remove <f32>

We can remove this since its already a f32 value

.ci/templates/build-single.yml

@@ -1,20 +1,22 @@
 parameters:
   artifactSource: 'true'
+  cache: 'false'
 
 steps:
 - task: DockerInstaller@0
   displayName: 'Prepare Environment'
   inputs:
     dockerVersion: '17.09.0-ce'
-- task: CacheBeta@0
-  displayName: 'Cache Build System'
-  inputs:
-    key: yuzu-v1-$(BuildName)-$(BuildSuffix)-$(CacheSuffix)
-    path: $(System.DefaultWorkingDirectory)/ccache
-    cacheHitVar: CACHE_RESTORED
+- ${{ if eq(parameters.cache, 'true') }}:
+  - task: CacheBeta@0
+    displayName: 'Cache Build System'
+    inputs:
+      key: yuzu-v1-$(BuildName)-$(BuildSuffix)-$(CacheSuffix)
+      path: $(System.DefaultWorkingDirectory)/ccache
+      cacheHitVar: CACHE_RESTORED
 - script: chmod a+x ./.ci/scripts/$(ScriptFolder)/exec.sh && ./.ci/scripts/$(ScriptFolder)/exec.sh
   displayName: 'Build'
-- script: chmod a+x ./.ci/scripts/$(ScriptFolder)/upload.sh && ./.ci/scripts/$(ScriptFolder)/upload.sh
+- script: chmod a+x ./.ci/scripts/$(ScriptFolder)/upload.sh && RELEASE_NAME=$(BuildName) ./.ci/scripts/$(ScriptFolder)/upload.sh
   displayName: 'Package Artifacts'
 - publish: artifacts
   artifact: 'yuzu-$(BuildName)-$(BuildSuffix)'

.ci/templates/build-standard.yml

@@ -19,4 +19,5 @@ jobs:
       needSubmodules: 'true'
   - template: ./build-single.yml
     parameters:
-      artifactSource: 'false'
+      artifactSource: 'false'
+      cache: $(parameters.cache)

.ci/templates/build-testing.yml

@@ -4,18 +4,20 @@ jobs:
   pool:
     vmImage: ubuntu-latest
   strategy:
-    maxParallel: 10
+    maxParallel: 5
     matrix:
       windows:
         BuildSuffix: 'windows-testing'
         ScriptFolder: 'windows'
   steps:
+  - script: sudo apt upgrade python3-pip && pip install requests urllib3
+    displayName: 'Prepare Environment'
   - task: PythonScript@0
     condition: eq(variables['Build.Reason'], 'PullRequest')
     displayName: 'Determine Testing Status'
     inputs:
       scriptSource: 'filePath'
-      scriptPath: '../scripts/merge/check-label-presence.py'
+      scriptPath: '.ci/scripts/merge/check-label-presence.py'
       arguments: '$(System.PullRequest.PullRequestNumber) create-testing-build'
   - ${{ if eq(variables.enabletesting, 'true') }}:
     - template: ./sync-source.yml
@@ -27,4 +29,5 @@ jobs:
         matchLabel: 'testing-merge'
     - template: ./build-single.yml
       parameters:
-        artifactSource: 'false'
+        artifactSource: 'false'
+        cache: 'false'

.ci/yuzu-mainline.yml

@@ -21,3 +21,5 @@ stages:
   dependsOn: format
   jobs:
   - template: ./templates/build-standard.yml
+    parameters:
+      cache: 'true'

.ci/yuzu-verify.yml

@@ -15,4 +15,6 @@ stages:
   dependsOn: format
   jobs:
   - template: ./templates/build-standard.yml
-  - template: ./templates/build-testing.yml
+    parameters:
+      cache: 'false'
+  - template: ./templates/build-testing.yml

CMakeModules/GenerateSCMRev.cmake

@@ -81,6 +81,7 @@ set(HASH_FILES
     "${VIDEO_CORE}/shader/decode/register_set_predicate.cpp"
     "${VIDEO_CORE}/shader/decode/shift.cpp"
     "${VIDEO_CORE}/shader/decode/video.cpp"
+    "${VIDEO_CORE}/shader/decode/warp.cpp"
     "${VIDEO_CORE}/shader/decode/xmad.cpp"
     "${VIDEO_CORE}/shader/control_flow.cpp"
     "${VIDEO_CORE}/shader/control_flow.h"

src/common/CMakeLists.txt

@@ -55,6 +55,7 @@ add_custom_command(OUTPUT scm_rev.cpp
       "${VIDEO_CORE}/shader/decode/register_set_predicate.cpp"
       "${VIDEO_CORE}/shader/decode/shift.cpp"
       "${VIDEO_CORE}/shader/decode/video.cpp"
+      "${VIDEO_CORE}/shader/decode/warp.cpp"
       "${VIDEO_CORE}/shader/decode/xmad.cpp"
       "${VIDEO_CORE}/shader/control_flow.cpp"
       "${VIDEO_CORE}/shader/control_flow.h"

src/common/alignment.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include <cstddef>
+#include <memory>
 #include <type_traits>
 
 namespace Common {
@@ -37,4 +38,63 @@ constexpr bool IsWordAligned(T value) {
     return (value & 0b11) == 0;
 }
 
+template <typename T, std::size_t Align = 16>
+class AlignmentAllocator {
+public:
+    using value_type = T;
+    using size_type = std::size_t;
+    using difference_type = std::ptrdiff_t;
+
+    using pointer = T*;
+    using const_pointer = const T*;
+
+    using reference = T&;
+    using const_reference = const T&;
+
+public:
+    pointer address(reference r) noexcept {
+        return std::addressof(r);
+    }
+
+    const_pointer address(const_reference r) const noexcept {
+        return std::addressof(r);
+    }
+
+    pointer allocate(size_type n) {
+        return static_cast<pointer>(::operator new (n, std::align_val_t{Align}));
+    }
+
+    void deallocate(pointer p, size_type) {
+        ::operator delete (p, std::align_val_t{Align});
+    }
+
+    void construct(pointer p, const value_type& wert) {
+        new (p) value_type(wert);
+    }
+
+    void destroy(pointer p) {
+        p->~value_type();
+    }
+
+    size_type max_size() const noexcept {
+        return size_type(-1) / sizeof(value_type);
+    }
+
+    template <typename T2>
+    struct rebind {
+        using other = AlignmentAllocator<T2, Align>;
+    };
+
+    bool operator!=(const AlignmentAllocator<T, Align>& other) const noexcept {
+        return !(*this == other);
+    }
+
+    // Returns true if and only if storage allocated from *this
+    // can be deallocated from other, and vice versa.
+    // Always returns true for stateless allocators.
+    bool operator==(const AlignmentAllocator<T, Align>& other) const noexcept {
+        return true;
+    }
+};
+
 } // namespace Common

src/core/CMakeLists.txt

@@ -111,6 +111,8 @@ add_library(core STATIC
     frontend/scope_acquire_window_context.h
     gdbstub/gdbstub.cpp
     gdbstub/gdbstub.h
+    hardware_interrupt_manager.cpp
+    hardware_interrupt_manager.h
     hle/ipc.h
     hle/ipc_helpers.h
     hle/kernel/address_arbiter.cpp
@@ -372,6 +374,7 @@ add_library(core STATIC
     hle/service/nvdrv/devices/nvmap.h
     hle/service/nvdrv/interface.cpp
     hle/service/nvdrv/interface.h
+    hle/service/nvdrv/nvdata.h
     hle/service/nvdrv/nvdrv.cpp
     hle/service/nvdrv/nvdrv.h
     hle/service/nvdrv/nvmemp.cpp

src/core/core.cpp

@@ -19,6 +19,7 @@
 #include "core/file_sys/vfs_concat.h"
 #include "core/file_sys/vfs_real.h"
 #include "core/gdbstub/gdbstub.h"
+#include "core/hardware_interrupt_manager.h"
 #include "core/hle/kernel/client_port.h"
 #include "core/hle/kernel/kernel.h"
 #include "core/hle/kernel/process.h"
@@ -151,7 +152,7 @@ struct System::Impl {
         if (!renderer->Init()) {
             return ResultStatus::ErrorVideoCore;
         }
-
+        interrupt_manager = std::make_unique<Core::Hardware::InterruptManager>(system);
         gpu_core = VideoCore::CreateGPU(system);
 
         is_powered_on = true;
@@ -298,6 +299,7 @@ struct System::Impl {
     std::unique_ptr<VideoCore::RendererBase> renderer;
     std::unique_ptr<Tegra::GPU> gpu_core;
     std::shared_ptr<Tegra::DebugContext> debug_context;
+    std::unique_ptr<Core::Hardware::InterruptManager> interrupt_manager;
     CpuCoreManager cpu_core_manager;
     bool is_powered_on = false;
 
@@ -444,6 +446,14 @@ const Tegra::GPU& System::GPU() const {
     return *impl->gpu_core;
 }
 
+Core::Hardware::InterruptManager& System::InterruptManager() {
+    return *impl->interrupt_manager;
+}
+
+const Core::Hardware::InterruptManager& System::InterruptManager() const {
+    return *impl->interrupt_manager;
+}
+
 VideoCore::RendererBase& System::Renderer() {
     return *impl->renderer;
 }

src/core/core.h

@@ -70,6 +70,10 @@ namespace Core::Timing {
 class CoreTiming;
 }
 
+namespace Core::Hardware {
+class InterruptManager;
+}
+
 namespace Core {
 
 class ARM_Interface;
@@ -234,6 +238,12 @@ public:
     /// Provides a constant reference to the core timing instance.
     const Timing::CoreTiming& CoreTiming() const;
 
+    /// Provides a reference to the interrupt manager instance.
+    Core::Hardware::InterruptManager& InterruptManager();
+
+    /// Provides a constant reference to the interrupt manager instance.
+    const Core::Hardware::InterruptManager& InterruptManager() const;
+
     /// Provides a reference to the kernel instance.
     Kernel::KernelCore& Kernel();
 

src/core/hardware_interrupt_manager.cpp

@@ -0,0 +1,30 @@
+// Copyright 2019 Yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "core/core.h"
+#include "core/core_timing.h"
+#include "core/hardware_interrupt_manager.h"
+#include "core/hle/service/nvdrv/interface.h"
+#include "core/hle/service/sm/sm.h"
+
+namespace Core::Hardware {
+
+InterruptManager::InterruptManager(Core::System& system_in) : system(system_in) {
+    gpu_interrupt_event =
+        system.CoreTiming().RegisterEvent("GPUInterrupt", [this](u64 message, s64) {
+            auto nvdrv = system.ServiceManager().GetService<Service::Nvidia::NVDRV>("nvdrv");
+            const u32 syncpt = static_cast<u32>(message >> 32);
+            const u32 value = static_cast<u32>(message);
+            nvdrv->SignalGPUInterruptSyncpt(syncpt, value);
+        });
+}
+
+InterruptManager::~InterruptManager() = default;
+
+void InterruptManager::GPUInterruptSyncpt(const u32 syncpoint_id, const u32 value) {
+    const u64 msg = (static_cast<u64>(syncpoint_id) << 32ULL) | value;
+    system.CoreTiming().ScheduleEvent(10, gpu_interrupt_event, msg);
+}
+
+} // namespace Core::Hardware

src/core/hardware_interrupt_manager.h

@@ -0,0 +1,31 @@
+// Copyright 2019 Yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/common_types.h"
+
+namespace Core {
+class System;
+}
+
+namespace Core::Timing {
+struct EventType;
+}
+
+namespace Core::Hardware {
+
+class InterruptManager {
+public:
+    explicit InterruptManager(Core::System& system);
+    ~InterruptManager();
+
+    void GPUInterruptSyncpt(u32 syncpoint_id, u32 value);
+
+private:
+    Core::System& system;
+    Core::Timing::EventType* gpu_interrupt_event{};
+};
+
+} // namespace Core::Hardware

src/core/hle/kernel/code_set.h

@@ -8,6 +8,7 @@
 #include <vector>
 
 #include "common/common_types.h"
+#include "core/hle/kernel/physical_memory.h"
 
 namespace Kernel {
 
@@ -77,7 +78,7 @@ struct CodeSet final {
     }
 
     /// The overall data that backs this code set.
-    std::vector<u8> memory;
+    Kernel::PhysicalMemory memory;
 
     /// The segments that comprise this code set.
     std::array<Segment, 3> segments;

src/core/hle/kernel/physical_memory.h

@@ -0,0 +1,19 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/alignment.h"
+
+namespace Kernel {
+
+// This encapsulation serves 2 purposes:
+// - First, to encapsulate host physical memory under a single type and set an
+// standard for managing it.
+// - Second to ensure all host backing memory used is aligned to 256 bytes due
+// to strict alignment restrictions on GPU memory.
+
+using PhysicalMemory = std::vector<u8, Common::AlignmentAllocator<u8, 256>>;
+
+} // namespace Kernel

src/core/hle/kernel/process.cpp

@@ -247,7 +247,7 @@ VAddr Process::CreateTLSRegion() {
         ASSERT(region_address.Succeeded());
 
         const auto map_result = vm_manager.MapMemoryBlock(
-            *region_address, std::make_shared<std::vector<u8>>(Memory::PAGE_SIZE), 0,
+            *region_address, std::make_shared<PhysicalMemory>(Memory::PAGE_SIZE), 0,
             Memory::PAGE_SIZE, MemoryState::ThreadLocal);
         ASSERT(map_result.Succeeded());
 
@@ -277,7 +277,7 @@ void Process::FreeTLSRegion(VAddr tls_address) {
 }
 
 void Process::LoadModule(CodeSet module_, VAddr base_addr) {
-    const auto memory = std::make_shared<std::vector<u8>>(std::move(module_.memory));
+    const auto memory = std::make_shared<PhysicalMemory>(std::move(module_.memory));
 
     const auto MapSegment = [&](const CodeSet::Segment& segment, VMAPermission permissions,
                                 MemoryState memory_state) {
@@ -327,7 +327,7 @@ void Process::AllocateMainThreadStack(u64 stack_size) {
     // Allocate and map the main thread stack
     const VAddr mapping_address = vm_manager.GetTLSIORegionEndAddress() - main_thread_stack_size;
     vm_manager
-        .MapMemoryBlock(mapping_address, std::make_shared<std::vector<u8>>(main_thread_stack_size),
+        .MapMemoryBlock(mapping_address, std::make_shared<PhysicalMemory>(main_thread_stack_size),
                         0, main_thread_stack_size, MemoryState::Stack)
         .Unwrap();
 }

src/core/hle/kernel/shared_memory.cpp

@@ -28,7 +28,7 @@ SharedPtr<SharedMemory> SharedMemory::Create(KernelCore& kernel, Process* owner_
     shared_memory->other_permissions = other_permissions;
 
     if (address == 0) {
-        shared_memory->backing_block = std::make_shared<std::vector<u8>>(size);
+        shared_memory->backing_block = std::make_shared<Kernel::PhysicalMemory>(size);
         shared_memory->backing_block_offset = 0;
 
         // Refresh the address mappings for the current process.
@@ -59,8 +59,8 @@ SharedPtr<SharedMemory> SharedMemory::Create(KernelCore& kernel, Process* owner_
 }
 
 SharedPtr<SharedMemory> SharedMemory::CreateForApplet(
-    KernelCore& kernel, std::shared_ptr<std::vector<u8>> heap_block, std::size_t offset, u64 size,
-    MemoryPermission permissions, MemoryPermission other_permissions, std::string name) {
+    KernelCore& kernel, std::shared_ptr<Kernel::PhysicalMemory> heap_block, std::size_t offset,
+    u64 size, MemoryPermission permissions, MemoryPermission other_permissions, std::string name) {
     SharedPtr<SharedMemory> shared_memory(new SharedMemory(kernel));
 
     shared_memory->owner_process = nullptr;

src/core/hle/kernel/shared_memory.h

@@ -10,6 +10,7 @@
 
 #include "common/common_types.h"
 #include "core/hle/kernel/object.h"
+#include "core/hle/kernel/physical_memory.h"
 #include "core/hle/kernel/process.h"
 #include "core/hle/result.h"
 
@@ -62,12 +63,10 @@ public:
      * block.
      * @param name Optional object name, used for debugging purposes.
      */
-    static SharedPtr<SharedMemory> CreateForApplet(KernelCore& kernel,
-                                                   std::shared_ptr<std::vector<u8>> heap_block,
-                                                   std::size_t offset, u64 size,
-                                                   MemoryPermission permissions,
-                                                   MemoryPermission other_permissions,
-                                                   std::string name = "Unknown Applet");
+    static SharedPtr<SharedMemory> CreateForApplet(
+        KernelCore& kernel, std::shared_ptr<Kernel::PhysicalMemory> heap_block, std::size_t offset,
+        u64 size, MemoryPermission permissions, MemoryPermission other_permissions,
+        std::string name = "Unknown Applet");
 
     std::string GetTypeName() const override {
         return "SharedMemory";
@@ -135,7 +134,7 @@ private:
     ~SharedMemory() override;
 
     /// Backing memory for this shared memory block.
-    std::shared_ptr<std::vector<u8>> backing_block;
+    std::shared_ptr<PhysicalMemory> backing_block;
     /// Offset into the backing block for this shared memory.
     std::size_t backing_block_offset = 0;
     /// Size of the memory block. Page-aligned.

src/core/hle/kernel/transfer_memory.cpp

@@ -47,7 +47,7 @@ ResultCode TransferMemory::MapMemory(VAddr address, u64 size, MemoryPermission p
         return ERR_INVALID_STATE;
     }
 
-    backing_block = std::make_shared<std::vector<u8>>(size);
+    backing_block = std::make_shared<PhysicalMemory>(size);
 
     const auto map_state = owner_permissions == MemoryPermission::None
                                ? MemoryState::TransferMemoryIsolated

src/core/hle/kernel/transfer_memory.h

@@ -8,6 +8,7 @@
 #include <vector>
 
 #include "core/hle/kernel/object.h"
+#include "core/hle/kernel/physical_memory.h"
 
 union ResultCode;
 
@@ -82,7 +83,7 @@ private:
     ~TransferMemory() override;
 
     /// Memory block backing this instance.
-    std::shared_ptr<std::vector<u8>> backing_block;
+    std::shared_ptr<PhysicalMemory> backing_block;
 
     /// The base address for the memory managed by this instance.
     VAddr base_address = 0;

src/core/hle/kernel/vm_manager.cpp

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <iterator>
 #include <utility>
+#include "common/alignment.h"
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "common/memory_hook.h"
@@ -103,7 +104,7 @@ bool VMManager::IsValidHandle(VMAHandle handle) const {
 }
 
 ResultVal<VMManager::VMAHandle> VMManager::MapMemoryBlock(VAddr target,
-                                                          std::shared_ptr<std::vector<u8>> block,
+                                                          std::shared_ptr<PhysicalMemory> block,
                                                           std::size_t offset, u64 size,
                                                           MemoryState state, VMAPermission perm) {
     ASSERT(block != nullptr);
@@ -260,7 +261,7 @@ ResultVal<VAddr> VMManager::SetHeapSize(u64 size) {
 
     if (heap_memory == nullptr) {
         // Initialize heap
-        heap_memory = std::make_shared<std::vector<u8>>(size);
+        heap_memory = std::make_shared<PhysicalMemory>(size);
         heap_end = heap_region_base + size;
     } else {
         UnmapRange(heap_region_base, GetCurrentHeapSize());
@@ -341,7 +342,7 @@ ResultCode VMManager::MapPhysicalMemory(VAddr target, u64 size) {
             const auto map_size = std::min(end_addr - cur_addr, vma_end - cur_addr);
             if (vma.state == MemoryState::Unmapped) {
                 const auto map_res =
-                    MapMemoryBlock(cur_addr, std::make_shared<std::vector<u8>>(map_size, 0), 0,
+                    MapMemoryBlock(cur_addr, std::make_shared<PhysicalMemory>(map_size, 0), 0,
                                    map_size, MemoryState::Heap, VMAPermission::ReadWrite);
                 result = map_res.Code();
                 if (result.IsError()) {
@@ -442,7 +443,7 @@ ResultCode VMManager::UnmapPhysicalMemory(VAddr target, u64 size) {
     if (result.IsError()) {
         for (const auto [map_address, map_size] : unmapped_regions) {
             const auto remap_res =
-                MapMemoryBlock(map_address, std::make_shared<std::vector<u8>>(map_size, 0), 0,
+                MapMemoryBlock(map_address, std::make_shared<PhysicalMemory>(map_size, 0), 0,
                                map_size, MemoryState::Heap, VMAPermission::None);
             ASSERT_MSG(remap_res.Succeeded(), "UnmapPhysicalMemory re-map on error");
         }
@@ -593,7 +594,7 @@ ResultCode VMManager::MirrorMemory(VAddr dst_addr, VAddr src_addr, u64 size, Mem
     ASSERT_MSG(vma_offset + size <= vma->second.size,
                "Shared memory exceeds bounds of mapped block");
 
-    const std::shared_ptr<std::vector<u8>>& backing_block = vma->second.backing_block;
+    const std::shared_ptr<PhysicalMemory>& backing_block = vma->second.backing_block;
     const std::size_t backing_block_offset = vma->second.offset + vma_offset;
 
     CASCADE_RESULT(auto new_vma,
@@ -606,7 +607,7 @@ ResultCode VMManager::MirrorMemory(VAddr dst_addr, VAddr src_addr, u64 size, Mem
     return RESULT_SUCCESS;
 }
 
-void VMManager::RefreshMemoryBlockMappings(const std::vector<u8>* block) {
+void VMManager::RefreshMemoryBlockMappings(const PhysicalMemory* block) {
     // If this ever proves to have a noticeable performance impact, allow users of the function to
     // specify a specific range of addresses to limit the scan to.
     for (const auto& p : vma_map) {
@@ -764,7 +765,7 @@ void VMManager::MergeAdjacentVMA(VirtualMemoryArea& left, const VirtualMemoryAre
                                        right.backing_block->begin() + right.offset + right.size);
         } else {
             // Slow case: make a new memory block for left and right.
-            auto new_memory = std::make_shared<std::vector<u8>>();
+            auto new_memory = std::make_shared<PhysicalMemory>();
             new_memory->insert(new_memory->end(), left.backing_block->begin() + left.offset,
                                left.backing_block->begin() + left.offset + left.size);
             new_memory->insert(new_memory->end(), right.backing_block->begin() + right.offset,

src/core/hle/kernel/vm_manager.h

@@ -11,6 +11,7 @@
 #include "common/common_types.h"
 #include "common/memory_hook.h"
 #include "common/page_table.h"
+#include "core/hle/kernel/physical_memory.h"
 #include "core/hle/result.h"
 #include "core/memory.h"
 
@@ -290,7 +291,7 @@ struct VirtualMemoryArea {
 
     // Settings for type = AllocatedMemoryBlock
     /// Memory block backing this VMA.
-    std::shared_ptr<std::vector<u8>> backing_block = nullptr;
+    std::shared_ptr<PhysicalMemory> backing_block = nullptr;
     /// Offset into the backing_memory the mapping starts from.
     std::size_t offset = 0;
 
@@ -348,7 +349,7 @@ public:
      * @param size Size of the mapping.
      * @param state MemoryState tag to attach to the VMA.
      */
-    ResultVal<VMAHandle> MapMemoryBlock(VAddr target, std::shared_ptr<std::vector<u8>> block,
+    ResultVal<VMAHandle> MapMemoryBlock(VAddr target, std::shared_ptr<PhysicalMemory> block,
                                         std::size_t offset, u64 size, MemoryState state,
                                         VMAPermission perm = VMAPermission::ReadWrite);
 
@@ -547,7 +548,7 @@ public:
      * Scans all VMAs and updates the page table range of any that use the given vector as backing
      * memory. This should be called after any operation that causes reallocation of the vector.
      */
-    void RefreshMemoryBlockMappings(const std::vector<u8>* block);
+    void RefreshMemoryBlockMappings(const PhysicalMemory* block);
 
     /// Dumps the address space layout to the log, for debugging
     void LogLayout() const;
@@ -777,7 +778,7 @@ private:
     // the entire virtual address space extents that bound the allocations, including any holes.
     // This makes deallocation and reallocation of holes fast and keeps process memory contiguous
     // in the emulator address space, allowing Memory::GetPointer to be reasonably safe.
-    std::shared_ptr<std::vector<u8>> heap_memory;
+    std::shared_ptr<PhysicalMemory> heap_memory;
 
     // The end of the currently allocated heap. This is not an inclusive
     // end of the range. This is essentially 'base_address + current_size'.

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

@@ -19,16 +19,16 @@
 
 namespace Service::Audio {
 
-void InstallInterfaces(SM::ServiceManager& service_manager) {
+void InstallInterfaces(SM::ServiceManager& service_manager, Core::System& system) {
     std::make_shared<AudCtl>()->InstallAsService(service_manager);
     std::make_shared<AudOutA>()->InstallAsService(service_manager);
-    std::make_shared<AudOutU>()->InstallAsService(service_manager);
+    std::make_shared<AudOutU>(system)->InstallAsService(service_manager);
     std::make_shared<AudInA>()->InstallAsService(service_manager);
     std::make_shared<AudInU>()->InstallAsService(service_manager);
     std::make_shared<AudRecA>()->InstallAsService(service_manager);
     std::make_shared<AudRecU>()->InstallAsService(service_manager);
     std::make_shared<AudRenA>()->InstallAsService(service_manager);
-    std::make_shared<AudRenU>()->InstallAsService(service_manager);
+    std::make_shared<AudRenU>(system)->InstallAsService(service_manager);
     std::make_shared<CodecCtl>()->InstallAsService(service_manager);
     std::make_shared<HwOpus>()->InstallAsService(service_manager);
 

src/core/hle/service/audio/audio.h

@@ -4,6 +4,10 @@
 
 #pragma once
 
+namespace Core {
+class System;
+}
+
 namespace Service::SM {
 class ServiceManager;
 }
@@ -11,6 +15,6 @@ class ServiceManager;
 namespace Service::Audio {
 
 /// Registers all Audio services with the specified service manager.
-void InstallInterfaces(SM::ServiceManager& service_manager);
+void InstallInterfaces(SM::ServiceManager& service_manager, Core::System& system);
 
 } // namespace Service::Audio

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

@@ -40,8 +40,8 @@ enum class AudioState : u32 {
 
 class IAudioOut final : public ServiceFramework<IAudioOut> {
 public:
-    IAudioOut(AudoutParams audio_params, AudioCore::AudioOut& audio_core, std::string&& device_name,
-              std::string&& unique_name)
+    IAudioOut(Core::System& system, AudoutParams audio_params, AudioCore::AudioOut& audio_core,
+              std::string&& device_name, std::string&& unique_name)
         : ServiceFramework("IAudioOut"), audio_core(audio_core),
           device_name(std::move(device_name)), audio_params(audio_params) {
         // clang-format off
@@ -65,7 +65,6 @@ public:
         RegisterHandlers(functions);
 
         // This is the event handle used to check if the audio buffer was released
-        auto& system = Core::System::GetInstance();
         buffer_event = Kernel::WritableEvent::CreateEventPair(
             system.Kernel(), Kernel::ResetType::Manual, "IAudioOutBufferReleased");
 
@@ -212,6 +211,22 @@ private:
     Kernel::EventPair buffer_event;
 };
 
+AudOutU::AudOutU(Core::System& system_) : ServiceFramework("audout:u"), system{system_} {
+    // clang-format off
+    static const FunctionInfo functions[] = {
+        {0, &AudOutU::ListAudioOutsImpl, "ListAudioOuts"},
+        {1, &AudOutU::OpenAudioOutImpl, "OpenAudioOut"},
+        {2, &AudOutU::ListAudioOutsImpl, "ListAudioOutsAuto"},
+        {3, &AudOutU::OpenAudioOutImpl, "OpenAudioOutAuto"},
+    };
+    // clang-format on
+
+    RegisterHandlers(functions);
+    audio_core = std::make_unique<AudioCore::AudioOut>();
+}
+
+AudOutU::~AudOutU() = default;
+
 void AudOutU::ListAudioOutsImpl(Kernel::HLERequestContext& ctx) {
     LOG_DEBUG(Service_Audio, "called");
 
@@ -248,7 +263,7 @@ void AudOutU::OpenAudioOutImpl(Kernel::HLERequestContext& ctx) {
 
     std::string unique_name{fmt::format("{}-{}", device_name, audio_out_interfaces.size())};
     auto audio_out_interface = std::make_shared<IAudioOut>(
-        params, *audio_core, std::move(device_name), std::move(unique_name));
+        system, params, *audio_core, std::move(device_name), std::move(unique_name));
 
     IPC::ResponseBuilder rb{ctx, 6, 0, 1};
     rb.Push(RESULT_SUCCESS);
@@ -256,20 +271,9 @@ void AudOutU::OpenAudioOutImpl(Kernel::HLERequestContext& ctx) {
     rb.Push<u32>(params.channel_count);
     rb.Push<u32>(static_cast<u32>(AudioCore::Codec::PcmFormat::Int16));
     rb.Push<u32>(static_cast<u32>(AudioState::Stopped));
-    rb.PushIpcInterface<Audio::IAudioOut>(audio_out_interface);
+    rb.PushIpcInterface<IAudioOut>(audio_out_interface);
 
     audio_out_interfaces.push_back(std::move(audio_out_interface));
 }
 
-AudOutU::AudOutU() : ServiceFramework("audout:u") {
-    static const FunctionInfo functions[] = {{0, &AudOutU::ListAudioOutsImpl, "ListAudioOuts"},
-                                             {1, &AudOutU::OpenAudioOutImpl, "OpenAudioOut"},
-                                             {2, &AudOutU::ListAudioOutsImpl, "ListAudioOutsAuto"},
-                                             {3, &AudOutU::OpenAudioOutImpl, "OpenAudioOutAuto"}};
-    RegisterHandlers(functions);
-    audio_core = std::make_unique<AudioCore::AudioOut>();
-}
-
-AudOutU::~AudOutU() = default;
-
 } // namespace Service::Audio

src/core/hle/service/audio/audout_u.h

@@ -11,6 +11,10 @@ namespace AudioCore {
 class AudioOut;
 }
 
+namespace Core {
+class System;
+}
+
 namespace Kernel {
 class HLERequestContext;
 }
@@ -21,15 +25,17 @@ class IAudioOut;
 
 class AudOutU final : public ServiceFramework<AudOutU> {
 public:
-    AudOutU();
+    explicit AudOutU(Core::System& system_);
     ~AudOutU() override;
 
 private:
+    void ListAudioOutsImpl(Kernel::HLERequestContext& ctx);
+    void OpenAudioOutImpl(Kernel::HLERequestContext& ctx);
+
     std::vector<std::shared_ptr<IAudioOut>> audio_out_interfaces;
     std::unique_ptr<AudioCore::AudioOut> audio_core;
 
-    void ListAudioOutsImpl(Kernel::HLERequestContext& ctx);
-    void OpenAudioOutImpl(Kernel::HLERequestContext& ctx);
+    Core::System& system;
 };
 
 } // namespace Service::Audio

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

@@ -5,6 +5,7 @@
 #include <algorithm>
 #include <array>
 #include <memory>
+#include <string_view>
 
 #include "audio_core/audio_renderer.h"
 #include "common/alignment.h"
@@ -25,7 +26,7 @@ namespace Service::Audio {
 
 class IAudioRenderer final : public ServiceFramework<IAudioRenderer> {
 public:
-    explicit IAudioRenderer(AudioCore::AudioRendererParameter audren_params,
+    explicit IAudioRenderer(Core::System& system, AudioCore::AudioRendererParameter audren_params,
                             const std::size_t instance_number)
         : ServiceFramework("IAudioRenderer") {
         // clang-format off
@@ -46,7 +47,6 @@ public:
         // clang-format on
         RegisterHandlers(functions);
 
-        auto& system = Core::System::GetInstance();
         system_event = Kernel::WritableEvent::CreateEventPair(
             system.Kernel(), Kernel::ResetType::Manual, "IAudioRenderer:SystemEvent");
         renderer = std::make_unique<AudioCore::AudioRenderer>(
@@ -160,17 +160,18 @@ private:
 
 class IAudioDevice final : public ServiceFramework<IAudioDevice> {
 public:
-    IAudioDevice() : ServiceFramework("IAudioDevice") {
+    explicit IAudioDevice(Core::System& system, u32_le revision_num)
+        : ServiceFramework("IAudioDevice"), revision{revision_num} {
         static const FunctionInfo functions[] = {
             {0, &IAudioDevice::ListAudioDeviceName, "ListAudioDeviceName"},
             {1, &IAudioDevice::SetAudioDeviceOutputVolume, "SetAudioDeviceOutputVolume"},
-            {2, nullptr, "GetAudioDeviceOutputVolume"},
+            {2, &IAudioDevice::GetAudioDeviceOutputVolume, "GetAudioDeviceOutputVolume"},
             {3, &IAudioDevice::GetActiveAudioDeviceName, "GetActiveAudioDeviceName"},
             {4, &IAudioDevice::QueryAudioDeviceSystemEvent, "QueryAudioDeviceSystemEvent"},
             {5, &IAudioDevice::GetActiveChannelCount, "GetActiveChannelCount"},
             {6, &IAudioDevice::ListAudioDeviceName, "ListAudioDeviceNameAuto"},
             {7, &IAudioDevice::SetAudioDeviceOutputVolume, "SetAudioDeviceOutputVolumeAuto"},
-            {8, nullptr, "GetAudioDeviceOutputVolumeAuto"},
+            {8, &IAudioDevice::GetAudioDeviceOutputVolume, "GetAudioDeviceOutputVolumeAuto"},
             {10, &IAudioDevice::GetActiveAudioDeviceName, "GetActiveAudioDeviceNameAuto"},
             {11, nullptr, "QueryAudioDeviceInputEvent"},
             {12, &IAudioDevice::QueryAudioDeviceOutputEvent, "QueryAudioDeviceOutputEvent"},
@@ -178,7 +179,7 @@ public:
         };
         RegisterHandlers(functions);
 
-        auto& kernel = Core::System::GetInstance().Kernel();
+        auto& kernel = system.Kernel();
         buffer_event = Kernel::WritableEvent::CreateEventPair(kernel, Kernel::ResetType::Automatic,
                                                               "IAudioOutBufferReleasedEvent");
 
@@ -189,15 +190,47 @@ public:
     }
 
 private:
-    void ListAudioDeviceName(Kernel::HLERequestContext& ctx) {
-        LOG_WARNING(Service_Audio, "(STUBBED) called");
+    using AudioDeviceName = std::array<char, 256>;
+    static constexpr std::array<std::string_view, 4> audio_device_names{{
+        "AudioStereoJackOutput",
+        "AudioBuiltInSpeakerOutput",
+        "AudioTvOutput",
+        "AudioUsbDeviceOutput",
+    }};
+    enum class DeviceType {
+        AHUBHeadphones,
+        AHUBSpeakers,
+        HDA,
+        USBOutput,
+    };
 
-        constexpr std::array<char, 15> audio_interface{{"AudioInterface"}};
-        ctx.WriteBuffer(audio_interface);
+    void ListAudioDeviceName(Kernel::HLERequestContext& ctx) {
+        LOG_DEBUG(Service_Audio, "called");
+
+        const bool usb_output_supported =
+            IsFeatureSupported(AudioFeatures::AudioUSBDeviceOutput, revision);
+        const std::size_t count = ctx.GetWriteBufferSize() / sizeof(AudioDeviceName);
+
+        std::vector<AudioDeviceName> name_buffer;
+        name_buffer.reserve(audio_device_names.size());
+
+        for (std::size_t i = 0; i < count && i < audio_device_names.size(); i++) {
+            const auto type = static_cast<DeviceType>(i);
+
+            if (!usb_output_supported && type == DeviceType::USBOutput) {
+                continue;
+            }
+
+            const auto& device_name = audio_device_names[i];
+            auto& entry = name_buffer.emplace_back();
+            device_name.copy(entry.data(), device_name.size());
+        }
+
+        ctx.WriteBuffer(name_buffer);
 
         IPC::ResponseBuilder rb{ctx, 3};
         rb.Push(RESULT_SUCCESS);
-        rb.Push<u32>(1);
+        rb.Push(static_cast<u32>(name_buffer.size()));
     }
 
     void SetAudioDeviceOutputVolume(Kernel::HLERequestContext& ctx) {
@@ -213,15 +246,32 @@ private:
         rb.Push(RESULT_SUCCESS);
     }
 
-    void GetActiveAudioDeviceName(Kernel::HLERequestContext& ctx) {
-        LOG_WARNING(Service_Audio, "(STUBBED) called");
+    void GetAudioDeviceOutputVolume(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
 
-        constexpr std::array<char, 12> audio_interface{{"AudioDevice"}};
-        ctx.WriteBuffer(audio_interface);
+        const auto device_name_buffer = ctx.ReadBuffer();
+        const std::string name = Common::StringFromBuffer(device_name_buffer);
+
+        LOG_WARNING(Service_Audio, "(STUBBED) called. name={}", name);
 
         IPC::ResponseBuilder rb{ctx, 3};
         rb.Push(RESULT_SUCCESS);
-        rb.Push<u32>(1);
+        rb.Push(1.0f);
+    }
+
+    void GetActiveAudioDeviceName(Kernel::HLERequestContext& ctx) {
+        LOG_WARNING(Service_Audio, "(STUBBED) called");
+
+        // Currently set to always be TV audio output.
+        const auto& device_name = audio_device_names[2];
+
+        AudioDeviceName out_device_name{};
+        device_name.copy(out_device_name.data(), device_name.size());
+
+        ctx.WriteBuffer(out_device_name);
+
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(RESULT_SUCCESS);
     }
 
     void QueryAudioDeviceSystemEvent(Kernel::HLERequestContext& ctx) {
@@ -250,12 +300,13 @@ private:
         rb.PushCopyObjects(audio_output_device_switch_event.readable);
     }
 
+    u32_le revision = 0;
     Kernel::EventPair buffer_event;
     Kernel::EventPair audio_output_device_switch_event;
 
 }; // namespace Audio
 
-AudRenU::AudRenU() : ServiceFramework("audren:u") {
+AudRenU::AudRenU(Core::System& system_) : ServiceFramework("audren:u"), system{system_} {
     // clang-format off
     static const FunctionInfo functions[] = {
         {0, &AudRenU::OpenAudioRenderer, "OpenAudioRenderer"},
@@ -328,7 +379,7 @@ void AudRenU::GetAudioRendererWorkBufferSize(Kernel::HLERequestContext& ctx) {
     };
 
     // Calculates the portion of the size related to the mix data (and the sorting thereof).
-    const auto calculate_mix_info_size = [this](const AudioCore::AudioRendererParameter& params) {
+    const auto calculate_mix_info_size = [](const AudioCore::AudioRendererParameter& params) {
         // The size of the mixing info data structure.
         constexpr u64 mix_info_size = 0x940;
 
@@ -400,7 +451,7 @@ void AudRenU::GetAudioRendererWorkBufferSize(Kernel::HLERequestContext& ctx) {
 
     // Calculates the part of the size related to the splitter context.
     const auto calculate_splitter_context_size =
-        [this](const AudioCore::AudioRendererParameter& params) -> u64 {
+        [](const AudioCore::AudioRendererParameter& params) -> u64 {
         if (!IsFeatureSupported(AudioFeatures::Splitter, params.revision)) {
             return 0;
         }
@@ -447,7 +498,7 @@ void AudRenU::GetAudioRendererWorkBufferSize(Kernel::HLERequestContext& ctx) {
     };
 
     // Calculates the part of the size related to performance statistics.
-    const auto calculate_perf_size = [this](const AudioCore::AudioRendererParameter& params) {
+    const auto calculate_perf_size = [](const AudioCore::AudioRendererParameter& params) {
         // Extra size value appended to the end of the calculation.
         constexpr u64 appended = 128;
 
@@ -474,78 +525,76 @@ void AudRenU::GetAudioRendererWorkBufferSize(Kernel::HLERequestContext& ctx) {
     };
 
     // Calculates the part of the size that relates to the audio command buffer.
-    const auto calculate_command_buffer_size =
-        [this](const AudioCore::AudioRendererParameter& params) {
-            constexpr u64 alignment = (buffer_alignment_size - 1) * 2;
+    const auto calculate_command_buffer_size = [](const AudioCore::AudioRendererParameter& params) {
+        constexpr u64 alignment = (buffer_alignment_size - 1) * 2;
 
-            if (!IsFeatureSupported(AudioFeatures::VariadicCommandBuffer, params.revision)) {
-                constexpr u64 command_buffer_size = 0x18000;
+        if (!IsFeatureSupported(AudioFeatures::VariadicCommandBuffer, params.revision)) {
+            constexpr u64 command_buffer_size = 0x18000;
 
-                return command_buffer_size + alignment;
-            }
+            return command_buffer_size + alignment;
+        }
 
-            // When the variadic command buffer is supported, this means
-            // the command generator for the audio renderer can issue commands
-            // that are (as one would expect), variable in size. So what we need to do
-            // is determine the maximum possible size for a few command data structures
-            // then multiply them by the amount of present commands indicated by the given
-            // respective audio parameters.
+        // When the variadic command buffer is supported, this means
+        // the command generator for the audio renderer can issue commands
+        // that are (as one would expect), variable in size. So what we need to do
+        // is determine the maximum possible size for a few command data structures
+        // then multiply them by the amount of present commands indicated by the given
+        // respective audio parameters.
 
-            constexpr u64 max_biquad_filters = 2;
-            constexpr u64 max_mix_buffers = 24;
+        constexpr u64 max_biquad_filters = 2;
+        constexpr u64 max_mix_buffers = 24;
 
-            constexpr u64 biquad_filter_command_size = 0x2C;
+        constexpr u64 biquad_filter_command_size = 0x2C;
 
-            constexpr u64 depop_mix_command_size = 0x24;
-            constexpr u64 depop_setup_command_size = 0x50;
+        constexpr u64 depop_mix_command_size = 0x24;
+        constexpr u64 depop_setup_command_size = 0x50;
 
-            constexpr u64 effect_command_max_size = 0x540;
+        constexpr u64 effect_command_max_size = 0x540;
 
-            constexpr u64 mix_command_size = 0x1C;
-            constexpr u64 mix_ramp_command_size = 0x24;
-            constexpr u64 mix_ramp_grouped_command_size = 0x13C;
+        constexpr u64 mix_command_size = 0x1C;
+        constexpr u64 mix_ramp_command_size = 0x24;
+        constexpr u64 mix_ramp_grouped_command_size = 0x13C;
 
-            constexpr u64 perf_command_size = 0x28;
+        constexpr u64 perf_command_size = 0x28;
 
-            constexpr u64 sink_command_size = 0x130;
+        constexpr u64 sink_command_size = 0x130;
 
-            constexpr u64 submix_command_max_size =
-                depop_mix_command_size + (mix_command_size * max_mix_buffers) * max_mix_buffers;
+        constexpr u64 submix_command_max_size =
+            depop_mix_command_size + (mix_command_size * max_mix_buffers) * max_mix_buffers;
 
-            constexpr u64 volume_command_size = 0x1C;
-            constexpr u64 volume_ramp_command_size = 0x20;
+        constexpr u64 volume_command_size = 0x1C;
+        constexpr u64 volume_ramp_command_size = 0x20;
 
-            constexpr u64 voice_biquad_filter_command_size =
-                biquad_filter_command_size * max_biquad_filters;
-            constexpr u64 voice_data_command_size = 0x9C;
-            const u64 voice_command_max_size =
-                (params.splitter_count * depop_setup_command_size) +
-                (voice_data_command_size + voice_biquad_filter_command_size +
-                 volume_ramp_command_size + mix_ramp_grouped_command_size);
+        constexpr u64 voice_biquad_filter_command_size =
+            biquad_filter_command_size * max_biquad_filters;
+        constexpr u64 voice_data_command_size = 0x9C;
+        const u64 voice_command_max_size =
+            (params.splitter_count * depop_setup_command_size) +
+            (voice_data_command_size + voice_biquad_filter_command_size + volume_ramp_command_size +
+             mix_ramp_grouped_command_size);
 
-            // Now calculate the individual elements that comprise the size and add them together.
-            const u64 effect_commands_size = params.effect_count * effect_command_max_size;
+        // Now calculate the individual elements that comprise the size and add them together.
+        const u64 effect_commands_size = params.effect_count * effect_command_max_size;
 
-            const u64 final_mix_commands_size =
-                depop_mix_command_size + volume_command_size * max_mix_buffers;
+        const u64 final_mix_commands_size =
+            depop_mix_command_size + volume_command_size * max_mix_buffers;
 
-            const u64 perf_commands_size =
-                perf_command_size *
-                (CalculateNumPerformanceEntries(params) + max_perf_detail_entries);
+        const u64 perf_commands_size =
+            perf_command_size * (CalculateNumPerformanceEntries(params) + max_perf_detail_entries);
 
-            const u64 sink_commands_size = params.sink_count * sink_command_size;
+        const u64 sink_commands_size = params.sink_count * sink_command_size;
 
-            const u64 splitter_commands_size =
-                params.num_splitter_send_channels * max_mix_buffers * mix_ramp_command_size;
+        const u64 splitter_commands_size =
+            params.num_splitter_send_channels * max_mix_buffers * mix_ramp_command_size;
 
-            const u64 submix_commands_size = params.submix_count * submix_command_max_size;
+        const u64 submix_commands_size = params.submix_count * submix_command_max_size;
 
-            const u64 voice_commands_size = params.voice_count * voice_command_max_size;
+        const u64 voice_commands_size = params.voice_count * voice_command_max_size;
 
-            return effect_commands_size + final_mix_commands_size + perf_commands_size +
-                   sink_commands_size + splitter_commands_size + submix_commands_size +
-                   voice_commands_size + alignment;
-        };
+        return effect_commands_size + final_mix_commands_size + perf_commands_size +
+               sink_commands_size + splitter_commands_size + submix_commands_size +
+               voice_commands_size + alignment;
+    };
 
     IPC::RequestParser rp{ctx};
     const auto params = rp.PopRaw<AudioCore::AudioRendererParameter>();
@@ -578,12 +627,16 @@ void AudRenU::GetAudioRendererWorkBufferSize(Kernel::HLERequestContext& ctx) {
 }
 
 void AudRenU::GetAudioDeviceService(Kernel::HLERequestContext& ctx) {
-    LOG_DEBUG(Service_Audio, "called");
+    IPC::RequestParser rp{ctx};
+    const u64 aruid = rp.Pop<u64>();
 
+    LOG_DEBUG(Service_Audio, "called. aruid={:016X}", aruid);
+
+    // Revisionless variant of GetAudioDeviceServiceWithRevisionInfo that
+    // always assumes the initial release revision (REV1).
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
-
     rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<Audio::IAudioDevice>();
+    rb.PushIpcInterface<IAudioDevice>(system, Common::MakeMagic('R', 'E', 'V', '1'));
 }
 
 void AudRenU::OpenAudioRendererAuto(Kernel::HLERequestContext& ctx) {
@@ -593,13 +646,19 @@ void AudRenU::OpenAudioRendererAuto(Kernel::HLERequestContext& ctx) {
 }
 
 void AudRenU::GetAudioDeviceServiceWithRevisionInfo(Kernel::HLERequestContext& ctx) {
-    LOG_WARNING(Service_Audio, "(STUBBED) called");
+    struct Parameters {
+        u32 revision;
+        u64 aruid;
+    };
+
+    IPC::RequestParser rp{ctx};
+    const auto [revision, aruid] = rp.PopRaw<Parameters>();
+
+    LOG_DEBUG(Service_Audio, "called. revision={:08X}, aruid={:016X}", revision, aruid);
 
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
-
     rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<Audio::IAudioDevice>(); // TODO(ogniK): Figure out what is different
-                                                // based on the current revision
+    rb.PushIpcInterface<IAudioDevice>(system, revision);
 }
 
 void AudRenU::OpenAudioRendererImpl(Kernel::HLERequestContext& ctx) {
@@ -608,14 +667,16 @@ void AudRenU::OpenAudioRendererImpl(Kernel::HLERequestContext& ctx) {
     IPC::ResponseBuilder rb{ctx, 2, 0, 1};
 
     rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<IAudioRenderer>(params, audren_instance_count++);
+    rb.PushIpcInterface<IAudioRenderer>(system, params, audren_instance_count++);
 }
 
-bool AudRenU::IsFeatureSupported(AudioFeatures feature, u32_le revision) const {
+bool IsFeatureSupported(AudioFeatures feature, u32_le revision) {
     // Byte swap
     const u32_be version_num = revision - Common::MakeMagic('R', 'E', 'V', '0');
 
     switch (feature) {
+    case AudioFeatures::AudioUSBDeviceOutput:
+        return version_num >= 4U;
     case AudioFeatures::Splitter:
         return version_num >= 2U;
     case AudioFeatures::PerformanceMetricsVersion2:

src/core/hle/service/audio/audren_u.h

@@ -6,6 +6,10 @@
 
 #include "core/hle/service/service.h"
 
+namespace Core {
+class System;
+}
+
 namespace Kernel {
 class HLERequestContext;
 }
@@ -14,7 +18,7 @@ namespace Service::Audio {
 
 class AudRenU final : public ServiceFramework<AudRenU> {
 public:
-    explicit AudRenU();
+    explicit AudRenU(Core::System& system_);
     ~AudRenU() override;
 
 private:
@@ -26,14 +30,19 @@ private:
 
     void OpenAudioRendererImpl(Kernel::HLERequestContext& ctx);
 
-    enum class AudioFeatures : u32 {
-        Splitter,
-        PerformanceMetricsVersion2,
-        VariadicCommandBuffer,
-    };
-
-    bool IsFeatureSupported(AudioFeatures feature, u32_le revision) const;
     std::size_t audren_instance_count = 0;
+    Core::System& system;
 };
 
+// Describes a particular audio feature that may be supported in a particular revision.
+enum class AudioFeatures : u32 {
+    AudioUSBDeviceOutput,
+    Splitter,
+    PerformanceMetricsVersion2,
+    VariadicCommandBuffer,
+};
+
+// Tests if a particular audio feature is supported with a given audio revision.
+bool IsFeatureSupported(AudioFeatures feature, u32_le revision);
+
 } // namespace Service::Audio

src/core/hle/service/ns/pl_u.cpp

@@ -77,7 +77,7 @@ enum class LoadState : u32 {
     Done = 1,
 };
 
-static void DecryptSharedFont(const std::vector<u32>& input, std::vector<u8>& output,
+static void DecryptSharedFont(const std::vector<u32>& input, Kernel::PhysicalMemory& output,
                               std::size_t& offset) {
     ASSERT_MSG(offset + (input.size() * sizeof(u32)) < SHARED_FONT_MEM_SIZE,
                "Shared fonts exceeds 17mb!");
@@ -94,7 +94,7 @@ static void DecryptSharedFont(const std::vector<u32>& input, std::vector<u8>& ou
     offset += transformed_font.size() * sizeof(u32);
 }
 
-static void EncryptSharedFont(const std::vector<u8>& input, std::vector<u8>& output,
+static void EncryptSharedFont(const std::vector<u8>& input, Kernel::PhysicalMemory& output,
                               std::size_t& offset) {
     ASSERT_MSG(offset + input.size() + 8 < SHARED_FONT_MEM_SIZE, "Shared fonts exceeds 17mb!");
     const u32 KEY = EXPECTED_MAGIC ^ EXPECTED_RESULT;
@@ -121,7 +121,7 @@ struct PL_U::Impl {
         return shared_font_regions.at(index);
     }
 
-    void BuildSharedFontsRawRegions(const std::vector<u8>& input) {
+    void BuildSharedFontsRawRegions(const Kernel::PhysicalMemory& input) {
         // As we can derive the xor key we can just populate the offsets
         // based on the shared memory dump
         unsigned cur_offset = 0;
@@ -144,7 +144,7 @@ struct PL_U::Impl {
     Kernel::SharedPtr<Kernel::SharedMemory> shared_font_mem;
 
     /// Backing memory for the shared font data
-    std::shared_ptr<std::vector<u8>> shared_font;
+    std::shared_ptr<Kernel::PhysicalMemory> shared_font;
 
     // Automatically populated based on shared_fonts dump or system archives.
     std::vector<FontRegion> shared_font_regions;
@@ -166,7 +166,7 @@ PL_U::PL_U() : ServiceFramework("pl:u"), impl{std::make_unique<Impl>()} {
     // Rebuild shared fonts from data ncas
     if (nand->HasEntry(static_cast<u64>(FontArchives::Standard),
                        FileSys::ContentRecordType::Data)) {
-        impl->shared_font = std::make_shared<std::vector<u8>>(SHARED_FONT_MEM_SIZE);
+        impl->shared_font = std::make_shared<Kernel::PhysicalMemory>(SHARED_FONT_MEM_SIZE);
         for (auto font : SHARED_FONTS) {
             const auto nca =
                 nand->GetEntry(static_cast<u64>(font.first), FileSys::ContentRecordType::Data);
@@ -207,7 +207,7 @@ PL_U::PL_U() : ServiceFramework("pl:u"), impl{std::make_unique<Impl>()} {
         }
 
     } else {
-        impl->shared_font = std::make_shared<std::vector<u8>>(
+        impl->shared_font = std::make_shared<Kernel::PhysicalMemory>(
             SHARED_FONT_MEM_SIZE); // Shared memory needs to always be allocated and a fixed size
 
         const std::string user_path = FileUtil::GetUserPath(FileUtil::UserPath::SysDataDir);

src/core/hle/service/nvdrv/devices/nvdevice.h

@@ -8,6 +8,11 @@
 #include "common/bit_field.h"
 #include "common/common_types.h"
 #include "common/swap.h"
+#include "core/hle/service/nvdrv/nvdata.h"
+
+namespace Core {
+class System;
+}
 
 namespace Service::Nvidia::Devices {
 
@@ -15,7 +20,7 @@ namespace Service::Nvidia::Devices {
 /// implement the ioctl interface.
 class nvdevice {
 public:
-    nvdevice() = default;
+    explicit nvdevice(Core::System& system) : system{system} {};
     virtual ~nvdevice() = default;
     union Ioctl {
         u32_le raw;
@@ -33,7 +38,11 @@ public:
      * @param output A buffer where the output data will be written to.
      * @returns The result code of the ioctl.
      */
-    virtual u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) = 0;
+    virtual u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                      IoctlCtrl& ctrl) = 0;
+
+protected:
+    Core::System& system;
 };
 
 } // namespace Service::Nvidia::Devices

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

@@ -13,10 +13,12 @@
 
 namespace Service::Nvidia::Devices {
 
-nvdisp_disp0::nvdisp_disp0(std::shared_ptr<nvmap> nvmap_dev) : nvmap_dev(std::move(nvmap_dev)) {}
+nvdisp_disp0::nvdisp_disp0(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
+    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
 nvdisp_disp0 ::~nvdisp_disp0() = default;
 
-u32 nvdisp_disp0::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvdisp_disp0::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                        IoctlCtrl& ctrl) {
     UNIMPLEMENTED_MSG("Unimplemented ioctl");
     return 0;
 }
@@ -34,9 +36,8 @@ void nvdisp_disp0::flip(u32 buffer_handle, u32 offset, u32 format, u32 width, u3
         addr,      offset,   width, height, stride, static_cast<PixelFormat>(format),
         transform, crop_rect};
 
-    auto& instance = Core::System::GetInstance();
-    instance.GetPerfStats().EndGameFrame();
-    instance.GPU().SwapBuffers(framebuffer);
+    system.GetPerfStats().EndGameFrame();
+    system.GPU().SwapBuffers(framebuffer);
 }
 
 } // namespace Service::Nvidia::Devices

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

@@ -17,10 +17,11 @@ class nvmap;
 
 class nvdisp_disp0 final : public nvdevice {
 public:
-    explicit nvdisp_disp0(std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvdisp_disp0(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
     ~nvdisp_disp0() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
     /// Performs a screen flip, drawing the buffer pointed to by the handle.
     void flip(u32 buffer_handle, u32 offset, u32 format, u32 width, u32 height, u32 stride,

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

@@ -22,10 +22,12 @@ enum {
 };
 }
 
-nvhost_as_gpu::nvhost_as_gpu(std::shared_ptr<nvmap> nvmap_dev) : nvmap_dev(std::move(nvmap_dev)) {}
+nvhost_as_gpu::nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
+    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
 nvhost_as_gpu::~nvhost_as_gpu() = default;
 
-u32 nvhost_as_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_as_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                         IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 
@@ -65,7 +67,7 @@ u32 nvhost_as_gpu::AllocateSpace(const std::vector<u8>& input, std::vector<u8>&
     LOG_DEBUG(Service_NVDRV, "called, pages={:X}, page_size={:X}, flags={:X}", params.pages,
               params.page_size, params.flags);
 
-    auto& gpu = Core::System::GetInstance().GPU();
+    auto& gpu = system.GPU();
     const u64 size{static_cast<u64>(params.pages) * static_cast<u64>(params.page_size)};
     if (params.flags & 1) {
         params.offset = gpu.MemoryManager().AllocateSpace(params.offset, size, 1);
@@ -85,7 +87,7 @@ u32 nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& output)
     std::vector<IoctlRemapEntry> entries(num_entries);
     std::memcpy(entries.data(), input.data(), input.size());
 
-    auto& gpu = Core::System::GetInstance().GPU();
+    auto& gpu = system.GPU();
     for (const auto& entry : entries) {
         LOG_WARNING(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}",
                     entry.offset, entry.nvmap_handle, entry.pages);
@@ -136,7 +138,7 @@ u32 nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8>& ou
     // case to prevent unexpected behavior.
     ASSERT(object->id == params.nvmap_handle);
 
-    auto& gpu = Core::System::GetInstance().GPU();
+    auto& gpu = system.GPU();
 
     if (params.flags & 1) {
         params.offset = gpu.MemoryManager().MapBufferEx(object->addr, params.offset, object->size);
@@ -173,8 +175,7 @@ u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& ou
         return 0;
     }
 
-    params.offset = Core::System::GetInstance().GPU().MemoryManager().UnmapBuffer(params.offset,
-                                                                                  itr->second.size);
+    params.offset = system.GPU().MemoryManager().UnmapBuffer(params.offset, itr->second.size);
     buffer_mappings.erase(itr->second.offset);
 
     std::memcpy(output.data(), &params, output.size());

src/core/hle/service/nvdrv/devices/nvhost_as_gpu.h

@@ -17,10 +17,11 @@ class nvmap;
 
 class nvhost_as_gpu final : public nvdevice {
 public:
-    explicit nvhost_as_gpu(std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
     ~nvhost_as_gpu() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {

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

@@ -7,14 +7,20 @@
 
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "core/core.h"
+#include "core/hle/kernel/readable_event.h"
+#include "core/hle/kernel/writable_event.h"
 #include "core/hle/service/nvdrv/devices/nvhost_ctrl.h"
+#include "video_core/gpu.h"
 
 namespace Service::Nvidia::Devices {
 
-nvhost_ctrl::nvhost_ctrl() = default;
+nvhost_ctrl::nvhost_ctrl(Core::System& system, EventInterface& events_interface)
+    : nvdevice(system), events_interface{events_interface} {}
 nvhost_ctrl::~nvhost_ctrl() = default;
 
-u32 nvhost_ctrl::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_ctrl::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                       IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 
@@ -22,11 +28,15 @@ u32 nvhost_ctrl::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<
     case IoctlCommand::IocGetConfigCommand:
         return NvOsGetConfigU32(input, output);
     case IoctlCommand::IocCtrlEventWaitCommand:
-        return IocCtrlEventWait(input, output, false);
+        return IocCtrlEventWait(input, output, false, ctrl);
     case IoctlCommand::IocCtrlEventWaitAsyncCommand:
-        return IocCtrlEventWait(input, output, true);
+        return IocCtrlEventWait(input, output, true, ctrl);
     case IoctlCommand::IocCtrlEventRegisterCommand:
         return IocCtrlEventRegister(input, output);
+    case IoctlCommand::IocCtrlEventUnregisterCommand:
+        return IocCtrlEventUnregister(input, output);
+    case IoctlCommand::IocCtrlEventSignalCommand:
+        return IocCtrlEventSignal(input, output);
     }
     UNIMPLEMENTED_MSG("Unimplemented ioctl");
     return 0;
@@ -41,23 +51,137 @@ u32 nvhost_ctrl::NvOsGetConfigU32(const std::vector<u8>& input, std::vector<u8>&
 }
 
 u32 nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output,
-                                  bool is_async) {
+                                  bool is_async, IoctlCtrl& ctrl) {
     IocCtrlEventWaitParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
-    LOG_WARNING(Service_NVDRV,
-                "(STUBBED) called, syncpt_id={}, threshold={}, timeout={}, is_async={}",
-                params.syncpt_id, params.threshold, params.timeout, is_async);
+    LOG_DEBUG(Service_NVDRV, "syncpt_id={}, threshold={}, timeout={}, is_async={}",
+              params.syncpt_id, params.threshold, params.timeout, is_async);
 
-    // TODO(Subv): Implement actual syncpt waiting.
-    params.value = 0;
+    if (params.syncpt_id >= MaxSyncPoints) {
+        return NvResult::BadParameter;
+    }
+
+    auto& gpu = system.GPU();
+    // This is mostly to take into account unimplemented features. As synced
+    // gpu is always synced.
+    if (!gpu.IsAsync()) {
+        return NvResult::Success;
+    }
+    auto lock = gpu.LockSync();
+    const u32 current_syncpoint_value = gpu.GetSyncpointValue(params.syncpt_id);
+    const s32 diff = current_syncpoint_value - params.threshold;
+    if (diff >= 0) {
+        params.value = current_syncpoint_value;
+        std::memcpy(output.data(), &params, sizeof(params));
+        return NvResult::Success;
+    }
+    const u32 target_value = current_syncpoint_value - diff;
+
+    if (!is_async) {
+        params.value = 0;
+    }
+
+    if (params.timeout == 0) {
+        std::memcpy(output.data(), &params, sizeof(params));
+        return NvResult::Timeout;
+    }
+
+    u32 event_id;
+    if (is_async) {
+        event_id = params.value & 0x00FF;
+        if (event_id >= MaxNvEvents) {
+            std::memcpy(output.data(), &params, sizeof(params));
+            return NvResult::BadParameter;
+        }
+    } else {
+        if (ctrl.fresh_call) {
+            const auto result = events_interface.GetFreeEvent();
+            if (result) {
+                event_id = *result;
+            } else {
+                LOG_CRITICAL(Service_NVDRV, "No Free Events available!");
+                event_id = params.value & 0x00FF;
+            }
+        } else {
+            event_id = ctrl.event_id;
+        }
+    }
+
+    EventState status = events_interface.status[event_id];
+    if (event_id < MaxNvEvents || status == EventState::Free || status == EventState::Registered) {
+        events_interface.SetEventStatus(event_id, EventState::Waiting);
+        events_interface.assigned_syncpt[event_id] = params.syncpt_id;
+        events_interface.assigned_value[event_id] = target_value;
+        if (is_async) {
+            params.value = params.syncpt_id << 4;
+        } else {
+            params.value = ((params.syncpt_id & 0xfff) << 16) | 0x10000000;
+        }
+        params.value |= event_id;
+        events_interface.events[event_id].writable->Clear();
+        gpu.RegisterSyncptInterrupt(params.syncpt_id, target_value);
+        if (!is_async && ctrl.fresh_call) {
+            ctrl.must_delay = true;
+            ctrl.timeout = params.timeout;
+            ctrl.event_id = event_id;
+            return NvResult::Timeout;
+        }
+        std::memcpy(output.data(), &params, sizeof(params));
+        return NvResult::Timeout;
+    }
     std::memcpy(output.data(), &params, sizeof(params));
-    return 0;
+    return NvResult::BadParameter;
 }
 
 u32 nvhost_ctrl::IocCtrlEventRegister(const std::vector<u8>& input, std::vector<u8>& output) {
-    LOG_WARNING(Service_NVDRV, "(STUBBED) called");
-    // TODO(bunnei): Implement this.
-    return 0;
+    IocCtrlEventRegisterParams params{};
+    std::memcpy(&params, input.data(), sizeof(params));
+    const u32 event_id = params.user_event_id & 0x00FF;
+    LOG_DEBUG(Service_NVDRV, " called, user_event_id: {:X}", event_id);
+    if (event_id >= MaxNvEvents) {
+        return NvResult::BadParameter;
+    }
+    if (events_interface.registered[event_id]) {
+        return NvResult::BadParameter;
+    }
+    events_interface.RegisterEvent(event_id);
+    return NvResult::Success;
+}
+
+u32 nvhost_ctrl::IocCtrlEventUnregister(const std::vector<u8>& input, std::vector<u8>& output) {
+    IocCtrlEventUnregisterParams params{};
+    std::memcpy(&params, input.data(), sizeof(params));
+    const u32 event_id = params.user_event_id & 0x00FF;
+    LOG_DEBUG(Service_NVDRV, " called, user_event_id: {:X}", event_id);
+    if (event_id >= MaxNvEvents) {
+        return NvResult::BadParameter;
+    }
+    if (!events_interface.registered[event_id]) {
+        return NvResult::BadParameter;
+    }
+    events_interface.UnregisterEvent(event_id);
+    return NvResult::Success;
+}
+
+u32 nvhost_ctrl::IocCtrlEventSignal(const std::vector<u8>& input, std::vector<u8>& output) {
+    IocCtrlEventSignalParams params{};
+    std::memcpy(&params, input.data(), sizeof(params));
+    // TODO(Blinkhawk): This is normally called when an NvEvents timeout on WaitSynchronization
+    // It is believed from RE to cancel the GPU Event. However, better research is required
+    u32 event_id = params.user_event_id & 0x00FF;
+    LOG_WARNING(Service_NVDRV, "(STUBBED) called, user_event_id: {:X}", event_id);
+    if (event_id >= MaxNvEvents) {
+        return NvResult::BadParameter;
+    }
+    if (events_interface.status[event_id] == EventState::Waiting) {
+        auto& gpu = system.GPU();
+        if (gpu.CancelSyncptInterrupt(events_interface.assigned_syncpt[event_id],
+                                      events_interface.assigned_value[event_id])) {
+            events_interface.LiberateEvent(event_id);
+            events_interface.events[event_id].writable->Signal();
+        }
+    }
+    return NvResult::Success;
 }
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_ctrl.h

@@ -8,15 +8,17 @@
 #include <vector>
 #include "common/common_types.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
+#include "core/hle/service/nvdrv/nvdrv.h"
 
 namespace Service::Nvidia::Devices {
 
 class nvhost_ctrl final : public nvdevice {
 public:
-    nvhost_ctrl();
+    explicit nvhost_ctrl(Core::System& system, EventInterface& events_interface);
     ~nvhost_ctrl() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {
@@ -132,9 +134,16 @@ private:
 
     u32 NvOsGetConfigU32(const std::vector<u8>& input, std::vector<u8>& output);
 
-    u32 IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output, bool is_async);
+    u32 IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output, bool is_async,
+                         IoctlCtrl& ctrl);
 
     u32 IocCtrlEventRegister(const std::vector<u8>& input, std::vector<u8>& output);
+
+    u32 IocCtrlEventUnregister(const std::vector<u8>& input, std::vector<u8>& output);
+
+    u32 IocCtrlEventSignal(const std::vector<u8>& input, std::vector<u8>& output);
+
+    EventInterface& events_interface;
 };
 
 } // namespace Service::Nvidia::Devices

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

@@ -12,10 +12,11 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_ctrl_gpu::nvhost_ctrl_gpu() = default;
+nvhost_ctrl_gpu::nvhost_ctrl_gpu(Core::System& system) : nvdevice(system) {}
 nvhost_ctrl_gpu::~nvhost_ctrl_gpu() = default;
 
-u32 nvhost_ctrl_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_ctrl_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                           IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 
@@ -185,7 +186,7 @@ u32 nvhost_ctrl_gpu::GetGpuTime(const std::vector<u8>& input, std::vector<u8>& o
 
     IoctlGetGpuTime params{};
     std::memcpy(&params, input.data(), input.size());
-    const auto ns = Core::Timing::CyclesToNs(Core::System::GetInstance().CoreTiming().GetTicks());
+    const auto ns = Core::Timing::CyclesToNs(system.CoreTiming().GetTicks());
     params.gpu_time = static_cast<u64_le>(ns.count());
     std::memcpy(output.data(), &params, output.size());
     return 0;

src/core/hle/service/nvdrv/devices/nvhost_ctrl_gpu.h

@@ -13,10 +13,11 @@ namespace Service::Nvidia::Devices {
 
 class nvhost_ctrl_gpu final : public nvdevice {
 public:
-    nvhost_ctrl_gpu();
+    explicit nvhost_ctrl_gpu(Core::System& system);
     ~nvhost_ctrl_gpu() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {

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

@@ -13,10 +13,12 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_gpu::nvhost_gpu(std::shared_ptr<nvmap> nvmap_dev) : nvmap_dev(std::move(nvmap_dev)) {}
+nvhost_gpu::nvhost_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
+    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
 nvhost_gpu::~nvhost_gpu() = default;
 
-u32 nvhost_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_gpu::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                      IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 
@@ -119,8 +121,10 @@ u32 nvhost_gpu::AllocGPFIFOEx2(const std::vector<u8>& input, std::vector<u8>& ou
                 params.num_entries, params.flags, params.unk0, params.unk1, params.unk2,
                 params.unk3);
 
-    params.fence_out.id = 0;
-    params.fence_out.value = 0;
+    auto& gpu = system.GPU();
+    params.fence_out.id = assigned_syncpoints;
+    params.fence_out.value = gpu.GetSyncpointValue(assigned_syncpoints);
+    assigned_syncpoints++;
     std::memcpy(output.data(), &params, output.size());
     return 0;
 }
@@ -143,7 +147,7 @@ u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& outp
     IoctlSubmitGpfifo params{};
     std::memcpy(&params, input.data(), sizeof(IoctlSubmitGpfifo));
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, gpfifo={:X}, num_entries={:X}, flags={:X}",
-                params.address, params.num_entries, params.flags);
+                params.address, params.num_entries, params.flags.raw);
 
     ASSERT_MSG(input.size() == sizeof(IoctlSubmitGpfifo) +
                                    params.num_entries * sizeof(Tegra::CommandListHeader),
@@ -153,10 +157,18 @@ 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));
 
-    Core::System::GetInstance().GPU().PushGPUEntries(std::move(entries));
+    UNIMPLEMENTED_IF(params.flags.add_wait.Value() != 0);
+    UNIMPLEMENTED_IF(params.flags.add_increment.Value() != 0);
+
+    auto& gpu = system.GPU();
+    u32 current_syncpoint_value = gpu.GetSyncpointValue(params.fence_out.id);
+    if (params.flags.increment.Value()) {
+        params.fence_out.value += current_syncpoint_value;
+    } else {
+        params.fence_out.value = current_syncpoint_value;
+    }
+    gpu.PushGPUEntries(std::move(entries));
 
-    params.fence_out.id = 0;
-    params.fence_out.value = 0;
     std::memcpy(output.data(), &params, sizeof(IoctlSubmitGpfifo));
     return 0;
 }
@@ -168,16 +180,24 @@ u32 nvhost_gpu::KickoffPB(const std::vector<u8>& input, std::vector<u8>& output)
     IoctlSubmitGpfifo params{};
     std::memcpy(&params, input.data(), sizeof(IoctlSubmitGpfifo));
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, gpfifo={:X}, num_entries={:X}, flags={:X}",
-                params.address, params.num_entries, params.flags);
+                params.address, params.num_entries, params.flags.raw);
 
     Tegra::CommandList entries(params.num_entries);
     Memory::ReadBlock(params.address, entries.data(),
                       params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    Core::System::GetInstance().GPU().PushGPUEntries(std::move(entries));
+    UNIMPLEMENTED_IF(params.flags.add_wait.Value() != 0);
+    UNIMPLEMENTED_IF(params.flags.add_increment.Value() != 0);
+
+    auto& gpu = system.GPU();
+    u32 current_syncpoint_value = gpu.GetSyncpointValue(params.fence_out.id);
+    if (params.flags.increment.Value()) {
+        params.fence_out.value += current_syncpoint_value;
+    } else {
+        params.fence_out.value = current_syncpoint_value;
+    }
+    gpu.PushGPUEntries(std::move(entries));
 
-    params.fence_out.id = 0;
-    params.fence_out.value = 0;
     std::memcpy(output.data(), &params, output.size());
     return 0;
 }

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

@@ -10,6 +10,7 @@
 #include "common/common_types.h"
 #include "common/swap.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 
 namespace Service::Nvidia::Devices {
 
@@ -20,10 +21,11 @@ constexpr u32 NVGPU_IOCTL_CHANNEL_KICKOFF_PB(0x1b);
 
 class nvhost_gpu final : public nvdevice {
 public:
-    explicit nvhost_gpu(std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvhost_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
     ~nvhost_gpu() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {
@@ -113,11 +115,7 @@ private:
     static_assert(sizeof(IoctlGetErrorNotification) == 16,
                   "IoctlGetErrorNotification is incorrect size");
 
-    struct IoctlFence {
-        u32_le id;
-        u32_le value;
-    };
-    static_assert(sizeof(IoctlFence) == 8, "IoctlFence is incorrect size");
+    static_assert(sizeof(Fence) == 8, "Fence is incorrect size");
 
     struct IoctlAllocGpfifoEx {
         u32_le num_entries;
@@ -132,13 +130,13 @@ private:
     static_assert(sizeof(IoctlAllocGpfifoEx) == 32, "IoctlAllocGpfifoEx is incorrect size");
 
     struct IoctlAllocGpfifoEx2 {
-        u32_le num_entries;   // in
-        u32_le flags;         // in
-        u32_le unk0;          // in (1 works)
-        IoctlFence fence_out; // out
-        u32_le unk1;          // in
-        u32_le unk2;          // in
-        u32_le unk3;          // in
+        u32_le num_entries; // in
+        u32_le flags;       // in
+        u32_le unk0;        // in (1 works)
+        Fence fence_out;    // out
+        u32_le unk1;        // in
+        u32_le unk2;        // in
+        u32_le unk3;        // in
     };
     static_assert(sizeof(IoctlAllocGpfifoEx2) == 32, "IoctlAllocGpfifoEx2 is incorrect size");
 
@@ -153,10 +151,16 @@ private:
     struct IoctlSubmitGpfifo {
         u64_le address;     // pointer to gpfifo entry structs
         u32_le num_entries; // number of fence objects being submitted
-        u32_le flags;
-        IoctlFence fence_out; // returned new fence object for others to wait on
+        union {
+            u32_le raw;
+            BitField<0, 1, u32_le> add_wait;      // append a wait sync_point to the list
+            BitField<1, 1, u32_le> add_increment; // append an increment to the list
+            BitField<2, 1, u32_le> new_hw_format; // Mostly ignored
+            BitField<8, 1, u32_le> increment;     // increment the returned fence
+        } flags;
+        Fence fence_out; // returned new fence object for others to wait on
     };
-    static_assert(sizeof(IoctlSubmitGpfifo) == 16 + sizeof(IoctlFence),
+    static_assert(sizeof(IoctlSubmitGpfifo) == 16 + sizeof(Fence),
                   "IoctlSubmitGpfifo is incorrect size");
 
     struct IoctlGetWaitbase {
@@ -184,6 +188,7 @@ private:
     u32 ChannelSetTimeout(const std::vector<u8>& input, std::vector<u8>& output);
 
     std::shared_ptr<nvmap> nvmap_dev;
+    u32 assigned_syncpoints{};
 };
 
 } // namespace Service::Nvidia::Devices

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

@@ -10,10 +10,11 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_nvdec::nvhost_nvdec() = default;
+nvhost_nvdec::nvhost_nvdec(Core::System& system) : nvdevice(system) {}
 nvhost_nvdec::~nvhost_nvdec() = default;
 
-u32 nvhost_nvdec::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_nvdec::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                        IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 

src/core/hle/service/nvdrv/devices/nvhost_nvdec.h

@@ -13,10 +13,11 @@ namespace Service::Nvidia::Devices {
 
 class nvhost_nvdec final : public nvdevice {
 public:
-    nvhost_nvdec();
+    explicit nvhost_nvdec(Core::System& system);
     ~nvhost_nvdec() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {

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

@@ -10,10 +10,11 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_nvjpg::nvhost_nvjpg() = default;
+nvhost_nvjpg::nvhost_nvjpg(Core::System& system) : nvdevice(system) {}
 nvhost_nvjpg::~nvhost_nvjpg() = default;
 
-u32 nvhost_nvjpg::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_nvjpg::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                        IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 

src/core/hle/service/nvdrv/devices/nvhost_nvjpg.h

@@ -13,10 +13,11 @@ namespace Service::Nvidia::Devices {
 
 class nvhost_nvjpg final : public nvdevice {
 public:
-    nvhost_nvjpg();
+    explicit nvhost_nvjpg(Core::System& system);
     ~nvhost_nvjpg() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {

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

@@ -10,10 +10,11 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_vic::nvhost_vic() = default;
+nvhost_vic::nvhost_vic(Core::System& system) : nvdevice(system) {}
 nvhost_vic::~nvhost_vic() = default;
 
-u32 nvhost_vic::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_vic::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                      IoctlCtrl& ctrl) {
     LOG_DEBUG(Service_NVDRV, "called, command=0x{:08X}, input_size=0x{:X}, output_size=0x{:X}",
               command.raw, input.size(), output.size());
 

src/core/hle/service/nvdrv/devices/nvhost_vic.h

@@ -13,10 +13,11 @@ namespace Service::Nvidia::Devices {
 
 class nvhost_vic final : public nvdevice {
 public:
-    nvhost_vic();
+    explicit nvhost_vic(Core::System& system);
     ~nvhost_vic() override;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
 private:
     enum class IoctlCommand : u32_le {

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

@@ -18,7 +18,7 @@ enum {
 };
 }
 
-nvmap::nvmap() = default;
+nvmap::nvmap(Core::System& system) : nvdevice(system) {}
 nvmap::~nvmap() = default;
 
 VAddr nvmap::GetObjectAddress(u32 handle) const {
@@ -28,7 +28,8 @@ VAddr nvmap::GetObjectAddress(u32 handle) const {
     return object->addr;
 }
 
-u32 nvmap::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvmap::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+                 IoctlCtrl& ctrl) {
     switch (static_cast<IoctlCommand>(command.raw)) {
     case IoctlCommand::Create:
         return IocCreate(input, output);

src/core/hle/service/nvdrv/devices/nvmap.h

@@ -16,13 +16,14 @@ namespace Service::Nvidia::Devices {
 
 class nvmap final : public nvdevice {
 public:
-    nvmap();
+    explicit nvmap(Core::System& system);
     ~nvmap() override;
 
     /// Returns the allocated address of an nvmap object given its handle.
     VAddr GetObjectAddress(u32 handle) const;
 
-    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;
+    u32 ioctl(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl) override;
 
     /// Represents an nvmap object.
     struct Object {

src/core/hle/service/nvdrv/interface.cpp

@@ -8,12 +8,18 @@
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/kernel.h"
 #include "core/hle/kernel/readable_event.h"
+#include "core/hle/kernel/thread.h"
 #include "core/hle/kernel/writable_event.h"
 #include "core/hle/service/nvdrv/interface.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/nvdrv/nvdrv.h"
 
 namespace Service::Nvidia {
 
+void NVDRV::SignalGPUInterruptSyncpt(const u32 syncpoint_id, const u32 value) {
+    nvdrv->SignalSyncpt(syncpoint_id, value);
+}
+
 void NVDRV::Open(Kernel::HLERequestContext& ctx) {
     LOG_DEBUG(Service_NVDRV, "called");
 
@@ -36,11 +42,31 @@ void NVDRV::Ioctl(Kernel::HLERequestContext& ctx) {
 
     std::vector<u8> output(ctx.GetWriteBufferSize());
 
+    IoctlCtrl ctrl{};
+
+    u32 result = nvdrv->Ioctl(fd, command, ctx.ReadBuffer(), output, ctrl);
+
+    if (ctrl.must_delay) {
+        ctrl.fresh_call = false;
+        ctx.SleepClientThread(
+            "NVServices::DelayedResponse", ctrl.timeout,
+            [=](Kernel::SharedPtr<Kernel::Thread> thread, Kernel::HLERequestContext& ctx,
+                Kernel::ThreadWakeupReason reason) {
+                IoctlCtrl ctrl2{ctrl};
+                std::vector<u8> output2 = output;
+                u32 result = nvdrv->Ioctl(fd, command, ctx.ReadBuffer(), output2, ctrl2);
+                ctx.WriteBuffer(output2);
+                IPC::ResponseBuilder rb{ctx, 3};
+                rb.Push(RESULT_SUCCESS);
+                rb.Push(result);
+            },
+            nvdrv->GetEventWriteable(ctrl.event_id));
+    } else {
+        ctx.WriteBuffer(output);
+    }
     IPC::ResponseBuilder rb{ctx, 3};
     rb.Push(RESULT_SUCCESS);
-    rb.Push(nvdrv->Ioctl(fd, command, ctx.ReadBuffer(), output));
-
-    ctx.WriteBuffer(output);
+    rb.Push(result);
 }
 
 void NVDRV::Close(Kernel::HLERequestContext& ctx) {
@@ -66,13 +92,19 @@ void NVDRV::Initialize(Kernel::HLERequestContext& ctx) {
 void NVDRV::QueryEvent(Kernel::HLERequestContext& ctx) {
     IPC::RequestParser rp{ctx};
     u32 fd = rp.Pop<u32>();
-    u32 event_id = rp.Pop<u32>();
+    // TODO(Blinkhawk): Figure the meaning of the flag at bit 16
+    u32 event_id = rp.Pop<u32>() & 0x000000FF;
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, fd={:X}, event_id={:X}", fd, event_id);
 
     IPC::ResponseBuilder rb{ctx, 3, 1};
     rb.Push(RESULT_SUCCESS);
-    rb.PushCopyObjects(query_event.readable);
-    rb.Push<u32>(0);
+    if (event_id < MaxNvEvents) {
+        rb.PushCopyObjects(nvdrv->GetEvent(event_id));
+        rb.Push<u32>(NvResult::Success);
+    } else {
+        rb.Push<u32>(0);
+        rb.Push<u32>(NvResult::BadParameter);
+    }
 }
 
 void NVDRV::SetClientPID(Kernel::HLERequestContext& ctx) {
@@ -127,10 +159,6 @@ NVDRV::NVDRV(std::shared_ptr<Module> nvdrv, const char* name)
         {13, &NVDRV::FinishInitialize, "FinishInitialize"},
     };
     RegisterHandlers(functions);
-
-    auto& kernel = Core::System::GetInstance().Kernel();
-    query_event = Kernel::WritableEvent::CreateEventPair(kernel, Kernel::ResetType::Automatic,
-                                                         "NVDRV::query_event");
 }
 
 NVDRV::~NVDRV() = default;

src/core/hle/service/nvdrv/interface.h

@@ -19,6 +19,8 @@ public:
     NVDRV(std::shared_ptr<Module> nvdrv, const char* name);
     ~NVDRV() override;
 
+    void SignalGPUInterruptSyncpt(const u32 syncpoint_id, const u32 value);
+
 private:
     void Open(Kernel::HLERequestContext& ctx);
     void Ioctl(Kernel::HLERequestContext& ctx);
@@ -33,8 +35,6 @@ private:
     std::shared_ptr<Module> nvdrv;
 
     u64 pid{};
-
-    Kernel::EventPair query_event;
 };
 
 } // namespace Service::Nvidia

src/core/hle/service/nvdrv/nvdata.h

@@ -0,0 +1,48 @@
+#pragma once
+
+#include <array>
+#include "common/common_types.h"
+
+namespace Service::Nvidia {
+
+constexpr u32 MaxSyncPoints = 192;
+constexpr u32 MaxNvEvents = 64;
+
+struct Fence {
+    s32 id;
+    u32 value;
+};
+
+static_assert(sizeof(Fence) == 8, "Fence has wrong size");
+
+struct MultiFence {
+    u32 num_fences;
+    std::array<Fence, 4> fences;
+};
+
+enum NvResult : u32 {
+    Success = 0,
+    BadParameter = 4,
+    Timeout = 5,
+    ResourceError = 15,
+};
+
+enum class EventState {
+    Free = 0,
+    Registered = 1,
+    Waiting = 2,
+    Busy = 3,
+};
+
+struct IoctlCtrl {
+    // First call done to the servioce for services that call itself again after a call.
+    bool fresh_call{true};
+    // Tells the Ioctl Wrapper that it must delay the IPC response and send the thread to sleep
+    bool must_delay{};
+    // Timeout for the delay
+    s64 timeout{};
+    // NV Event Id
+    s32 event_id{-1};
+};
+
+} // namespace Service::Nvidia

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

@@ -4,7 +4,10 @@
 
 #include <utility>
 
+#include <fmt/format.h>
 #include "core/hle/ipc_helpers.h"
+#include "core/hle/kernel/readable_event.h"
+#include "core/hle/kernel/writable_event.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
 #include "core/hle/service/nvdrv/devices/nvdisp_disp0.h"
 #include "core/hle/service/nvdrv/devices/nvhost_as_gpu.h"
@@ -22,8 +25,9 @@
 
 namespace Service::Nvidia {
 
-void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger) {
-    auto module_ = std::make_shared<Module>();
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger,
+                       Core::System& system) {
+    auto module_ = std::make_shared<Module>(system);
     std::make_shared<NVDRV>(module_, "nvdrv")->InstallAsService(service_manager);
     std::make_shared<NVDRV>(module_, "nvdrv:a")->InstallAsService(service_manager);
     std::make_shared<NVDRV>(module_, "nvdrv:s")->InstallAsService(service_manager);
@@ -32,17 +36,25 @@ void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger
     nvflinger.SetNVDrvInstance(module_);
 }
 
-Module::Module() {
-    auto nvmap_dev = std::make_shared<Devices::nvmap>();
-    devices["/dev/nvhost-as-gpu"] = std::make_shared<Devices::nvhost_as_gpu>(nvmap_dev);
-    devices["/dev/nvhost-gpu"] = std::make_shared<Devices::nvhost_gpu>(nvmap_dev);
-    devices["/dev/nvhost-ctrl-gpu"] = std::make_shared<Devices::nvhost_ctrl_gpu>();
+Module::Module(Core::System& system) {
+    auto& kernel = system.Kernel();
+    for (u32 i = 0; i < MaxNvEvents; i++) {
+        std::string event_label = fmt::format("NVDRV::NvEvent_{}", i);
+        events_interface.events[i] = Kernel::WritableEvent::CreateEventPair(
+            kernel, Kernel::ResetType::Automatic, event_label);
+        events_interface.status[i] = EventState::Free;
+        events_interface.registered[i] = false;
+    }
+    auto nvmap_dev = std::make_shared<Devices::nvmap>(system);
+    devices["/dev/nvhost-as-gpu"] = std::make_shared<Devices::nvhost_as_gpu>(system, nvmap_dev);
+    devices["/dev/nvhost-gpu"] = std::make_shared<Devices::nvhost_gpu>(system, nvmap_dev);
+    devices["/dev/nvhost-ctrl-gpu"] = std::make_shared<Devices::nvhost_ctrl_gpu>(system);
     devices["/dev/nvmap"] = nvmap_dev;
-    devices["/dev/nvdisp_disp0"] = std::make_shared<Devices::nvdisp_disp0>(nvmap_dev);
-    devices["/dev/nvhost-ctrl"] = std::make_shared<Devices::nvhost_ctrl>();
-    devices["/dev/nvhost-nvdec"] = std::make_shared<Devices::nvhost_nvdec>();
-    devices["/dev/nvhost-nvjpg"] = std::make_shared<Devices::nvhost_nvjpg>();
-    devices["/dev/nvhost-vic"] = std::make_shared<Devices::nvhost_vic>();
+    devices["/dev/nvdisp_disp0"] = std::make_shared<Devices::nvdisp_disp0>(system, nvmap_dev);
+    devices["/dev/nvhost-ctrl"] = std::make_shared<Devices::nvhost_ctrl>(system, events_interface);
+    devices["/dev/nvhost-nvdec"] = std::make_shared<Devices::nvhost_nvdec>(system);
+    devices["/dev/nvhost-nvjpg"] = std::make_shared<Devices::nvhost_nvjpg>(system);
+    devices["/dev/nvhost-vic"] = std::make_shared<Devices::nvhost_vic>(system);
 }
 
 Module::~Module() = default;
@@ -59,12 +71,13 @@ u32 Module::Open(const std::string& device_name) {
     return fd;
 }
 
-u32 Module::Ioctl(u32 fd, u32 command, const std::vector<u8>& input, std::vector<u8>& output) {
+u32 Module::Ioctl(u32 fd, u32 command, const std::vector<u8>& input, std::vector<u8>& output,
+                  IoctlCtrl& ctrl) {
     auto itr = open_files.find(fd);
     ASSERT_MSG(itr != open_files.end(), "Tried to talk to an invalid device");
 
     auto& device = itr->second;
-    return device->ioctl({command}, input, output);
+    return device->ioctl({command}, input, output, ctrl);
 }
 
 ResultCode Module::Close(u32 fd) {
@@ -77,4 +90,22 @@ ResultCode Module::Close(u32 fd) {
     return RESULT_SUCCESS;
 }
 
+void Module::SignalSyncpt(const u32 syncpoint_id, const u32 value) {
+    for (u32 i = 0; i < MaxNvEvents; i++) {
+        if (events_interface.assigned_syncpt[i] == syncpoint_id &&
+            events_interface.assigned_value[i] == value) {
+            events_interface.LiberateEvent(i);
+            events_interface.events[i].writable->Signal();
+        }
+    }
+}
+
+Kernel::SharedPtr<Kernel::ReadableEvent> Module::GetEvent(const u32 event_id) const {
+    return events_interface.events[event_id].readable;
+}
+
+Kernel::SharedPtr<Kernel::WritableEvent> Module::GetEventWriteable(const u32 event_id) const {
+    return events_interface.events[event_id].writable;
+}
+
 } // namespace Service::Nvidia

src/core/hle/service/nvdrv/nvdrv.h

@@ -8,8 +8,14 @@
 #include <unordered_map>
 #include <vector>
 #include "common/common_types.h"
+#include "core/hle/kernel/writable_event.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/service.h"
 
+namespace Core {
+class System;
+}
+
 namespace Service::NVFlinger {
 class NVFlinger;
 }
@@ -20,16 +26,72 @@ namespace Devices {
 class nvdevice;
 }
 
-struct IoctlFence {
-    u32 id;
-    u32 value;
+struct EventInterface {
+    // Mask representing currently busy events
+    u64 events_mask{};
+    // Each kernel event associated to an NV event
+    std::array<Kernel::EventPair, MaxNvEvents> events;
+    // The status of the current NVEvent
+    std::array<EventState, MaxNvEvents> status{};
+    // Tells if an NVEvent is registered or not
+    std::array<bool, MaxNvEvents> registered{};
+    // When an NVEvent is waiting on GPU interrupt, this is the sync_point
+    // associated with it.
+    std::array<u32, MaxNvEvents> assigned_syncpt{};
+    // This is the value of the GPU interrupt for which the NVEvent is waiting
+    // for.
+    std::array<u32, MaxNvEvents> assigned_value{};
+    // Constant to denote an unasigned syncpoint.
+    static constexpr u32 unassigned_syncpt = 0xFFFFFFFF;
+    std::optional<u32> GetFreeEvent() const {
+        u64 mask = events_mask;
+        for (u32 i = 0; i < MaxNvEvents; i++) {
+            const bool is_free = (mask & 0x1) == 0;
+            if (is_free) {
+                if (status[i] == EventState::Registered || status[i] == EventState::Free) {
+                    return {i};
+                }
+            }
+            mask = mask >> 1;
+        }
+        return {};
+    }
+    void SetEventStatus(const u32 event_id, EventState new_status) {
+        EventState old_status = status[event_id];
+        if (old_status == new_status) {
+            return;
+        }
+        status[event_id] = new_status;
+        if (new_status == EventState::Registered) {
+            registered[event_id] = true;
+        }
+        if (new_status == EventState::Waiting || new_status == EventState::Busy) {
+            events_mask |= (1ULL << event_id);
+        }
+    }
+    void RegisterEvent(const u32 event_id) {
+        registered[event_id] = true;
+        if (status[event_id] == EventState::Free) {
+            status[event_id] = EventState::Registered;
+        }
+    }
+    void UnregisterEvent(const u32 event_id) {
+        registered[event_id] = false;
+        if (status[event_id] == EventState::Registered) {
+            status[event_id] = EventState::Free;
+        }
+    }
+    void LiberateEvent(const u32 event_id) {
+        status[event_id] = registered[event_id] ? EventState::Registered : EventState::Free;
+        events_mask &= ~(1ULL << event_id);
+        assigned_syncpt[event_id] = unassigned_syncpt;
+        assigned_value[event_id] = 0;
+    }
 };
 
-static_assert(sizeof(IoctlFence) == 8, "IoctlFence has wrong size");
-
 class Module final {
 public:
-    Module();
+    Module(Core::System& system);
     ~Module();
 
     /// Returns a pointer to one of the available devices, identified by its name.
@@ -44,10 +106,17 @@ public:
     /// Opens a device node and returns a file descriptor to it.
     u32 Open(const std::string& device_name);
     /// Sends an ioctl command to the specified file descriptor.
-    u32 Ioctl(u32 fd, u32 command, const std::vector<u8>& input, std::vector<u8>& output);
+    u32 Ioctl(u32 fd, u32 command, const std::vector<u8>& input, std::vector<u8>& output,
+              IoctlCtrl& ctrl);
     /// Closes a device file descriptor and returns operation success.
     ResultCode Close(u32 fd);
 
+    void SignalSyncpt(const u32 syncpoint_id, const u32 value);
+
+    Kernel::SharedPtr<Kernel::ReadableEvent> GetEvent(u32 event_id) const;
+
+    Kernel::SharedPtr<Kernel::WritableEvent> GetEventWriteable(u32 event_id) const;
+
 private:
     /// Id to use for the next open file descriptor.
     u32 next_fd = 1;
@@ -57,9 +126,12 @@ private:
 
     /// Mapping of device node names to their implementation.
     std::unordered_map<std::string, std::shared_ptr<Devices::nvdevice>> devices;
+
+    EventInterface events_interface;
 };
 
 /// Registers all NVDRV services with the specified service manager.
-void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger);
+void InstallInterfaces(SM::ServiceManager& service_manager, NVFlinger::NVFlinger& nvflinger,
+                       Core::System& system);
 
 } // namespace Service::Nvidia

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

@@ -34,7 +34,8 @@ void BufferQueue::SetPreallocatedBuffer(u32 slot, const IGBPBuffer& igbp_buffer)
     buffer_wait_event.writable->Signal();
 }
 
-std::optional<u32> BufferQueue::DequeueBuffer(u32 width, u32 height) {
+std::optional<std::pair<u32, Service::Nvidia::MultiFence*>> BufferQueue::DequeueBuffer(u32 width,
+                                                                                       u32 height) {
     auto itr = std::find_if(queue.begin(), queue.end(), [&](const Buffer& buffer) {
         // Only consider free buffers. Buffers become free once again after they've been Acquired
         // and Released by the compositor, see the NVFlinger::Compose method.
@@ -51,7 +52,7 @@ std::optional<u32> BufferQueue::DequeueBuffer(u32 width, u32 height) {
     }
 
     itr->status = Buffer::Status::Dequeued;
-    return itr->slot;
+    return {{itr->slot, &itr->multi_fence}};
 }
 
 const IGBPBuffer& BufferQueue::RequestBuffer(u32 slot) const {
@@ -63,7 +64,8 @@ const IGBPBuffer& BufferQueue::RequestBuffer(u32 slot) const {
 }
 
 void BufferQueue::QueueBuffer(u32 slot, BufferTransformFlags transform,
-                              const Common::Rectangle<int>& crop_rect) {
+                              const Common::Rectangle<int>& crop_rect, u32 swap_interval,
+                              Service::Nvidia::MultiFence& multi_fence) {
     auto itr = std::find_if(queue.begin(), queue.end(),
                             [&](const Buffer& buffer) { return buffer.slot == slot; });
     ASSERT(itr != queue.end());
@@ -71,12 +73,21 @@ void BufferQueue::QueueBuffer(u32 slot, BufferTransformFlags transform,
     itr->status = Buffer::Status::Queued;
     itr->transform = transform;
     itr->crop_rect = crop_rect;
+    itr->swap_interval = swap_interval;
+    itr->multi_fence = multi_fence;
+    queue_sequence.push_back(slot);
 }
 
 std::optional<std::reference_wrapper<const BufferQueue::Buffer>> BufferQueue::AcquireBuffer() {
-    auto itr = std::find_if(queue.begin(), queue.end(), [](const Buffer& buffer) {
-        return buffer.status == Buffer::Status::Queued;
-    });
+    auto itr = queue.end();
+    // Iterate to find a queued buffer matching the requested slot.
+    while (itr == queue.end() && !queue_sequence.empty()) {
+        u32 slot = queue_sequence.front();
+        itr = std::find_if(queue.begin(), queue.end(), [&slot](const Buffer& buffer) {
+            return buffer.status == Buffer::Status::Queued && buffer.slot == slot;
+        });
+        queue_sequence.pop_front();
+    }
     if (itr == queue.end())
         return {};
     itr->status = Buffer::Status::Acquired;

src/core/hle/service/nvflinger/buffer_queue.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <list>
 #include <optional>
 #include <vector>
 
@@ -12,6 +13,7 @@
 #include "common/swap.h"
 #include "core/hle/kernel/object.h"
 #include "core/hle/kernel/writable_event.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 
 namespace Service::NVFlinger {
 
@@ -68,13 +70,17 @@ public:
         IGBPBuffer igbp_buffer;
         BufferTransformFlags transform;
         Common::Rectangle<int> crop_rect;
+        u32 swap_interval;
+        Service::Nvidia::MultiFence multi_fence;
     };
 
     void SetPreallocatedBuffer(u32 slot, const IGBPBuffer& igbp_buffer);
-    std::optional<u32> DequeueBuffer(u32 width, u32 height);
+    std::optional<std::pair<u32, Service::Nvidia::MultiFence*>> DequeueBuffer(u32 width,
+                                                                              u32 height);
     const IGBPBuffer& RequestBuffer(u32 slot) const;
     void QueueBuffer(u32 slot, BufferTransformFlags transform,
-                     const Common::Rectangle<int>& crop_rect);
+                     const Common::Rectangle<int>& crop_rect, u32 swap_interval,
+                     Service::Nvidia::MultiFence& multi_fence);
     std::optional<std::reference_wrapper<const Buffer>> AcquireBuffer();
     void ReleaseBuffer(u32 slot);
     u32 Query(QueryType type);
@@ -92,6 +98,7 @@ private:
     u64 layer_id;
 
     std::vector<Buffer> queue;
+    std::list<u32> queue_sequence;
     Kernel::EventPair buffer_wait_event;
 };
 

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

@@ -37,15 +37,14 @@ NVFlinger::NVFlinger(Core::Timing::CoreTiming& core_timing) : core_timing{core_t
     displays.emplace_back(4, "Null");
 
     // Schedule the screen composition events
-    const auto ticks = Settings::values.force_30fps_mode ? frame_ticks_30fps : frame_ticks;
+    composition_event = core_timing.RegisterEvent("ScreenComposition", [this](u64 userdata,
+                                                                              s64 cycles_late) {
+        Compose();
+        const auto ticks = Settings::values.force_30fps_mode ? frame_ticks_30fps : GetNextTicks();
+        this->core_timing.ScheduleEvent(std::max<s64>(0LL, ticks - cycles_late), composition_event);
+    });
 
-    composition_event = core_timing.RegisterEvent(
-        "ScreenComposition", [this, ticks](u64 userdata, s64 cycles_late) {
-            Compose();
-            this->core_timing.ScheduleEvent(ticks - cycles_late, composition_event);
-        });
-
-    core_timing.ScheduleEvent(ticks, composition_event);
+    core_timing.ScheduleEvent(frame_ticks, composition_event);
 }
 
 NVFlinger::~NVFlinger() {
@@ -206,8 +205,14 @@ void NVFlinger::Compose() {
                      igbp_buffer.width, igbp_buffer.height, igbp_buffer.stride,
                      buffer->get().transform, buffer->get().crop_rect);
 
+        swap_interval = buffer->get().swap_interval;
         buffer_queue.ReleaseBuffer(buffer->get().slot);
     }
 }
 
+s64 NVFlinger::GetNextTicks() const {
+    constexpr s64 max_hertz = 120LL;
+    return (Core::Timing::BASE_CLOCK_RATE * (1LL << swap_interval)) / max_hertz;
+}
+
 } // namespace Service::NVFlinger

src/core/hle/service/nvflinger/nvflinger.h

@@ -74,6 +74,8 @@ public:
     /// finished.
     void Compose();
 
+    s64 GetNextTicks() const;
+
 private:
     /// Finds the display identified by the specified ID.
     VI::Display* FindDisplay(u64 display_id);
@@ -98,6 +100,8 @@ private:
     /// layers.
     u32 next_buffer_queue_id = 1;
 
+    u32 swap_interval = 1;
+
     /// Event that handles screen composition.
     Core::Timing::EventType* composition_event;
 

src/core/hle/service/service.cpp

@@ -206,7 +206,7 @@ void Init(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system) {
     AM::InstallInterfaces(*sm, nv_flinger, system);
     AOC::InstallInterfaces(*sm);
     APM::InstallInterfaces(system);
-    Audio::InstallInterfaces(*sm);
+    Audio::InstallInterfaces(*sm, system);
     BCAT::InstallInterfaces(*sm);
     BPC::InstallInterfaces(*sm);
     BtDrv::InstallInterfaces(*sm);
@@ -236,7 +236,7 @@ void Init(std::shared_ptr<SM::ServiceManager>& sm, Core::System& system) {
     NIM::InstallInterfaces(*sm);
     NPNS::InstallInterfaces(*sm);
     NS::InstallInterfaces(*sm);
-    Nvidia::InstallInterfaces(*sm, *nv_flinger);
+    Nvidia::InstallInterfaces(*sm, *nv_flinger, system);
     PCIe::InstallInterfaces(*sm);
     PCTL::InstallInterfaces(*sm);
     PCV::InstallInterfaces(*sm);

src/core/hle/service/vi/vi.cpp

@@ -21,6 +21,7 @@
 #include "core/hle/kernel/readable_event.h"
 #include "core/hle/kernel/thread.h"
 #include "core/hle/kernel/writable_event.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/nvdrv/nvdrv.h"
 #include "core/hle/service/nvflinger/buffer_queue.h"
 #include "core/hle/service/nvflinger/nvflinger.h"
@@ -328,32 +329,22 @@ public:
     Data data;
 };
 
-struct BufferProducerFence {
-    u32 is_valid;
-    std::array<Nvidia::IoctlFence, 4> fences;
-};
-static_assert(sizeof(BufferProducerFence) == 36, "BufferProducerFence has wrong size");
-
 class IGBPDequeueBufferResponseParcel : public Parcel {
 public:
-    explicit IGBPDequeueBufferResponseParcel(u32 slot) : slot(slot) {}
+    explicit IGBPDequeueBufferResponseParcel(u32 slot, Service::Nvidia::MultiFence& multi_fence)
+        : slot(slot), multi_fence(multi_fence) {}
     ~IGBPDequeueBufferResponseParcel() override = default;
 
 protected:
     void SerializeData() override {
-        // TODO(Subv): Find out how this Fence is used.
-        BufferProducerFence fence = {};
-        fence.is_valid = 1;
-        for (auto& fence_ : fence.fences)
-            fence_.id = -1;
-
         Write(slot);
         Write<u32_le>(1);
-        WriteObject(fence);
+        WriteObject(multi_fence);
         Write<u32_le>(0);
     }
 
     u32_le slot;
+    Service::Nvidia::MultiFence multi_fence;
 };
 
 class IGBPRequestBufferRequestParcel : public Parcel {
@@ -400,12 +391,6 @@ public:
         data = Read<Data>();
     }
 
-    struct Fence {
-        u32_le id;
-        u32_le value;
-    };
-    static_assert(sizeof(Fence) == 8, "Fence has wrong size");
-
     struct Data {
         u32_le slot;
         INSERT_PADDING_WORDS(3);
@@ -418,15 +403,15 @@ public:
         s32_le scaling_mode;
         NVFlinger::BufferQueue::BufferTransformFlags transform;
         u32_le sticky_transform;
-        INSERT_PADDING_WORDS(2);
-        u32_le fence_is_valid;
-        std::array<Fence, 2> fences;
+        INSERT_PADDING_WORDS(1);
+        u32_le swap_interval;
+        Service::Nvidia::MultiFence multi_fence;
 
         Common::Rectangle<int> GetCropRect() const {
             return {crop_left, crop_top, crop_right, crop_bottom};
         }
     };
-    static_assert(sizeof(Data) == 80, "ParcelData has wrong size");
+    static_assert(sizeof(Data) == 96, "ParcelData has wrong size");
 
     Data data;
 };
@@ -547,11 +532,11 @@ private:
             IGBPDequeueBufferRequestParcel request{ctx.ReadBuffer()};
             const u32 width{request.data.width};
             const u32 height{request.data.height};
-            std::optional<u32> slot = buffer_queue.DequeueBuffer(width, height);
+            auto result = buffer_queue.DequeueBuffer(width, height);
 
-            if (slot) {
+            if (result) {
                 // Buffer is available
-                IGBPDequeueBufferResponseParcel response{*slot};
+                IGBPDequeueBufferResponseParcel response{result->first, *result->second};
                 ctx.WriteBuffer(response.Serialize());
             } else {
                 // Wait the current thread until a buffer becomes available
@@ -561,10 +546,10 @@ private:
                         Kernel::ThreadWakeupReason reason) {
                         // Repeat TransactParcel DequeueBuffer when a buffer is available
                         auto& buffer_queue = nv_flinger->FindBufferQueue(id);
-                        std::optional<u32> slot = buffer_queue.DequeueBuffer(width, height);
-                        ASSERT_MSG(slot != std::nullopt, "Could not dequeue buffer.");
+                        auto result = buffer_queue.DequeueBuffer(width, height);
+                        ASSERT_MSG(result != std::nullopt, "Could not dequeue buffer.");
 
-                        IGBPDequeueBufferResponseParcel response{*slot};
+                        IGBPDequeueBufferResponseParcel response{result->first, *result->second};
                         ctx.WriteBuffer(response.Serialize());
                         IPC::ResponseBuilder rb{ctx, 2};
                         rb.Push(RESULT_SUCCESS);
@@ -582,7 +567,8 @@ private:
             IGBPQueueBufferRequestParcel request{ctx.ReadBuffer()};
 
             buffer_queue.QueueBuffer(request.data.slot, request.data.transform,
-                                     request.data.GetCropRect());
+                                     request.data.GetCropRect(), request.data.swap_interval,
+                                     request.data.multi_fence);
 
             IGBPQueueBufferResponseParcel response{1280, 720};
             ctx.WriteBuffer(response.Serialize());

src/core/loader/elf.cpp

@@ -295,7 +295,7 @@ Kernel::CodeSet ElfReader::LoadInto(VAddr vaddr) {
         }
     }
 
-    std::vector<u8> program_image(total_image_size);
+    Kernel::PhysicalMemory program_image(total_image_size);
     std::size_t current_image_position = 0;
 
     Kernel::CodeSet codeset;

src/core/loader/kip.cpp

@@ -69,7 +69,7 @@ AppLoader::LoadResult AppLoader_KIP::Load(Kernel::Process& process) {
 
     const VAddr base_address = process.VMManager().GetCodeRegionBaseAddress();
     Kernel::CodeSet codeset;
-    std::vector<u8> program_image;
+    Kernel::PhysicalMemory program_image;
 
     const auto load_segment = [&program_image](Kernel::CodeSet::Segment& segment,
                                                const std::vector<u8>& data, u32 offset) {

src/core/loader/nro.cpp

@@ -143,7 +143,7 @@ static bool LoadNroImpl(Kernel::Process& process, const std::vector<u8>& data,
     }
 
     // Build program image
-    std::vector<u8> program_image(PageAlignSize(nro_header.file_size));
+    Kernel::PhysicalMemory program_image(PageAlignSize(nro_header.file_size));
     std::memcpy(program_image.data(), data.data(), program_image.size());
     if (program_image.size() != PageAlignSize(nro_header.file_size)) {
         return {};

src/core/loader/nso.cpp

@@ -89,7 +89,7 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::Process& process,
 
     // Build program image
     Kernel::CodeSet codeset;
-    std::vector<u8> program_image;
+    Kernel::PhysicalMemory program_image;
     for (std::size_t i = 0; i < nso_header.segments.size(); ++i) {
         std::vector<u8> data =
             file.ReadBytes(nso_header.segments_compressed_size[i], nso_header.segments[i].offset);

src/video_core/CMakeLists.txt

@@ -1,5 +1,7 @@
 add_library(video_core STATIC
-    buffer_cache.h
+    buffer_cache/buffer_block.h
+    buffer_cache/buffer_cache.h
+    buffer_cache/map_interval.h
     dma_pusher.cpp
     dma_pusher.h
     debug_utils/debug_utils.cpp
@@ -100,6 +102,7 @@ add_library(video_core STATIC
     shader/decode/integer_set.cpp
     shader/decode/half_set.cpp
     shader/decode/video.cpp
+    shader/decode/warp.cpp
     shader/decode/xmad.cpp
     shader/decode/other.cpp
     shader/control_flow.cpp

src/video_core/buffer_cache.h

@@ -1,299 +0,0 @@
-// Copyright 2019 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include <array>
-#include <memory>
-#include <mutex>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "common/alignment.h"
-#include "common/common_types.h"
-#include "core/core.h"
-#include "video_core/memory_manager.h"
-#include "video_core/rasterizer_cache.h"
-
-namespace VideoCore {
-class RasterizerInterface;
-}
-
-namespace VideoCommon {
-
-template <typename BufferStorageType>
-class CachedBuffer final : public RasterizerCacheObject {
-public:
-    explicit CachedBuffer(VAddr cpu_addr, u8* host_ptr)
-        : RasterizerCacheObject{host_ptr}, host_ptr{host_ptr}, cpu_addr{cpu_addr} {}
-    ~CachedBuffer() override = default;
-
-    VAddr GetCpuAddr() const override {
-        return cpu_addr;
-    }
-
-    std::size_t GetSizeInBytes() const override {
-        return size;
-    }
-
-    u8* GetWritableHostPtr() const {
-        return host_ptr;
-    }
-
-    std::size_t GetSize() const {
-        return size;
-    }
-
-    std::size_t GetCapacity() const {
-        return capacity;
-    }
-
-    bool IsInternalized() const {
-        return is_internal;
-    }
-
-    const BufferStorageType& GetBuffer() const {
-        return buffer;
-    }
-
-    void SetSize(std::size_t new_size) {
-        size = new_size;
-    }
-
-    void SetInternalState(bool is_internal_) {
-        is_internal = is_internal_;
-    }
-
-    BufferStorageType ExchangeBuffer(BufferStorageType buffer_, std::size_t new_capacity) {
-        capacity = new_capacity;
-        std::swap(buffer, buffer_);
-        return buffer_;
-    }
-
-private:
-    u8* host_ptr{};
-    VAddr cpu_addr{};
-    std::size_t size{};
-    std::size_t capacity{};
-    bool is_internal{};
-    BufferStorageType buffer;
-};
-
-template <typename BufferStorageType, typename BufferType, typename StreamBuffer>
-class BufferCache : public RasterizerCache<std::shared_ptr<CachedBuffer<BufferStorageType>>> {
-public:
-    using Buffer = std::shared_ptr<CachedBuffer<BufferStorageType>>;
-    using BufferInfo = std::pair<const BufferType*, u64>;
-
-    explicit BufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
-                         std::unique_ptr<StreamBuffer> stream_buffer)
-        : RasterizerCache<Buffer>{rasterizer}, system{system},
-          stream_buffer{std::move(stream_buffer)}, stream_buffer_handle{
-                                                       this->stream_buffer->GetHandle()} {}
-    ~BufferCache() = default;
-
-    void Unregister(const Buffer& entry) override {
-        std::lock_guard lock{RasterizerCache<Buffer>::mutex};
-        if (entry->IsInternalized()) {
-            internalized_entries.erase(entry->GetCacheAddr());
-        }
-        ReserveBuffer(entry);
-        RasterizerCache<Buffer>::Unregister(entry);
-    }
-
-    void TickFrame() {
-        marked_for_destruction_index =
-            (marked_for_destruction_index + 1) % marked_for_destruction_ring_buffer.size();
-        MarkedForDestruction().clear();
-    }
-
-    BufferInfo UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4,
-                            bool internalize = false, bool is_written = false) {
-        std::lock_guard lock{RasterizerCache<Buffer>::mutex};
-
-        auto& memory_manager = system.GPU().MemoryManager();
-        const auto host_ptr = memory_manager.GetPointer(gpu_addr);
-        if (!host_ptr) {
-            return {GetEmptyBuffer(size), 0};
-        }
-        const auto cache_addr = ToCacheAddr(host_ptr);
-
-        // Cache management is a big overhead, so only cache entries with a given size.
-        // TODO: Figure out which size is the best for given games.
-        constexpr std::size_t max_stream_size = 0x800;
-        if (!internalize && size < max_stream_size &&
-            internalized_entries.find(cache_addr) == internalized_entries.end()) {
-            return StreamBufferUpload(host_ptr, size, alignment);
-        }
-
-        auto entry = RasterizerCache<Buffer>::TryGet(cache_addr);
-        if (!entry) {
-            return FixedBufferUpload(gpu_addr, host_ptr, size, internalize, is_written);
-        }
-
-        if (entry->GetSize() < size) {
-            IncreaseBufferSize(entry, size);
-        }
-        if (is_written) {
-            entry->MarkAsModified(true, *this);
-        }
-        return {ToHandle(entry->GetBuffer()), 0};
-    }
-
-    /// Uploads from a host memory. Returns the OpenGL buffer where it's located and its offset.
-    BufferInfo UploadHostMemory(const void* raw_pointer, std::size_t size,
-                                std::size_t alignment = 4) {
-        std::lock_guard lock{RasterizerCache<Buffer>::mutex};
-        return StreamBufferUpload(raw_pointer, size, alignment);
-    }
-
-    void Map(std::size_t max_size) {
-        std::tie(buffer_ptr, buffer_offset_base, invalidated) = stream_buffer->Map(max_size, 4);
-        buffer_offset = buffer_offset_base;
-    }
-
-    /// Finishes the upload stream, returns true on bindings invalidation.
-    bool Unmap() {
-        stream_buffer->Unmap(buffer_offset - buffer_offset_base);
-        return std::exchange(invalidated, false);
-    }
-
-    virtual const BufferType* GetEmptyBuffer(std::size_t size) = 0;
-
-protected:
-    void FlushObjectInner(const Buffer& entry) override {
-        DownloadBufferData(entry->GetBuffer(), 0, entry->GetSize(), entry->GetWritableHostPtr());
-    }
-
-    virtual BufferStorageType CreateBuffer(std::size_t size) = 0;
-
-    virtual const BufferType* ToHandle(const BufferStorageType& storage) = 0;
-
-    virtual void UploadBufferData(const BufferStorageType& buffer, std::size_t offset,
-                                  std::size_t size, const u8* data) = 0;
-
-    virtual void DownloadBufferData(const BufferStorageType& buffer, std::size_t offset,
-                                    std::size_t size, u8* data) = 0;
-
-    virtual void CopyBufferData(const BufferStorageType& src, const BufferStorageType& dst,
-                                std::size_t src_offset, std::size_t dst_offset,
-                                std::size_t size) = 0;
-
-private:
-    BufferInfo StreamBufferUpload(const void* raw_pointer, std::size_t size,
-                                  std::size_t alignment) {
-        AlignBuffer(alignment);
-        const std::size_t uploaded_offset = buffer_offset;
-        std::memcpy(buffer_ptr, raw_pointer, size);
-
-        buffer_ptr += size;
-        buffer_offset += size;
-        return {&stream_buffer_handle, uploaded_offset};
-    }
-
-    BufferInfo FixedBufferUpload(GPUVAddr gpu_addr, u8* host_ptr, std::size_t size,
-                                 bool internalize, bool is_written) {
-        auto& memory_manager = Core::System::GetInstance().GPU().MemoryManager();
-        const auto cpu_addr = memory_manager.GpuToCpuAddress(gpu_addr);
-        ASSERT(cpu_addr);
-
-        auto entry = GetUncachedBuffer(*cpu_addr, host_ptr);
-        entry->SetSize(size);
-        entry->SetInternalState(internalize);
-        RasterizerCache<Buffer>::Register(entry);
-
-        if (internalize) {
-            internalized_entries.emplace(ToCacheAddr(host_ptr));
-        }
-        if (is_written) {
-            entry->MarkAsModified(true, *this);
-        }
-
-        if (entry->GetCapacity() < size) {
-            MarkedForDestruction().push_back(entry->ExchangeBuffer(CreateBuffer(size), size));
-        }
-
-        UploadBufferData(entry->GetBuffer(), 0, size, host_ptr);
-        return {ToHandle(entry->GetBuffer()), 0};
-    }
-
-    void IncreaseBufferSize(Buffer& entry, std::size_t new_size) {
-        const std::size_t old_size = entry->GetSize();
-        if (entry->GetCapacity() < new_size) {
-            const auto& old_buffer = entry->GetBuffer();
-            auto new_buffer = CreateBuffer(new_size);
-
-            // Copy bits from the old buffer to the new buffer.
-            CopyBufferData(old_buffer, new_buffer, 0, 0, old_size);
-            MarkedForDestruction().push_back(
-                entry->ExchangeBuffer(std::move(new_buffer), new_size));
-
-            // This buffer could have been used
-            invalidated = true;
-        }
-        // Upload the new bits.
-        const std::size_t size_diff = new_size - old_size;
-        UploadBufferData(entry->GetBuffer(), old_size, size_diff, entry->GetHostPtr() + old_size);
-
-        // Update entry's size in the object and in the cache.
-        Unregister(entry);
-
-        entry->SetSize(new_size);
-        RasterizerCache<Buffer>::Register(entry);
-    }
-
-    Buffer GetUncachedBuffer(VAddr cpu_addr, u8* host_ptr) {
-        if (auto entry = TryGetReservedBuffer(host_ptr)) {
-            return entry;
-        }
-        return std::make_shared<CachedBuffer<BufferStorageType>>(cpu_addr, host_ptr);
-    }
-
-    Buffer TryGetReservedBuffer(u8* host_ptr) {
-        const auto it = buffer_reserve.find(ToCacheAddr(host_ptr));
-        if (it == buffer_reserve.end()) {
-            return {};
-        }
-        auto& reserve = it->second;
-        auto entry = reserve.back();
-        reserve.pop_back();
-        return entry;
-    }
-
-    void ReserveBuffer(Buffer entry) {
-        buffer_reserve[entry->GetCacheAddr()].push_back(std::move(entry));
-    }
-
-    void AlignBuffer(std::size_t alignment) {
-        // Align the offset, not the mapped pointer
-        const std::size_t offset_aligned = Common::AlignUp(buffer_offset, alignment);
-        buffer_ptr += offset_aligned - buffer_offset;
-        buffer_offset = offset_aligned;
-    }
-
-    std::vector<BufferStorageType>& MarkedForDestruction() {
-        return marked_for_destruction_ring_buffer[marked_for_destruction_index];
-    }
-
-    Core::System& system;
-
-    std::unique_ptr<StreamBuffer> stream_buffer;
-    BufferType stream_buffer_handle{};
-
-    bool invalidated = false;
-
-    u8* buffer_ptr = nullptr;
-    u64 buffer_offset = 0;
-    u64 buffer_offset_base = 0;
-
-    std::size_t marked_for_destruction_index = 0;
-    std::array<std::vector<BufferStorageType>, 4> marked_for_destruction_ring_buffer;
-
-    std::unordered_set<CacheAddr> internalized_entries;
-    std::unordered_map<CacheAddr, std::vector<Buffer>> buffer_reserve;
-};
-
-} // namespace VideoCommon

src/video_core/buffer_cache/buffer_block.h

@@ -0,0 +1,76 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <unordered_set>
+#include <utility>
+
+#include "common/alignment.h"
+#include "common/common_types.h"
+#include "video_core/gpu.h"
+
+namespace VideoCommon {
+
+class BufferBlock {
+public:
+    bool Overlaps(const CacheAddr start, const CacheAddr end) const {
+        return (cache_addr < end) && (cache_addr_end > start);
+    }
+
+    bool IsInside(const CacheAddr other_start, const CacheAddr other_end) const {
+        return cache_addr <= other_start && other_end <= cache_addr_end;
+    }
+
+    u8* GetWritableHostPtr() const {
+        return FromCacheAddr(cache_addr);
+    }
+
+    u8* GetWritableHostPtr(std::size_t offset) const {
+        return FromCacheAddr(cache_addr + offset);
+    }
+
+    std::size_t GetOffset(const CacheAddr in_addr) {
+        return static_cast<std::size_t>(in_addr - cache_addr);
+    }
+
+    CacheAddr GetCacheAddr() const {
+        return cache_addr;
+    }
+
+    CacheAddr GetCacheAddrEnd() const {
+        return cache_addr_end;
+    }
+
+    void SetCacheAddr(const CacheAddr new_addr) {
+        cache_addr = new_addr;
+        cache_addr_end = new_addr + size;
+    }
+
+    std::size_t GetSize() const {
+        return size;
+    }
+
+    void SetEpoch(u64 new_epoch) {
+        epoch = new_epoch;
+    }
+
+    u64 GetEpoch() {
+        return epoch;
+    }
+
+protected:
+    explicit BufferBlock(CacheAddr cache_addr, const std::size_t size) : size{size} {
+        SetCacheAddr(cache_addr);
+    }
+    ~BufferBlock() = default;
+
+private:
+    CacheAddr cache_addr{};
+    CacheAddr cache_addr_end{};
+    std::size_t size{};
+    u64 epoch{};
+};
+
+} // namespace VideoCommon

src/video_core/buffer_cache/buffer_cache.h

@@ -0,0 +1,447 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "common/alignment.h"
+#include "common/common_types.h"
+#include "core/core.h"
+#include "video_core/buffer_cache/buffer_block.h"
+#include "video_core/buffer_cache/map_interval.h"
+#include "video_core/memory_manager.h"
+#include "video_core/rasterizer_interface.h"
+
+namespace VideoCommon {
+
+using MapInterval = std::shared_ptr<MapIntervalBase>;
+
+template <typename TBuffer, typename TBufferType, typename StreamBuffer>
+class BufferCache {
+public:
+    using BufferInfo = std::pair<const TBufferType*, u64>;
+
+    BufferInfo UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4,
+                            bool is_written = false) {
+        std::lock_guard lock{mutex};
+
+        auto& memory_manager = system.GPU().MemoryManager();
+        const auto host_ptr = memory_manager.GetPointer(gpu_addr);
+        if (!host_ptr) {
+            return {GetEmptyBuffer(size), 0};
+        }
+        const auto cache_addr = ToCacheAddr(host_ptr);
+
+        // Cache management is a big overhead, so only cache entries with a given size.
+        // TODO: Figure out which size is the best for given games.
+        constexpr std::size_t max_stream_size = 0x800;
+        if (size < max_stream_size) {
+            if (!is_written && !IsRegionWritten(cache_addr, cache_addr + size - 1)) {
+                return StreamBufferUpload(host_ptr, size, alignment);
+            }
+        }
+
+        auto block = GetBlock(cache_addr, size);
+        auto map = MapAddress(block, gpu_addr, cache_addr, size);
+        if (is_written) {
+            map->MarkAsModified(true, GetModifiedTicks());
+            if (!map->IsWritten()) {
+                map->MarkAsWritten(true);
+                MarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
+            }
+        } else {
+            if (map->IsWritten()) {
+                WriteBarrier();
+            }
+        }
+
+        const u64 offset = static_cast<u64>(block->GetOffset(cache_addr));
+
+        return {ToHandle(block), offset};
+    }
+
+    /// Uploads from a host memory. Returns the OpenGL buffer where it's located and its offset.
+    BufferInfo UploadHostMemory(const void* raw_pointer, std::size_t size,
+                                std::size_t alignment = 4) {
+        std::lock_guard lock{mutex};
+        return StreamBufferUpload(raw_pointer, size, alignment);
+    }
+
+    void Map(std::size_t max_size) {
+        std::lock_guard lock{mutex};
+
+        std::tie(buffer_ptr, buffer_offset_base, invalidated) = stream_buffer->Map(max_size, 4);
+        buffer_offset = buffer_offset_base;
+    }
+
+    /// Finishes the upload stream, returns true on bindings invalidation.
+    bool Unmap() {
+        std::lock_guard lock{mutex};
+
+        stream_buffer->Unmap(buffer_offset - buffer_offset_base);
+        return std::exchange(invalidated, false);
+    }
+
+    void TickFrame() {
+        ++epoch;
+        while (!pending_destruction.empty()) {
+            if (pending_destruction.front()->GetEpoch() + 1 > epoch) {
+                break;
+            }
+            pending_destruction.pop_front();
+        }
+    }
+
+    /// Write any cached resources overlapping the specified region back to memory
+    void FlushRegion(CacheAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        std::vector<MapInterval> objects = GetMapsInRange(addr, size);
+        std::sort(objects.begin(), objects.end(), [](const MapInterval& a, const MapInterval& b) {
+            return a->GetModificationTick() < b->GetModificationTick();
+        });
+        for (auto& object : objects) {
+            if (object->IsModified() && object->IsRegistered()) {
+                FlushMap(object);
+            }
+        }
+    }
+
+    /// Mark the specified region as being invalidated
+    void InvalidateRegion(CacheAddr addr, u64 size) {
+        std::lock_guard lock{mutex};
+
+        std::vector<MapInterval> objects = GetMapsInRange(addr, size);
+        for (auto& object : objects) {
+            if (object->IsRegistered()) {
+                Unregister(object);
+            }
+        }
+    }
+
+    virtual const TBufferType* GetEmptyBuffer(std::size_t size) = 0;
+
+protected:
+    explicit BufferCache(VideoCore::RasterizerInterface& rasterizer, Core::System& system,
+                         std::unique_ptr<StreamBuffer> stream_buffer)
+        : rasterizer{rasterizer}, system{system}, stream_buffer{std::move(stream_buffer)},
+          stream_buffer_handle{this->stream_buffer->GetHandle()} {}
+
+    ~BufferCache() = default;
+
+    virtual const TBufferType* ToHandle(const TBuffer& storage) = 0;
+
+    virtual void WriteBarrier() = 0;
+
+    virtual TBuffer CreateBlock(CacheAddr cache_addr, std::size_t size) = 0;
+
+    virtual void UploadBlockData(const TBuffer& buffer, std::size_t offset, std::size_t size,
+                                 const u8* data) = 0;
+
+    virtual void DownloadBlockData(const TBuffer& buffer, std::size_t offset, std::size_t size,
+                                   u8* data) = 0;
+
+    virtual void CopyBlock(const TBuffer& src, const TBuffer& dst, std::size_t src_offset,
+                           std::size_t dst_offset, std::size_t size) = 0;
+
+    /// Register an object into the cache
+    void Register(const MapInterval& new_map, bool inherit_written = false) {
+        const CacheAddr cache_ptr = new_map->GetStart();
+        const std::optional<VAddr> cpu_addr =
+            system.GPU().MemoryManager().GpuToCpuAddress(new_map->GetGpuAddress());
+        if (!cache_ptr || !cpu_addr) {
+            LOG_CRITICAL(HW_GPU, "Failed to register buffer with unmapped gpu_address 0x{:016x}",
+                         new_map->GetGpuAddress());
+            return;
+        }
+        const std::size_t size = new_map->GetEnd() - new_map->GetStart();
+        new_map->SetCpuAddress(*cpu_addr);
+        new_map->MarkAsRegistered(true);
+        const IntervalType interval{new_map->GetStart(), new_map->GetEnd()};
+        mapped_addresses.insert({interval, new_map});
+        rasterizer.UpdatePagesCachedCount(*cpu_addr, size, 1);
+        if (inherit_written) {
+            MarkRegionAsWritten(new_map->GetStart(), new_map->GetEnd() - 1);
+            new_map->MarkAsWritten(true);
+        }
+    }
+
+    /// Unregisters an object from the cache
+    void Unregister(MapInterval& map) {
+        const std::size_t size = map->GetEnd() - map->GetStart();
+        rasterizer.UpdatePagesCachedCount(map->GetCpuAddress(), size, -1);
+        map->MarkAsRegistered(false);
+        if (map->IsWritten()) {
+            UnmarkRegionAsWritten(map->GetStart(), map->GetEnd() - 1);
+        }
+        const IntervalType delete_interval{map->GetStart(), map->GetEnd()};
+        mapped_addresses.erase(delete_interval);
+    }
+
+private:
+    MapInterval CreateMap(const CacheAddr start, const CacheAddr end, const GPUVAddr gpu_addr) {
+        return std::make_shared<MapIntervalBase>(start, end, gpu_addr);
+    }
+
+    MapInterval MapAddress(const TBuffer& block, const GPUVAddr gpu_addr,
+                           const CacheAddr cache_addr, const std::size_t size) {
+
+        std::vector<MapInterval> overlaps = GetMapsInRange(cache_addr, size);
+        if (overlaps.empty()) {
+            const CacheAddr cache_addr_end = cache_addr + size;
+            MapInterval new_map = CreateMap(cache_addr, cache_addr_end, gpu_addr);
+            u8* host_ptr = FromCacheAddr(cache_addr);
+            UploadBlockData(block, block->GetOffset(cache_addr), size, host_ptr);
+            Register(new_map);
+            return new_map;
+        }
+
+        const CacheAddr cache_addr_end = cache_addr + size;
+        if (overlaps.size() == 1) {
+            MapInterval& current_map = overlaps[0];
+            if (current_map->IsInside(cache_addr, cache_addr_end)) {
+                return current_map;
+            }
+        }
+        CacheAddr new_start = cache_addr;
+        CacheAddr new_end = cache_addr_end;
+        bool write_inheritance = false;
+        bool modified_inheritance = false;
+        // Calculate new buffer parameters
+        for (auto& overlap : overlaps) {
+            new_start = std::min(overlap->GetStart(), new_start);
+            new_end = std::max(overlap->GetEnd(), new_end);
+            write_inheritance |= overlap->IsWritten();
+            modified_inheritance |= overlap->IsModified();
+        }
+        GPUVAddr new_gpu_addr = gpu_addr + new_start - cache_addr;
+        for (auto& overlap : overlaps) {
+            Unregister(overlap);
+        }
+        UpdateBlock(block, new_start, new_end, overlaps);
+        MapInterval new_map = CreateMap(new_start, new_end, new_gpu_addr);
+        if (modified_inheritance) {
+            new_map->MarkAsModified(true, GetModifiedTicks());
+        }
+        Register(new_map, write_inheritance);
+        return new_map;
+    }
+
+    void UpdateBlock(const TBuffer& block, CacheAddr start, CacheAddr end,
+                     std::vector<MapInterval>& overlaps) {
+        const IntervalType base_interval{start, end};
+        IntervalSet interval_set{};
+        interval_set.add(base_interval);
+        for (auto& overlap : overlaps) {
+            const IntervalType subtract{overlap->GetStart(), overlap->GetEnd()};
+            interval_set.subtract(subtract);
+        }
+        for (auto& interval : interval_set) {
+            std::size_t size = interval.upper() - interval.lower();
+            if (size > 0) {
+                u8* host_ptr = FromCacheAddr(interval.lower());
+                UploadBlockData(block, block->GetOffset(interval.lower()), size, host_ptr);
+            }
+        }
+    }
+
+    std::vector<MapInterval> GetMapsInRange(CacheAddr addr, std::size_t size) {
+        if (size == 0) {
+            return {};
+        }
+
+        std::vector<MapInterval> objects{};
+        const IntervalType interval{addr, addr + size};
+        for (auto& pair : boost::make_iterator_range(mapped_addresses.equal_range(interval))) {
+            objects.push_back(pair.second);
+        }
+
+        return objects;
+    }
+
+    /// Returns a ticks counter used for tracking when cached objects were last modified
+    u64 GetModifiedTicks() {
+        return ++modified_ticks;
+    }
+
+    void FlushMap(MapInterval map) {
+        std::size_t size = map->GetEnd() - map->GetStart();
+        TBuffer block = blocks[map->GetStart() >> block_page_bits];
+        u8* host_ptr = FromCacheAddr(map->GetStart());
+        DownloadBlockData(block, block->GetOffset(map->GetStart()), size, host_ptr);
+        map->MarkAsModified(false, 0);
+    }
+
+    BufferInfo StreamBufferUpload(const void* raw_pointer, std::size_t size,
+                                  std::size_t alignment) {
+        AlignBuffer(alignment);
+        const std::size_t uploaded_offset = buffer_offset;
+        std::memcpy(buffer_ptr, raw_pointer, size);
+
+        buffer_ptr += size;
+        buffer_offset += size;
+        return {&stream_buffer_handle, uploaded_offset};
+    }
+
+    void AlignBuffer(std::size_t alignment) {
+        // Align the offset, not the mapped pointer
+        const std::size_t offset_aligned = Common::AlignUp(buffer_offset, alignment);
+        buffer_ptr += offset_aligned - buffer_offset;
+        buffer_offset = offset_aligned;
+    }
+
+    TBuffer EnlargeBlock(TBuffer buffer) {
+        const std::size_t old_size = buffer->GetSize();
+        const std::size_t new_size = old_size + block_page_size;
+        const CacheAddr cache_addr = buffer->GetCacheAddr();
+        TBuffer new_buffer = CreateBlock(cache_addr, new_size);
+        CopyBlock(buffer, new_buffer, 0, 0, old_size);
+        buffer->SetEpoch(epoch);
+        pending_destruction.push_back(buffer);
+        const CacheAddr cache_addr_end = cache_addr + new_size - 1;
+        u64 page_start = cache_addr >> block_page_bits;
+        const u64 page_end = cache_addr_end >> block_page_bits;
+        while (page_start <= page_end) {
+            blocks[page_start] = new_buffer;
+            ++page_start;
+        }
+        return new_buffer;
+    }
+
+    TBuffer MergeBlocks(TBuffer first, TBuffer second) {
+        const std::size_t size_1 = first->GetSize();
+        const std::size_t size_2 = second->GetSize();
+        const CacheAddr first_addr = first->GetCacheAddr();
+        const CacheAddr second_addr = second->GetCacheAddr();
+        const CacheAddr new_addr = std::min(first_addr, second_addr);
+        const std::size_t new_size = size_1 + size_2;
+        TBuffer new_buffer = CreateBlock(new_addr, new_size);
+        CopyBlock(first, new_buffer, 0, new_buffer->GetOffset(first_addr), size_1);
+        CopyBlock(second, new_buffer, 0, new_buffer->GetOffset(second_addr), size_2);
+        first->SetEpoch(epoch);
+        second->SetEpoch(epoch);
+        pending_destruction.push_back(first);
+        pending_destruction.push_back(second);
+        const CacheAddr cache_addr_end = new_addr + new_size - 1;
+        u64 page_start = new_addr >> block_page_bits;
+        const u64 page_end = cache_addr_end >> block_page_bits;
+        while (page_start <= page_end) {
+            blocks[page_start] = new_buffer;
+            ++page_start;
+        }
+        return new_buffer;
+    }
+
+    TBuffer GetBlock(const CacheAddr cache_addr, const std::size_t size) {
+        TBuffer found{};
+        const CacheAddr cache_addr_end = cache_addr + size - 1;
+        u64 page_start = cache_addr >> block_page_bits;
+        const u64 page_end = cache_addr_end >> block_page_bits;
+        while (page_start <= page_end) {
+            auto it = blocks.find(page_start);
+            if (it == blocks.end()) {
+                if (found) {
+                    found = EnlargeBlock(found);
+                } else {
+                    const CacheAddr start_addr = (page_start << block_page_bits);
+                    found = CreateBlock(start_addr, block_page_size);
+                    blocks[page_start] = found;
+                }
+            } else {
+                if (found) {
+                    if (found == it->second) {
+                        ++page_start;
+                        continue;
+                    }
+                    found = MergeBlocks(found, it->second);
+                } else {
+                    found = it->second;
+                }
+            }
+            ++page_start;
+        }
+        return found;
+    }
+
+    void MarkRegionAsWritten(const CacheAddr start, const CacheAddr end) {
+        u64 page_start = start >> write_page_bit;
+        const u64 page_end = end >> write_page_bit;
+        while (page_start <= page_end) {
+            auto it = written_pages.find(page_start);
+            if (it != written_pages.end()) {
+                it->second = it->second + 1;
+            } else {
+                written_pages[page_start] = 1;
+            }
+            page_start++;
+        }
+    }
+
+    void UnmarkRegionAsWritten(const CacheAddr start, const CacheAddr end) {
+        u64 page_start = start >> write_page_bit;
+        const u64 page_end = end >> write_page_bit;
+        while (page_start <= page_end) {
+            auto it = written_pages.find(page_start);
+            if (it != written_pages.end()) {
+                if (it->second > 1) {
+                    it->second = it->second - 1;
+                } else {
+                    written_pages.erase(it);
+                }
+            }
+            page_start++;
+        }
+    }
+
+    bool IsRegionWritten(const CacheAddr start, const CacheAddr end) const {
+        u64 page_start = start >> write_page_bit;
+        const u64 page_end = end >> write_page_bit;
+        while (page_start <= page_end) {
+            if (written_pages.count(page_start) > 0) {
+                return true;
+            }
+            page_start++;
+        }
+        return false;
+    }
+
+    VideoCore::RasterizerInterface& rasterizer;
+    Core::System& system;
+    std::unique_ptr<StreamBuffer> stream_buffer;
+
+    TBufferType stream_buffer_handle{};
+
+    bool invalidated = false;
+
+    u8* buffer_ptr = nullptr;
+    u64 buffer_offset = 0;
+    u64 buffer_offset_base = 0;
+
+    using IntervalSet = boost::icl::interval_set<CacheAddr>;
+    using IntervalCache = boost::icl::interval_map<CacheAddr, MapInterval>;
+    using IntervalType = typename IntervalCache::interval_type;
+    IntervalCache mapped_addresses{};
+
+    static constexpr u64 write_page_bit{11};
+    std::unordered_map<u64, u32> written_pages{};
+
+    static constexpr u64 block_page_bits{21};
+    static constexpr u64 block_page_size{1 << block_page_bits};
+    std::unordered_map<u64, TBuffer> blocks{};
+
+    std::list<TBuffer> pending_destruction{};
+    u64 epoch{};
+    u64 modified_ticks{};
+
+    std::recursive_mutex mutex;
+};
+
+} // namespace VideoCommon

src/video_core/buffer_cache/map_interval.h

@@ -0,0 +1,89 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/common_types.h"
+#include "video_core/gpu.h"
+
+namespace VideoCommon {
+
+class MapIntervalBase {
+public:
+    MapIntervalBase(const CacheAddr start, const CacheAddr end, const GPUVAddr gpu_addr)
+        : start{start}, end{end}, gpu_addr{gpu_addr} {}
+
+    void SetCpuAddress(VAddr new_cpu_addr) {
+        cpu_addr = new_cpu_addr;
+    }
+
+    VAddr GetCpuAddress() const {
+        return cpu_addr;
+    }
+
+    GPUVAddr GetGpuAddress() const {
+        return gpu_addr;
+    }
+
+    bool IsInside(const CacheAddr other_start, const CacheAddr other_end) const {
+        return (start <= other_start && other_end <= end);
+    }
+
+    bool operator==(const MapIntervalBase& rhs) const {
+        return std::tie(start, end) == std::tie(rhs.start, rhs.end);
+    }
+
+    bool operator!=(const MapIntervalBase& rhs) const {
+        return !operator==(rhs);
+    }
+
+    void MarkAsRegistered(const bool registered) {
+        is_registered = registered;
+    }
+
+    bool IsRegistered() const {
+        return is_registered;
+    }
+
+    CacheAddr GetStart() const {
+        return start;
+    }
+
+    CacheAddr GetEnd() const {
+        return end;
+    }
+
+    void MarkAsModified(const bool is_modified_, const u64 tick) {
+        is_modified = is_modified_;
+        ticks = tick;
+    }
+
+    bool IsModified() const {
+        return is_modified;
+    }
+
+    u64 GetModificationTick() const {
+        return ticks;
+    }
+
+    void MarkAsWritten(const bool is_written_) {
+        is_written = is_written_;
+    }
+
+    bool IsWritten() const {
+        return is_written;
+    }
+
+private:
+    CacheAddr start;
+    CacheAddr end;
+    GPUVAddr gpu_addr;
+    VAddr cpu_addr{};
+    bool is_written{};
+    bool is_modified{};
+    bool is_registered{};
+    u64 ticks{};
+};
+
+} // namespace VideoCommon

src/video_core/dma_pusher.cpp

@@ -22,7 +22,7 @@ void DmaPusher::DispatchCalls() {
     MICROPROFILE_SCOPE(DispatchCalls);
 
     // On entering GPU code, assume all memory may be touched by the ARM core.
-    gpu.Maxwell3D().dirty_flags.OnMemoryWrite();
+    gpu.Maxwell3D().dirty.OnMemoryWrite();
 
     dma_pushbuffer_subindex = 0;
 
@@ -31,6 +31,7 @@ void DmaPusher::DispatchCalls() {
             break;
         }
     }
+    gpu.FlushCommands();
 }
 
 bool DmaPusher::Step() {

src/video_core/engines/fermi_2d.cpp

@@ -10,8 +10,7 @@
 
 namespace Tegra::Engines {
 
-Fermi2D::Fermi2D(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager)
-    : rasterizer{rasterizer}, memory_manager{memory_manager} {}
+Fermi2D::Fermi2D(VideoCore::RasterizerInterface& rasterizer) : rasterizer{rasterizer} {}
 
 void Fermi2D::CallMethod(const GPU::MethodCall& method_call) {
     ASSERT_MSG(method_call.method < Regs::NUM_REGS,

src/video_core/engines/fermi_2d.h

@@ -33,7 +33,7 @@ namespace Tegra::Engines {
 
 class Fermi2D final {
 public:
-    explicit Fermi2D(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager);
+    explicit Fermi2D(VideoCore::RasterizerInterface& rasterizer);
     ~Fermi2D() = default;
 
     /// Write the value to the register identified by method.
@@ -145,7 +145,6 @@ public:
 
 private:
     VideoCore::RasterizerInterface& rasterizer;
-    MemoryManager& memory_manager;
 
     /// Performs the copy from the source surface to the destination surface as configured in the
     /// registers.

src/video_core/engines/kepler_compute.cpp

@@ -37,7 +37,7 @@ void KeplerCompute::CallMethod(const GPU::MethodCall& method_call) {
         const bool is_last_call = method_call.IsLastCall();
         upload_state.ProcessData(method_call.argument, is_last_call);
         if (is_last_call) {
-            system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
+            system.GPU().Maxwell3D().dirty.OnMemoryWrite();
         }
         break;
     }
@@ -50,13 +50,14 @@ void KeplerCompute::CallMethod(const GPU::MethodCall& method_call) {
 }
 
 void KeplerCompute::ProcessLaunch() {
-
     const GPUVAddr launch_desc_loc = regs.launch_desc_loc.Address();
     memory_manager.ReadBlockUnsafe(launch_desc_loc, &launch_description,
                                    LaunchParams::NUM_LAUNCH_PARAMETERS * sizeof(u32));
 
-    const GPUVAddr code_loc = regs.code_loc.Address() + launch_description.program_start;
-    LOG_WARNING(HW_GPU, "Compute Kernel Execute at Address 0x{:016x}, STUBBED", code_loc);
+    const GPUVAddr code_addr = regs.code_loc.Address() + launch_description.program_start;
+    LOG_TRACE(HW_GPU, "Compute invocation launched at address 0x{:016x}", code_addr);
+
+    rasterizer.DispatchCompute(code_addr);
 }
 
 } // namespace Tegra::Engines

src/video_core/engines/kepler_memory.cpp

@@ -15,7 +15,7 @@
 namespace Tegra::Engines {
 
 KeplerMemory::KeplerMemory(Core::System& system, MemoryManager& memory_manager)
-    : system{system}, memory_manager{memory_manager}, upload_state{memory_manager, regs.upload} {}
+    : system{system}, upload_state{memory_manager, regs.upload} {}
 
 KeplerMemory::~KeplerMemory() = default;
 
@@ -34,7 +34,7 @@ void KeplerMemory::CallMethod(const GPU::MethodCall& method_call) {
         const bool is_last_call = method_call.IsLastCall();
         upload_state.ProcessData(method_call.argument, is_last_call);
         if (is_last_call) {
-            system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
+            system.GPU().Maxwell3D().dirty.OnMemoryWrite();
         }
         break;
     }

src/video_core/engines/kepler_memory.h

@@ -65,7 +65,6 @@ public:
 
 private:
     Core::System& system;
-    MemoryManager& memory_manager;
     Upload::State upload_state;
 };
 

src/video_core/engines/maxwell_3d.cpp

@@ -22,6 +22,7 @@ Maxwell3D::Maxwell3D(Core::System& system, VideoCore::RasterizerInterface& raste
                      MemoryManager& memory_manager)
     : system{system}, rasterizer{rasterizer}, memory_manager{memory_manager},
       macro_interpreter{*this}, upload_state{memory_manager, regs.upload} {
+    InitDirtySettings();
     InitializeRegisterDefaults();
 }
 
@@ -69,6 +70,10 @@ void Maxwell3D::InitializeRegisterDefaults() {
     regs.stencil_back_func_mask = 0xFFFFFFFF;
     regs.stencil_back_mask = 0xFFFFFFFF;
 
+    regs.depth_test_func = Regs::ComparisonOp::Always;
+    regs.cull.front_face = Regs::Cull::FrontFace::CounterClockWise;
+    regs.cull.cull_face = Regs::Cull::CullFace::Back;
+
     // TODO(Rodrigo): Most games do not set a point size. I think this is a case of a
     // register carrying a default value. Assume it's OpenGL's default (1).
     regs.point_size = 1.0f;
@@ -86,6 +91,159 @@ void Maxwell3D::InitializeRegisterDefaults() {
     regs.rt_separate_frag_data = 1;
 }
 
+#define DIRTY_REGS_POS(field_name) (offsetof(Maxwell3D::DirtyRegs, field_name))
+
+void Maxwell3D::InitDirtySettings() {
+    const auto set_block = [this](const u32 start, const u32 range, const u8 position) {
+        const auto start_itr = dirty_pointers.begin() + start;
+        const auto end_itr = start_itr + range;
+        std::fill(start_itr, end_itr, position);
+    };
+    dirty.regs.fill(true);
+
+    // Init Render Targets
+    constexpr u32 registers_per_rt = sizeof(regs.rt[0]) / sizeof(u32);
+    constexpr u32 rt_start_reg = MAXWELL3D_REG_INDEX(rt);
+    constexpr u32 rt_end_reg = rt_start_reg + registers_per_rt * 8;
+    u32 rt_dirty_reg = DIRTY_REGS_POS(render_target);
+    for (u32 rt_reg = rt_start_reg; rt_reg < rt_end_reg; rt_reg += registers_per_rt) {
+        set_block(rt_reg, registers_per_rt, rt_dirty_reg);
+        rt_dirty_reg++;
+    }
+    constexpr u32 depth_buffer_flag = DIRTY_REGS_POS(depth_buffer);
+    dirty_pointers[MAXWELL3D_REG_INDEX(zeta_enable)] = depth_buffer_flag;
+    dirty_pointers[MAXWELL3D_REG_INDEX(zeta_width)] = depth_buffer_flag;
+    dirty_pointers[MAXWELL3D_REG_INDEX(zeta_height)] = depth_buffer_flag;
+    constexpr u32 registers_in_zeta = sizeof(regs.zeta) / sizeof(u32);
+    constexpr u32 zeta_reg = MAXWELL3D_REG_INDEX(zeta);
+    set_block(zeta_reg, registers_in_zeta, depth_buffer_flag);
+
+    // Init Vertex Arrays
+    constexpr u32 vertex_array_start = MAXWELL3D_REG_INDEX(vertex_array);
+    constexpr u32 vertex_array_size = sizeof(regs.vertex_array[0]) / sizeof(u32);
+    constexpr u32 vertex_array_end = vertex_array_start + vertex_array_size * Regs::NumVertexArrays;
+    u32 va_reg = DIRTY_REGS_POS(vertex_array);
+    u32 vi_reg = DIRTY_REGS_POS(vertex_instance);
+    for (u32 vertex_reg = vertex_array_start; vertex_reg < vertex_array_end;
+         vertex_reg += vertex_array_size) {
+        set_block(vertex_reg, 3, va_reg);
+        // The divisor concerns vertex array instances
+        dirty_pointers[vertex_reg + 3] = vi_reg;
+        va_reg++;
+        vi_reg++;
+    }
+    constexpr u32 vertex_limit_start = MAXWELL3D_REG_INDEX(vertex_array_limit);
+    constexpr u32 vertex_limit_size = sizeof(regs.vertex_array_limit[0]) / sizeof(u32);
+    constexpr u32 vertex_limit_end = vertex_limit_start + vertex_limit_size * Regs::NumVertexArrays;
+    va_reg = DIRTY_REGS_POS(vertex_array);
+    for (u32 vertex_reg = vertex_limit_start; vertex_reg < vertex_limit_end;
+         vertex_reg += vertex_limit_size) {
+        set_block(vertex_reg, vertex_limit_size, va_reg);
+        va_reg++;
+    }
+    constexpr u32 vertex_instance_start = MAXWELL3D_REG_INDEX(instanced_arrays);
+    constexpr u32 vertex_instance_size =
+        sizeof(regs.instanced_arrays.is_instanced[0]) / sizeof(u32);
+    constexpr u32 vertex_instance_end =
+        vertex_instance_start + vertex_instance_size * Regs::NumVertexArrays;
+    vi_reg = DIRTY_REGS_POS(vertex_instance);
+    for (u32 vertex_reg = vertex_instance_start; vertex_reg < vertex_instance_end;
+         vertex_reg += vertex_instance_size) {
+        set_block(vertex_reg, vertex_instance_size, vi_reg);
+        vi_reg++;
+    }
+    set_block(MAXWELL3D_REG_INDEX(vertex_attrib_format), regs.vertex_attrib_format.size(),
+              DIRTY_REGS_POS(vertex_attrib_format));
+
+    // Init Shaders
+    constexpr u32 shader_registers_count =
+        sizeof(regs.shader_config[0]) * Regs::MaxShaderProgram / sizeof(u32);
+    set_block(MAXWELL3D_REG_INDEX(shader_config[0]), shader_registers_count,
+              DIRTY_REGS_POS(shaders));
+
+    // State
+
+    // Viewport
+    constexpr u32 viewport_dirty_reg = DIRTY_REGS_POS(viewport);
+    constexpr u32 viewport_start = MAXWELL3D_REG_INDEX(viewports);
+    constexpr u32 viewport_size = sizeof(regs.viewports) / sizeof(u32);
+    set_block(viewport_start, viewport_size, viewport_dirty_reg);
+    constexpr u32 view_volume_start = MAXWELL3D_REG_INDEX(view_volume_clip_control);
+    constexpr u32 view_volume_size = sizeof(regs.view_volume_clip_control) / sizeof(u32);
+    set_block(view_volume_start, view_volume_size, viewport_dirty_reg);
+
+    // Viewport transformation
+    constexpr u32 viewport_trans_start = MAXWELL3D_REG_INDEX(viewport_transform);
+    constexpr u32 viewport_trans_size = sizeof(regs.viewport_transform) / sizeof(u32);
+    set_block(viewport_trans_start, viewport_trans_size, DIRTY_REGS_POS(viewport_transform));
+
+    // Cullmode
+    constexpr u32 cull_mode_start = MAXWELL3D_REG_INDEX(cull);
+    constexpr u32 cull_mode_size = sizeof(regs.cull) / sizeof(u32);
+    set_block(cull_mode_start, cull_mode_size, DIRTY_REGS_POS(cull_mode));
+
+    // Screen y control
+    dirty_pointers[MAXWELL3D_REG_INDEX(screen_y_control)] = DIRTY_REGS_POS(screen_y_control);
+
+    // Primitive Restart
+    constexpr u32 primitive_restart_start = MAXWELL3D_REG_INDEX(primitive_restart);
+    constexpr u32 primitive_restart_size = sizeof(regs.primitive_restart) / sizeof(u32);
+    set_block(primitive_restart_start, primitive_restart_size, DIRTY_REGS_POS(primitive_restart));
+
+    // Depth Test
+    constexpr u32 depth_test_dirty_reg = DIRTY_REGS_POS(depth_test);
+    dirty_pointers[MAXWELL3D_REG_INDEX(depth_test_enable)] = depth_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(depth_write_enabled)] = depth_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(depth_test_func)] = depth_test_dirty_reg;
+
+    // Stencil Test
+    constexpr u32 stencil_test_dirty_reg = DIRTY_REGS_POS(stencil_test);
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_enable)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_func_func)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_func_ref)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_func_mask)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_op_fail)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_op_zfail)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_op_zpass)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_front_mask)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_two_side_enable)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_func_func)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_func_ref)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_func_mask)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_op_fail)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_op_zfail)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_op_zpass)] = stencil_test_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(stencil_back_mask)] = stencil_test_dirty_reg;
+
+    // Color Mask
+    constexpr u32 color_mask_dirty_reg = DIRTY_REGS_POS(color_mask);
+    dirty_pointers[MAXWELL3D_REG_INDEX(color_mask_common)] = color_mask_dirty_reg;
+    set_block(MAXWELL3D_REG_INDEX(color_mask), sizeof(regs.color_mask) / sizeof(u32),
+              color_mask_dirty_reg);
+    // Blend State
+    constexpr u32 blend_state_dirty_reg = DIRTY_REGS_POS(blend_state);
+    set_block(MAXWELL3D_REG_INDEX(blend_color), sizeof(regs.blend_color) / sizeof(u32),
+              blend_state_dirty_reg);
+    dirty_pointers[MAXWELL3D_REG_INDEX(independent_blend_enable)] = blend_state_dirty_reg;
+    set_block(MAXWELL3D_REG_INDEX(blend), sizeof(regs.blend) / sizeof(u32), blend_state_dirty_reg);
+    set_block(MAXWELL3D_REG_INDEX(independent_blend), sizeof(regs.independent_blend) / sizeof(u32),
+              blend_state_dirty_reg);
+
+    // Scissor State
+    constexpr u32 scissor_test_dirty_reg = DIRTY_REGS_POS(scissor_test);
+    set_block(MAXWELL3D_REG_INDEX(scissor_test), sizeof(regs.scissor_test) / sizeof(u32),
+              scissor_test_dirty_reg);
+
+    // Polygon Offset
+    constexpr u32 polygon_offset_dirty_reg = DIRTY_REGS_POS(polygon_offset);
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_fill_enable)] = polygon_offset_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_line_enable)] = polygon_offset_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_point_enable)] = polygon_offset_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_units)] = polygon_offset_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_factor)] = polygon_offset_dirty_reg;
+    dirty_pointers[MAXWELL3D_REG_INDEX(polygon_offset_clamp)] = polygon_offset_dirty_reg;
+}
+
 void Maxwell3D::CallMacroMethod(u32 method, std::vector<u32> parameters) {
     // Reset the current macro.
     executing_macro = 0;
@@ -108,6 +266,14 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
 
     const u32 method = method_call.method;
 
+    if (method == cb_data_state.current) {
+        regs.reg_array[method] = method_call.argument;
+        ProcessCBData(method_call.argument);
+        return;
+    } else if (cb_data_state.current != null_cb_data) {
+        FinishCBData();
+    }
+
     // It is an error to write to a register other than the current macro's ARG register before it
     // has finished execution.
     if (executing_macro != 0) {
@@ -143,49 +309,19 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
 
     if (regs.reg_array[method] != method_call.argument) {
         regs.reg_array[method] = method_call.argument;
-        // Color buffers
-        constexpr u32 first_rt_reg = MAXWELL3D_REG_INDEX(rt);
-        constexpr u32 registers_per_rt = sizeof(regs.rt[0]) / sizeof(u32);
-        if (method >= first_rt_reg &&
-            method < first_rt_reg + registers_per_rt * Regs::NumRenderTargets) {
-            const std::size_t rt_index = (method - first_rt_reg) / registers_per_rt;
-            dirty_flags.color_buffer.set(rt_index);
-        }
-
-        // Zeta buffer
-        constexpr u32 registers_in_zeta = sizeof(regs.zeta) / sizeof(u32);
-        if (method == MAXWELL3D_REG_INDEX(zeta_enable) ||
-            method == MAXWELL3D_REG_INDEX(zeta_width) ||
-            method == MAXWELL3D_REG_INDEX(zeta_height) ||
-            (method >= MAXWELL3D_REG_INDEX(zeta) &&
-             method < MAXWELL3D_REG_INDEX(zeta) + registers_in_zeta)) {
-            dirty_flags.zeta_buffer = true;
-        }
-
-        // Shader
-        constexpr u32 shader_registers_count =
-            sizeof(regs.shader_config[0]) * Regs::MaxShaderProgram / sizeof(u32);
-        if (method >= MAXWELL3D_REG_INDEX(shader_config[0]) &&
-            method < MAXWELL3D_REG_INDEX(shader_config[0]) + shader_registers_count) {
-            dirty_flags.shaders = true;
-        }
-
-        // Vertex format
-        if (method >= MAXWELL3D_REG_INDEX(vertex_attrib_format) &&
-            method < MAXWELL3D_REG_INDEX(vertex_attrib_format) + regs.vertex_attrib_format.size()) {
-            dirty_flags.vertex_attrib_format = true;
-        }
-
-        // Vertex buffer
-        if (method >= MAXWELL3D_REG_INDEX(vertex_array) &&
-            method < MAXWELL3D_REG_INDEX(vertex_array) + 4 * Regs::NumVertexArrays) {
-            dirty_flags.vertex_array.set((method - MAXWELL3D_REG_INDEX(vertex_array)) >> 2);
-        } else if (method >= MAXWELL3D_REG_INDEX(vertex_array_limit) &&
-                   method < MAXWELL3D_REG_INDEX(vertex_array_limit) + 2 * Regs::NumVertexArrays) {
-            dirty_flags.vertex_array.set((method - MAXWELL3D_REG_INDEX(vertex_array_limit)) >> 1);
-        } else if (method >= MAXWELL3D_REG_INDEX(instanced_arrays) &&
-                   method < MAXWELL3D_REG_INDEX(instanced_arrays) + Regs::NumVertexArrays) {
-            dirty_flags.vertex_array.set(method - MAXWELL3D_REG_INDEX(instanced_arrays));
+        const std::size_t dirty_reg = dirty_pointers[method];
+        if (dirty_reg) {
+            dirty.regs[dirty_reg] = true;
+            if (dirty_reg >= DIRTY_REGS_POS(vertex_array) &&
+                dirty_reg < DIRTY_REGS_POS(vertex_array_buffers)) {
+                dirty.vertex_array_buffers = true;
+            } else if (dirty_reg >= DIRTY_REGS_POS(vertex_instance) &&
+                       dirty_reg < DIRTY_REGS_POS(vertex_instances)) {
+                dirty.vertex_instances = true;
+            } else if (dirty_reg >= DIRTY_REGS_POS(render_target) &&
+                       dirty_reg < DIRTY_REGS_POS(render_settings)) {
+                dirty.render_settings = true;
+            }
         }
     }
 
@@ -214,7 +350,7 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[13]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[14]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[15]): {
-        ProcessCBData(method_call.argument);
+        StartCBData(method);
         break;
     }
     case MAXWELL3D_REG_INDEX(cb_bind[0].raw_config): {
@@ -249,6 +385,10 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
         ProcessQueryGet();
         break;
     }
+    case MAXWELL3D_REG_INDEX(condition.mode): {
+        ProcessQueryCondition();
+        break;
+    }
     case MAXWELL3D_REG_INDEX(sync_info): {
         ProcessSyncPoint();
         break;
@@ -261,7 +401,7 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
         const bool is_last_call = method_call.IsLastCall();
         upload_state.ProcessData(method_call.argument, is_last_call);
         if (is_last_call) {
-            dirty_flags.OnMemoryWrite();
+            dirty.OnMemoryWrite();
         }
         break;
     }
@@ -302,6 +442,7 @@ void Maxwell3D::ProcessQueryGet() {
         result = regs.query.query_sequence;
         break;
     default:
+        result = 1;
         UNIMPLEMENTED_MSG("Unimplemented query select type {}",
                           static_cast<u32>(regs.query.query_get.select.Value()));
     }
@@ -333,7 +474,6 @@ void Maxwell3D::ProcessQueryGet() {
             query_result.timestamp = system.CoreTiming().GetTicks();
             memory_manager.WriteBlock(sequence_address, &query_result, sizeof(query_result));
         }
-        dirty_flags.OnMemoryWrite();
         break;
     }
     default:
@@ -342,12 +482,52 @@ void Maxwell3D::ProcessQueryGet() {
     }
 }
 
+void Maxwell3D::ProcessQueryCondition() {
+    const GPUVAddr condition_address{regs.condition.Address()};
+    switch (regs.condition.mode) {
+    case Regs::ConditionMode::Always: {
+        execute_on = true;
+        break;
+    }
+    case Regs::ConditionMode::Never: {
+        execute_on = false;
+        break;
+    }
+    case Regs::ConditionMode::ResNonZero: {
+        Regs::QueryCompare cmp;
+        memory_manager.ReadBlockUnsafe(condition_address, &cmp, sizeof(cmp));
+        execute_on = cmp.initial_sequence != 0U && cmp.initial_mode != 0U;
+        break;
+    }
+    case Regs::ConditionMode::Equal: {
+        Regs::QueryCompare cmp;
+        memory_manager.ReadBlockUnsafe(condition_address, &cmp, sizeof(cmp));
+        execute_on =
+            cmp.initial_sequence == cmp.current_sequence && cmp.initial_mode == cmp.current_mode;
+        break;
+    }
+    case Regs::ConditionMode::NotEqual: {
+        Regs::QueryCompare cmp;
+        memory_manager.ReadBlockUnsafe(condition_address, &cmp, sizeof(cmp));
+        execute_on =
+            cmp.initial_sequence != cmp.current_sequence || cmp.initial_mode != cmp.current_mode;
+        break;
+    }
+    default: {
+        UNIMPLEMENTED_MSG("Uninplemented Condition Mode!");
+        execute_on = true;
+        break;
+    }
+    }
+}
+
 void Maxwell3D::ProcessSyncPoint() {
     const u32 sync_point = regs.sync_info.sync_point.Value();
     const u32 increment = regs.sync_info.increment.Value();
-    const u32 cache_flush = regs.sync_info.unknown.Value();
-    LOG_DEBUG(HW_GPU, "Syncpoint set {}, increment: {}, unk: {}", sync_point, increment,
-              cache_flush);
+    [[maybe_unused]] const u32 cache_flush = regs.sync_info.unknown.Value();
+    if (increment) {
+        system.GPU().IncrementSyncPoint(sync_point);
+    }
 }
 
 void Maxwell3D::DrawArrays() {
@@ -405,23 +585,39 @@ void Maxwell3D::ProcessCBBind(Regs::ShaderStage stage) {
 }
 
 void Maxwell3D::ProcessCBData(u32 value) {
+    const u32 id = cb_data_state.id;
+    cb_data_state.buffer[id][cb_data_state.counter] = value;
+    // Increment the current buffer position.
+    regs.const_buffer.cb_pos = regs.const_buffer.cb_pos + 4;
+    cb_data_state.counter++;
+}
+
+void Maxwell3D::StartCBData(u32 method) {
+    constexpr u32 first_cb_data = MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]);
+    cb_data_state.start_pos = regs.const_buffer.cb_pos;
+    cb_data_state.id = method - first_cb_data;
+    cb_data_state.current = method;
+    cb_data_state.counter = 0;
+    ProcessCBData(regs.const_buffer.cb_data[cb_data_state.id]);
+}
+
+void Maxwell3D::FinishCBData() {
     // Write the input value to the current const buffer at the current position.
     const GPUVAddr buffer_address = regs.const_buffer.BufferAddress();
     ASSERT(buffer_address != 0);
 
     // Don't allow writing past the end of the buffer.
-    ASSERT(regs.const_buffer.cb_pos + sizeof(u32) <= regs.const_buffer.cb_size);
+    ASSERT(regs.const_buffer.cb_pos <= regs.const_buffer.cb_size);
 
-    const GPUVAddr address{buffer_address + regs.const_buffer.cb_pos};
+    const GPUVAddr address{buffer_address + cb_data_state.start_pos};
+    const std::size_t size = regs.const_buffer.cb_pos - cb_data_state.start_pos;
 
-    u8* ptr{memory_manager.GetPointer(address)};
-    rasterizer.InvalidateRegion(ToCacheAddr(ptr), sizeof(u32));
-    memory_manager.Write<u32>(address, value);
+    const u32 id = cb_data_state.id;
+    memory_manager.WriteBlock(address, cb_data_state.buffer[id].data(), size);
+    dirty.OnMemoryWrite();
 
-    dirty_flags.OnMemoryWrite();
-
-    // Increment the current buffer position.
-    regs.const_buffer.cb_pos = regs.const_buffer.cb_pos + 4;
+    cb_data_state.id = null_cb_data;
+    cb_data_state.current = null_cb_data;
 }
 
 Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
@@ -430,10 +626,10 @@ Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
     Texture::TICEntry tic_entry;
     memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
 
-    const auto r_type{tic_entry.r_type.Value()};
-    const auto g_type{tic_entry.g_type.Value()};
-    const auto b_type{tic_entry.b_type.Value()};
-    const auto a_type{tic_entry.a_type.Value()};
+    [[maybe_unused]] const auto r_type{tic_entry.r_type.Value()};
+    [[maybe_unused]] const auto g_type{tic_entry.g_type.Value()};
+    [[maybe_unused]] const auto b_type{tic_entry.b_type.Value()};
+    [[maybe_unused]] const auto a_type{tic_entry.a_type.Value()};
 
     // TODO(Subv): Different data types for separate components are not supported
     DEBUG_ASSERT(r_type == g_type && r_type == b_type && r_type == a_type);

src/video_core/engines/maxwell_3d.h

@@ -90,6 +90,20 @@ public:
 
         enum class QuerySelect : u32 {
             Zero = 0,
+            TimeElapsed = 2,
+            TransformFeedbackPrimitivesGenerated = 11,
+            PrimitivesGenerated = 18,
+            SamplesPassed = 21,
+            TransformFeedbackUnknown = 26,
+        };
+
+        struct QueryCompare {
+            u32 initial_sequence;
+            u32 initial_mode;
+            u32 unknown1;
+            u32 unknown2;
+            u32 current_sequence;
+            u32 current_mode;
         };
 
         enum class QuerySyncCondition : u32 {
@@ -97,6 +111,14 @@ public:
             GreaterThan = 1,
         };
 
+        enum class ConditionMode : u32 {
+            Never = 0,
+            Always = 1,
+            ResNonZero = 2,
+            Equal = 3,
+            NotEqual = 4,
+        };
+
         enum class ShaderProgram : u32 {
             VertexA = 0,
             VertexB = 1,
@@ -815,7 +837,18 @@ public:
                     BitField<4, 1, u32> alpha_to_one;
                 } multisample_control;
 
-                INSERT_PADDING_WORDS(0x7);
+                INSERT_PADDING_WORDS(0x4);
+
+                struct {
+                    u32 address_high;
+                    u32 address_low;
+                    ConditionMode mode;
+
+                    GPUVAddr Address() const {
+                        return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
+                                                     address_low);
+                    }
+                } condition;
 
                 struct {
                     u32 tsc_address_high;
@@ -1124,23 +1157,77 @@ public:
 
     State state{};
 
-    struct DirtyFlags {
-        std::bitset<8> color_buffer{0xFF};
-        std::bitset<32> vertex_array{0xFFFFFFFF};
+    struct DirtyRegs {
+        static constexpr std::size_t NUM_REGS = 256;
+        union {
+            struct {
+                bool null_dirty;
 
-        bool vertex_attrib_format = true;
-        bool zeta_buffer = true;
-        bool shaders = true;
+                // Vertex Attributes
+                bool vertex_attrib_format;
+
+                // Vertex Arrays
+                std::array<bool, 32> vertex_array;
+
+                bool vertex_array_buffers;
+
+                // Vertex Instances
+                std::array<bool, 32> vertex_instance;
+
+                bool vertex_instances;
+
+                // Render Targets
+                std::array<bool, 8> render_target;
+                bool depth_buffer;
+
+                bool render_settings;
+
+                // Shaders
+                bool shaders;
+
+                // Rasterizer State
+                bool viewport;
+                bool clip_coefficient;
+                bool cull_mode;
+                bool primitive_restart;
+                bool depth_test;
+                bool stencil_test;
+                bool blend_state;
+                bool scissor_test;
+                bool transform_feedback;
+                bool color_mask;
+                bool polygon_offset;
+
+                // Complementary
+                bool viewport_transform;
+                bool screen_y_control;
+
+                bool memory_general;
+            };
+            std::array<bool, NUM_REGS> regs;
+        };
+
+        void ResetVertexArrays() {
+            vertex_array.fill(true);
+            vertex_array_buffers = true;
+        }
+
+        void ResetRenderTargets() {
+            depth_buffer = true;
+            render_target.fill(true);
+            render_settings = true;
+        }
 
         void OnMemoryWrite() {
-            zeta_buffer = true;
             shaders = true;
-            color_buffer.set();
-            vertex_array.set();
+            memory_general = true;
+            ResetRenderTargets();
+            ResetVertexArrays();
         }
-    };
 
-    DirtyFlags dirty_flags;
+    } dirty{};
+
+    std::array<u8, Regs::NUM_REGS> dirty_pointers{};
 
     /// Reads a register value located at the input method address
     u32 GetRegisterValue(u32 method) const;
@@ -1169,6 +1256,10 @@ public:
         return macro_memory;
     }
 
+    bool ShouldExecute() const {
+        return execute_on;
+    }
+
 private:
     void InitializeRegisterDefaults();
 
@@ -1192,14 +1283,27 @@ private:
     /// Interpreter for the macro codes uploaded to the GPU.
     MacroInterpreter macro_interpreter;
 
+    static constexpr u32 null_cb_data = 0xFFFFFFFF;
+    struct {
+        std::array<std::array<u32, 0x4000>, 16> buffer;
+        u32 current{null_cb_data};
+        u32 id{null_cb_data};
+        u32 start_pos{};
+        u32 counter{};
+    } cb_data_state;
+
     Upload::State upload_state;
 
+    bool execute_on{true};
+
     /// Retrieves information about a specific TIC entry from the TIC buffer.
     Texture::TICEntry GetTICEntry(u32 tic_index) const;
 
     /// Retrieves information about a specific TSC entry from the TSC buffer.
     Texture::TSCEntry GetTSCEntry(u32 tsc_index) const;
 
+    void InitDirtySettings();
+
     /**
      * Call a macro on this engine.
      * @param method Method to call
@@ -1219,11 +1323,16 @@ private:
     /// Handles a write to the QUERY_GET register.
     void ProcessQueryGet();
 
+    // Handles Conditional Rendering
+    void ProcessQueryCondition();
+
     /// Handles writes to syncing register.
     void ProcessSyncPoint();
 
     /// Handles a write to the CB_DATA[i] register.
+    void StartCBData(u32 method);
     void ProcessCBData(u32 value);
+    void FinishCBData();
 
     /// Handles a write to the CB_BIND register.
     void ProcessCBBind(Regs::ShaderStage stage);
@@ -1290,6 +1399,7 @@ ASSERT_REG_POSITION(clip_distance_enabled, 0x544);
 ASSERT_REG_POSITION(point_size, 0x546);
 ASSERT_REG_POSITION(zeta_enable, 0x54E);
 ASSERT_REG_POSITION(multisample_control, 0x54F);
+ASSERT_REG_POSITION(condition, 0x554);
 ASSERT_REG_POSITION(tsc, 0x557);
 ASSERT_REG_POSITION(polygon_offset_factor, 0x55b);
 ASSERT_REG_POSITION(tic, 0x55D);

src/video_core/engines/maxwell_dma.cpp

@@ -5,18 +5,17 @@
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "core/core.h"
+#include "core/settings.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/memory_manager.h"
-#include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 #include "video_core/textures/decoders.h"
 
 namespace Tegra::Engines {
 
-MaxwellDMA::MaxwellDMA(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
-                       MemoryManager& memory_manager)
-    : system{system}, rasterizer{rasterizer}, memory_manager{memory_manager} {}
+MaxwellDMA::MaxwellDMA(Core::System& system, MemoryManager& memory_manager)
+    : system{system}, memory_manager{memory_manager} {}
 
 void MaxwellDMA::CallMethod(const GPU::MethodCall& method_call) {
     ASSERT_MSG(method_call.method < Regs::NUM_REGS,
@@ -38,7 +37,7 @@ void MaxwellDMA::CallMethod(const GPU::MethodCall& method_call) {
 }
 
 void MaxwellDMA::HandleCopy() {
-    LOG_WARNING(HW_GPU, "Requested a DMA copy");
+    LOG_TRACE(HW_GPU, "Requested a DMA copy");
 
     const GPUVAddr source = regs.src_address.Address();
     const GPUVAddr dest = regs.dst_address.Address();
@@ -58,7 +57,7 @@ void MaxwellDMA::HandleCopy() {
     }
 
     // All copies here update the main memory, so mark all rasterizer states as invalid.
-    system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
+    system.GPU().Maxwell3D().dirty.OnMemoryWrite();
 
     if (regs.exec.is_dst_linear && regs.exec.is_src_linear) {
         // When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D
@@ -84,13 +83,17 @@ void MaxwellDMA::HandleCopy() {
     ASSERT(regs.exec.enable_2d == 1);
 
     if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
-        ASSERT(regs.src_params.size_z == 1);
+        ASSERT(regs.src_params.BlockDepth() == 0);
         // If the input is tiled and the output is linear, deswizzle the input and copy it over.
-        const u32 src_bytes_per_pixel = regs.src_pitch / regs.src_params.size_x;
+        const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
         const std::size_t src_size = Texture::CalculateSize(
-            true, src_bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y,
+            true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y,
             regs.src_params.size_z, regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
 
+        const std::size_t src_layer_size = Texture::CalculateSize(
+            true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y, 1,
+            regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
+
         const std::size_t dst_size = regs.dst_pitch * regs.y_count;
 
         if (read_buffer.size() < src_size) {
@@ -104,23 +107,23 @@ void MaxwellDMA::HandleCopy() {
         memory_manager.ReadBlock(source, read_buffer.data(), src_size);
         memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
 
-        Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch,
-                                  regs.src_params.size_x, src_bytes_per_pixel, read_buffer.data(),
-                                  write_buffer.data(), regs.src_params.BlockHeight(),
-                                  regs.src_params.pos_x, regs.src_params.pos_y);
+        Texture::UnswizzleSubrect(
+            regs.x_count, regs.y_count, regs.dst_pitch, regs.src_params.size_x, bytes_per_pixel,
+            read_buffer.data() + src_layer_size * regs.src_params.pos_z, write_buffer.data(),
+            regs.src_params.BlockHeight(), regs.src_params.pos_x, regs.src_params.pos_y);
 
         memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
     } else {
         ASSERT(regs.dst_params.BlockDepth() == 0);
 
-        const u32 src_bytes_per_pixel = regs.src_pitch / regs.x_count;
+        const u32 bytes_per_pixel = regs.src_pitch / regs.x_count;
 
         const std::size_t dst_size = Texture::CalculateSize(
-            true, src_bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y,
+            true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y,
             regs.dst_params.size_z, regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
 
         const std::size_t dst_layer_size = Texture::CalculateSize(
-            true, src_bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y, 1,
+            true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y, 1,
             regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
 
         const std::size_t src_size = regs.src_pitch * regs.y_count;
@@ -133,14 +136,19 @@ void MaxwellDMA::HandleCopy() {
             write_buffer.resize(dst_size);
         }
 
-        memory_manager.ReadBlock(source, read_buffer.data(), src_size);
-        memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+        if (Settings::values.use_accurate_gpu_emulation) {
+            memory_manager.ReadBlock(source, read_buffer.data(), src_size);
+            memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+        } else {
+            memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
+            memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
+        }
 
         // If the input is linear and the output is tiled, swizzle the input and copy it over.
-        Texture::SwizzleSubrect(regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x,
-                                src_bytes_per_pixel,
-                                write_buffer.data() + dst_layer_size * regs.dst_params.pos_z,
-                                read_buffer.data(), regs.dst_params.BlockHeight());
+        Texture::SwizzleSubrect(
+            regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x, bytes_per_pixel,
+            write_buffer.data() + dst_layer_size * regs.dst_params.pos_z, read_buffer.data(),
+            regs.dst_params.BlockHeight(), regs.dst_params.pos_x, regs.dst_params.pos_y);
 
         memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
     }

src/video_core/engines/maxwell_dma.h

@@ -20,10 +20,6 @@ namespace Tegra {
 class MemoryManager;
 }
 
-namespace VideoCore {
-class RasterizerInterface;
-}
-
 namespace Tegra::Engines {
 
 /**
@@ -33,8 +29,7 @@ namespace Tegra::Engines {
 
 class MaxwellDMA final {
 public:
-    explicit MaxwellDMA(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
-                        MemoryManager& memory_manager);
+    explicit MaxwellDMA(Core::System& system, MemoryManager& memory_manager);
     ~MaxwellDMA() = default;
 
     /// Write the value to the register identified by method.
@@ -180,8 +175,6 @@ public:
 private:
     Core::System& system;
 
-    VideoCore::RasterizerInterface& rasterizer;
-
     MemoryManager& memory_manager;
 
     std::vector<u8> read_buffer;

src/video_core/engines/shader_bytecode.h

@@ -538,6 +538,12 @@ enum class PhysicalAttributeDirection : u64 {
     Output = 1,
 };
 
+enum class VoteOperation : u64 {
+    All = 0, // allThreadsNV
+    Any = 1, // anyThreadNV
+    Eq = 2,  // allThreadsEqualNV
+};
+
 union Instruction {
     Instruction& operator=(const Instruction& instr) {
         value = instr.value;
@@ -559,6 +565,18 @@ union Instruction {
     BitField<39, 8, Register> gpr39;
     BitField<48, 16, u64> opcode;
 
+    union {
+        BitField<8, 5, ConditionCode> cc;
+        BitField<13, 1, u64> trigger;
+    } nop;
+
+    union {
+        BitField<48, 2, VoteOperation> operation;
+        BitField<45, 3, u64> dest_pred;
+        BitField<39, 3, u64> value;
+        BitField<42, 1, u64> negate_value;
+    } vote;
+
     union {
         BitField<8, 8, Register> gpr;
         BitField<20, 24, s64> offset;
@@ -1482,6 +1500,7 @@ public:
         SYNC,
         BRK,
         DEPBAR,
+        VOTE,
         BFE_C,
         BFE_R,
         BFE_IMM,
@@ -1514,6 +1533,7 @@ public:
         TMML,   // Texture Mip Map Level
         SUST,   // Surface Store
         EXIT,
+        NOP,
         IPA,
         OUT_R, // Emit vertex/primitive
         ISBERD,
@@ -1643,6 +1663,7 @@ public:
         Hfma2,
         Flow,
         Synch,
+        Warp,
         Memory,
         Texture,
         Image,
@@ -1769,6 +1790,7 @@ private:
             INST("111000110100---", Id::BRK, Type::Flow, "BRK"),
             INST("111000110000----", Id::EXIT, Type::Flow, "EXIT"),
             INST("1111000011110---", Id::DEPBAR, Type::Synch, "DEPBAR"),
+            INST("0101000011011---", Id::VOTE, Type::Warp, "VOTE"),
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111101001---", Id::LD_S, Type::Memory, "LD_S"),
             INST("1110111101000---", Id::LD_L, Type::Memory, "LD_L"),
@@ -1793,6 +1815,7 @@ private:
             INST("110111110110----", Id::TMML_B, Type::Texture, "TMML_B"),
             INST("1101111101011---", Id::TMML, Type::Texture, "TMML"),
             INST("11101011001-----", Id::SUST, Type::Image, "SUST"),
+            INST("0101000010110---", Id::NOP, Type::Trivial, "NOP"),
             INST("11100000--------", Id::IPA, Type::Trivial, "IPA"),
             INST("1111101111100---", Id::OUT_R, Type::Trivial, "OUT_R"),
             INST("1110111111010---", Id::ISBERD, Type::Trivial, "ISBERD"),

src/video_core/gpu.cpp

@@ -29,14 +29,15 @@ u32 FramebufferConfig::BytesPerPixel(PixelFormat format) {
     UNREACHABLE();
 }
 
-GPU::GPU(Core::System& system, VideoCore::RendererBase& renderer) : renderer{renderer} {
+GPU::GPU(Core::System& system, VideoCore::RendererBase& renderer, bool is_async)
+    : system{system}, renderer{renderer}, is_async{is_async} {
     auto& rasterizer{renderer.Rasterizer()};
     memory_manager = std::make_unique<Tegra::MemoryManager>(system, rasterizer);
     dma_pusher = std::make_unique<Tegra::DmaPusher>(*this);
     maxwell_3d = std::make_unique<Engines::Maxwell3D>(system, rasterizer, *memory_manager);
-    fermi_2d = std::make_unique<Engines::Fermi2D>(rasterizer, *memory_manager);
+    fermi_2d = std::make_unique<Engines::Fermi2D>(rasterizer);
     kepler_compute = std::make_unique<Engines::KeplerCompute>(system, rasterizer, *memory_manager);
-    maxwell_dma = std::make_unique<Engines::MaxwellDMA>(system, rasterizer, *memory_manager);
+    maxwell_dma = std::make_unique<Engines::MaxwellDMA>(system, *memory_manager);
     kepler_memory = std::make_unique<Engines::KeplerMemory>(system, *memory_manager);
 }
 
@@ -50,6 +51,14 @@ const Engines::Maxwell3D& GPU::Maxwell3D() const {
     return *maxwell_3d;
 }
 
+Engines::KeplerCompute& GPU::KeplerCompute() {
+    return *kepler_compute;
+}
+
+const Engines::KeplerCompute& GPU::KeplerCompute() const {
+    return *kepler_compute;
+}
+
 MemoryManager& GPU::MemoryManager() {
     return *memory_manager;
 }
@@ -66,6 +75,55 @@ const DmaPusher& GPU::DmaPusher() const {
     return *dma_pusher;
 }
 
+void GPU::IncrementSyncPoint(const u32 syncpoint_id) {
+    syncpoints[syncpoint_id]++;
+    std::lock_guard lock{sync_mutex};
+    if (!syncpt_interrupts[syncpoint_id].empty()) {
+        u32 value = syncpoints[syncpoint_id].load();
+        auto it = syncpt_interrupts[syncpoint_id].begin();
+        while (it != syncpt_interrupts[syncpoint_id].end()) {
+            if (value >= *it) {
+                TriggerCpuInterrupt(syncpoint_id, *it);
+                it = syncpt_interrupts[syncpoint_id].erase(it);
+                continue;
+            }
+            it++;
+        }
+    }
+}
+
+u32 GPU::GetSyncpointValue(const u32 syncpoint_id) const {
+    return syncpoints[syncpoint_id].load();
+}
+
+void GPU::RegisterSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
+    auto& interrupt = syncpt_interrupts[syncpoint_id];
+    bool contains = std::any_of(interrupt.begin(), interrupt.end(),
+                                [value](u32 in_value) { return in_value == value; });
+    if (contains) {
+        return;
+    }
+    syncpt_interrupts[syncpoint_id].emplace_back(value);
+}
+
+bool GPU::CancelSyncptInterrupt(const u32 syncpoint_id, const u32 value) {
+    std::lock_guard lock{sync_mutex};
+    auto& interrupt = syncpt_interrupts[syncpoint_id];
+    const auto iter =
+        std::find_if(interrupt.begin(), interrupt.end(),
+                     [value](u32 interrupt_value) { return value == interrupt_value; });
+
+    if (iter == interrupt.end()) {
+        return false;
+    }
+    interrupt.erase(iter);
+    return true;
+}
+
+void GPU::FlushCommands() {
+    renderer.Rasterizer().FlushCommands();
+}
+
 u32 RenderTargetBytesPerPixel(RenderTargetFormat format) {
     ASSERT(format != RenderTargetFormat::NONE);
 
@@ -143,12 +201,12 @@ enum class BufferMethods {
     NotifyIntr = 0x8,
     WrcacheFlush = 0x9,
     Unk28 = 0xA,
-    Unk2c = 0xB,
+    UnkCacheFlush = 0xB,
     RefCnt = 0x14,
     SemaphoreAcquire = 0x1A,
     SemaphoreRelease = 0x1B,
-    Unk70 = 0x1C,
-    Unk74 = 0x1D,
+    FenceValue = 0x1C,
+    FenceAction = 0x1D,
     Unk78 = 0x1E,
     Unk7c = 0x1F,
     Yield = 0x20,
@@ -194,6 +252,10 @@ void GPU::CallPullerMethod(const MethodCall& method_call) {
     case BufferMethods::SemaphoreAddressLow:
     case BufferMethods::SemaphoreSequence:
     case BufferMethods::RefCnt:
+    case BufferMethods::UnkCacheFlush:
+    case BufferMethods::WrcacheFlush:
+    case BufferMethods::FenceValue:
+    case BufferMethods::FenceAction:
         break;
     case BufferMethods::SemaphoreTrigger: {
         ProcessSemaphoreTriggerMethod();
@@ -204,21 +266,11 @@ void GPU::CallPullerMethod(const MethodCall& method_call) {
         LOG_ERROR(HW_GPU, "Special puller engine method NotifyIntr not implemented");
         break;
     }
-    case BufferMethods::WrcacheFlush: {
-        // TODO(Kmather73): Research and implement this method.
-        LOG_ERROR(HW_GPU, "Special puller engine method WrcacheFlush not implemented");
-        break;
-    }
     case BufferMethods::Unk28: {
         // TODO(Kmather73): Research and implement this method.
         LOG_ERROR(HW_GPU, "Special puller engine method Unk28 not implemented");
         break;
     }
-    case BufferMethods::Unk2c: {
-        // TODO(Kmather73): Research and implement this method.
-        LOG_ERROR(HW_GPU, "Special puller engine method Unk2c not implemented");
-        break;
-    }
     case BufferMethods::SemaphoreAcquire: {
         ProcessSemaphoreAcquire();
         break;

src/video_core/gpu.h

@@ -5,8 +5,12 @@
 #pragma once
 
 #include <array>
+#include <atomic>
+#include <list>
 #include <memory>
+#include <mutex>
 #include "common/common_types.h"
+#include "core/hle/service/nvdrv/nvdata.h"
 #include "core/hle/service/nvflinger/buffer_queue.h"
 #include "video_core/dma_pusher.h"
 
@@ -15,6 +19,10 @@ inline CacheAddr ToCacheAddr(const void* host_ptr) {
     return reinterpret_cast<CacheAddr>(host_ptr);
 }
 
+inline u8* FromCacheAddr(CacheAddr cache_addr) {
+    return reinterpret_cast<u8*>(cache_addr);
+}
+
 namespace Core {
 class System;
 }
@@ -127,7 +135,7 @@ class MemoryManager;
 
 class GPU {
 public:
-    explicit GPU(Core::System& system, VideoCore::RendererBase& renderer);
+    explicit GPU(Core::System& system, VideoCore::RendererBase& renderer, bool is_async);
 
     virtual ~GPU();
 
@@ -149,12 +157,20 @@ public:
     /// Calls a GPU method.
     void CallMethod(const MethodCall& method_call);
 
+    void FlushCommands();
+
     /// Returns a reference to the Maxwell3D GPU engine.
     Engines::Maxwell3D& Maxwell3D();
 
     /// Returns a const reference to the Maxwell3D GPU engine.
     const Engines::Maxwell3D& Maxwell3D() const;
 
+    /// Returns a reference to the KeplerCompute GPU engine.
+    Engines::KeplerCompute& KeplerCompute();
+
+    /// Returns a reference to the KeplerCompute GPU engine.
+    const Engines::KeplerCompute& KeplerCompute() const;
+
     /// Returns a reference to the GPU memory manager.
     Tegra::MemoryManager& MemoryManager();
 
@@ -164,6 +180,22 @@ public:
     /// Returns a reference to the GPU DMA pusher.
     Tegra::DmaPusher& DmaPusher();
 
+    void IncrementSyncPoint(u32 syncpoint_id);
+
+    u32 GetSyncpointValue(u32 syncpoint_id) const;
+
+    void RegisterSyncptInterrupt(u32 syncpoint_id, u32 value);
+
+    bool CancelSyncptInterrupt(u32 syncpoint_id, u32 value);
+
+    std::unique_lock<std::mutex> LockSync() {
+        return std::unique_lock{sync_mutex};
+    }
+
+    bool IsAsync() const {
+        return is_async;
+    }
+
     /// Returns a const reference to the GPU DMA pusher.
     const Tegra::DmaPusher& DmaPusher() const;
 
@@ -194,7 +226,12 @@ public:
 
                 u32 semaphore_acquire;
                 u32 semaphore_release;
-                INSERT_PADDING_WORDS(0xE4);
+                u32 fence_value;
+                union {
+                    BitField<4, 4, u32> operation;
+                    BitField<8, 8, u32> id;
+                } fence_action;
+                INSERT_PADDING_WORDS(0xE2);
 
                 // Puller state
                 u32 acquire_mode;
@@ -228,6 +265,9 @@ public:
     /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
     virtual void FlushAndInvalidateRegion(CacheAddr addr, u64 size) = 0;
 
+protected:
+    virtual void TriggerCpuInterrupt(u32 syncpoint_id, u32 value) const = 0;
+
 private:
     void ProcessBindMethod(const MethodCall& method_call);
     void ProcessSemaphoreTriggerMethod();
@@ -245,6 +285,7 @@ private:
 
 protected:
     std::unique_ptr<Tegra::DmaPusher> dma_pusher;
+    Core::System& system;
     VideoCore::RendererBase& renderer;
 
 private:
@@ -262,6 +303,14 @@ private:
     std::unique_ptr<Engines::MaxwellDMA> maxwell_dma;
     /// Inline memory engine
     std::unique_ptr<Engines::KeplerMemory> kepler_memory;
+
+    std::array<std::atomic<u32>, Service::Nvidia::MaxSyncPoints> syncpoints{};
+
+    std::array<std::list<u32>, Service::Nvidia::MaxSyncPoints> syncpt_interrupts;
+
+    std::mutex sync_mutex;
+
+    const bool is_async;
 };
 
 #define ASSERT_REG_POSITION(field_name, position)                                                  \
@@ -274,6 +323,8 @@ ASSERT_REG_POSITION(semaphore_trigger, 0x7);
 ASSERT_REG_POSITION(reference_count, 0x14);
 ASSERT_REG_POSITION(semaphore_acquire, 0x1A);
 ASSERT_REG_POSITION(semaphore_release, 0x1B);
+ASSERT_REG_POSITION(fence_value, 0x1C);
+ASSERT_REG_POSITION(fence_action, 0x1D);
 
 ASSERT_REG_POSITION(acquire_mode, 0x100);
 ASSERT_REG_POSITION(acquire_source, 0x101);

src/video_core/gpu_asynch.cpp

@@ -2,6 +2,8 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "core/core.h"
+#include "core/hardware_interrupt_manager.h"
 #include "video_core/gpu_asynch.h"
 #include "video_core/gpu_thread.h"
 #include "video_core/renderer_base.h"
@@ -9,7 +11,7 @@
 namespace VideoCommon {
 
 GPUAsynch::GPUAsynch(Core::System& system, VideoCore::RendererBase& renderer)
-    : GPU(system, renderer), gpu_thread{system} {}
+    : GPU(system, renderer, true), gpu_thread{system} {}
 
 GPUAsynch::~GPUAsynch() = default;
 
@@ -38,4 +40,9 @@ void GPUAsynch::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
     gpu_thread.FlushAndInvalidateRegion(addr, size);
 }
 
+void GPUAsynch::TriggerCpuInterrupt(const u32 syncpoint_id, const u32 value) const {
+    auto& interrupt_manager = system.InterruptManager();
+    interrupt_manager.GPUInterruptSyncpt(syncpoint_id, value);
+}
+
 } // namespace VideoCommon

src/video_core/gpu_asynch.h

@@ -27,6 +27,9 @@ public:
     void InvalidateRegion(CacheAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
 
+protected:
+    void TriggerCpuInterrupt(u32 syncpoint_id, u32 value) const override;
+
 private:
     GPUThread::ThreadManager gpu_thread;
 };

src/video_core/gpu_synch.cpp

@@ -8,7 +8,7 @@
 namespace VideoCommon {
 
 GPUSynch::GPUSynch(Core::System& system, VideoCore::RendererBase& renderer)
-    : GPU(system, renderer) {}
+    : GPU(system, renderer, false) {}
 
 GPUSynch::~GPUSynch() = default;
 

src/video_core/gpu_synch.h

@@ -25,6 +25,10 @@ public:
     void FlushRegion(CacheAddr addr, u64 size) override;
     void InvalidateRegion(CacheAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
+
+protected:
+    void TriggerCpuInterrupt([[maybe_unused]] u32 syncpoint_id,
+                             [[maybe_unused]] u32 value) const override {}
 };
 
 } // namespace VideoCommon

src/video_core/gpu_thread.cpp

@@ -21,7 +21,8 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
     MicroProfileOnThreadCreate("GpuThread");
 
     // Wait for first GPU command before acquiring the window context
-    state.WaitForCommands();
+    while (state.queue.Empty())
+        ;
 
     // If emulation was stopped during disk shader loading, abort before trying to acquire context
     if (!state.is_running) {
@@ -32,7 +33,6 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
 
     CommandDataContainer next;
     while (state.is_running) {
-        state.WaitForCommands();
         while (!state.queue.Empty()) {
             state.queue.Pop(next);
             if (const auto submit_list = std::get_if<SubmitListCommand>(&next.data)) {
@@ -49,8 +49,7 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
             } else {
                 UNREACHABLE();
             }
-            state.signaled_fence = next.fence;
-            state.TrySynchronize();
+            state.signaled_fence.store(next.fence);
         }
     }
 }
@@ -89,12 +88,7 @@ void ThreadManager::FlushRegion(CacheAddr addr, u64 size) {
 }
 
 void ThreadManager::InvalidateRegion(CacheAddr addr, u64 size) {
-    if (state.queue.Empty()) {
-        // It's quicker to invalidate a single region on the CPU if the queue is already empty
-        system.Renderer().Rasterizer().InvalidateRegion(addr, size);
-    } else {
-        PushCommand(InvalidateRegionCommand(addr, size));
-    }
+    system.Renderer().Rasterizer().InvalidateRegion(addr, size);
 }
 
 void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
@@ -105,22 +99,13 @@ void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
 u64 ThreadManager::PushCommand(CommandData&& command_data) {
     const u64 fence{++state.last_fence};
     state.queue.Push(CommandDataContainer(std::move(command_data), fence));
-    state.SignalCommands();
     return fence;
 }
 
 MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
 void SynchState::WaitForSynchronization(u64 fence) {
-    if (signaled_fence >= fence) {
-        return;
-    }
-
-    // Wait for the GPU to be idle (all commands to be executed)
-    {
-        MICROPROFILE_SCOPE(GPU_wait);
-        std::unique_lock lock{synchronization_mutex};
-        synchronization_condition.wait(lock, [this, fence] { return signaled_fence >= fence; });
-    }
+    while (signaled_fence.load() < fence)
+        ;
 }
 
 } // namespace VideoCommon::GPUThread

src/video_core/gpu_thread.h

@@ -88,41 +88,9 @@ struct CommandDataContainer {
 /// Struct used to synchronize the GPU thread
 struct SynchState final {
     std::atomic_bool is_running{true};
-    std::atomic_int queued_frame_count{};
-    std::mutex synchronization_mutex;
-    std::mutex commands_mutex;
-    std::condition_variable commands_condition;
-    std::condition_variable synchronization_condition;
-
-    /// Returns true if the gap in GPU commands is small enough that we can consider the CPU and GPU
-    /// synchronized. This is entirely empirical.
-    bool IsSynchronized() const {
-        constexpr std::size_t max_queue_gap{5};
-        return queue.Size() <= max_queue_gap;
-    }
-
-    void TrySynchronize() {
-        if (IsSynchronized()) {
-            std::lock_guard lock{synchronization_mutex};
-            synchronization_condition.notify_one();
-        }
-    }
 
     void WaitForSynchronization(u64 fence);
 
-    void SignalCommands() {
-        if (queue.Empty()) {
-            return;
-        }
-
-        commands_condition.notify_one();
-    }
-
-    void WaitForCommands() {
-        std::unique_lock lock{commands_mutex};
-        commands_condition.wait(lock, [this] { return !queue.Empty(); });
-    }
-
     using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
     CommandQueue queue;
     u64 last_fence{};

src/video_core/rasterizer_interface.h

@@ -34,6 +34,9 @@ public:
     /// Clear the current framebuffer
     virtual void Clear() = 0;
 
+    /// Dispatches a compute shader invocation
+    virtual void DispatchCompute(GPUVAddr code_addr) = 0;
+
     /// Notify rasterizer that all caches should be flushed to Switch memory
     virtual void FlushAll() = 0;
 
@@ -47,6 +50,9 @@ public:
     /// and invalidated
     virtual void FlushAndInvalidateRegion(CacheAddr addr, u64 size) = 0;
 
+    /// Notify the rasterizer to send all written commands to the host GPU.
+    virtual void FlushCommands() = 0;
+
     /// Notify rasterizer that a frame is about to finish
     virtual void TickFrame() = 0;
 

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -7,28 +7,41 @@
 #include <glad/glad.h>
 
 #include "common/assert.h"
+#include "common/microprofile.h"
+#include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_opengl/gl_buffer_cache.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
 namespace OpenGL {
 
+MICROPROFILE_DEFINE(OpenGL_Buffer_Download, "OpenGL", "Buffer Download", MP_RGB(192, 192, 128));
+
+CachedBufferBlock::CachedBufferBlock(CacheAddr cache_addr, const std::size_t size)
+    : VideoCommon::BufferBlock{cache_addr, size} {
+    gl_buffer.Create();
+    glNamedBufferData(gl_buffer.handle, static_cast<GLsizeiptr>(size), nullptr, GL_DYNAMIC_DRAW);
+}
+
+CachedBufferBlock::~CachedBufferBlock() = default;
+
 OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
                                std::size_t stream_size)
-    : VideoCommon::BufferCache<OGLBuffer, GLuint, OGLStreamBuffer>{
+    : VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer>{
           rasterizer, system, std::make_unique<OGLStreamBuffer>(stream_size, true)} {}
 
 OGLBufferCache::~OGLBufferCache() = default;
 
-OGLBuffer OGLBufferCache::CreateBuffer(std::size_t size) {
-    OGLBuffer buffer;
-    buffer.Create();
-    glNamedBufferData(buffer.handle, static_cast<GLsizeiptr>(size), nullptr, GL_DYNAMIC_DRAW);
-    return buffer;
+Buffer OGLBufferCache::CreateBlock(CacheAddr cache_addr, std::size_t size) {
+    return std::make_shared<CachedBufferBlock>(cache_addr, size);
 }
 
-const GLuint* OGLBufferCache::ToHandle(const OGLBuffer& buffer) {
-    return &buffer.handle;
+void OGLBufferCache::WriteBarrier() {
+    glMemoryBarrier(GL_ALL_BARRIER_BITS);
+}
+
+const GLuint* OGLBufferCache::ToHandle(const Buffer& buffer) {
+    return buffer->GetHandle();
 }
 
 const GLuint* OGLBufferCache::GetEmptyBuffer(std::size_t) {
@@ -36,23 +49,24 @@ const GLuint* OGLBufferCache::GetEmptyBuffer(std::size_t) {
     return &null_buffer;
 }
 
-void OGLBufferCache::UploadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
-                                      const u8* data) {
-    glNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+void OGLBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                                     const u8* data) {
+    glNamedBufferSubData(*buffer->GetHandle(), static_cast<GLintptr>(offset),
                          static_cast<GLsizeiptr>(size), data);
 }
 
-void OGLBufferCache::DownloadBufferData(const OGLBuffer& buffer, std::size_t offset,
-                                        std::size_t size, u8* data) {
-    glGetNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+void OGLBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                                       u8* data) {
+    MICROPROFILE_SCOPE(OpenGL_Buffer_Download);
+    glGetNamedBufferSubData(*buffer->GetHandle(), static_cast<GLintptr>(offset),
                             static_cast<GLsizeiptr>(size), data);
 }
 
-void OGLBufferCache::CopyBufferData(const OGLBuffer& src, const OGLBuffer& dst,
-                                    std::size_t src_offset, std::size_t dst_offset,
-                                    std::size_t size) {
-    glCopyNamedBufferSubData(src.handle, dst.handle, static_cast<GLintptr>(src_offset),
-                             static_cast<GLintptr>(dst_offset), static_cast<GLsizeiptr>(size));
+void OGLBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
+                               std::size_t dst_offset, std::size_t size) {
+    glCopyNamedBufferSubData(*src->GetHandle(), *dst->GetHandle(),
+                             static_cast<GLintptr>(src_offset), static_cast<GLintptr>(dst_offset),
+                             static_cast<GLsizeiptr>(size));
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -7,7 +7,7 @@
 #include <memory>
 
 #include "common/common_types.h"
-#include "video_core/buffer_cache.h"
+#include "video_core/buffer_cache/buffer_cache.h"
 #include "video_core/rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_stream_buffer.h"
@@ -21,7 +21,24 @@ namespace OpenGL {
 class OGLStreamBuffer;
 class RasterizerOpenGL;
 
-class OGLBufferCache final : public VideoCommon::BufferCache<OGLBuffer, GLuint, OGLStreamBuffer> {
+class CachedBufferBlock;
+
+using Buffer = std::shared_ptr<CachedBufferBlock>;
+
+class CachedBufferBlock : public VideoCommon::BufferBlock {
+public:
+    explicit CachedBufferBlock(CacheAddr cache_addr, const std::size_t size);
+    ~CachedBufferBlock();
+
+    const GLuint* GetHandle() const {
+        return &gl_buffer.handle;
+    }
+
+private:
+    OGLBuffer gl_buffer{};
+};
+
+class OGLBufferCache final : public VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer> {
 public:
     explicit OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
                             std::size_t stream_size);
@@ -30,18 +47,20 @@ public:
     const GLuint* GetEmptyBuffer(std::size_t) override;
 
 protected:
-    OGLBuffer CreateBuffer(std::size_t size) override;
+    Buffer CreateBlock(CacheAddr cache_addr, std::size_t size) override;
 
-    const GLuint* ToHandle(const OGLBuffer& buffer) override;
+    void WriteBarrier() override;
 
-    void UploadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
-                          const u8* data) override;
+    const GLuint* ToHandle(const Buffer& buffer) override;
 
-    void DownloadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
-                            u8* data) override;
+    void UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                         const u8* data) override;
 
-    void CopyBufferData(const OGLBuffer& src, const OGLBuffer& dst, std::size_t src_offset,
-                        std::size_t dst_offset, std::size_t size) override;
+    void DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+                           u8* data) override;
+
+    void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
+                   std::size_t dst_offset, std::size_t size) override;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_device.cpp

@@ -27,6 +27,8 @@ Device::Device() {
     shader_storage_alignment = GetInteger<std::size_t>(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT);
     max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS);
     max_varyings = GetInteger<u32>(GL_MAX_VARYING_VECTORS);
+    has_warp_intrinsics = GLAD_GL_NV_gpu_shader5 && GLAD_GL_NV_shader_thread_group &&
+                          GLAD_GL_NV_shader_thread_shuffle;
     has_vertex_viewport_layer = GLAD_GL_ARB_shader_viewport_layer_array;
     has_variable_aoffi = TestVariableAoffi();
     has_component_indexing_bug = TestComponentIndexingBug();
@@ -36,6 +38,7 @@ Device::Device(std::nullptr_t) {
     uniform_buffer_alignment = 0;
     max_vertex_attributes = 16;
     max_varyings = 15;
+    has_warp_intrinsics = true;
     has_vertex_viewport_layer = true;
     has_variable_aoffi = true;
     has_component_indexing_bug = false;

src/video_core/renderer_opengl/gl_device.h

@@ -30,6 +30,10 @@ public:
         return max_varyings;
     }
 
+    bool HasWarpIntrinsics() const {
+        return has_warp_intrinsics;
+    }
+
     bool HasVertexViewportLayer() const {
         return has_vertex_viewport_layer;
     }
@@ -50,6 +54,7 @@ private:
     std::size_t shader_storage_alignment{};
     u32 max_vertex_attributes{};
     u32 max_varyings{};
+    bool has_warp_intrinsics{};
     bool has_vertex_viewport_layer{};
     bool has_variable_aoffi{};
     bool has_component_indexing_bug{};

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -4,6 +4,7 @@
 
 #include <algorithm>
 #include <array>
+#include <bitset>
 #include <memory>
 #include <string>
 #include <string_view>
@@ -19,6 +20,7 @@
 #include "core/core.h"
 #include "core/hle/kernel/process.h"
 #include "core/settings.h"
+#include "video_core/engines/kepler_compute.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
@@ -105,6 +107,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
     shader_program_manager = std::make_unique<GLShader::ProgramManager>();
     state.draw.shader_program = 0;
     state.Apply();
+    clear_framebuffer.Create();
 
     LOG_DEBUG(Render_OpenGL, "Sync fixed function OpenGL state here");
     CheckExtensions();
@@ -124,10 +127,10 @@ GLuint RasterizerOpenGL::SetupVertexFormat() {
     auto& gpu = system.GPU().Maxwell3D();
     const auto& regs = gpu.regs;
 
-    if (!gpu.dirty_flags.vertex_attrib_format) {
+    if (!gpu.dirty.vertex_attrib_format) {
         return state.draw.vertex_array;
     }
-    gpu.dirty_flags.vertex_attrib_format = false;
+    gpu.dirty.vertex_attrib_format = false;
 
     MICROPROFILE_SCOPE(OpenGL_VAO);
 
@@ -181,7 +184,7 @@ GLuint RasterizerOpenGL::SetupVertexFormat() {
     }
 
     // Rebinding the VAO invalidates the vertex buffer bindings.
-    gpu.dirty_flags.vertex_array.set();
+    gpu.dirty.ResetVertexArrays();
 
     state.draw.vertex_array = vao_entry.handle;
     return vao_entry.handle;
@@ -189,17 +192,20 @@ GLuint RasterizerOpenGL::SetupVertexFormat() {
 
 void RasterizerOpenGL::SetupVertexBuffer(GLuint vao) {
     auto& gpu = system.GPU().Maxwell3D();
-    const auto& regs = gpu.regs;
-
-    if (gpu.dirty_flags.vertex_array.none())
+    if (!gpu.dirty.vertex_array_buffers)
         return;
+    gpu.dirty.vertex_array_buffers = false;
+
+    const auto& regs = gpu.regs;
 
     MICROPROFILE_SCOPE(OpenGL_VB);
 
     // Upload all guest vertex arrays sequentially to our buffer
     for (u32 index = 0; index < Maxwell::NumVertexArrays; ++index) {
-        if (!gpu.dirty_flags.vertex_array[index])
+        if (!gpu.dirty.vertex_array[index])
             continue;
+        gpu.dirty.vertex_array[index] = false;
+        gpu.dirty.vertex_instance[index] = false;
 
         const auto& vertex_array = regs.vertex_array[index];
         if (!vertex_array.IsEnabled())
@@ -224,8 +230,32 @@ void RasterizerOpenGL::SetupVertexBuffer(GLuint vao) {
             glVertexArrayBindingDivisor(vao, index, 0);
         }
     }
+}
 
-    gpu.dirty_flags.vertex_array.reset();
+void RasterizerOpenGL::SetupVertexInstances(GLuint vao) {
+    auto& gpu = system.GPU().Maxwell3D();
+
+    if (!gpu.dirty.vertex_instances)
+        return;
+    gpu.dirty.vertex_instances = false;
+
+    const auto& regs = gpu.regs;
+    // Upload all guest vertex arrays sequentially to our buffer
+    for (u32 index = 0; index < Maxwell::NumVertexArrays; ++index) {
+        if (!gpu.dirty.vertex_instance[index])
+            continue;
+
+        gpu.dirty.vertex_instance[index] = false;
+
+        if (regs.instanced_arrays.IsInstancingEnabled(index) &&
+            regs.vertex_array[index].divisor != 0) {
+            // Enable vertex buffer instancing with the specified divisor.
+            glVertexArrayBindingDivisor(vao, index, regs.vertex_array[index].divisor);
+        } else {
+            // Disable the vertex buffer instancing.
+            glVertexArrayBindingDivisor(vao, index, 0);
+        }
+    }
 }
 
 GLintptr RasterizerOpenGL::SetupIndexBuffer() {
@@ -298,9 +328,9 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
 
         Shader shader{shader_cache.GetStageProgram(program)};
 
-        const auto stage_enum{static_cast<Maxwell::ShaderStage>(stage)};
+        const auto stage_enum = static_cast<Maxwell::ShaderStage>(stage);
         SetupDrawConstBuffers(stage_enum, shader);
-        SetupGlobalRegions(stage_enum, shader);
+        SetupDrawGlobalMemory(stage_enum, shader);
         const auto texture_buffer_usage{SetupTextures(stage_enum, shader, base_bindings)};
 
         const ProgramVariant variant{base_bindings, primitive_mode, texture_buffer_usage};
@@ -341,7 +371,7 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
 
     SyncClipEnabled(clip_distances);
 
-    gpu.dirty_flags.shaders = false;
+    gpu.dirty.shaders = false;
 }
 
 std::size_t RasterizerOpenGL::CalculateVertexArraysSize() const {
@@ -424,13 +454,13 @@ std::pair<bool, bool> RasterizerOpenGL::ConfigureFramebuffers(
 
     const FramebufferConfigState fb_config_state{using_color_fb, using_depth_fb, preserve_contents,
                                                  single_color_target};
-    if (fb_config_state == current_framebuffer_config_state &&
-        gpu.dirty_flags.color_buffer.none() && !gpu.dirty_flags.zeta_buffer) {
+    if (fb_config_state == current_framebuffer_config_state && !gpu.dirty.render_settings) {
         // Only skip if the previous ConfigureFramebuffers call was from the same kind (multiple or
         // single color targets). This is done because the guest registers may not change but the
         // host framebuffer may contain different attachments
         return current_depth_stencil_usage;
     }
+    gpu.dirty.render_settings = false;
     current_framebuffer_config_state = fb_config_state;
 
     texture_cache.GuardRenderTargets(true);
@@ -519,13 +549,71 @@ std::pair<bool, bool> RasterizerOpenGL::ConfigureFramebuffers(
     return current_depth_stencil_usage = {static_cast<bool>(depth_surface), fbkey.stencil_enable};
 }
 
+void RasterizerOpenGL::ConfigureClearFramebuffer(OpenGLState& current_state, bool using_color_fb,
+                                                 bool using_depth_fb, bool using_stencil_fb) {
+    auto& gpu = system.GPU().Maxwell3D();
+    const auto& regs = gpu.regs;
+
+    texture_cache.GuardRenderTargets(true);
+    View color_surface{};
+    if (using_color_fb) {
+        color_surface = texture_cache.GetColorBufferSurface(regs.clear_buffers.RT, false);
+    }
+    View depth_surface{};
+    if (using_depth_fb || using_stencil_fb) {
+        depth_surface = texture_cache.GetDepthBufferSurface(false);
+    }
+    texture_cache.GuardRenderTargets(false);
+
+    current_state.draw.draw_framebuffer = clear_framebuffer.handle;
+    current_state.ApplyFramebufferState();
+
+    if (color_surface) {
+        color_surface->Attach(GL_COLOR_ATTACHMENT0, GL_DRAW_FRAMEBUFFER);
+    } else {
+        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+    }
+
+    if (depth_surface) {
+        const auto& params = depth_surface->GetSurfaceParams();
+        switch (params.type) {
+        case VideoCore::Surface::SurfaceType::Depth: {
+            depth_surface->Attach(GL_DEPTH_ATTACHMENT, GL_DRAW_FRAMEBUFFER);
+            glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+            break;
+        }
+        case VideoCore::Surface::SurfaceType::DepthStencil: {
+            depth_surface->Attach(GL_DEPTH_ATTACHMENT, GL_DRAW_FRAMEBUFFER);
+            break;
+        }
+        default: { UNIMPLEMENTED(); }
+        }
+    } else {
+        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0,
+                               0);
+    }
+}
+
 void RasterizerOpenGL::Clear() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
+    const auto& maxwell3d = system.GPU().Maxwell3D();
+
+    if (!maxwell3d.ShouldExecute()) {
+        return;
+    }
+
+    const auto& regs = maxwell3d.regs;
     bool use_color{};
     bool use_depth{};
     bool use_stencil{};
 
-    OpenGLState clear_state;
+    OpenGLState prev_state{OpenGLState::GetCurState()};
+    SCOPE_EXIT({
+        prev_state.AllDirty();
+        prev_state.Apply();
+    });
+
+    OpenGLState clear_state{OpenGLState::GetCurState()};
+    clear_state.SetDefaultViewports();
     if (regs.clear_buffers.R || regs.clear_buffers.G || regs.clear_buffers.B ||
         regs.clear_buffers.A) {
         use_color = true;
@@ -545,6 +633,7 @@ void RasterizerOpenGL::Clear() {
         // true.
         clear_state.depth.test_enabled = true;
         clear_state.depth.test_func = GL_ALWAYS;
+        clear_state.depth.write_mask = GL_TRUE;
     }
     if (regs.clear_buffers.S) {
         ASSERT_MSG(regs.zeta_enable != 0, "Tried to clear stencil but buffer is not enabled!");
@@ -581,8 +670,9 @@ void RasterizerOpenGL::Clear() {
         return;
     }
 
-    const auto [clear_depth, clear_stencil] = ConfigureFramebuffers(
-        clear_state, use_color, use_depth || use_stencil, false, regs.clear_buffers.RT.Value());
+    ConfigureClearFramebuffer(clear_state, use_color, use_depth, use_stencil);
+
+    SyncViewport(clear_state);
     if (regs.clear_flags.scissor) {
         SyncScissorTest(clear_state);
     }
@@ -591,21 +681,18 @@ void RasterizerOpenGL::Clear() {
         clear_state.EmulateViewportWithScissor();
     }
 
-    clear_state.ApplyColorMask();
-    clear_state.ApplyDepth();
-    clear_state.ApplyStencilTest();
-    clear_state.ApplyViewport();
-    clear_state.ApplyFramebufferState();
+    clear_state.AllDirty();
+    clear_state.Apply();
 
     if (use_color) {
-        glClearBufferfv(GL_COLOR, regs.clear_buffers.RT, regs.clear_color);
+        glClearBufferfv(GL_COLOR, 0, regs.clear_color);
     }
 
-    if (clear_depth && clear_stencil) {
+    if (use_depth && use_stencil) {
         glClearBufferfi(GL_DEPTH_STENCIL, 0, regs.clear_depth, regs.clear_stencil);
-    } else if (clear_depth) {
+    } else if (use_depth) {
         glClearBufferfv(GL_DEPTH, 0, &regs.clear_depth);
-    } else if (clear_stencil) {
+    } else if (use_stencil) {
         glClearBufferiv(GL_STENCIL, 0, &regs.clear_stencil);
     }
 }
@@ -616,7 +703,10 @@ void RasterizerOpenGL::DrawArrays() {
 
     MICROPROFILE_SCOPE(OpenGL_Drawing);
     auto& gpu = system.GPU().Maxwell3D();
-    const auto& regs = gpu.regs;
+
+    if (!gpu.ShouldExecute()) {
+        return;
+    }
 
     SyncColorMask();
     SyncFragmentColorClampState();
@@ -661,6 +751,7 @@ void RasterizerOpenGL::DrawArrays() {
 
     // Upload vertex and index data.
     SetupVertexBuffer(vao);
+    SetupVertexInstances(vao);
     const GLintptr index_buffer_offset = SetupIndexBuffer();
 
     // Setup draw parameters. It will automatically choose what glDraw* method to use.
@@ -687,7 +778,7 @@ void RasterizerOpenGL::DrawArrays() {
 
     if (invalidate) {
         // As all cached buffers are invalidated, we need to recheck their state.
-        gpu.dirty_flags.vertex_array.set();
+        gpu.dirty.ResetVertexArrays();
     }
 
     shader_program_manager->ApplyTo(state);
@@ -700,6 +791,46 @@ void RasterizerOpenGL::DrawArrays() {
     params.DispatchDraw();
 
     accelerate_draw = AccelDraw::Disabled;
+    gpu.dirty.memory_general = false;
+}
+
+void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
+    if (!GLAD_GL_ARB_compute_variable_group_size) {
+        LOG_ERROR(Render_OpenGL, "Compute is currently not supported on this device due to the "
+                                 "lack of GL_ARB_compute_variable_group_size");
+        return;
+    }
+
+    auto kernel = shader_cache.GetComputeKernel(code_addr);
+    const auto [program, next_bindings] = kernel->GetProgramHandle({});
+    state.draw.shader_program = program;
+    state.draw.program_pipeline = 0;
+
+    const std::size_t buffer_size =
+        Tegra::Engines::KeplerCompute::NumConstBuffers *
+        (Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
+    buffer_cache.Map(buffer_size);
+
+    bind_ubo_pushbuffer.Setup(0);
+    bind_ssbo_pushbuffer.Setup(0);
+
+    SetupComputeConstBuffers(kernel);
+    SetupComputeGlobalMemory(kernel);
+
+    // TODO(Rodrigo): Bind images and samplers
+
+    buffer_cache.Unmap();
+
+    bind_ubo_pushbuffer.Bind();
+    bind_ssbo_pushbuffer.Bind();
+
+    state.ApplyShaderProgram();
+    state.ApplyProgramPipeline();
+
+    const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
+    glDispatchComputeGroupSizeARB(launch_desc.grid_dim_x, launch_desc.grid_dim_y,
+                                  launch_desc.grid_dim_z, launch_desc.block_dim_x,
+                                  launch_desc.block_dim_y, launch_desc.block_dim_z);
 }
 
 void RasterizerOpenGL::FlushAll() {}
@@ -730,6 +861,10 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
     InvalidateRegion(addr, size);
 }
 
+void RasterizerOpenGL::FlushCommands() {
+    glFlush();
+}
+
 void RasterizerOpenGL::TickFrame() {
     buffer_cache.TickFrame();
 }
@@ -775,12 +910,25 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
 void RasterizerOpenGL::SetupDrawConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
                                              const Shader& shader) {
     MICROPROFILE_SCOPE(OpenGL_UBO);
-    const auto stage_index = static_cast<std::size_t>(stage);
-    const auto& shader_stage = system.GPU().Maxwell3D().state.shader_stages[stage_index];
-
-    // Upload only the enabled buffers from the 16 constbuffers of each shader stage
+    const auto& stages = system.GPU().Maxwell3D().state.shader_stages;
+    const auto& shader_stage = stages[static_cast<std::size_t>(stage)];
     for (const auto& entry : shader->GetShaderEntries().const_buffers) {
-        SetupConstBuffer(shader_stage.const_buffers[entry.GetIndex()], entry);
+        const auto& buffer = shader_stage.const_buffers[entry.GetIndex()];
+        SetupConstBuffer(buffer, entry);
+    }
+}
+
+void RasterizerOpenGL::SetupComputeConstBuffers(const Shader& kernel) {
+    MICROPROFILE_SCOPE(OpenGL_UBO);
+    const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
+    for (const auto& entry : kernel->GetShaderEntries().const_buffers) {
+        const auto& config = launch_desc.const_buffer_config[entry.GetIndex()];
+        const std::bitset<8> mask = launch_desc.memory_config.const_buffer_enable_mask.Value();
+        Tegra::Engines::ConstBufferInfo buffer;
+        buffer.address = config.Address();
+        buffer.size = config.size;
+        buffer.enabled = mask[entry.GetIndex()];
+        SetupConstBuffer(buffer, entry);
     }
 }
 
@@ -801,24 +949,39 @@ void RasterizerOpenGL::SetupConstBuffer(const Tegra::Engines::ConstBufferInfo& b
     bind_ubo_pushbuffer.Push(cbuf, offset, size);
 }
 
-void RasterizerOpenGL::SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
-                                          const Shader& shader) {
+void RasterizerOpenGL::SetupDrawGlobalMemory(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
+                                             const Shader& shader) {
     auto& gpu{system.GPU()};
     auto& memory_manager{gpu.MemoryManager()};
     const auto cbufs{gpu.Maxwell3D().state.shader_stages[static_cast<std::size_t>(stage)]};
-    const auto alignment{device.GetShaderStorageBufferAlignment()};
-
     for (const auto& entry : shader->GetShaderEntries().global_memory_entries) {
         const auto addr{cbufs.const_buffers[entry.GetCbufIndex()].address + entry.GetCbufOffset()};
-        const auto actual_addr{memory_manager.Read<u64>(addr)};
+        const auto gpu_addr{memory_manager.Read<u64>(addr)};
         const auto size{memory_manager.Read<u32>(addr + 8)};
-
-        const auto [ssbo, buffer_offset] =
-            buffer_cache.UploadMemory(actual_addr, size, alignment, true, entry.IsWritten());
-        bind_ssbo_pushbuffer.Push(ssbo, buffer_offset, static_cast<GLsizeiptr>(size));
+        SetupGlobalMemory(entry, gpu_addr, size);
     }
 }
 
+void RasterizerOpenGL::SetupComputeGlobalMemory(const Shader& kernel) {
+    auto& gpu{system.GPU()};
+    auto& memory_manager{gpu.MemoryManager()};
+    const auto cbufs{gpu.KeplerCompute().launch_description.const_buffer_config};
+    for (const auto& entry : kernel->GetShaderEntries().global_memory_entries) {
+        const auto addr{cbufs[entry.GetCbufIndex()].Address() + entry.GetCbufOffset()};
+        const auto gpu_addr{memory_manager.Read<u64>(addr)};
+        const auto size{memory_manager.Read<u32>(addr + 8)};
+        SetupGlobalMemory(entry, gpu_addr, size);
+    }
+}
+
+void RasterizerOpenGL::SetupGlobalMemory(const GLShader::GlobalMemoryEntry& entry,
+                                         GPUVAddr gpu_addr, std::size_t size) {
+    const auto alignment{device.GetShaderStorageBufferAlignment()};
+    const auto [ssbo, buffer_offset] =
+        buffer_cache.UploadMemory(gpu_addr, size, alignment, entry.IsWritten());
+    bind_ssbo_pushbuffer.Push(ssbo, buffer_offset, static_cast<GLsizeiptr>(size));
+}
+
 TextureBufferUsage RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, const Shader& shader,
                                                    BaseBindings base_bindings) {
     MICROPROFILE_SCOPE(OpenGL_Texture);
@@ -907,10 +1070,11 @@ void RasterizerOpenGL::SyncClipCoef() {
 }
 
 void RasterizerOpenGL::SyncCullMode() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
+    auto& maxwell3d = system.GPU().Maxwell3D();
+
+    const auto& regs = maxwell3d.regs;
 
     state.cull.enabled = regs.cull.enabled != 0;
-
     if (state.cull.enabled) {
         state.cull.front_face = MaxwellToGL::FrontFace(regs.cull.front_face);
         state.cull.mode = MaxwellToGL::CullFace(regs.cull.cull_face);
@@ -943,16 +1107,21 @@ void RasterizerOpenGL::SyncDepthTestState() {
     state.depth.test_enabled = regs.depth_test_enable != 0;
     state.depth.write_mask = regs.depth_write_enabled ? GL_TRUE : GL_FALSE;
 
-    if (!state.depth.test_enabled)
+    if (!state.depth.test_enabled) {
         return;
+    }
 
     state.depth.test_func = MaxwellToGL::ComparisonOp(regs.depth_test_func);
 }
 
 void RasterizerOpenGL::SyncStencilTestState() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
-    state.stencil.test_enabled = regs.stencil_enable != 0;
+    auto& maxwell3d = system.GPU().Maxwell3D();
+    if (!maxwell3d.dirty.stencil_test) {
+        return;
+    }
+    const auto& regs = maxwell3d.regs;
 
+    state.stencil.test_enabled = regs.stencil_enable != 0;
     if (!regs.stencil_enable) {
         return;
     }
@@ -981,10 +1150,17 @@ void RasterizerOpenGL::SyncStencilTestState() {
         state.stencil.back.action_depth_fail = GL_KEEP;
         state.stencil.back.action_depth_pass = GL_KEEP;
     }
+    state.MarkDirtyStencilState();
+    maxwell3d.dirty.stencil_test = false;
 }
 
 void RasterizerOpenGL::SyncColorMask() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
+    auto& maxwell3d = system.GPU().Maxwell3D();
+    if (!maxwell3d.dirty.color_mask) {
+        return;
+    }
+    const auto& regs = maxwell3d.regs;
+
     const std::size_t count =
         regs.independent_blend_enable ? Tegra::Engines::Maxwell3D::Regs::NumRenderTargets : 1;
     for (std::size_t i = 0; i < count; i++) {
@@ -995,6 +1171,9 @@ void RasterizerOpenGL::SyncColorMask() {
         dest.blue_enabled = (source.B == 0) ? GL_FALSE : GL_TRUE;
         dest.alpha_enabled = (source.A == 0) ? GL_FALSE : GL_TRUE;
     }
+
+    state.MarkDirtyColorMask();
+    maxwell3d.dirty.color_mask = false;
 }
 
 void RasterizerOpenGL::SyncMultiSampleState() {
@@ -1009,7 +1188,11 @@ void RasterizerOpenGL::SyncFragmentColorClampState() {
 }
 
 void RasterizerOpenGL::SyncBlendState() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
+    auto& maxwell3d = system.GPU().Maxwell3D();
+    if (!maxwell3d.dirty.blend_state) {
+        return;
+    }
+    const auto& regs = maxwell3d.regs;
 
     state.blend_color.red = regs.blend_color.r;
     state.blend_color.green = regs.blend_color.g;
@@ -1032,6 +1215,8 @@ void RasterizerOpenGL::SyncBlendState() {
         for (std::size_t i = 1; i < Tegra::Engines::Maxwell3D::Regs::NumRenderTargets; i++) {
             state.blend[i].enabled = false;
         }
+        maxwell3d.dirty.blend_state = false;
+        state.MarkDirtyBlendState();
         return;
     }
 
@@ -1048,6 +1233,9 @@ void RasterizerOpenGL::SyncBlendState() {
         blend.src_a_func = MaxwellToGL::BlendFunc(src.factor_source_a);
         blend.dst_a_func = MaxwellToGL::BlendFunc(src.factor_dest_a);
     }
+
+    state.MarkDirtyBlendState();
+    maxwell3d.dirty.blend_state = false;
 }
 
 void RasterizerOpenGL::SyncLogicOpState() {
@@ -1099,13 +1287,21 @@ void RasterizerOpenGL::SyncPointState() {
 }
 
 void RasterizerOpenGL::SyncPolygonOffset() {
-    const auto& regs = system.GPU().Maxwell3D().regs;
+    auto& maxwell3d = system.GPU().Maxwell3D();
+    if (!maxwell3d.dirty.polygon_offset) {
+        return;
+    }
+    const auto& regs = maxwell3d.regs;
+
     state.polygon_offset.fill_enable = regs.polygon_offset_fill_enable != 0;
     state.polygon_offset.line_enable = regs.polygon_offset_line_enable != 0;
     state.polygon_offset.point_enable = regs.polygon_offset_point_enable != 0;
     state.polygon_offset.units = regs.polygon_offset_units;
     state.polygon_offset.factor = regs.polygon_offset_factor;
     state.polygon_offset.clamp = regs.polygon_offset_clamp;
+
+    state.MarkDirtyPolygonOffset();
+    maxwell3d.dirty.polygon_offset = false;
 }
 
 void RasterizerOpenGL::SyncAlphaTest() {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -58,10 +58,12 @@ public:
 
     void DrawArrays() override;
     void Clear() override;
+    void DispatchCompute(GPUVAddr code_addr) override;
     void FlushAll() override;
     void FlushRegion(CacheAddr addr, u64 size) override;
     void InvalidateRegion(CacheAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
+    void FlushCommands() override;
     void TickFrame() override;
     bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src,
                                const Tegra::Engines::Fermi2D::Regs::Surface& dst,
@@ -108,17 +110,30 @@ private:
         OpenGLState& current_state, bool using_color_fb = true, bool using_depth_fb = true,
         bool preserve_contents = true, std::optional<std::size_t> single_color_target = {});
 
+    void ConfigureClearFramebuffer(OpenGLState& current_state, bool using_color_fb,
+                                   bool using_depth_fb, bool using_stencil_fb);
+
     /// Configures the current constbuffers to use for the draw command.
     void SetupDrawConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
                                const Shader& shader);
 
+    /// Configures the current constbuffers to use for the kernel invocation.
+    void SetupComputeConstBuffers(const Shader& kernel);
+
     /// Configures a constant buffer.
     void SetupConstBuffer(const Tegra::Engines::ConstBufferInfo& buffer,
                           const GLShader::ConstBufferEntry& entry);
 
     /// Configures the current global memory entries to use for the draw command.
-    void SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
-                            const Shader& shader);
+    void SetupDrawGlobalMemory(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
+                               const Shader& shader);
+
+    /// Configures the current global memory entries to use for the kernel invocation.
+    void SetupComputeGlobalMemory(const Shader& kernel);
+
+    /// Configures a constant buffer.
+    void SetupGlobalMemory(const GLShader::GlobalMemoryEntry& entry, GPUVAddr gpu_addr,
+                           std::size_t size);
 
     /// Configures the current textures to use for the draw command. Returns shaders texture buffer
     /// usage.
@@ -216,6 +231,7 @@ private:
     GLuint SetupVertexFormat();
 
     void SetupVertexBuffer(GLuint vao);
+    void SetupVertexInstances(GLuint vao);
 
     GLintptr SetupIndexBuffer();
 
@@ -226,6 +242,8 @@ private:
     enum class AccelDraw { Disabled, Arrays, Indexed };
     AccelDraw accelerate_draw = AccelDraw::Disabled;
 
+    OGLFramebuffer clear_framebuffer;
+
     using CachedPageMap = boost::icl::interval_map<u64, int>;
     CachedPageMap cached_pages;
 };

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -23,13 +23,13 @@ namespace OpenGL {
 
 using VideoCommon::Shader::ProgramCode;
 
-// One UBO is always reserved for emulation values
-constexpr u32 RESERVED_UBOS = 1;
+// One UBO is always reserved for emulation values on staged shaders
+constexpr u32 STAGE_RESERVED_UBOS = 1;
 
 struct UnspecializedShader {
     std::string code;
     GLShader::ShaderEntries entries;
-    Maxwell::ShaderProgram program_type;
+    ProgramType program_type;
 };
 
 namespace {
@@ -55,15 +55,17 @@ ProgramCode GetShaderCode(Tegra::MemoryManager& memory_manager, const GPUVAddr g
 }
 
 /// Gets the shader type from a Maxwell program type
-constexpr GLenum GetShaderType(Maxwell::ShaderProgram program_type) {
+constexpr GLenum GetShaderType(ProgramType program_type) {
     switch (program_type) {
-    case Maxwell::ShaderProgram::VertexA:
-    case Maxwell::ShaderProgram::VertexB:
+    case ProgramType::VertexA:
+    case ProgramType::VertexB:
         return GL_VERTEX_SHADER;
-    case Maxwell::ShaderProgram::Geometry:
+    case ProgramType::Geometry:
         return GL_GEOMETRY_SHADER;
-    case Maxwell::ShaderProgram::Fragment:
+    case ProgramType::Fragment:
         return GL_FRAGMENT_SHADER;
+    case ProgramType::Compute:
+        return GL_COMPUTE_SHADER;
     default:
         return GL_NONE;
     }
@@ -100,6 +102,25 @@ constexpr std::tuple<const char*, const char*, u32> GetPrimitiveDescription(GLen
     }
 }
 
+ProgramType GetProgramType(Maxwell::ShaderProgram program) {
+    switch (program) {
+    case Maxwell::ShaderProgram::VertexA:
+        return ProgramType::VertexA;
+    case Maxwell::ShaderProgram::VertexB:
+        return ProgramType::VertexB;
+    case Maxwell::ShaderProgram::TesselationControl:
+        return ProgramType::TessellationControl;
+    case Maxwell::ShaderProgram::TesselationEval:
+        return ProgramType::TessellationEval;
+    case Maxwell::ShaderProgram::Geometry:
+        return ProgramType::Geometry;
+    case Maxwell::ShaderProgram::Fragment:
+        return ProgramType::Fragment;
+    }
+    UNREACHABLE();
+    return {};
+}
+
 /// Calculates the size of a program stream
 std::size_t CalculateProgramSize(const GLShader::ProgramCode& program) {
     constexpr std::size_t start_offset = 10;
@@ -128,13 +149,13 @@ std::size_t CalculateProgramSize(const GLShader::ProgramCode& program) {
 }
 
 /// Hashes one (or two) program streams
-u64 GetUniqueIdentifier(Maxwell::ShaderProgram program_type, const ProgramCode& code,
+u64 GetUniqueIdentifier(ProgramType program_type, const ProgramCode& code,
                         const ProgramCode& code_b, std::size_t size_a = 0, std::size_t size_b = 0) {
     if (size_a == 0) {
         size_a = CalculateProgramSize(code);
     }
     u64 unique_identifier = Common::CityHash64(reinterpret_cast<const char*>(code.data()), size_a);
-    if (program_type != Maxwell::ShaderProgram::VertexA) {
+    if (program_type != ProgramType::VertexA) {
         return unique_identifier;
     }
     // VertexA programs include two programs
@@ -152,12 +173,12 @@ u64 GetUniqueIdentifier(Maxwell::ShaderProgram program_type, const ProgramCode&
 }
 
 /// Creates an unspecialized program from code streams
-GLShader::ProgramResult CreateProgram(const Device& device, Maxwell::ShaderProgram program_type,
+GLShader::ProgramResult CreateProgram(const Device& device, ProgramType program_type,
                                       ProgramCode program_code, ProgramCode program_code_b) {
     GLShader::ShaderSetup setup(program_code);
     setup.program.size_a = CalculateProgramSize(program_code);
     setup.program.size_b = 0;
-    if (program_type == Maxwell::ShaderProgram::VertexA) {
+    if (program_type == ProgramType::VertexA) {
         // VertexB is always enabled, so when VertexA is enabled, we have two vertex shaders.
         // Conventional HW does not support this, so we combine VertexA and VertexB into one
         // stage here.
@@ -168,33 +189,43 @@ GLShader::ProgramResult CreateProgram(const Device& device, Maxwell::ShaderProgr
         program_type, program_code, program_code_b, setup.program.size_a, setup.program.size_b);
 
     switch (program_type) {
-    case Maxwell::ShaderProgram::VertexA:
-    case Maxwell::ShaderProgram::VertexB:
+    case ProgramType::VertexA:
+    case ProgramType::VertexB:
         return GLShader::GenerateVertexShader(device, setup);
-    case Maxwell::ShaderProgram::Geometry:
+    case ProgramType::Geometry:
         return GLShader::GenerateGeometryShader(device, setup);
-    case Maxwell::ShaderProgram::Fragment:
+    case ProgramType::Fragment:
         return GLShader::GenerateFragmentShader(device, setup);
+    case ProgramType::Compute:
+        return GLShader::GenerateComputeShader(device, setup);
     default:
-        LOG_CRITICAL(HW_GPU, "Unimplemented program_type={}", static_cast<u32>(program_type));
-        UNREACHABLE();
+        UNIMPLEMENTED_MSG("Unimplemented program_type={}", static_cast<u32>(program_type));
         return {};
     }
 }
 
 CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEntries& entries,
-                               Maxwell::ShaderProgram program_type, const ProgramVariant& variant,
+                               ProgramType program_type, const ProgramVariant& variant,
                                bool hint_retrievable = false) {
     auto base_bindings{variant.base_bindings};
     const auto primitive_mode{variant.primitive_mode};
     const auto texture_buffer_usage{variant.texture_buffer_usage};
 
     std::string source = "#version 430 core\n"
-                         "#extension GL_ARB_separate_shader_objects : enable\n";
+                         "#extension GL_ARB_separate_shader_objects : enable\n"
+                         "#extension GL_NV_gpu_shader5 : enable\n"
+                         "#extension GL_NV_shader_thread_group : enable\n";
     if (entries.shader_viewport_layer_array) {
         source += "#extension GL_ARB_shader_viewport_layer_array : enable\n";
     }
-    source += fmt::format("\n#define EMULATION_UBO_BINDING {}\n", base_bindings.cbuf++);
+    if (program_type == ProgramType::Compute) {
+        source += "#extension GL_ARB_compute_variable_group_size : require\n";
+    }
+    source += '\n';
+
+    if (program_type != ProgramType::Compute) {
+        source += fmt::format("#define EMULATION_UBO_BINDING {}\n", base_bindings.cbuf++);
+    }
 
     for (const auto& cbuf : entries.const_buffers) {
         source +=
@@ -218,17 +249,24 @@ CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEn
         if (!texture_buffer_usage.test(i)) {
             continue;
         }
-        source += fmt::format("#define SAMPLER_{}_IS_BUFFER", i);
+        source += fmt::format("#define SAMPLER_{}_IS_BUFFER\n", i);
+    }
+    if (texture_buffer_usage.any()) {
+        source += '\n';
     }
 
-    if (program_type == Maxwell::ShaderProgram::Geometry) {
+    if (program_type == ProgramType::Geometry) {
         const auto [glsl_topology, debug_name, max_vertices] =
             GetPrimitiveDescription(primitive_mode);
 
-        source += "layout (" + std::string(glsl_topology) + ") in;\n";
+        source += "layout (" + std::string(glsl_topology) + ") in;\n\n";
         source += "#define MAX_VERTEX_INPUT " + std::to_string(max_vertices) + '\n';
     }
+    if (program_type == ProgramType::Compute) {
+        source += "layout (local_size_variable) in;\n";
+    }
 
+    source += '\n';
     source += code;
 
     OGLShader shader;
@@ -255,9 +293,9 @@ std::set<GLenum> GetSupportedFormats() {
 
 } // Anonymous namespace
 
-CachedShader::CachedShader(const ShaderParameters& params, Maxwell::ShaderProgram program_type,
+CachedShader::CachedShader(const ShaderParameters& params, ProgramType program_type,
                            GLShader::ProgramResult result)
-    : RasterizerCacheObject{params.host_ptr}, host_ptr{params.host_ptr}, cpu_addr{params.cpu_addr},
+    : RasterizerCacheObject{params.host_ptr}, cpu_addr{params.cpu_addr},
       unique_identifier{params.unique_identifier}, program_type{program_type},
       disk_cache{params.disk_cache}, precompiled_programs{params.precompiled_programs},
       entries{result.second}, code{std::move(result.first)}, shader_length{entries.shader_length} {}
@@ -268,29 +306,50 @@ Shader CachedShader::CreateStageFromMemory(const ShaderParameters& params,
                                            ProgramCode&& program_code_b) {
     const auto code_size{CalculateProgramSize(program_code)};
     const auto code_size_b{CalculateProgramSize(program_code_b)};
-    auto result{CreateProgram(params.device, program_type, program_code, program_code_b)};
+    auto result{
+        CreateProgram(params.device, GetProgramType(program_type), program_code, program_code_b)};
     if (result.first.empty()) {
         // TODO(Rodrigo): Unimplemented shader stages hit here, avoid using these for now
         return {};
     }
 
     params.disk_cache.SaveRaw(ShaderDiskCacheRaw(
-        params.unique_identifier, program_type, static_cast<u32>(code_size / sizeof(u64)),
-        static_cast<u32>(code_size_b / sizeof(u64)), std::move(program_code),
-        std::move(program_code_b)));
+        params.unique_identifier, GetProgramType(program_type),
+        static_cast<u32>(code_size / sizeof(u64)), static_cast<u32>(code_size_b / sizeof(u64)),
+        std::move(program_code), std::move(program_code_b)));
 
-    return std::shared_ptr<CachedShader>(new CachedShader(params, program_type, std::move(result)));
+    return std::shared_ptr<CachedShader>(
+        new CachedShader(params, GetProgramType(program_type), std::move(result)));
 }
 
 Shader CachedShader::CreateStageFromCache(const ShaderParameters& params,
                                           Maxwell::ShaderProgram program_type,
                                           GLShader::ProgramResult result) {
-    return std::shared_ptr<CachedShader>(new CachedShader(params, program_type, std::move(result)));
+    return std::shared_ptr<CachedShader>(
+        new CachedShader(params, GetProgramType(program_type), std::move(result)));
+}
+
+Shader CachedShader::CreateKernelFromMemory(const ShaderParameters& params, ProgramCode&& code) {
+    auto result{CreateProgram(params.device, ProgramType::Compute, code, {})};
+
+    const auto code_size{CalculateProgramSize(code)};
+    params.disk_cache.SaveRaw(ShaderDiskCacheRaw(params.unique_identifier, ProgramType::Compute,
+                                                 static_cast<u32>(code_size / sizeof(u64)), 0,
+                                                 std::move(code), {}));
+
+    return std::shared_ptr<CachedShader>(
+        new CachedShader(params, ProgramType::Compute, std::move(result)));
+}
+
+Shader CachedShader::CreateKernelFromCache(const ShaderParameters& params,
+                                           GLShader::ProgramResult result) {
+    return std::shared_ptr<CachedShader>(
+        new CachedShader(params, ProgramType::Compute, std::move(result)));
 }
 
 std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(const ProgramVariant& variant) {
     GLuint handle{};
-    if (program_type == Maxwell::ShaderProgram::Geometry) {
+    if (program_type == ProgramType::Geometry) {
         handle = GetGeometryShader(variant);
     } else {
         const auto [entry, is_cache_miss] = programs.try_emplace(variant);
@@ -308,8 +367,11 @@ std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(const ProgramVar
         handle = program->handle;
     }
 
-    auto base_bindings{variant.base_bindings};
-    base_bindings.cbuf += static_cast<u32>(entries.const_buffers.size()) + RESERVED_UBOS;
+    auto base_bindings = variant.base_bindings;
+    base_bindings.cbuf += static_cast<u32>(entries.const_buffers.size());
+    if (program_type != ProgramType::Compute) {
+        base_bindings.cbuf += STAGE_RESERVED_UBOS;
+    }
     base_bindings.gmem += static_cast<u32>(entries.global_memory_entries.size());
     base_bindings.sampler += static_cast<u32>(entries.samplers.size());
 
@@ -572,7 +634,7 @@ std::unordered_map<u64, UnspecializedShader> ShaderCacheOpenGL::GenerateUnspecia
 }
 
 Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
-    if (!system.GPU().Maxwell3D().dirty_flags.shaders) {
+    if (!system.GPU().Maxwell3D().dirty.shaders) {
         return last_shaders[static_cast<std::size_t>(program)];
     }
 
@@ -589,13 +651,15 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
     // No shader found - create a new one
     ProgramCode program_code{GetShaderCode(memory_manager, program_addr, host_ptr)};
     ProgramCode program_code_b;
-    if (program == Maxwell::ShaderProgram::VertexA) {
+    const bool is_program_a{program == Maxwell::ShaderProgram::VertexA};
+    if (is_program_a) {
         const GPUVAddr program_addr_b{GetShaderAddress(system, Maxwell::ShaderProgram::VertexB)};
         program_code_b = GetShaderCode(memory_manager, program_addr_b,
                                        memory_manager.GetPointer(program_addr_b));
     }
 
-    const auto unique_identifier = GetUniqueIdentifier(program, program_code, program_code_b);
+    const auto unique_identifier =
+        GetUniqueIdentifier(GetProgramType(program), program_code, program_code_b);
     const auto cpu_addr{*memory_manager.GpuToCpuAddress(program_addr)};
     const ShaderParameters params{disk_cache, precompiled_programs, device, cpu_addr,
                                   host_ptr,   unique_identifier};
@@ -612,4 +676,30 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
     return last_shaders[static_cast<std::size_t>(program)] = shader;
 }
 
+Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
+    auto& memory_manager{system.GPU().MemoryManager()};
+    const auto host_ptr{memory_manager.GetPointer(code_addr)};
+    auto kernel = TryGet(host_ptr);
+    if (kernel) {
+        return kernel;
+    }
+
+    // No kernel found - create a new one
+    auto code{GetShaderCode(memory_manager, code_addr, host_ptr)};
+    const auto unique_identifier{GetUniqueIdentifier(ProgramType::Compute, code, {})};
+    const auto cpu_addr{*memory_manager.GpuToCpuAddress(code_addr)};
+    const ShaderParameters params{disk_cache, precompiled_programs, device, cpu_addr,
+                                  host_ptr,   unique_identifier};
+
+    const auto found = precompiled_shaders.find(unique_identifier);
+    if (found == precompiled_shaders.end()) {
+        kernel = CachedShader::CreateKernelFromMemory(params, std::move(code));
+    } else {
+        kernel = CachedShader::CreateKernelFromCache(params, found->second);
+    }
+
+    Register(kernel);
+    return kernel;
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -61,6 +61,11 @@ public:
                                        Maxwell::ShaderProgram program_type,
                                        GLShader::ProgramResult result);
 
+    static Shader CreateKernelFromMemory(const ShaderParameters& params, ProgramCode&& code);
+
+    static Shader CreateKernelFromCache(const ShaderParameters& params,
+                                        GLShader::ProgramResult result);
+
     VAddr GetCpuAddr() const override {
         return cpu_addr;
     }
@@ -78,7 +83,7 @@ public:
     std::tuple<GLuint, BaseBindings> GetProgramHandle(const ProgramVariant& variant);
 
 private:
-    explicit CachedShader(const ShaderParameters& params, Maxwell::ShaderProgram program_type,
+    explicit CachedShader(const ShaderParameters& params, ProgramType program_type,
                           GLShader::ProgramResult result);
 
     // Geometry programs. These are needed because GLSL needs an input topology but it's not
@@ -101,10 +106,9 @@ private:
 
     ShaderDiskCacheUsage GetUsage(const ProgramVariant& variant) const;
 
-    u8* host_ptr{};
     VAddr cpu_addr{};
     u64 unique_identifier{};
-    Maxwell::ShaderProgram program_type{};
+    ProgramType program_type{};
     ShaderDiskCacheOpenGL& disk_cache;
     const PrecompiledPrograms& precompiled_programs;
 
@@ -132,6 +136,9 @@ public:
     /// Gets the current specified shader stage program
     Shader GetStageProgram(Maxwell::ShaderProgram program);
 
+    /// Gets a compute kernel in the passed address
+    Shader GetComputeKernel(GPUVAddr code_addr);
+
 protected:
     // We do not have to flush this cache as things in it are never modified by us.
     void FlushObjectInner(const Shader& object) override {}

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -37,7 +37,6 @@ using namespace std::string_literals;
 using namespace VideoCommon::Shader;
 
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
-using ShaderStage = Tegra::Engines::Maxwell3D::Regs::ShaderStage;
 using Operation = const OperationNode&;
 
 enum class Type { Bool, Bool2, Float, Int, Uint, HalfFloat };
@@ -162,9 +161,13 @@ std::string FlowStackTopName(MetaStackClass stack) {
     return fmt::format("{}_flow_stack_top", GetFlowStackPrefix(stack));
 }
 
+constexpr bool IsVertexShader(ProgramType stage) {
+    return stage == ProgramType::VertexA || stage == ProgramType::VertexB;
+}
+
 class GLSLDecompiler final {
 public:
-    explicit GLSLDecompiler(const Device& device, const ShaderIR& ir, ShaderStage stage,
+    explicit GLSLDecompiler(const Device& device, const ShaderIR& ir, ProgramType stage,
                             std::string suffix)
         : device{device}, ir{ir}, stage{stage}, suffix{suffix}, header{ir.GetHeader()} {}
 
@@ -248,21 +251,21 @@ public:
         }
         entries.clip_distances = ir.GetClipDistances();
         entries.shader_viewport_layer_array =
-            stage == ShaderStage::Vertex && (ir.UsesLayer() || ir.UsesViewportIndex());
+            IsVertexShader(stage) && (ir.UsesLayer() || ir.UsesViewportIndex());
         entries.shader_length = ir.GetLength();
         return entries;
     }
 
 private:
     void DeclareVertex() {
-        if (stage != ShaderStage::Vertex)
+        if (!IsVertexShader(stage))
             return;
 
         DeclareVertexRedeclarations();
     }
 
     void DeclareGeometry() {
-        if (stage != ShaderStage::Geometry) {
+        if (stage != ProgramType::Geometry) {
             return;
         }
 
@@ -293,14 +296,14 @@ private:
                 break;
             }
         }
-        if (stage != ShaderStage::Vertex || device.HasVertexViewportLayer()) {
+        if (!IsVertexShader(stage) || device.HasVertexViewportLayer()) {
             if (ir.UsesLayer()) {
                 code.AddLine("int gl_Layer;");
             }
             if (ir.UsesViewportIndex()) {
                 code.AddLine("int gl_ViewportIndex;");
             }
-        } else if ((ir.UsesLayer() || ir.UsesViewportIndex()) && stage == ShaderStage::Vertex &&
+        } else if ((ir.UsesLayer() || ir.UsesViewportIndex()) && IsVertexShader(stage) &&
                    !device.HasVertexViewportLayer()) {
             LOG_ERROR(
                 Render_OpenGL,
@@ -337,11 +340,16 @@ private:
     }
 
     void DeclareLocalMemory() {
-        if (const u64 local_memory_size = header.GetLocalMemorySize(); local_memory_size > 0) {
-            const auto element_count = Common::AlignUp(local_memory_size, 4) / 4;
-            code.AddLine("float {}[{}];", GetLocalMemory(), element_count);
-            code.AddNewLine();
+        // TODO(Rodrigo): Unstub kernel local memory size and pass it from a register at
+        // specialization time.
+        const u64 local_memory_size =
+            stage == ProgramType::Compute ? 0x400 : header.GetLocalMemorySize();
+        if (local_memory_size == 0) {
+            return;
         }
+        const auto element_count = Common::AlignUp(local_memory_size, 4) / 4;
+        code.AddLine("float {}[{}];", GetLocalMemory(), element_count);
+        code.AddNewLine();
     }
 
     void DeclareInternalFlags() {
@@ -395,12 +403,12 @@ private:
         const u32 location{GetGenericAttributeIndex(index)};
 
         std::string name{GetInputAttribute(index)};
-        if (stage == ShaderStage::Geometry) {
+        if (stage == ProgramType::Geometry) {
             name = "gs_" + name + "[]";
         }
 
         std::string suffix;
-        if (stage == ShaderStage::Fragment) {
+        if (stage == ProgramType::Fragment) {
             const auto input_mode{header.ps.GetAttributeUse(location)};
             if (skip_unused && input_mode == AttributeUse::Unused) {
                 return;
@@ -412,7 +420,7 @@ private:
     }
 
     void DeclareOutputAttributes() {
-        if (ir.HasPhysicalAttributes() && stage != ShaderStage::Fragment) {
+        if (ir.HasPhysicalAttributes() && stage != ProgramType::Fragment) {
             for (u32 i = 0; i < GetNumPhysicalVaryings(); ++i) {
                 DeclareOutputAttribute(ToGenericAttribute(i));
             }
@@ -534,7 +542,7 @@ private:
                 constexpr u32 element_stride{4};
                 const u32 address{generic_base + index * generic_stride + element * element_stride};
 
-                const bool declared{stage != ShaderStage::Fragment ||
+                const bool declared{stage != ProgramType::Fragment ||
                                     header.ps.GetAttributeUse(index) != AttributeUse::Unused};
                 const std::string value{declared ? ReadAttribute(attribute, element) : "0"};
                 code.AddLine("case 0x{:x}: return {};", address, value);
@@ -557,7 +565,7 @@ private:
                 case Tegra::Shader::ImageType::Texture1D:
                     return "image1D";
                 case Tegra::Shader::ImageType::TextureBuffer:
-                    return "bufferImage";
+                    return "imageBuffer";
                 case Tegra::Shader::ImageType::Texture1DArray:
                     return "image1DArray";
                 case Tegra::Shader::ImageType::Texture2D:
@@ -638,7 +646,7 @@ private:
         }
 
         if (const auto abuf = std::get_if<AbufNode>(&*node)) {
-            UNIMPLEMENTED_IF_MSG(abuf->IsPhysicalBuffer() && stage == ShaderStage::Geometry,
+            UNIMPLEMENTED_IF_MSG(abuf->IsPhysicalBuffer() && stage == ProgramType::Geometry,
                                  "Physical attributes in geometry shaders are not implemented");
             if (abuf->IsPhysicalBuffer()) {
                 return fmt::format("readPhysicalAttribute(ftou({}))",
@@ -693,6 +701,9 @@ private:
         }
 
         if (const auto lmem = std::get_if<LmemNode>(&*node)) {
+            if (stage == ProgramType::Compute) {
+                LOG_WARNING(Render_OpenGL, "Local memory is stubbed on compute shaders");
+            }
             return fmt::format("{}[ftou({}) / 4]", GetLocalMemory(), Visit(lmem->GetAddress()));
         }
 
@@ -722,7 +733,7 @@ private:
 
     std::string ReadAttribute(Attribute::Index attribute, u32 element, const Node& buffer = {}) {
         const auto GeometryPass = [&](std::string_view name) {
-            if (stage == ShaderStage::Geometry && buffer) {
+            if (stage == ProgramType::Geometry && buffer) {
                 // TODO(Rodrigo): Guard geometry inputs against out of bound reads. Some games
                 // set an 0x80000000 index for those and the shader fails to build. Find out why
                 // this happens and what's its intent.
@@ -734,10 +745,10 @@ private:
         switch (attribute) {
         case Attribute::Index::Position:
             switch (stage) {
-            case ShaderStage::Geometry:
+            case ProgramType::Geometry:
                 return fmt::format("gl_in[ftou({})].gl_Position{}", Visit(buffer),
                                    GetSwizzle(element));
-            case ShaderStage::Fragment:
+            case ProgramType::Fragment:
                 return element == 3 ? "1.0f" : ("gl_FragCoord"s + GetSwizzle(element));
             default:
                 UNREACHABLE();
@@ -758,7 +769,7 @@ private:
             // TODO(Subv): Find out what the values are for the first two elements when inside a
             // vertex shader, and what's the value of the fourth element when inside a Tess Eval
             // shader.
-            ASSERT(stage == ShaderStage::Vertex);
+            ASSERT(IsVertexShader(stage));
             switch (element) {
             case 2:
                 // Config pack's first value is instance_id.
@@ -770,7 +781,7 @@ private:
             return "0";
         case Attribute::Index::FrontFacing:
             // TODO(Subv): Find out what the values are for the other elements.
-            ASSERT(stage == ShaderStage::Fragment);
+            ASSERT(stage == ProgramType::Fragment);
             switch (element) {
             case 3:
                 return "itof(gl_FrontFacing ? -1 : 0)";
@@ -792,7 +803,7 @@ private:
             return value;
         }
         // There's a bug in NVidia's proprietary drivers that makes precise fail on fragment shaders
-        const std::string precise = stage != ShaderStage::Fragment ? "precise " : "";
+        const std::string precise = stage != ProgramType::Fragment ? "precise " : "";
 
         const std::string temporary = code.GenerateTemporary();
         code.AddLine("{}float {} = {};", precise, temporary, value);
@@ -827,12 +838,12 @@ private:
                 UNIMPLEMENTED();
                 return {};
             case 1:
-                if (stage == ShaderStage::Vertex && !device.HasVertexViewportLayer()) {
+                if (IsVertexShader(stage) && !device.HasVertexViewportLayer()) {
                     return {};
                 }
                 return std::make_pair("gl_Layer", true);
             case 2:
-                if (stage == ShaderStage::Vertex && !device.HasVertexViewportLayer()) {
+                if (IsVertexShader(stage) && !device.HasVertexViewportLayer()) {
                     return {};
                 }
                 return std::make_pair("gl_ViewportIndex", true);
@@ -1069,6 +1080,9 @@ private:
             target = result->first;
             is_integer = result->second;
         } else if (const auto lmem = std::get_if<LmemNode>(&*dest)) {
+            if (stage == ProgramType::Compute) {
+                LOG_WARNING(Render_OpenGL, "Local memory is stubbed on compute shaders");
+            }
             target = fmt::format("{}[ftou({}) / 4]", GetLocalMemory(), Visit(lmem->GetAddress()));
         } else if (const auto gmem = std::get_if<GmemNode>(&*dest)) {
             const std::string real = Visit(gmem->GetRealAddress());
@@ -1637,7 +1651,7 @@ private:
     }
 
     std::string Exit(Operation operation) {
-        if (stage != ShaderStage::Fragment) {
+        if (stage != ProgramType::Fragment) {
             code.AddLine("return;");
             return {};
         }
@@ -1688,7 +1702,7 @@ private:
     }
 
     std::string EmitVertex(Operation operation) {
-        ASSERT_MSG(stage == ShaderStage::Geometry,
+        ASSERT_MSG(stage == ProgramType::Geometry,
                    "EmitVertex is expected to be used in a geometry shader.");
 
         // If a geometry shader is attached, it will always flip (it's the last stage before
@@ -1699,7 +1713,7 @@ private:
     }
 
     std::string EndPrimitive(Operation operation) {
-        ASSERT_MSG(stage == ShaderStage::Geometry,
+        ASSERT_MSG(stage == ProgramType::Geometry,
                    "EndPrimitive is expected to be used in a geometry shader.");
 
         code.AddLine("EndPrimitive();");
@@ -1721,6 +1735,48 @@ private:
         return "utof(gl_WorkGroupID"s + GetSwizzle(element) + ')';
     }
 
+    std::string BallotThread(Operation operation) {
+        const std::string value = VisitOperand(operation, 0, Type::Bool);
+        if (!device.HasWarpIntrinsics()) {
+            LOG_ERROR(Render_OpenGL,
+                      "Nvidia warp intrinsics are not available and its required by a shader");
+            // Stub on non-Nvidia devices by simulating all threads voting the same as the active
+            // one.
+            return fmt::format("utof({} ? 0xFFFFFFFFU : 0U)", value);
+        }
+        return fmt::format("utof(ballotThreadNV({}))", value);
+    }
+
+    std::string Vote(Operation operation, const char* func) {
+        const std::string value = VisitOperand(operation, 0, Type::Bool);
+        if (!device.HasWarpIntrinsics()) {
+            LOG_ERROR(Render_OpenGL,
+                      "Nvidia vote intrinsics are not available and its required by a shader");
+            // Stub with a warp size of one.
+            return value;
+        }
+        return fmt::format("{}({})", func, value);
+    }
+
+    std::string VoteAll(Operation operation) {
+        return Vote(operation, "allThreadsNV");
+    }
+
+    std::string VoteAny(Operation operation) {
+        return Vote(operation, "anyThreadNV");
+    }
+
+    std::string VoteEqual(Operation operation) {
+        if (!device.HasWarpIntrinsics()) {
+            LOG_ERROR(Render_OpenGL,
+                      "Nvidia vote intrinsics are not available and its required by a shader");
+            // We must return true here since a stub for a theoretical warp size of 1 will always
+            // return an equal result for all its votes.
+            return "true";
+        }
+        return Vote(operation, "allThreadsEqualNV");
+    }
+
     static constexpr std::array operation_decompilers = {
         &GLSLDecompiler::Assign,
 
@@ -1871,6 +1927,11 @@ private:
         &GLSLDecompiler::WorkGroupId<0>,
         &GLSLDecompiler::WorkGroupId<1>,
         &GLSLDecompiler::WorkGroupId<2>,
+
+        &GLSLDecompiler::BallotThread,
+        &GLSLDecompiler::VoteAll,
+        &GLSLDecompiler::VoteAny,
+        &GLSLDecompiler::VoteEqual,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
 
@@ -1937,7 +1998,7 @@ private:
     }
 
     u32 GetNumPhysicalInputAttributes() const {
-        return stage == ShaderStage::Vertex ? GetNumPhysicalAttributes() : GetNumPhysicalVaryings();
+        return IsVertexShader(stage) ? GetNumPhysicalAttributes() : GetNumPhysicalVaryings();
     }
 
     u32 GetNumPhysicalAttributes() const {
@@ -1950,7 +2011,7 @@ private:
 
     const Device& device;
     const ShaderIR& ir;
-    const ShaderStage stage;
+    const ProgramType stage;
     const std::string suffix;
     const Header header;
 
@@ -1981,7 +2042,7 @@ std::string GetCommonDeclarations() {
         MAX_CONSTBUFFER_ELEMENTS);
 }
 
-ProgramResult Decompile(const Device& device, const ShaderIR& ir, Maxwell::ShaderStage stage,
+ProgramResult Decompile(const Device& device, const ShaderIR& ir, ProgramType stage,
                         const std::string& suffix) {
     GLSLDecompiler decompiler(device, ir, stage, suffix);
     decompiler.Decompile();

src/video_core/renderer_opengl/gl_shader_decompiler.h

@@ -12,14 +12,26 @@
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/shader/shader_ir.h"
 
-namespace OpenGL {
-class Device;
-}
-
 namespace VideoCommon::Shader {
 class ShaderIR;
 }
 
+namespace OpenGL {
+
+class Device;
+
+enum class ProgramType : u32 {
+    VertexA = 0,
+    VertexB = 1,
+    TessellationControl = 2,
+    TessellationEval = 3,
+    Geometry = 4,
+    Fragment = 5,
+    Compute = 6
+};
+
+} // namespace OpenGL
+
 namespace OpenGL::GLShader {
 
 struct ShaderEntries;
@@ -85,6 +97,6 @@ struct ShaderEntries {
 std::string GetCommonDeclarations();
 
 ProgramResult Decompile(const Device& device, const VideoCommon::Shader::ShaderIR& ir,
-                        Maxwell::ShaderStage stage, const std::string& suffix);
+                        ProgramType stage, const std::string& suffix);
 
 } // namespace OpenGL::GLShader

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -51,7 +51,7 @@ ShaderCacheVersionHash GetShaderCacheVersionHash() {
 
 } // namespace
 
-ShaderDiskCacheRaw::ShaderDiskCacheRaw(u64 unique_identifier, Maxwell::ShaderProgram program_type,
+ShaderDiskCacheRaw::ShaderDiskCacheRaw(u64 unique_identifier, ProgramType program_type,
                                        u32 program_code_size, u32 program_code_size_b,
                                        ProgramCode program_code, ProgramCode program_code_b)
     : unique_identifier{unique_identifier}, program_type{program_type},

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -18,7 +18,6 @@
 #include "common/assert.h"
 #include "common/common_types.h"
 #include "core/file_sys/vfs_vector.h"
-#include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_opengl/gl_shader_gen.h"
 
 namespace Core {
@@ -34,14 +33,11 @@ namespace OpenGL {
 struct ShaderDiskCacheUsage;
 struct ShaderDiskCacheDump;
 
-using ShaderDumpsMap = std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump>;
-
 using ProgramCode = std::vector<u64>;
-using Maxwell = Tegra::Engines::Maxwell3D::Regs;
-
+using ShaderDumpsMap = std::unordered_map<ShaderDiskCacheUsage, ShaderDiskCacheDump>;
 using TextureBufferUsage = std::bitset<64>;
 
-/// Allocated bindings used by an OpenGL shader program.
+/// Allocated bindings used by an OpenGL shader program
 struct BaseBindings {
     u32 cbuf{};
     u32 gmem{};
@@ -126,7 +122,7 @@ namespace OpenGL {
 /// Describes a shader how it's used by the guest GPU
 class ShaderDiskCacheRaw {
 public:
-    explicit ShaderDiskCacheRaw(u64 unique_identifier, Maxwell::ShaderProgram program_type,
+    explicit ShaderDiskCacheRaw(u64 unique_identifier, ProgramType program_type,
                                 u32 program_code_size, u32 program_code_size_b,
                                 ProgramCode program_code, ProgramCode program_code_b);
     ShaderDiskCacheRaw();
@@ -141,30 +137,13 @@ public:
     }
 
     bool HasProgramA() const {
-        return program_type == Maxwell::ShaderProgram::VertexA;
+        return program_type == ProgramType::VertexA;
     }
 
-    Maxwell::ShaderProgram GetProgramType() const {
+    ProgramType GetProgramType() const {
         return program_type;
     }
 
-    Maxwell::ShaderStage GetProgramStage() const {
-        switch (program_type) {
-        case Maxwell::ShaderProgram::VertexA:
-        case Maxwell::ShaderProgram::VertexB:
-            return Maxwell::ShaderStage::Vertex;
-        case Maxwell::ShaderProgram::TesselationControl:
-            return Maxwell::ShaderStage::TesselationControl;
-        case Maxwell::ShaderProgram::TesselationEval:
-            return Maxwell::ShaderStage::TesselationEval;
-        case Maxwell::ShaderProgram::Geometry:
-            return Maxwell::ShaderStage::Geometry;
-        case Maxwell::ShaderProgram::Fragment:
-            return Maxwell::ShaderStage::Fragment;
-        }
-        UNREACHABLE();
-    }
-
     const ProgramCode& GetProgramCode() const {
         return program_code;
     }
@@ -175,7 +154,7 @@ public:
 
 private:
     u64 unique_identifier{};
-    Maxwell::ShaderProgram program_type{};
+    ProgramType program_type{};
     u32 program_code_size{};
     u32 program_code_size_b{};