gpu_thread: Fix deadlock with threading idle state check.

src/core/core.cpp

@@ -36,7 +36,8 @@
 #include "frontend/applets/software_keyboard.h"
 #include "frontend/applets/web_browser.h"
 #include "video_core/debug_utils/debug_utils.h"
-#include "video_core/gpu.h"
+#include "video_core/gpu_asynch.h"
+#include "video_core/gpu_synch.h"
 #include "video_core/renderer_base.h"
 #include "video_core/video_core.h"
 
@@ -129,10 +130,16 @@ struct System::Impl {
             return ResultStatus::ErrorVideoCore;
         }
 
-        gpu_core = std::make_unique<Tegra::GPU>(system, renderer->Rasterizer());
+        is_powered_on = true;
+
+        if (Settings::values.use_asynchronous_gpu_emulation) {
+            gpu_core = std::make_unique<VideoCommon::GPUAsynch>(system, *renderer);
+        } else {
+            gpu_core = std::make_unique<VideoCommon::GPUSynch>(system, *renderer);
+        }
 
         cpu_core_manager.Initialize(system);
-        is_powered_on = true;
+
         LOG_DEBUG(Core, "Initialized OK");
 
         // Reset counters and set time origin to current frame

src/core/hle/service/audio/hwopus.cpp

@@ -9,32 +9,43 @@
 
 #include <opus.h>
 
-#include "common/assert.h"
+#include "common/common_funcs.h"
 #include "common/logging/log.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/hle_ipc.h"
 #include "core/hle/service/audio/hwopus.h"
 
 namespace Service::Audio {
-namespace {
+
 struct OpusDeleter {
     void operator()(void* ptr) const {
         operator delete(ptr);
     }
 };
 
-using OpusDecoderPtr = std::unique_ptr<OpusDecoder, OpusDeleter>;
-
-struct OpusPacketHeader {
-    // Packet size in bytes.
-    u32_be size;
-    // Indicates the final range of the codec's entropy coder.
-    u32_be final_range;
-};
-static_assert(sizeof(OpusPacketHeader) == 0x8, "OpusHeader is an invalid size");
-
-class OpusDecoderStateBase {
+class IHardwareOpusDecoderManager final : public ServiceFramework<IHardwareOpusDecoderManager> {
 public:
+    IHardwareOpusDecoderManager(std::unique_ptr<OpusDecoder, OpusDeleter> decoder, u32 sample_rate,
+                                u32 channel_count)
+        : ServiceFramework("IHardwareOpusDecoderManager"), decoder(std::move(decoder)),
+          sample_rate(sample_rate), channel_count(channel_count) {
+        // clang-format off
+        static const FunctionInfo functions[] = {
+            {0, &IHardwareOpusDecoderManager::DecodeInterleavedOld, "DecodeInterleavedOld"},
+            {1, nullptr, "SetContext"},
+            {2, nullptr, "DecodeInterleavedForMultiStreamOld"},
+            {3, nullptr, "SetContextForMultiStream"},
+            {4, &IHardwareOpusDecoderManager::DecodeInterleavedWithPerfOld, "DecodeInterleavedWithPerfOld"},
+            {5, nullptr, "DecodeInterleavedForMultiStreamWithPerfOld"},
+            {6, &IHardwareOpusDecoderManager::DecodeInterleaved, "DecodeInterleaved"},
+            {7, nullptr, "DecodeInterleavedForMultiStream"},
+        };
+        // clang-format on
+
+        RegisterHandlers(functions);
+    }
+
+private:
     /// Describes extra behavior that may be asked of the decoding context.
     enum class ExtraBehavior {
         /// No extra behavior.
@@ -44,36 +55,30 @@ public:
         ResetContext,
     };
 
-    enum class PerfTime {
-        Disabled,
-        Enabled,
-    };
+    void DecodeInterleavedOld(Kernel::HLERequestContext& ctx) {
+        LOG_DEBUG(Audio, "called");
 
-    virtual ~OpusDecoderStateBase() = default;
-
-    // Decodes interleaved Opus packets. Optionally allows reporting time taken to
-    // perform the decoding, as well as any relevant extra behavior.
-    virtual void DecodeInterleaved(Kernel::HLERequestContext& ctx, PerfTime perf_time,
-                                   ExtraBehavior extra_behavior) = 0;
-};
-
-// Represents the decoder state for a non-multistream decoder.
-class OpusDecoderState final : public OpusDecoderStateBase {
-public:
-    explicit OpusDecoderState(OpusDecoderPtr decoder, u32 sample_rate, u32 channel_count)
-        : decoder{std::move(decoder)}, sample_rate{sample_rate}, channel_count{channel_count} {}
-
-    void DecodeInterleaved(Kernel::HLERequestContext& ctx, PerfTime perf_time,
-                           ExtraBehavior extra_behavior) override {
-        if (perf_time == PerfTime::Disabled) {
-            DecodeInterleavedHelper(ctx, nullptr, extra_behavior);
-        } else {
-            u64 performance = 0;
-            DecodeInterleavedHelper(ctx, &performance, extra_behavior);
-        }
+        DecodeInterleavedHelper(ctx, nullptr, ExtraBehavior::None);
+    }
+
+    void DecodeInterleavedWithPerfOld(Kernel::HLERequestContext& ctx) {
+        LOG_DEBUG(Audio, "called");
+
+        u64 performance = 0;
+        DecodeInterleavedHelper(ctx, &performance, ExtraBehavior::None);
+    }
+
+    void DecodeInterleaved(Kernel::HLERequestContext& ctx) {
+        LOG_DEBUG(Audio, "called");
+
+        IPC::RequestParser rp{ctx};
+        const auto extra_behavior =
+            rp.Pop<bool>() ? ExtraBehavior::ResetContext : ExtraBehavior::None;
+
+        u64 performance = 0;
+        DecodeInterleavedHelper(ctx, &performance, extra_behavior);
     }
 
-private:
     void DecodeInterleavedHelper(Kernel::HLERequestContext& ctx, u64* performance,
                                  ExtraBehavior extra_behavior) {
         u32 consumed = 0;
@@ -84,7 +89,8 @@ private:
             ResetDecoderContext();
         }
 
-        if (!DecodeOpusData(consumed, sample_count, ctx.ReadBuffer(), samples, performance)) {
+        if (!Decoder_DecodeInterleaved(consumed, sample_count, ctx.ReadBuffer(), samples,
+                                       performance)) {
             LOG_ERROR(Audio, "Failed to decode opus data");
             IPC::ResponseBuilder rb{ctx, 2};
             // TODO(ogniK): Use correct error code
@@ -103,27 +109,27 @@ private:
         ctx.WriteBuffer(samples.data(), samples.size() * sizeof(s16));
     }
 
-    bool DecodeOpusData(u32& consumed, u32& sample_count, const std::vector<u8>& input,
-                        std::vector<opus_int16>& output, u64* out_performance_time) const {
+    bool Decoder_DecodeInterleaved(u32& consumed, u32& sample_count, const std::vector<u8>& input,
+                                   std::vector<opus_int16>& output, u64* out_performance_time) {
         const auto start_time = std::chrono::high_resolution_clock::now();
         const std::size_t raw_output_sz = output.size() * sizeof(opus_int16);
-        if (sizeof(OpusPacketHeader) > input.size()) {
+        if (sizeof(OpusHeader) > input.size()) {
             LOG_ERROR(Audio, "Input is smaller than the header size, header_sz={}, input_sz={}",
-                      sizeof(OpusPacketHeader), input.size());
+                      sizeof(OpusHeader), input.size());
             return false;
         }
 
-        OpusPacketHeader hdr{};
-        std::memcpy(&hdr, input.data(), sizeof(OpusPacketHeader));
-        if (sizeof(OpusPacketHeader) + static_cast<u32>(hdr.size) > input.size()) {
+        OpusHeader hdr{};
+        std::memcpy(&hdr, input.data(), sizeof(OpusHeader));
+        if (sizeof(OpusHeader) + static_cast<u32>(hdr.sz) > input.size()) {
             LOG_ERROR(Audio, "Input does not fit in the opus header size. data_sz={}, input_sz={}",
-                      sizeof(OpusPacketHeader) + static_cast<u32>(hdr.size), input.size());
+                      sizeof(OpusHeader) + static_cast<u32>(hdr.sz), input.size());
             return false;
         }
 
-        const auto frame = input.data() + sizeof(OpusPacketHeader);
+        const auto frame = input.data() + sizeof(OpusHeader);
         const auto decoded_sample_count = opus_packet_get_nb_samples(
-            frame, static_cast<opus_int32>(input.size() - sizeof(OpusPacketHeader)),
+            frame, static_cast<opus_int32>(input.size() - sizeof(OpusHeader)),
             static_cast<opus_int32>(sample_rate));
         if (decoded_sample_count * channel_count * sizeof(u16) > raw_output_sz) {
             LOG_ERROR(
@@ -135,18 +141,18 @@ private:
 
         const int frame_size = (static_cast<int>(raw_output_sz / sizeof(s16) / channel_count));
         const auto out_sample_count =
-            opus_decode(decoder.get(), frame, hdr.size, output.data(), frame_size, 0);
+            opus_decode(decoder.get(), frame, hdr.sz, output.data(), frame_size, 0);
         if (out_sample_count < 0) {
             LOG_ERROR(Audio,
                       "Incorrect sample count received from opus_decode, "
                       "output_sample_count={}, frame_size={}, data_sz_from_hdr={}",
-                      out_sample_count, frame_size, static_cast<u32>(hdr.size));
+                      out_sample_count, frame_size, static_cast<u32>(hdr.sz));
             return false;
         }
 
         const auto end_time = std::chrono::high_resolution_clock::now() - start_time;
         sample_count = out_sample_count;
-        consumed = static_cast<u32>(sizeof(OpusPacketHeader) + hdr.size);
+        consumed = static_cast<u32>(sizeof(OpusHeader) + hdr.sz);
         if (out_performance_time != nullptr) {
             *out_performance_time =
                 std::chrono::duration_cast<std::chrono::milliseconds>(end_time).count();
@@ -161,66 +167,21 @@ private:
         opus_decoder_ctl(decoder.get(), OPUS_RESET_STATE);
     }
 
-    OpusDecoderPtr decoder;
+    struct OpusHeader {
+        u32_be sz; // Needs to be BE for some odd reason
+        INSERT_PADDING_WORDS(1);
+    };
+    static_assert(sizeof(OpusHeader) == 0x8, "OpusHeader is an invalid size");
+
+    std::unique_ptr<OpusDecoder, OpusDeleter> decoder;
     u32 sample_rate;
     u32 channel_count;
 };
 
-class IHardwareOpusDecoderManager final : public ServiceFramework<IHardwareOpusDecoderManager> {
-public:
-    explicit IHardwareOpusDecoderManager(std::unique_ptr<OpusDecoderStateBase> decoder_state)
-        : ServiceFramework("IHardwareOpusDecoderManager"), decoder_state{std::move(decoder_state)} {
-        // clang-format off
-        static const FunctionInfo functions[] = {
-            {0, &IHardwareOpusDecoderManager::DecodeInterleavedOld, "DecodeInterleavedOld"},
-            {1, nullptr, "SetContext"},
-            {2, nullptr, "DecodeInterleavedForMultiStreamOld"},
-            {3, nullptr, "SetContextForMultiStream"},
-            {4, &IHardwareOpusDecoderManager::DecodeInterleavedWithPerfOld, "DecodeInterleavedWithPerfOld"},
-            {5, nullptr, "DecodeInterleavedForMultiStreamWithPerfOld"},
-            {6, &IHardwareOpusDecoderManager::DecodeInterleaved, "DecodeInterleaved"},
-            {7, nullptr, "DecodeInterleavedForMultiStream"},
-        };
-        // clang-format on
-
-        RegisterHandlers(functions);
-    }
-
-private:
-    void DecodeInterleavedOld(Kernel::HLERequestContext& ctx) {
-        LOG_DEBUG(Audio, "called");
-
-        decoder_state->DecodeInterleaved(ctx, OpusDecoderStateBase::PerfTime::Disabled,
-                                         OpusDecoderStateBase::ExtraBehavior::None);
-    }
-
-    void DecodeInterleavedWithPerfOld(Kernel::HLERequestContext& ctx) {
-        LOG_DEBUG(Audio, "called");
-
-        decoder_state->DecodeInterleaved(ctx, OpusDecoderStateBase::PerfTime::Enabled,
-                                         OpusDecoderStateBase::ExtraBehavior::None);
-    }
-
-    void DecodeInterleaved(Kernel::HLERequestContext& ctx) {
-        LOG_DEBUG(Audio, "called");
-
-        IPC::RequestParser rp{ctx};
-        const auto extra_behavior = rp.Pop<bool>()
-                                        ? OpusDecoderStateBase::ExtraBehavior::ResetContext
-                                        : OpusDecoderStateBase::ExtraBehavior::None;
-
-        decoder_state->DecodeInterleaved(ctx, OpusDecoderStateBase::PerfTime::Enabled,
-                                         extra_behavior);
-    }
-
-    std::unique_ptr<OpusDecoderStateBase> decoder_state;
-};
-
-std::size_t WorkerBufferSize(u32 channel_count) {
+static std::size_t WorkerBufferSize(u32 channel_count) {
     ASSERT_MSG(channel_count == 1 || channel_count == 2, "Invalid channel count");
     return opus_decoder_get_size(static_cast<int>(channel_count));
 }
-} // Anonymous namespace
 
 void HwOpus::GetWorkBufferSize(Kernel::HLERequestContext& ctx) {
     IPC::RequestParser rp{ctx};
@@ -259,7 +220,8 @@ void HwOpus::OpenOpusDecoder(Kernel::HLERequestContext& ctx) {
     const std::size_t worker_sz = WorkerBufferSize(channel_count);
     ASSERT_MSG(buffer_sz >= worker_sz, "Worker buffer too large");
 
-    OpusDecoderPtr decoder{static_cast<OpusDecoder*>(operator new(worker_sz))};
+    std::unique_ptr<OpusDecoder, OpusDeleter> decoder{
+        static_cast<OpusDecoder*>(operator new(worker_sz))};
     if (const int err = opus_decoder_init(decoder.get(), sample_rate, channel_count)) {
         LOG_ERROR(Audio, "Failed to init opus decoder with error={}", err);
         IPC::ResponseBuilder rb{ctx, 2};
@@ -270,8 +232,8 @@ void HwOpus::OpenOpusDecoder(Kernel::HLERequestContext& ctx) {
 
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
     rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<IHardwareOpusDecoderManager>(
-        std::make_unique<OpusDecoderState>(std::move(decoder), sample_rate, channel_count));
+    rb.PushIpcInterface<IHardwareOpusDecoderManager>(std::move(decoder), sample_rate,
+                                                     channel_count);
 }
 
 HwOpus::HwOpus() : ServiceFramework("hwopus") {

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -36,7 +36,7 @@ void nvdisp_disp0::flip(u32 buffer_handle, u32 offset, u32 format, u32 width, u3
 
     auto& instance = Core::System::GetInstance();
     instance.GetPerfStats().EndGameFrame();
-    instance.Renderer().SwapBuffers(framebuffer);
+    instance.GPU().SwapBuffers(framebuffer);
 }
 
 } // namespace Service::Nvidia::Devices

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);
 

src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp

@@ -136,16 +136,6 @@ u32 nvhost_gpu::AllocateObjectContext(const std::vector<u8>& input, std::vector<
     return 0;
 }
 
-static void PushGPUEntries(Tegra::CommandList&& entries) {
-    if (entries.empty()) {
-        return;
-    }
-
-    auto& dma_pusher{Core::System::GetInstance().GPU().DmaPusher()};
-    dma_pusher.Push(std::move(entries));
-    dma_pusher.DispatchCalls();
-}
-
 u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& output) {
     if (input.size() < sizeof(IoctlSubmitGpfifo)) {
         UNIMPLEMENTED();
@@ -163,7 +153,7 @@ u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& outp
     std::memcpy(entries.data(), &input[sizeof(IoctlSubmitGpfifo)],
                 params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    PushGPUEntries(std::move(entries));
+    Core::System::GetInstance().GPU().PushGPUEntries(std::move(entries));
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;
@@ -184,7 +174,7 @@ u32 nvhost_gpu::KickoffPB(const std::vector<u8>& input, std::vector<u8>& output)
     Memory::ReadBlock(params.address, entries.data(),
                       params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    PushGPUEntries(std::move(entries));
+    Core::System::GetInstance().GPU().PushGPUEntries(std::move(entries));
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;

src/core/hle/service/nvflinger/nvflinger.cpp

@@ -186,7 +186,7 @@ void NVFlinger::Compose() {
 
             // There was no queued buffer to draw, render previous frame
             system_instance.GetPerfStats().EndGameFrame();
-            system_instance.Renderer().SwapBuffers({});
+            system_instance.GPU().SwapBuffers({});
             continue;
         }
 

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;
         }
     };

src/core/settings.h

@@ -393,6 +393,7 @@ struct Values {
     u16 frame_limit;
     bool use_disk_shader_cache;
     bool use_accurate_gpu_emulation;
+    bool use_asynchronous_gpu_emulation;
 
     float bg_red;
     float bg_green;

src/core/telemetry_session.cpp

@@ -162,6 +162,8 @@ TelemetrySession::TelemetrySession() {
              Settings::values.use_disk_shader_cache);
     AddField(Telemetry::FieldType::UserConfig, "Renderer_UseAccurateGpuEmulation",
              Settings::values.use_accurate_gpu_emulation);
+    AddField(Telemetry::FieldType::UserConfig, "Renderer_UseAsynchronousGpuEmulation",
+             Settings::values.use_asynchronous_gpu_emulation);
     AddField(Telemetry::FieldType::UserConfig, "System_UseDockedMode",
              Settings::values.use_docked_mode);
 }

src/video_core/CMakeLists.txt

@@ -17,6 +17,12 @@ add_library(video_core STATIC
     engines/shader_header.h
     gpu.cpp
     gpu.h
+    gpu_asynch.cpp
+    gpu_asynch.h
+    gpu_synch.cpp
+    gpu_synch.h
+    gpu_thread.cpp
+    gpu_thread.h
     macro_interpreter.cpp
     macro_interpreter.h
     memory_manager.cpp

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();

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) {

src/video_core/gpu.cpp

@@ -12,7 +12,7 @@
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/gpu.h"
-#include "video_core/rasterizer_interface.h"
+#include "video_core/renderer_base.h"
 
 namespace Tegra {
 
@@ -28,7 +28,8 @@ u32 FramebufferConfig::BytesPerPixel(PixelFormat format) {
     UNREACHABLE();
 }
 
-GPU::GPU(Core::System& system, VideoCore::RasterizerInterface& rasterizer) {
+GPU::GPU(Core::System& system, VideoCore::RendererBase& renderer) : renderer{renderer} {
+    auto& rasterizer{renderer.Rasterizer()};
     memory_manager = std::make_unique<Tegra::MemoryManager>();
     dma_pusher = std::make_unique<Tegra::DmaPusher>(*this);
     maxwell_3d = std::make_unique<Engines::Maxwell3D>(system, rasterizer, *memory_manager);

src/video_core/gpu.h

@@ -16,8 +16,8 @@ class System;
 }
 
 namespace VideoCore {
-class RasterizerInterface;
-}
+class RendererBase;
+} // namespace VideoCore
 
 namespace Tegra {
 
@@ -119,9 +119,10 @@ enum class EngineID {
     MAXWELL_DMA_COPY_A = 0xB0B5,
 };
 
-class GPU final {
+class GPU {
 public:
-    explicit GPU(Core::System& system, VideoCore::RasterizerInterface& rasterizer);
+    explicit GPU(Core::System& system, VideoCore::RendererBase& renderer);
+
     ~GPU();
 
     struct MethodCall {
@@ -200,8 +201,42 @@ public:
         };
     } regs{};
 
+    /// Push GPU command entries to be processed
+    virtual void PushGPUEntries(Tegra::CommandList&& entries) = 0;
+
+    /// Swap buffers (render frame)
+    virtual void SwapBuffers(
+        std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) = 0;
+
+    /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
+    virtual void FlushRegion(VAddr addr, u64 size) = 0;
+
+    /// Notify rasterizer that any caches of the specified region should be invalidated
+    virtual void InvalidateRegion(VAddr addr, u64 size) = 0;
+
+    /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
+    virtual void FlushAndInvalidateRegion(VAddr addr, u64 size) = 0;
+
 private:
+    void ProcessBindMethod(const MethodCall& method_call);
+    void ProcessSemaphoreTriggerMethod();
+    void ProcessSemaphoreRelease();
+    void ProcessSemaphoreAcquire();
+
+    /// 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);
+
+protected:
     std::unique_ptr<Tegra::DmaPusher> dma_pusher;
+    VideoCore::RendererBase& renderer;
+
+private:
     std::unique_ptr<Tegra::MemoryManager> memory_manager;
 
     /// Mapping of command subchannels to their bound engine ids.
@@ -217,18 +252,6 @@ private:
     std::unique_ptr<Engines::MaxwellDMA> maxwell_dma;
     /// Inline memory engine
     std::unique_ptr<Engines::KeplerMemory> kepler_memory;
-
-    void ProcessBindMethod(const MethodCall& method_call);
-    void ProcessSemaphoreTriggerMethod();
-    void ProcessSemaphoreRelease();
-    void ProcessSemaphoreAcquire();
-
-    // 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);
 };
 
 #define ASSERT_REG_POSITION(field_name, position)                                                  \

src/video_core/gpu_asynch.cpp

@@ -0,0 +1,37 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/gpu_asynch.h"
+#include "video_core/gpu_thread.h"
+#include "video_core/renderer_base.h"
+
+namespace VideoCommon {
+
+GPUAsynch::GPUAsynch(Core::System& system, VideoCore::RendererBase& renderer)
+    : Tegra::GPU(system, renderer), gpu_thread{renderer, *dma_pusher} {}
+
+GPUAsynch::~GPUAsynch() = default;
+
+void GPUAsynch::PushGPUEntries(Tegra::CommandList&& entries) {
+    gpu_thread.SubmitList(std::move(entries));
+}
+
+void GPUAsynch::SwapBuffers(
+    std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
+    gpu_thread.SwapBuffers(std::move(framebuffer));
+}
+
+void GPUAsynch::FlushRegion(VAddr addr, u64 size) {
+    gpu_thread.FlushRegion(addr, size);
+}
+
+void GPUAsynch::InvalidateRegion(VAddr addr, u64 size) {
+    gpu_thread.InvalidateRegion(addr, size);
+}
+
+void GPUAsynch::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    gpu_thread.FlushAndInvalidateRegion(addr, size);
+}
+
+} // namespace VideoCommon

src/video_core/gpu_asynch.h

@@ -0,0 +1,37 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "video_core/gpu.h"
+#include "video_core/gpu_thread.h"
+
+namespace VideoCore {
+class RendererBase;
+} // namespace VideoCore
+
+namespace VideoCommon {
+
+namespace GPUThread {
+class ThreadManager;
+} // namespace GPUThread
+
+/// Implementation of GPU interface that runs the GPU asynchronously
+class GPUAsynch : public Tegra::GPU {
+public:
+    explicit GPUAsynch(Core::System& system, VideoCore::RendererBase& renderer);
+    ~GPUAsynch();
+
+    void PushGPUEntries(Tegra::CommandList&& entries) override;
+    void SwapBuffers(
+        std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) override;
+    void FlushRegion(VAddr addr, u64 size) override;
+    void InvalidateRegion(VAddr addr, u64 size) override;
+    void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
+
+private:
+    GPUThread::ThreadManager gpu_thread;
+};
+
+} // namespace VideoCommon

src/video_core/gpu_synch.cpp

@@ -0,0 +1,37 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/gpu_synch.h"
+#include "video_core/renderer_base.h"
+
+namespace VideoCommon {
+
+GPUSynch::GPUSynch(Core::System& system, VideoCore::RendererBase& renderer)
+    : Tegra::GPU(system, renderer) {}
+
+GPUSynch::~GPUSynch() = default;
+
+void GPUSynch::PushGPUEntries(Tegra::CommandList&& entries) {
+    dma_pusher->Push(std::move(entries));
+    dma_pusher->DispatchCalls();
+}
+
+void GPUSynch::SwapBuffers(
+    std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
+    renderer.SwapBuffers(std::move(framebuffer));
+}
+
+void GPUSynch::FlushRegion(VAddr addr, u64 size) {
+    renderer.Rasterizer().FlushRegion(addr, size);
+}
+
+void GPUSynch::InvalidateRegion(VAddr addr, u64 size) {
+    renderer.Rasterizer().InvalidateRegion(addr, size);
+}
+
+void GPUSynch::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    renderer.Rasterizer().FlushAndInvalidateRegion(addr, size);
+}
+
+} // namespace VideoCommon

src/video_core/gpu_synch.h

@@ -0,0 +1,29 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "video_core/gpu.h"
+
+namespace VideoCore {
+class RendererBase;
+} // namespace VideoCore
+
+namespace VideoCommon {
+
+/// Implementation of GPU interface that runs the GPU synchronously
+class GPUSynch : public Tegra::GPU {
+public:
+    explicit GPUSynch(Core::System& system, VideoCore::RendererBase& renderer);
+    ~GPUSynch();
+
+    void PushGPUEntries(Tegra::CommandList&& entries) override;
+    void SwapBuffers(
+        std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) override;
+    void FlushRegion(VAddr addr, u64 size) override;
+    void InvalidateRegion(VAddr addr, u64 size) override;
+    void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
+};
+
+} // namespace VideoCommon

src/video_core/gpu_thread.cpp

@@ -0,0 +1,152 @@
+// 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.is_idle || !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();
+        }
+
+        state.UpdateIdleState();
+
+        // Signal that the GPU thread has finished processing commands
+        if (state.is_idle) {
+            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) {
+    // Block the CPU when using accurate emulation
+    PushCommand(FlushRegionCommand(addr, size), Settings::values.use_accurate_gpu_emulation, false);
+}
+
+void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
+    PushCommand(InvalidateRegionCommand(addr, size), true, true);
+}
+
+void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    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.is_idle) || IsGpuThread()) {
+            // Execute the command synchronously on the current thread
+            ExecuteCommand(&command_data, renderer, dma_pusher);
+            return;
+        }
+
+        // Push the command to the GPU thread
+        state.UpdateIdleState();
+        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 static_cast<bool>(state.is_idle); });
+    }
+}
+
+} // namespace VideoCommon::GPUThread

src/video_core/gpu_thread.h

@@ -0,0 +1,136 @@
+// 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::atomic<bool> is_idle{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]};
+
+    void UpdateIdleState() {
+        std::lock_guard<std::mutex> lock{idle_mutex};
+        is_idle = 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

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -749,11 +749,7 @@ void RasterizerOpenGL::FlushAll() {}
 
 void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
-
-    if (Settings::values.use_accurate_gpu_emulation) {
-        // Only flush if use_accurate_gpu_emulation is enabled, as it incurs a performance hit
-        res_cache.FlushRegion(addr, size);
-    }
+    res_cache.FlushRegion(addr, size);
 }
 
 void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {

src/yuzu/bootmanager.cpp

@@ -20,10 +20,7 @@
 EmuThread::EmuThread(GRenderWindow* render_window) : render_window(render_window) {}
 
 void EmuThread::run() {
-    if (!Settings::values.use_multi_core) {
-        // Single core mode must acquire OpenGL context for entire emulation session
-        render_window->MakeCurrent();
-    }
+    render_window->MakeCurrent();
 
     MicroProfileOnThreadCreate("EmuThread");
 
@@ -38,6 +35,11 @@ void EmuThread::run() {
 
     emit LoadProgress(VideoCore::LoadCallbackStage::Complete, 0, 0);
 
+    if (Settings::values.use_asynchronous_gpu_emulation) {
+        // Release OpenGL context for the GPU thread
+        render_window->DoneCurrent();
+    }
+
     // holds whether the cpu was running during the last iteration,
     // so that the DebugModeLeft signal can be emitted before the
     // next execution step

src/yuzu/configuration/config.cpp

@@ -374,6 +374,8 @@ void Config::ReadValues() {
         qt_config->value("use_disk_shader_cache", false).toBool();
     Settings::values.use_accurate_gpu_emulation =
         qt_config->value("use_accurate_gpu_emulation", false).toBool();
+    Settings::values.use_asynchronous_gpu_emulation =
+        qt_config->value("use_asynchronous_gpu_emulation", false).toBool();
 
     Settings::values.bg_red = qt_config->value("bg_red", 0.0).toFloat();
     Settings::values.bg_green = qt_config->value("bg_green", 0.0).toFloat();
@@ -633,6 +635,8 @@ void Config::SaveValues() {
     qt_config->setValue("frame_limit", Settings::values.frame_limit);
     qt_config->setValue("use_disk_shader_cache", Settings::values.use_disk_shader_cache);
     qt_config->setValue("use_accurate_gpu_emulation", Settings::values.use_accurate_gpu_emulation);
+    qt_config->setValue("use_asynchronous_gpu_emulation",
+                        Settings::values.use_asynchronous_gpu_emulation);
 
     // Cast to double because Qt's written float values are not human-readable
     qt_config->setValue("bg_red", (double)Settings::values.bg_red);

src/yuzu/configuration/configure_graphics.cpp

@@ -75,6 +75,8 @@ void ConfigureGraphics::setConfiguration() {
     ui->frame_limit->setValue(Settings::values.frame_limit);
     ui->use_disk_shader_cache->setChecked(Settings::values.use_disk_shader_cache);
     ui->use_accurate_gpu_emulation->setChecked(Settings::values.use_accurate_gpu_emulation);
+    ui->use_asynchronous_gpu_emulation->setEnabled(!Core::System::GetInstance().IsPoweredOn());
+    ui->use_asynchronous_gpu_emulation->setChecked(Settings::values.use_asynchronous_gpu_emulation);
     UpdateBackgroundColorButton(QColor::fromRgbF(Settings::values.bg_red, Settings::values.bg_green,
                                                  Settings::values.bg_blue));
 }
@@ -86,6 +88,8 @@ void ConfigureGraphics::applyConfiguration() {
     Settings::values.frame_limit = ui->frame_limit->value();
     Settings::values.use_disk_shader_cache = ui->use_disk_shader_cache->isChecked();
     Settings::values.use_accurate_gpu_emulation = ui->use_accurate_gpu_emulation->isChecked();
+    Settings::values.use_asynchronous_gpu_emulation =
+        ui->use_asynchronous_gpu_emulation->isChecked();
     Settings::values.bg_red = static_cast<float>(bg_color.redF());
     Settings::values.bg_green = static_cast<float>(bg_color.greenF());
     Settings::values.bg_blue = static_cast<float>(bg_color.blueF());

src/yuzu/configuration/configure_graphics.ui

@@ -63,6 +63,13 @@
           </property>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="use_asynchronous_gpu_emulation">
+          <property name="text">
+           <string>Use asynchronous GPU emulation</string>
+          </property>
+         </widget>
+        </item>
         <item>
          <layout class="QHBoxLayout" name="horizontalLayout">
           <item>

src/yuzu_cmd/config.cpp

@@ -354,6 +354,8 @@ void Config::ReadValues() {
         sdl2_config->GetBoolean("Renderer", "use_disk_shader_cache", false);
     Settings::values.use_accurate_gpu_emulation =
         sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
+    Settings::values.use_asynchronous_gpu_emulation =
+        sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
 
     Settings::values.bg_red = (float)sdl2_config->GetReal("Renderer", "bg_red", 0.0);
     Settings::values.bg_green = (float)sdl2_config->GetReal("Renderer", "bg_green", 0.0);

src/yuzu_cmd/default_ini.h

@@ -118,6 +118,10 @@ use_disk_shader_cache =
 # 0 (default): Off (fast), 1 : On (slow)
 use_accurate_gpu_emulation =
 
+# Whether to use asynchronous GPU emulation
+# 0 : Off (slow), 1 (default): On (fast)
+use_asynchronous_gpu_emulation =
+
 # The clear color for the renderer. What shows up on the sides of the bottom screen.
 # Must be in range of 0.0-1.0. Defaults to 1.0 for all.
 bg_red =