9 files changed, 358 insertions, 15 deletions

diff --git a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
index 466db7ccd..a34b9e753 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
+++ b/src/core/hle/service/nvdrv/devices/nvhost_as_gpu.cpp
@@ -178,7 +178,7 @@ u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& ou
     auto& gpu = system_instance.GPU();
     auto cpu_addr = gpu.MemoryManager().GpuToCpuAddress(params.offset);
     ASSERT(cpu_addr);
-    system_instance.Renderer().Rasterizer().FlushAndInvalidateRegion(*cpu_addr, itr->second.size);
+    gpu.FlushAndInvalidateRegion(*cpu_addr, itr->second.size);
 
     params.offset = gpu.MemoryManager().UnmapBuffer(params.offset, itr->second.size);
 
diff --git a/src/core/memory.cpp b/src/core/memory.cpp
index ec279cef8..6591c45d2 100644
--- a/src/core/memory.cpp
+++ b/src/core/memory.cpp
@@ -356,16 +356,16 @@ void RasterizerFlushVirtualRegion(VAddr start, u64 size, FlushMode mode) {
         const VAddr overlap_end = std::min(end, region_end);
         const VAddr overlap_size = overlap_end - overlap_start;
 
-        auto& rasterizer = system_instance.Renderer().Rasterizer();
+        auto& gpu = system_instance.GPU();
         switch (mode) {
         case FlushMode::Flush:
-            rasterizer.FlushRegion(overlap_start, overlap_size);
+            gpu.FlushRegion(overlap_start, overlap_size);
             break;
         case FlushMode::Invalidate:
-            rasterizer.InvalidateRegion(overlap_start, overlap_size);
+            gpu.InvalidateRegion(overlap_start, overlap_size);
             break;
         case FlushMode::FlushAndInvalidate:
-            rasterizer.FlushAndInvalidateRegion(overlap_start, overlap_size);
+            gpu.FlushAndInvalidateRegion(overlap_start, overlap_size);
             break;
         }
     };
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 3e9d2b3be..3bb5d0ed7 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -17,6 +17,8 @@ add_library(video_core STATIC
     engines/shader_header.h
     gpu.cpp
     gpu.h
+    gpu_thread.cpp
+    gpu_thread.h
     macro_interpreter.cpp
     macro_interpreter.h
     memory_manager.cpp
diff --git a/src/video_core/engines/kepler_memory.cpp b/src/video_core/engines/kepler_memory.cpp
index 4f6126116..aae2a4019 100644
--- a/src/video_core/engines/kepler_memory.cpp
+++ b/src/video_core/engines/kepler_memory.cpp
@@ -48,7 +48,7 @@ void KeplerMemory::ProcessData(u32 data) {
     // We have to invalidate the destination region to evict any outdated surfaces from the cache.
     // We do this before actually writing the new data because the destination address might contain
     // a dirty surface that will have to be written back to memory.
-    rasterizer.InvalidateRegion(*dest_address, sizeof(u32));
+    Core::System::GetInstance().GPU().InvalidateRegion(*dest_address, sizeof(u32));
 
     Memory::Write32(*dest_address, data);
     system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 0474c7ba3..9dfea5999 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -92,12 +92,12 @@ void MaxwellDMA::HandleCopy() {
     const auto FlushAndInvalidate = [&](u32 src_size, u64 dst_size) {
         // TODO(Subv): For now, manually flush the regions until we implement GPU-accelerated
         // copying.
-        rasterizer.FlushRegion(*source_cpu, src_size);
+        Core::System::GetInstance().GPU().FlushRegion(*source_cpu, src_size);
 
         // We have to invalidate the destination region to evict any outdated surfaces from the
         // cache. We do this before actually writing the new data because the destination address
         // might contain a dirty surface that will have to be written back to memory.
-        rasterizer.InvalidateRegion(*dest_cpu, dst_size);
+        Core::System::GetInstance().GPU().InvalidateRegion(*dest_cpu, dst_size);
     };
 
     if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index b0f3310e5..0d7a052dd 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -6,12 +6,14 @@
 #include "core/core.h"
 #include "core/core_timing.h"
 #include "core/memory.h"
+#include "core/settings.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/kepler_compute.h"
 #include "video_core/engines/kepler_memory.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/gpu.h"
+#include "video_core/gpu_thread.h"
 #include "video_core/renderer_base.h"
 
 namespace Tegra {
@@ -37,6 +39,10 @@ GPU::GPU(Core::System& system, VideoCore::RendererBase& renderer) : renderer{ren
     kepler_compute = std::make_unique<Engines::KeplerCompute>(*memory_manager);
     maxwell_dma = std::make_unique<Engines::MaxwellDMA>(system, rasterizer, *memory_manager);
     kepler_memory = std::make_unique<Engines::KeplerMemory>(system, rasterizer, *memory_manager);
+
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread = std::make_unique<VideoCommon::GPUThread::ThreadManager>(renderer, *dma_pusher);
+    }
 }
 
 GPU::~GPU() = default;
@@ -66,13 +72,45 @@ const DmaPusher& GPU::DmaPusher() const {
 }
 
 void GPU::PushGPUEntries(Tegra::CommandList&& entries) {
-    dma_pusher->Push(std::move(entries));
-    dma_pusher->DispatchCalls();
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread->SubmitList(std::move(entries));
+    } else {
+        dma_pusher->Push(std::move(entries));
+        dma_pusher->DispatchCalls();
+    }
 }
 
 void GPU::SwapBuffers(
     std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
-    renderer.SwapBuffers(std::move(framebuffer));
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread->SwapBuffers(std::move(framebuffer));
+    } else {
+        renderer.SwapBuffers(std::move(framebuffer));
+    }
+}
+
+void GPU::FlushRegion(VAddr addr, u64 size) {
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread->FlushRegion(addr, size);
+    } else {
+        renderer.Rasterizer().FlushRegion(addr, size);
+    }
+}
+
+void GPU::InvalidateRegion(VAddr addr, u64 size) {
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread->InvalidateRegion(addr, size);
+    } else {
+        renderer.Rasterizer().InvalidateRegion(addr, size);
+    }
+}
+
+void GPU::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        gpu_thread->FlushAndInvalidateRegion(addr, size);
+    } else {
+        renderer.Rasterizer().FlushAndInvalidateRegion(addr, size);
+    }
 }
 
 u32 RenderTargetBytesPerPixel(RenderTargetFormat format) {
diff --git a/src/video_core/gpu.h b/src/video_core/gpu.h
index 62649bd6e..3f3098bf1 100644
--- a/src/video_core/gpu.h
+++ b/src/video_core/gpu.h
@@ -19,6 +19,10 @@ namespace VideoCore {
 class RendererBase;
 } // namespace VideoCore
 
+namespace VideoCommon::GPUThread {
+class ThreadManager;
+} // namespace VideoCommon::GPUThread
+
 namespace Tegra {
 
 enum class RenderTargetFormat : u32 {
@@ -200,7 +204,7 @@ public:
             std::array<u32, NUM_REGS> reg_array;
         };
     } regs{};
-    
+
     /// Push GPU command entries to be processed
     void PushGPUEntries(Tegra::CommandList&& entries);
 
@@ -208,6 +212,15 @@ public:
     void SwapBuffers(
         std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer);
 
+    /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
+    void FlushRegion(VAddr addr, u64 size);
+
+    /// Notify rasterizer that any caches of the specified region should be invalidated
+    void InvalidateRegion(VAddr addr, u64 size);
+
+    /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
+    void FlushAndInvalidateRegion(VAddr addr, u64 size);
+
 private:
     void ProcessBindMethod(const MethodCall& method_call);
     void ProcessSemaphoreTriggerMethod();
@@ -216,17 +229,18 @@ private:
 
     /// Calls a GPU puller method.
     void CallPullerMethod(const MethodCall& method_call);
-    
+
     /// Calls a GPU engine method.
     void CallEngineMethod(const MethodCall& method_call);
-    
+
     /// Determines where the method should be executed.
     bool ExecuteMethodOnEngine(const MethodCall& method_call);
 
 private:
     std::unique_ptr<Tegra::DmaPusher> dma_pusher;
     std::unique_ptr<Tegra::MemoryManager> memory_manager;
-    
+    std::unique_ptr<VideoCommon::GPUThread::ThreadManager> gpu_thread;
+
     VideoCore::RendererBase& renderer;
 
     /// Mapping of command subchannels to their bound engine ids.
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
new file mode 100644
index 000000000..22c4cca4d
--- /dev/null
+++ b/src/video_core/gpu_thread.cpp
@@ -0,0 +1,154 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+#include "common/microprofile.h"
+#include "core/frontend/scope_acquire_window_context.h"
+#include "core/settings.h"
+#include "video_core/dma_pusher.h"
+#include "video_core/gpu.h"
+#include "video_core/gpu_thread.h"
+#include "video_core/renderer_base.h"
+
+namespace VideoCommon::GPUThread {
+
+/// Executes a single GPU thread command
+static void ExecuteCommand(CommandData* command, VideoCore::RendererBase& renderer,
+                           Tegra::DmaPusher& dma_pusher) {
+    if (const auto submit_list = std::get_if<SubmitListCommand>(command)) {
+        dma_pusher.Push(std::move(submit_list->entries));
+        dma_pusher.DispatchCalls();
+    } else if (const auto data = std::get_if<SwapBuffersCommand>(command)) {
+        renderer.SwapBuffers(data->framebuffer);
+    } else if (const auto data = std::get_if<FlushRegionCommand>(command)) {
+        renderer.Rasterizer().FlushRegion(data->addr, data->size);
+    } else if (const auto data = std::get_if<InvalidateRegionCommand>(command)) {
+        renderer.Rasterizer().InvalidateRegion(data->addr, data->size);
+    } else if (const auto data = std::get_if<FlushAndInvalidateRegionCommand>(command)) {
+        renderer.Rasterizer().FlushAndInvalidateRegion(data->addr, data->size);
+    } else {
+        UNREACHABLE();
+    }
+}
+
+/// Runs the GPU thread
+static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher,
+                      SynchState& state) {
+
+    MicroProfileOnThreadCreate("GpuThread");
+
+    auto WaitForWakeup = [&]() {
+        std::unique_lock<std::mutex> lock{state.signal_mutex};
+        state.signal_condition.wait(lock, [&] { return !state.IsIdle() || !state.is_running; });
+    };
+
+    // Wait for first GPU command before acquiring the window context
+    WaitForWakeup();
+
+    // If emulation was stopped during disk shader loading, abort before trying to acquire context
+    if (!state.is_running) {
+        return;
+    }
+
+    Core::Frontend::ScopeAcquireWindowContext acquire_context{renderer.GetRenderWindow()};
+
+    while (state.is_running) {
+        if (!state.is_running) {
+            return;
+        }
+
+        {
+            // Thread has been woken up, so make the previous write queue the next read queue
+            std::lock_guard<std::mutex> lock{state.signal_mutex};
+            std::swap(state.push_queue, state.pop_queue);
+        }
+
+        // Execute all of the GPU commands
+        while (!state.pop_queue->empty()) {
+            ExecuteCommand(&state.pop_queue->front(), renderer, dma_pusher);
+            state.pop_queue->pop();
+        }
+
+        // Signal that the GPU thread has finished processing commands
+        if (state.IsIdle()) {
+            state.idle_condition.notify_one();
+        }
+
+        // Wait for CPU thread to send more GPU commands
+        WaitForWakeup();
+    }
+}
+
+ThreadManager::ThreadManager(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher)
+    : renderer{renderer}, dma_pusher{dma_pusher}, thread{RunThread, std::ref(renderer),
+                                                         std::ref(dma_pusher), std::ref(state)},
+      thread_id{thread.get_id()} {}
+
+ThreadManager::~ThreadManager() {
+    {
+        // Notify GPU thread that a shutdown is pending
+        std::lock_guard<std::mutex> lock{state.signal_mutex};
+        state.is_running = false;
+    }
+
+    state.signal_condition.notify_one();
+    thread.join();
+}
+
+void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
+    if (entries.empty()) {
+        return;
+    }
+
+    PushCommand(SubmitListCommand(std::move(entries)), false, false);
+}
+
+void ThreadManager::SwapBuffers(
+    std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
+    PushCommand(SwapBuffersCommand(std::move(framebuffer)), true, false);
+}
+
+void ThreadManager::FlushRegion(VAddr addr, u64 size) {
+    if (Settings::values.use_accurate_gpu_emulation) {
+        PushCommand(FlushRegionCommand(addr, size), true, false);
+    }
+}
+
+void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
+    PushCommand(InvalidateRegionCommand(addr, size), true, true);
+}
+
+void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    if (Settings::values.use_accurate_gpu_emulation) {
+        PushCommand(FlushAndInvalidateRegionCommand(addr, size), true, false);
+    } else {
+        InvalidateRegion(addr, size);
+    }
+}
+
+void ThreadManager::PushCommand(CommandData&& command_data, bool wait_for_idle, bool allow_on_cpu) {
+    {
+        std::lock_guard<std::mutex> lock{state.signal_mutex};
+
+        if ((allow_on_cpu && state.IsIdle()) || IsGpuThread()) {
+            // Execute the command synchronously on the current thread
+            ExecuteCommand(&command_data, renderer, dma_pusher);
+            return;
+        }
+
+        // Push the command to the GPU thread
+        state.push_queue->emplace(command_data);
+    }
+
+    // Signal the GPU thread that commands are pending
+    state.signal_condition.notify_one();
+
+    if (wait_for_idle) {
+        // Wait for the GPU to be idle (all commands to be executed)
+        std::unique_lock<std::mutex> lock{state.idle_mutex};
+        state.idle_condition.wait(lock, [this] { return state.IsIdle(); });
+    }
+}
+
+} // namespace VideoCommon::GPUThread
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
new file mode 100644
index 000000000..ad9f9462b
--- /dev/null
+++ b/src/video_core/gpu_thread.h
@@ -0,0 +1,135 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+#include <atomic>
+#include <condition_variable>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <thread>
+#include <variant>
+
+namespace Tegra {
+struct FramebufferConfig;
+class DmaPusher;
+} // namespace Tegra
+
+namespace VideoCore {
+class RendererBase;
+} // namespace VideoCore
+
+namespace VideoCommon::GPUThread {
+
+/// Command to signal to the GPU thread that a command list is ready for processing
+struct SubmitListCommand final {
+    explicit SubmitListCommand(Tegra::CommandList&& entries) : entries{std::move(entries)} {}
+
+    Tegra::CommandList entries;
+};
+
+/// Command to signal to the GPU thread that a swap buffers is pending
+struct SwapBuffersCommand final {
+    explicit SwapBuffersCommand(std::optional<const Tegra::FramebufferConfig> framebuffer)
+        : framebuffer{std::move(framebuffer)} {}
+
+    std::optional<const Tegra::FramebufferConfig> framebuffer;
+};
+
+/// Command to signal to the GPU thread to flush a region
+struct FlushRegionCommand final {
+    explicit constexpr FlushRegionCommand(VAddr addr, u64 size) : addr{addr}, size{size} {}
+
+    const VAddr addr;
+    const u64 size;
+};
+
+/// Command to signal to the GPU thread to invalidate a region
+struct InvalidateRegionCommand final {
+    explicit constexpr InvalidateRegionCommand(VAddr addr, u64 size) : addr{addr}, size{size} {}
+
+    const VAddr addr;
+    const u64 size;
+};
+
+/// Command to signal to the GPU thread to flush and invalidate a region
+struct FlushAndInvalidateRegionCommand final {
+    explicit constexpr FlushAndInvalidateRegionCommand(VAddr addr, u64 size)
+        : addr{addr}, size{size} {}
+
+    const VAddr addr;
+    const u64 size;
+};
+
+using CommandData = std::variant<SubmitListCommand, SwapBuffersCommand, FlushRegionCommand,
+                                 InvalidateRegionCommand, FlushAndInvalidateRegionCommand>;
+
+/// Struct used to synchronize the GPU thread
+struct SynchState final {
+    std::atomic<bool> is_running{true};
+    std::condition_variable signal_condition;
+    std::mutex signal_mutex;
+    std::condition_variable idle_condition;
+    std::mutex idle_mutex;
+
+    // We use two queues for sending commands to the GPU thread, one for writing (push_queue) to and
+    // one for reading from (pop_queue). These are swapped whenever the current pop_queue becomes
+    // empty. This allows for efficient thread-safe access, as it does not require any copies.
+
+    using CommandQueue = std::queue<CommandData>;
+    std::array<CommandQueue, 2> command_queues;
+    CommandQueue* push_queue{&command_queues[0]};
+    CommandQueue* pop_queue{&command_queues[1]};
+
+    /// Returns true if the GPU thread should be idle, meaning there are no commands to process
+    bool IsIdle() const {
+        return command_queues[0].empty() && command_queues[1].empty();
+    }
+};
+
+/// Class used to manage the GPU thread
+class ThreadManager final {
+public:
+    explicit ThreadManager(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher);
+    ~ThreadManager();
+
+    /// Push GPU command entries to be processed
+    void SubmitList(Tegra::CommandList&& entries);
+
+    /// Swap buffers (render frame)
+    void SwapBuffers(
+        std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer);
+
+    /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
+    void FlushRegion(VAddr addr, u64 size);
+
+    /// Notify rasterizer that any caches of the specified region should be invalidated
+    void InvalidateRegion(VAddr addr, u64 size);
+
+    /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
+    void FlushAndInvalidateRegion(VAddr addr, u64 size);
+
+    /// Waits the caller until the GPU thread is idle, used for synchronization
+    void WaitForIdle();
+
+private:
+    /// Pushes a command to be executed by the GPU thread
+    void PushCommand(CommandData&& command_data, bool wait_for_idle, bool allow_on_cpu);
+
+    /// Returns true if this is called by the GPU thread
+    bool IsGpuThread() const {
+        return std::this_thread::get_id() == thread_id;
+    }
+
+private:
+    SynchState state;
+    std::thread thread;
+    std::thread::id thread_id;
+    VideoCore::RendererBase& renderer;
+    Tegra::DmaPusher& dma_pusher;
+};
+
+} // namespace VideoCommon::GPUThread