y() / y_scale, ev->delta() / 120);
     ev->accept();
 }
 
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index f6a7566bf..12080a802 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -42,6 +42,7 @@ set(SRCS
             hle/kernel/timer.cpp
             hle/kernel/vm_manager.cpp
             hle/service/ac_u.cpp
+            hle/service/act_a.cpp
             hle/service/act_u.cpp
             hle/service/am/am.cpp
             hle/service/am/am_app.cpp
@@ -176,6 +177,7 @@ set(HEADERS
             hle/kernel/vm_manager.h
             hle/result.h
             hle/service/ac_u.h
+            hle/service/act_a.h
             hle/service/act_u.h
             hle/service/am/am.h
             hle/service/am/am_app.h
diff --git a/src/core/arm/arm_interface.h b/src/core/arm/arm_interface.h
index 533067d4f..d8abe5aeb 100644
--- a/src/core/arm/arm_interface.h
+++ b/src/core/arm/arm_interface.h
@@ -6,6 +6,7 @@
 
 #include "common/common_types.h"
 #include "core/arm/skyeye_common/arm_regformat.h"
+#include "core/arm/skyeye_common/vfp/asm_vfp.h"
 
 namespace Core {
     struct ThreadContext;
diff --git a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
index 8d4b26815..cfc67287f 100644
--- a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
+++ b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
@@ -5527,28 +5527,32 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
 
             // SMUAD and SMLAD
             if (BIT(op2, 1) == 0) {
-                RD = (product1 + product2);
+                u32 rd_val = (product1 + product2);
 
                 if (inst_cream->Ra != 15) {
-                    RD += cpu->Reg[inst_cream->Ra];
+                    rd_val += cpu->Reg[inst_cream->Ra];
 
                     if (ARMul_AddOverflowQ(product1 + product2, cpu->Reg[inst_cream->Ra]))
                         cpu->Cpsr |= (1 << 27);
                 }
 
+                RD = rd_val;
+
                 if (ARMul_AddOverflowQ(product1, product2))
                     cpu->Cpsr |= (1 << 27);
             }
             // SMUSD and SMLSD
             else {
-                RD = (product1 - product2);
+                u32 rd_val = (product1 - product2);
 
                 if (inst_cream->Ra != 15) {
-                    RD += cpu->Reg[inst_cream->Ra];
+                    rd_val += cpu->Reg[inst_cream->Ra];
 
                     if (ARMul_AddOverflowQ(product1 - product2, cpu->Reg[inst_cream->Ra]))
                         cpu->Cpsr |= (1 << 27);
                 }
+
+                RD = rd_val;
             }
         }
 
diff --git a/src/core/hle/kernel/thread.cpp b/src/core/hle/kernel/thread.cpp
index 68f026918..43def6146 100644
--- a/src/core/hle/kernel/thread.cpp
+++ b/src/core/hle/kernel/thread.cpp
@@ -526,6 +526,8 @@ SharedPtr<Thread> SetupMainThread(u32 entry_point, s32 priority) {
 
     SharedPtr<Thread> thread = thread_res.MoveFrom();
 
+    thread->context.fpscr = FPSCR_DEFAULT_NAN | FPSCR_FLUSH_TO_ZERO | FPSCR_ROUND_TOZERO | FPSCR_IXC; // 0x03C00010
+
     // Run new "main" thread
     SwitchContext(thread.get());
 
diff --git a/src/core/hle/service/act_a.cpp b/src/core/hle/service/act_a.cpp
new file mode 100644
index 000000000..3a775fa90
--- /dev/null
+++ b/src/core/hle/service/act_a.cpp
@@ -0,0 +1,26 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "core/hle/service/act_a.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Namespace ACT_A
+
+namespace ACT_A {
+
+const Interface::FunctionInfo FunctionTable[] = {
+    {0x041300C2, nullptr, "UpdateMiiImage"},
+    {0x041B0142, nullptr, "AgreeEula"},
+    {0x04210042, nullptr, "UploadMii"},
+    {0x04230082, nullptr, "ValidateMailAddress"},
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Interface class
+
+Interface::Interface() {
+    Register(FunctionTable);
+}
+
+} // namespace
diff --git a/src/core/hle/service/act_a.h b/src/core/hle/service/act_a.h
new file mode 100644
index 000000000..765cae644
--- /dev/null
+++ b/src/core/hle/service/act_a.h
@@ -0,0 +1,23 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "core/hle/service/service.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Namespace ACT_A
+
+namespace ACT_A {
+
+class Interface : public Service::Interface {
+public:
+    Interface();
+
+    std::string GetPortName() const override {
+        return "act:a";
+    }
+};
+
+} // namespace
diff --git a/src/core/hle/service/act_u.cpp b/src/core/hle/service/act_u.cpp
index b23d17fba..05de4d002 100644
--- a/src/core/hle/service/act_u.cpp
+++ b/src/core/hle/service/act_u.cpp
@@ -10,7 +10,10 @@
 namespace ACT_U {
 
 const Interface::FunctionInfo FunctionTable[] = {
+    {0x00010084, nullptr, "Initialize"},
+    {0x00020040, nullptr, "GetErrorCode"},
     {0x000600C2, nullptr, "GetAccountDataBlock"},
+    {0x000D0040, nullptr, "GenerateUuid"},
 };
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/src/core/hle/service/dsp_dsp.cpp b/src/core/hle/service/dsp_dsp.cpp
index 274fc751a..10730d7ac 100644
--- a/src/core/hle/service/dsp_dsp.cpp
+++ b/src/core/hle/service/dsp_dsp.cpp
@@ -440,9 +440,9 @@ static void GetHeadphoneStatus(Service::Interface* self) {
 
     cmd_buff[0] = IPC::MakeHeader(0x1F, 2, 0);
     cmd_buff[1] = RESULT_SUCCESS.raw; // No error
-    cmd_buff[2] = 0; // Not using headphones?
+    cmd_buff[2] = 0; // Not using headphones
 
-    LOG_WARNING(Service_DSP, "(STUBBED) called");
+    LOG_DEBUG(Service_DSP, "called");
 }
 
 /**
diff --git a/src/core/hle/service/service.cpp b/src/core/hle/service/service.cpp
index 0fe3a4d7a..d7e7d4fe3 100644
--- a/src/core/hle/service/service.cpp
+++ b/src/core/hle/service/service.cpp
@@ -7,6 +7,7 @@
 
 #include "core/hle/service/service.h"
 #include "core/hle/service/ac_u.h"
+#include "core/hle/service/act_a.h"
 #include "core/hle/service/act_u.h"
 #include "core/hle/service/csnd_snd.h"
 #include "core/hle/service/dlp_srvr.h"
@@ -119,6 +120,7 @@ void Init() {
     Service::PTM::Init();
 
     AddService(new AC_U::Interface);
+    AddService(new ACT_A::Interface);
     AddService(new ACT_U::Interface);
     AddService(new CSND_SND::Interface);
     AddService(new DLP_SRVR::Interface);
diff --git a/src/core/hle/svc.cpp b/src/core/hle/svc.cpp
index 3a53126c1..2bf122a6d 100644
--- a/src/core/hle/svc.cpp
+++ b/src/core/hle/svc.cpp
@@ -502,6 +502,9 @@ static ResultCode CreateThread(Handle* out_handle, s32 priority, u32 entry_point
 
     CASCADE_RESULT(SharedPtr<Thread> thread, Kernel::Thread::Create(
             name, entry_point, priority, arg, processor_id, stack_top));
+
+    thread->context.fpscr = FPSCR_DEFAULT_NAN | FPSCR_FLUSH_TO_ZERO | FPSCR_ROUND_TOZERO; // 0x03C00000
+
     CASCADE_RESULT(*out_handle, Kernel::g_handle_table.Create(std::move(thread)));
 
     LOG_TRACE(Kernel_SVC, "called entrypoint=0x%08X (%s), arg=0x%08X, stacktop=0x%08X, "
diff --git a/src/video_core/debug_utils/debug_utils.cpp b/src/video_core/debug_utils/debug_utils.cpp
index 2f645b441..871368323 100644
--- a/src/video_core/debug_utils/debug_utils.cpp
+++ b/src/video_core/debug_utils/debug_utils.cpp
@@ -696,106 +696,125 @@ finalise:
 #endif
 }
 
-void DumpTevStageConfig(const std::array<Pica::Regs::TevStageConfig,6>& stages)
-{
+static std::string ReplacePattern(const std::string& input, const std::string& pattern, const std::string& replacement) {
+    size_t start = input.find(pattern);
+    if (start == std::string::npos)
+        return input;
+
+    std::string ret = input;
+    ret.replace(start, pattern.length(), replacement);
+    return ret;
+}
+
+static std::string GetTevStageConfigSourceString(const Pica::Regs::TevStageConfig::Source& source) {
     using Source = Pica::Regs::TevStageConfig::Source;
+    static const std::map<Source, std::string> source_map = {
+        { Source::PrimaryColor,           "PrimaryColor" },
+        { Source::PrimaryFragmentColor,   "PrimaryFragmentColor" },
+        { Source::SecondaryFragmentColor, "SecondaryFragmentColor" },
+        { Source::Texture0,               "Texture0" },
+        { Source::Texture1,               "Texture1" },
+        { Source::Texture2,               "Texture2" },
+        { Source::Texture3,               "Texture3" },
+        { Source::PreviousBuffer,         "PreviousBuffer" },
+        { Source::Constant,               "Constant" },
+        { Source::Previous,               "Previous" },
+    };
+
+    const auto src_it = source_map.find(source);
+    if (src_it == source_map.end())
+        return "Unknown";
+
+    return src_it->second;
+}
+
+static std::string GetTevStageConfigColorSourceString(const Pica::Regs::TevStageConfig::Source& source, const Pica::Regs::TevStageConfig::ColorModifier modifier) {
     using ColorModifier = Pica::Regs::TevStageConfig::ColorModifier;
+    static const std::map<ColorModifier, std::string> color_modifier_map = {
+        { ColorModifier::SourceColor,         "%source.rgb" },
+        { ColorModifier::OneMinusSourceColor, "(1.0 - %source.rgb)" },
+        { ColorModifier::SourceAlpha,         "%source.aaa" },
+        { ColorModifier::OneMinusSourceAlpha, "(1.0 - %source.aaa)" },
+        { ColorModifier::SourceRed,           "%source.rrr" },
+        { ColorModifier::OneMinusSourceRed,   "(1.0 - %source.rrr)" },
+        { ColorModifier::SourceGreen,         "%source.ggg" },
+        { ColorModifier::OneMinusSourceGreen, "(1.0 - %source.ggg)" },
+        { ColorModifier::SourceBlue,          "%source.bbb" },
+        { ColorModifier::OneMinusSourceBlue,  "(1.0 - %source.bbb)" },
+    };
+
+    auto src_str = GetTevStageConfigSourceString(source);
+    auto modifier_it = color_modifier_map.find(modifier);
+    std::string modifier_str = "%source.????";
+    if (modifier_it != color_modifier_map.end())
+        modifier_str = modifier_it->second;
+
+    return ReplacePattern(modifier_str, "%source", src_str);
+}
+
+static std::string GetTevStageConfigAlphaSourceString(const Pica::Regs::TevStageConfig::Source& source, const Pica::Regs::TevStageConfig::AlphaModifier modifier) {
     using AlphaModifier = Pica::Regs::TevStageConfig::AlphaModifier;
+    static const std::map<AlphaModifier, std::string> alpha_modifier_map = {
+        { AlphaModifier::SourceAlpha,         "%source.a" },
+        { AlphaModifier::OneMinusSourceAlpha, "(1.0 - %source.a)" },
+        { AlphaModifier::SourceRed,           "%source.r" },
+        { AlphaModifier::OneMinusSourceRed,   "(1.0 - %source.r)" },
+        { AlphaModifier::SourceGreen,         "%source.g" },
+        { AlphaModifier::OneMinusSourceGreen, "(1.0 - %source.g)" },
+        { AlphaModifier::SourceBlue,          "%source.b" },
+        { AlphaModifier::OneMinusSourceBlue,  "(1.0 - %source.b)" },
+    };
+
+    auto src_str = GetTevStageConfigSourceString(source);
+    auto modifier_it = alpha_modifier_map.find(modifier);
+    std::string modifier_str = "%source.????";
+    if (modifier_it != alpha_modifier_map.end())
+        modifier_str = modifier_it->second;
+
+    return ReplacePattern(modifier_str, "%source", src_str);
+}
+
+static std::string GetTevStageConfigOperationString(const Pica::Regs::TevStageConfig::Operation& operation) {
     using Operation = Pica::Regs::TevStageConfig::Operation;
+    static const std::map<Operation, std::string> combiner_map = {
+        { Operation::Replace,         "%source1" },
+        { Operation::Modulate,        "(%source1 * %source2)" },
+        { Operation::Add,             "(%source1 + %source2)" },
+        { Operation::AddSigned,       "(%source1 + %source2) - 0.5" },
+        { Operation::Lerp,            "lerp(%source1, %source2, %source3)" },
+        { Operation::Subtract,        "(%source1 - %source2)" },
+        { Operation::Dot3_RGB,        "dot(%source1, %source2)" },
+        { Operation::MultiplyThenAdd, "((%source1 * %source2) + %source3)" },
+        { Operation::AddThenMultiply, "((%source1 + %source2) * %source3)" },
+    };
 
-    std::string stage_info = "Tev setup:\n";
-    for (size_t index = 0; index < stages.size(); ++index) {
-        const auto& tev_stage = stages[index];
+    const auto op_it = combiner_map.find(operation);
+    if (op_it == combiner_map.end())
+        return "Unknown op (%source1, %source2, %source3)";
 
-        static const std::map<Source, std::string> source_map = {
-            { Source::PrimaryColor, "PrimaryColor" },
-            { Source::Texture0, "Texture0" },
-            { Source::Texture1, "Texture1" },
-            { Source::Texture2, "Texture2" },
-            { Source::Constant, "Constant" },
-            { Source::Previous, "Previous" },
-        };
+    return op_it->second;
+}
 
-        static const std::map<ColorModifier, std::string> color_modifier_map = {
-            { ColorModifier::SourceColor, { "%source.rgb" } },
-            { ColorModifier::SourceAlpha, { "%source.aaa" } },
-        };
-        static const std::map<AlphaModifier, std::string> alpha_modifier_map = {
-            { AlphaModifier::SourceAlpha, "%source.a" },
-            { AlphaModifier::OneMinusSourceAlpha, "(255 - %source.a)" },
-        };
+std::string GetTevStageConfigColorCombinerString(const Pica::Regs::TevStageConfig& tev_stage) {
+    auto op_str = GetTevStageConfigOperationString(tev_stage.color_op);
+    op_str = ReplacePattern(op_str, "%source1", GetTevStageConfigColorSourceString(tev_stage.color_source1, tev_stage.color_modifier1));
+    op_str = ReplacePattern(op_str, "%source2", GetTevStageConfigColorSourceString(tev_stage.color_source2, tev_stage.color_modifier2));
+    return   ReplacePattern(op_str, "%source3", GetTevStageConfigColorSourceString(tev_stage.color_source3, tev_stage.color_modifier3));
+}
 
-        static const std::map<Operation, std::string> combiner_map = {
-            { Operation::Replace, "%source1" },
-            { Operation::Modulate, "(%source1 * %source2) / 255" },
-            { Operation::Add, "(%source1 + %source2)" },
-            { Operation::Lerp, "lerp(%source1, %source2, %source3)" },
-        };
+std::string GetTevStageConfigAlphaCombinerString(const Pica::Regs::TevStageConfig& tev_stage) {
+    auto op_str = GetTevStageConfigOperationString(tev_stage.alpha_op);
+    op_str = ReplacePattern(op_str, "%source1", GetTevStageConfigAlphaSourceString(tev_stage.alpha_source1, tev_stage.alpha_modifier1));
+    op_str = ReplacePattern(op_str, "%source2", GetTevStageConfigAlphaSourceString(tev_stage.alpha_source2, tev_stage.alpha_modifier2));
+    return   ReplacePattern(op_str, "%source3", GetTevStageConfigAlphaSourceString(tev_stage.alpha_source3, tev_stage.alpha_modifier3));
+}
 
-        static auto ReplacePattern =
-                [](const std::string& input, const std::string& pattern, const std::string& replacement) -> std::string {
-                    size_t start = input.find(pattern);
-                    if (start == std::string::npos)
-                        return input;
-
-                    std::string ret = input;
-                    ret.replace(start, pattern.length(), replacement);
-                    return ret;
-                };
-        static auto GetColorSourceStr =
-                [](const Source& src, const ColorModifier& modifier) {
-                    auto src_it = source_map.find(src);
-                    std::string src_str = "Unknown";
-                    if (src_it != source_map.end())
-                        src_str = src_it->second;
-
-                    auto modifier_it = color_modifier_map.find(modifier);
-                    std::string modifier_str = "%source.????";
-                    if (modifier_it != color_modifier_map.end())
-                        modifier_str = modifier_it->second;
-
-                    return ReplacePattern(modifier_str, "%source", src_str);
-                };
-        static auto GetColorCombinerStr =
-                [](const Regs::TevStageConfig& tev_stage) {
-                    auto op_it = combiner_map.find(tev_stage.color_op);
-                    std::string op_str = "Unknown op (%source1, %source2, %source3)";
-                    if (op_it != combiner_map.end())
-                        op_str = op_it->second;
-
-                    op_str = ReplacePattern(op_str, "%source1", GetColorSourceStr(tev_stage.color_source1, tev_stage.color_modifier1));
-                    op_str = ReplacePattern(op_str, "%source2", GetColorSourceStr(tev_stage.color_source2, tev_stage.color_modifier2));
-                    return   ReplacePattern(op_str, "%source3", GetColorSourceStr(tev_stage.color_source3, tev_stage.color_modifier3));
-                };
-        static auto GetAlphaSourceStr =
-                [](const Source& src, const AlphaModifier& modifier) {
-                    auto src_it = source_map.find(src);
-                    std::string src_str = "Unknown";
-                    if (src_it != source_map.end())
-                        src_str = src_it->second;
-
-                    auto modifier_it = alpha_modifier_map.find(modifier);
-                    std::string modifier_str = "%source.????";
-                    if (modifier_it != alpha_modifier_map.end())
-                        modifier_str = modifier_it->second;
-
-                    return ReplacePattern(modifier_str, "%source", src_str);
-                };
-        static auto GetAlphaCombinerStr =
-                [](const Regs::TevStageConfig& tev_stage) {
-                    auto op_it = combiner_map.find(tev_stage.alpha_op);
-                    std::string op_str = "Unknown op (%source1, %source2, %source3)";
-                    if (op_it != combiner_map.end())
-                        op_str = op_it->second;
-
-                    op_str = ReplacePattern(op_str, "%source1", GetAlphaSourceStr(tev_stage.alpha_source1, tev_stage.alpha_modifier1));
-                    op_str = ReplacePattern(op_str, "%source2", GetAlphaSourceStr(tev_stage.alpha_source2, tev_stage.alpha_modifier2));
-                    return   ReplacePattern(op_str, "%source3", GetAlphaSourceStr(tev_stage.alpha_source3, tev_stage.alpha_modifier3));
-                };
-
-        stage_info += "Stage " + std::to_string(index) + ": " + GetColorCombinerStr(tev_stage) + "   " + GetAlphaCombinerStr(tev_stage) + "\n";
+void DumpTevStageConfig(const std::array<Pica::Regs::TevStageConfig, 6>& stages) {
+    std::string stage_info = "Tev setup:\n";
+    for (size_t index = 0; index < stages.size(); ++index) {
+        const auto& tev_stage = stages[index];
+        stage_info += "Stage " + std::to_string(index) + ": " + GetTevStageConfigColorCombinerString(tev_stage) + "   " + GetTevStageConfigAlphaCombinerString(tev_stage) + "\n";
     }
-
     LOG_TRACE(HW_GPU, "%s", stage_info.c_str());
 }
 
diff --git a/src/video_core/debug_utils/debug_utils.h b/src/video_core/debug_utils/debug_utils.h
index f628292a4..92e9734ae 100644
--- a/src/video_core/debug_utils/debug_utils.h
+++ b/src/video_core/debug_utils/debug_utils.h
@@ -224,7 +224,11 @@ const Math::Vec4<u8> LookupTexture(const u8* source, int s, int t, const Texture
 
 void DumpTexture(const Pica::Regs::TextureConfig& texture_config, u8* data);
 
-void DumpTevStageConfig(const std::array<Pica::Regs::TevStageConfig,6>& stages);
+std::string GetTevStageConfigColorCombinerString(const Pica::Regs::TevStageConfig& tev_stage);
+std::string GetTevStageConfigAlphaCombinerString(const Pica::Regs::TevStageConfig& tev_stage);
+
+/// Dumps the Tev stage config to log at trace level
+void DumpTevStageConfig(const std::array<Pica::Regs::TevStageConfig, 6>& stages);
 
 /**
  * Used in the vertex loader to merge access records. TODO: Investigate if actually useful.
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index ed2e2f3ae..bcd1ae78d 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -104,7 +104,6 @@ RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
 
     // Sync fixed function OpenGL state
     SyncCullMode();
-    SyncDepthModifiers();
     SyncBlendEnabled();
     SyncBlendFuncs();
     SyncBlendColor();
@@ -259,8 +258,10 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
 
     // Depth modifiers
     case PICA_REG_INDEX(viewport_depth_range):
+        SyncDepthScale();
+        break;
     case PICA_REG_INDEX(viewport_depth_near_plane):
-        SyncDepthModifiers();
+        SyncDepthOffset();
         break;
 
     // Depth buffering
@@ -880,6 +881,8 @@ void RasterizerOpenGL::SetShader() {
         glUniformBlockBinding(current_shader->shader.handle, block_index, 0);
 
         // Update uniforms
+        SyncDepthScale();
+        SyncDepthOffset();
         SyncAlphaTest();
         SyncCombinerColor();
         auto& tev_stages = Pica::g_state.regs.GetTevStages();
@@ -922,13 +925,20 @@ void RasterizerOpenGL::SyncCullMode() {
     }
 }
 
-void RasterizerOpenGL::SyncDepthModifiers() {
+void RasterizerOpenGL::SyncDepthScale() {
     float depth_scale = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_range).ToFloat32();
-    float depth_offset = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_near_plane).ToFloat32();
+    if (depth_scale != uniform_block_data.data.depth_scale) {
+        uniform_block_data.data.depth_scale = depth_scale;
+        uniform_block_data.dirty = true;
+    }
+}
 
-    uniform_block_data.data.depth_scale = depth_scale;
-    uniform_block_data.data.depth_offset = depth_offset;
-    uniform_block_data.dirty = true;
+void RasterizerOpenGL::SyncDepthOffset() {
+    float depth_offset = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_near_plane).ToFloat32();
+    if (depth_offset != uniform_block_data.data.depth_offset) {
+        uniform_block_data.data.depth_offset = depth_offset;
+        uniform_block_data.dirty = true;
+    }
 }
 
 void RasterizerOpenGL::SyncBlendEnabled() {
@@ -937,6 +947,8 @@ void RasterizerOpenGL::SyncBlendEnabled() {
 
 void RasterizerOpenGL::SyncBlendFuncs() {
     const auto& regs = Pica::g_state.regs;
+    state.blend.rgb_equation = PicaToGL::BlendEquation(regs.output_merger.alpha_blending.blend_equation_rgb);
+    state.blend.a_equation = PicaToGL::BlendEquation(regs.output_merger.alpha_blending.blend_equation_a);
     state.blend.src_rgb_func = PicaToGL::BlendFunc(regs.output_merger.alpha_blending.factor_source_rgb);
     state.blend.dst_rgb_func = PicaToGL::BlendFunc(regs.output_merger.alpha_blending.factor_dest_rgb);
     state.blend.src_a_func = PicaToGL::BlendFunc(regs.output_merger.alpha_blending.factor_source_a);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index eed00011a..d70369400 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -339,8 +339,11 @@ private:
     /// Syncs the cull mode to match the PICA register
     void SyncCullMode();
 
-    /// Syncs the depth scale and offset to match the PICA registers
-    void SyncDepthModifiers();
+    /// Syncs the depth scale to match the PICA register
+    void SyncDepthScale();
+
+    /// Syncs the depth offset to match the PICA register
+    void SyncDepthOffset();
 
     /// Syncs the blend enabled status to match the PICA register
     void SyncBlendEnabled();
@@ -413,7 +416,7 @@ private:
         UniformData data;
         bool lut_dirty[6];
         bool dirty;
-    } uniform_block_data;
+    } uniform_block_data = {};
 
     std::array<SamplerInfo, 3> texture_samplers;
     OGLVertexArray vertex_array;
@@ -422,5 +425,5 @@ private:
     OGLFramebuffer framebuffer;
 
     std::array<OGLTexture, 6> lighting_luts;
-    std::array<std::array<GLvec4, 256>, 6> lighting_lut_data;
+    std::array<std::array<GLvec4, 256>, 6> lighting_lut_data{};
 };
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index 02cd9f417..fa141fc9a 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -36,6 +36,8 @@ OpenGLState::OpenGLState() {
     stencil.action_stencil_fail = GL_KEEP;
 
     blend.enabled = false;
+    blend.rgb_equation = GL_FUNC_ADD;
+    blend.a_equation = GL_FUNC_ADD;
     blend.src_rgb_func = GL_ONE;
     blend.dst_rgb_func = GL_ZERO;
     blend.src_a_func = GL_ONE;
@@ -165,6 +167,11 @@ void OpenGLState::Apply() const {
                             blend.src_a_func, blend.dst_a_func);
     }
 
+    if (blend.rgb_equation != cur_state.blend.rgb_equation ||
+            blend.a_equation != cur_state.blend.a_equation) {
+        glBlendEquationSeparate(blend.rgb_equation, blend.a_equation);
+    }
+
     if (logic_op != cur_state.logic_op) {
         glLogicOp(logic_op);
     }
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 24f20e47c..228727054 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -40,6 +40,8 @@ public:
 
     struct {
         bool enabled; // GL_BLEND
+        GLenum rgb_equation; // GL_BLEND_EQUATION_RGB
+        GLenum a_equation; // GL_BLEND_EQUATION_ALPHA
         GLenum src_rgb_func; // GL_BLEND_SRC_RGB
         GLenum dst_rgb_func; // GL_BLEND_DST_RGB
         GLenum src_a_func; // GL_BLEND_SRC_ALPHA
diff --git a/src/video_core/renderer_opengl/pica_to_gl.h b/src/video_core/renderer_opengl/pica_to_gl.h
index 976d1f364..6dc2758c5 100644
--- a/src/video_core/renderer_opengl/pica_to_gl.h
+++ b/src/video_core/renderer_opengl/pica_to_gl.h
@@ -78,6 +78,26 @@ inline GLenum WrapMode(Pica::Regs::TextureConfig::WrapMode mode) {
     return gl_mode;
 }
 
+inline GLenum BlendEquation(Pica::Regs::BlendEquation equation) {
+    static const GLenum blend_equation_table[] = {
+        GL_FUNC_ADD,              // BlendEquation::Add
+        GL_FUNC_SUBTRACT,         // BlendEquation::Subtract
+        GL_FUNC_REVERSE_SUBTRACT, // BlendEquation::ReverseSubtract
+        GL_MIN,                   // BlendEquation::Min
+        GL_MAX,                   // BlendEquation::Max
+    };
+
+    // Range check table for input
+    if (static_cast<size_t>(equation) >= ARRAY_SIZE(blend_equation_table)) {
+        LOG_CRITICAL(Render_OpenGL, "Unknown blend equation %d", equation);
+        UNREACHABLE();
+
+        return GL_FUNC_ADD;
+    }
+
+    return blend_equation_table[(unsigned)equation];
+}
+
 inline GLenum BlendFunc(Pica::Regs::BlendFactor factor) {
     static const GLenum blend_func_table[] = {
         GL_ZERO,                     // BlendFactor::Zero
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index e93a9d92a..161097610 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -64,6 +64,7 @@ MICROPROFILE_DEFINE(GPU_Shader, "GPU", "Shader", MP_RGB(50, 50, 240));
 
 OutputVertex ShaderSetup::Run(UnitState<false>& state, const InputVertex& input, int num_attributes) {
     auto& config = g_state.regs.vs;
+    auto& setup = g_state.vs;
 
     MICROPROFILE_SCOPE(GPU_Shader);
 
@@ -81,11 +82,11 @@ OutputVertex ShaderSetup::Run(UnitState<false>& state, const InputVertex& input,
 
 #ifdef ARCHITECTURE_x86_64
     if (VideoCore::g_shader_jit_enabled)
-        jit_shader->Run(&state.registers, g_state.regs.vs.main_offset);
+        jit_shader->Run(setup, state, config.main_offset);
     else
-        RunInterpreter(state);
+        RunInterpreter(setup, state, config.main_offset);
 #else
-    RunInterpreter(state);
+    RunInterpreter(setup, state, config.main_offset);
 #endif // ARCHITECTURE_x86_64
 
     // Setup output data
@@ -156,7 +157,7 @@ DebugData<true> ShaderSetup::ProduceDebugInfo(const InputVertex& input, int num_
     state.conditional_code[0] = false;
     state.conditional_code[1] = false;
 
-    RunInterpreter(state);
+    RunInterpreter(setup, state, config.main_offset);
     return state.debug;
 }
 
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index 983e4a967..84898f21c 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -283,10 +283,10 @@ struct UnitState {
     static size_t InputOffset(const SourceRegister& reg) {
         switch (reg.GetRegisterType()) {
         case RegisterType::Input:
-            return offsetof(UnitState::Registers, input) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
+            return offsetof(UnitState, registers.input) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
 
         case RegisterType::Temporary:
-            return offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
+            return offsetof(UnitState, registers.temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
 
         default:
             UNREACHABLE();
@@ -297,10 +297,10 @@ struct UnitState {
     static size_t OutputOffset(const DestRegister& reg) {
         switch (reg.GetRegisterType()) {
         case RegisterType::Output:
-            return offsetof(UnitState::Registers, output) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
+            return offsetof(UnitState, registers.output) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
 
         case RegisterType::Temporary:
-            return offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
+            return offsetof(UnitState, registers.temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>);
 
         default:
             UNREACHABLE();
@@ -323,6 +323,23 @@ struct ShaderSetup {
         std::array<Math::Vec4<u8>, 4> i;
     } uniforms;
 
+    static size_t UniformOffset(RegisterType type, unsigned index) {
+        switch (type) {
+        case RegisterType::FloatUniform:
+            return offsetof(ShaderSetup, uniforms.f) + index*sizeof(Math::Vec4<float24>);
+
+        case RegisterType::BoolUniform:
+            return offsetof(ShaderSetup, uniforms.b) + index*sizeof(bool);
+
+        case RegisterType::IntUniform:
+            return offsetof(ShaderSetup, uniforms.i) + index*sizeof(Math::Vec4<u8>);
+
+        default:
+            UNREACHABLE();
+            return 0;
+        }
+    }
+
     std::array<u32, 1024> program_code;
     std::array<u32, 1024> swizzle_data;
 
diff --git a/src/video_core/shader/shader_interpreter.cpp b/src/video_core/shader/shader_interpreter.cpp
index 3a827d11f..714e8bfd5 100644
--- a/src/video_core/shader/shader_interpreter.cpp
+++ b/src/video_core/shader/shader_interpreter.cpp
@@ -41,11 +41,11 @@ struct CallStackElement {
 };
 
 template<bool Debug>
-void RunInterpreter(UnitState<Debug>& state) {
+void RunInterpreter(const ShaderSetup& setup, UnitState<Debug>& state, unsigned offset) {
     // TODO: Is there a maximal size for this?
     boost::container::static_vector<CallStackElement, 16> call_stack;
 
-    u32 program_counter = g_state.regs.vs.main_offset;
+    u32 program_counter = offset;
 
     const auto& uniforms = g_state.vs.uniforms;
     const auto& swizzle_data = g_state.vs.swizzle_data;
@@ -647,8 +647,8 @@ void RunInterpreter(UnitState<Debug>& state) {
 }
 
 // Explicit instantiation
-template void RunInterpreter(UnitState<false>& state);
-template void RunInterpreter(UnitState<true>& state);
+template void RunInterpreter(const ShaderSetup& setup, UnitState<false>& state, unsigned offset);
+template void RunInterpreter(const ShaderSetup& setup, UnitState<true>& state, unsigned offset);
 
 } // namespace
 
diff --git a/src/video_core/shader/shader_interpreter.h b/src/video_core/shader/shader_interpreter.h
index 6048cdf3a..bb3ce1c6e 100644
--- a/src/video_core/shader/shader_interpreter.h
+++ b/src/video_core/shader/shader_interpreter.h
@@ -11,7 +11,7 @@ namespace Shader {
 template <bool Debug> struct UnitState;
 
 template<bool Debug>
-void RunInterpreter(UnitState<Debug>& state);
+void RunInterpreter(const ShaderSetup& setup, UnitState<Debug>& state, unsigned offset);
 
 } // namespace
 
diff --git a/src/video_core/shader/shader_jit_x64.cpp b/src/video_core/shader/shader_jit_x64.cpp
index 99f6c51eb..43e7e6b4c 100644
--- a/src/video_core/shader/shader_jit_x64.cpp
+++ b/src/video_core/shader/shader_jit_x64.cpp
@@ -102,7 +102,7 @@ const JitFunction instr_table[64] = {
 // purposes, as documented below:
 
 /// Pointer to the uniform memory
-static const X64Reg UNIFORMS = R9;
+static const X64Reg SETUP = R9;
 /// The two 32-bit VS address offset registers set by the MOVA instruction
 static const X64Reg ADDROFFS_REG_0 = R10;
 static const X64Reg ADDROFFS_REG_1 = R11;
@@ -117,7 +117,7 @@ static const X64Reg COND0 = R13;
 /// Result of the previous CMP instruction for the Y-component comparison
 static const X64Reg COND1 = R14;
 /// Pointer to the UnitState instance for the current VS unit
-static const X64Reg REGISTERS = R15;
+static const X64Reg STATE = R15;
 /// SIMD scratch register
 static const X64Reg SCRATCH = XMM0;
 /// Loaded with the first swizzled source register, otherwise can be used as a scratch register
@@ -136,7 +136,7 @@ static const X64Reg NEGBIT = XMM15;
 // State registers that must not be modified by external functions calls
 // Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed
 static const BitSet32 persistent_regs = {
-    UNIFORMS, REGISTERS, // Pointers to register blocks
+    SETUP, STATE, // Pointers to register blocks
     ADDROFFS_REG_0, ADDROFFS_REG_1, LOOPCOUNT_REG, COND0, COND1, // Cached registers
     ONE+16, NEGBIT+16, // Constants
 };
@@ -177,10 +177,10 @@ void JitShader::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRe
     size_t src_offset;
 
     if (src_reg.GetRegisterType() == RegisterType::FloatUniform) {
-        src_ptr = UNIFORMS;
-        src_offset = src_reg.GetIndex() * sizeof(float24) * 4;
+        src_ptr = SETUP;
+        src_offset = ShaderSetup::UniformOffset(RegisterType::FloatUniform, src_reg.GetIndex());
     } else {
-        src_ptr = REGISTERS;
+        src_ptr = STATE;
         src_offset = UnitState<false>::InputOffset(src_reg);
     }
 
@@ -264,11 +264,11 @@ void JitShader::Compile_DestEnable(Instruction instr,X64Reg src) {
     // If all components are enabled, write the result to the destination register
     if (swiz.dest_mask == NO_DEST_REG_MASK) {
         // Store dest back to memory
-        MOVAPS(MDisp(REGISTERS, dest_offset_disp), src);
+        MOVAPS(MDisp(STATE, dest_offset_disp), src);
 
     } else {
         // Not all components are enabled, so mask the result when storing to the destination register...
-        MOVAPS(SCRATCH, MDisp(REGISTERS, dest_offset_disp));
+        MOVAPS(SCRATCH, MDisp(STATE, dest_offset_disp));
 
         if (Common::GetCPUCaps().sse4_1) {
             u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) | ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
@@ -287,7 +287,7 @@ void JitShader::Compile_DestEnable(Instruction instr,X64Reg src) {
         }
 
         // Store dest back to memory
-        MOVAPS(MDisp(REGISTERS, dest_offset_disp), SCRATCH);
+        MOVAPS(MDisp(STATE, dest_offset_disp), SCRATCH);
     }
 }
 
@@ -336,8 +336,8 @@ void JitShader::Compile_EvaluateCondition(Instruction instr) {
 }
 
 void JitShader::Compile_UniformCondition(Instruction instr) {
-    int offset = offsetof(decltype(g_state.vs.uniforms), b) + (instr.flow_control.bool_uniform_id * sizeof(bool));
-    CMP(sizeof(bool) * 8, MDisp(UNIFORMS, offset), Imm8(0));
+    int offset = ShaderSetup::UniformOffset(RegisterType::BoolUniform, instr.flow_control.bool_uniform_id);
+    CMP(sizeof(bool) * 8, MDisp(SETUP, offset), Imm8(0));
 }
 
 BitSet32 JitShader::PersistentCallerSavedRegs() {
@@ -714,8 +714,8 @@ void JitShader::Compile_LOOP(Instruction instr) {
 
     looping = true;
 
-    int offset = offsetof(decltype(g_state.vs.uniforms), i) + (instr.flow_control.int_uniform_id * sizeof(Math::Vec4<u8>));
-    MOV(32, R(LOOPCOUNT), MDisp(UNIFORMS, offset));
+    int offset = ShaderSetup::UniformOffset(RegisterType::IntUniform, instr.flow_control.int_uniform_id);
+    MOV(32, R(LOOPCOUNT), MDisp(SETUP, offset));
     MOV(32, R(LOOPCOUNT_REG), R(LOOPCOUNT));
     SHR(32, R(LOOPCOUNT_REG), Imm8(8));
     AND(32, R(LOOPCOUNT_REG), Imm32(0xff)); // Y-component is the start
@@ -826,8 +826,8 @@ void JitShader::Compile() {
     // The stack pointer is 8 modulo 16 at the entry of a procedure
     ABI_PushRegistersAndAdjustStack(ABI_ALL_CALLEE_SAVED, 8);
 
-    MOV(PTRBITS, R(REGISTERS), R(ABI_PARAM1));
-    MOV(PTRBITS, R(UNIFORMS), ImmPtr(&g_state.vs.uniforms));
+    MOV(PTRBITS, R(SETUP), R(ABI_PARAM1));
+    MOV(PTRBITS, R(STATE), R(ABI_PARAM2));
 
     // Zero address/loop  registers
     XOR(64, R(ADDROFFS_REG_0), R(ADDROFFS_REG_0));
@@ -845,7 +845,7 @@ void JitShader::Compile() {
     MOVAPS(NEGBIT, MatR(RAX));
 
     // Jump to start of the shader program
-    JMPptr(R(ABI_PARAM2));
+    JMPptr(R(ABI_PARAM3));
 
     // Compile entire program
     Compile_Block(static_cast<unsigned>(g_state.vs.program_code.size()));
diff --git a/src/video_core/shader/shader_jit_x64.h b/src/video_core/shader/shader_jit_x64.h
index 30aa7ff30..5468459d4 100644
--- a/src/video_core/shader/shader_jit_x64.h
+++ b/src/video_core/shader/shader_jit_x64.h
@@ -36,8 +36,8 @@ class JitShader : public Gen::XCodeBlock {
 public:
     JitShader();
 
-    void Run(void* registers, unsigned offset) const {
-        program(registers, code_ptr[offset]);
+    void Run(const ShaderSetup& setup, UnitState<false>& state, unsigned offset) const {
+        program(&setup, &state, code_ptr[offset]);
     }
 
     void Compile();
@@ -117,7 +117,7 @@ private:
     /// Branches that need to be fixed up once the entire shader program is compiled
     std::vector<std::pair<Gen::FixupBranch, unsigned>> fixup_branches;
 
-    using CompiledShader = void(void* registers, const u8* start_addr);
+    using CompiledShader = void(const void* setup, void* state, const u8* start_addr);
     CompiledShader* program = nullptr;
 };