5 files changed, 202 insertions, 9 deletions

diff --git a/src/core/core.cpp b/src/core/core.cpp
index b74fd0a58..9e3eb3795 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -27,6 +27,7 @@
 #include "core/file_sys/savedata_factory.h"
 #include "core/file_sys/vfs_concat.h"
 #include "core/file_sys/vfs_real.h"
+#include "core/gpu_dirty_memory_manager.h"
 #include "core/hid/hid_core.h"
 #include "core/hle/kernel/k_memory_manager.h"
 #include "core/hle/kernel/k_process.h"
@@ -130,7 +131,10 @@ FileSys::VirtualFile GetGameFileFromPath(const FileSys::VirtualFilesystem& vfs,
 struct System::Impl {
     explicit Impl(System& system)
         : kernel{system}, fs_controller{system}, memory{system}, hid_core{}, room_network{},
-          cpu_manager{system}, reporter{system}, applet_manager{system}, time_manager{system} {}
+          cpu_manager{system}, reporter{system}, applet_manager{system}, time_manager{system},
+          gpu_dirty_memory_write_manager{} {
+        memory.SetGPUDirtyManagers(gpu_dirty_memory_write_manager);
+    }
 
     void Initialize(System& system) {
         device_memory = std::make_unique<Core::DeviceMemory>();
@@ -234,6 +238,8 @@ struct System::Impl {
         // Setting changes may require a full system reinitialization (e.g., disabling multicore).
         ReinitializeIfNecessary(system);
 
+        memory.SetGPUDirtyManagers(gpu_dirty_memory_write_manager);
+
         kernel.Initialize();
         cpu_manager.Initialize();
 
@@ -540,6 +546,9 @@ struct System::Impl {
 
     std::array<u64, Core::Hardware::NUM_CPU_CORES> dynarmic_ticks{};
     std::array<MicroProfileToken, Core::Hardware::NUM_CPU_CORES> microprofile_cpu{};
+
+    std::array<Core::GPUDirtyMemoryManager, Core::Hardware::NUM_CPU_CORES>
+        gpu_dirty_memory_write_manager{};
 };
 
 System::System() : impl{std::make_unique<Impl>(*this)} {}
@@ -629,10 +638,31 @@ void System::PrepareReschedule(const u32 core_index) {
     impl->kernel.PrepareReschedule(core_index);
 }
 
+Core::GPUDirtyMemoryManager& System::CurrentGPUDirtyMemoryManager() {
+    const std::size_t core = impl->kernel.GetCurrentHostThreadID();
+    return impl->gpu_dirty_memory_write_manager[core < Core::Hardware::NUM_CPU_CORES
+                                                    ? core
+                                                    : Core::Hardware::NUM_CPU_CORES - 1];
+}
+
+/// Provides a constant reference to the current gou dirty memory manager.
+const Core::GPUDirtyMemoryManager& System::CurrentGPUDirtyMemoryManager() const {
+    const std::size_t core = impl->kernel.GetCurrentHostThreadID();
+    return impl->gpu_dirty_memory_write_manager[core < Core::Hardware::NUM_CPU_CORES
+                                                    ? core
+                                                    : Core::Hardware::NUM_CPU_CORES - 1];
+}
+
 size_t System::GetCurrentHostThreadID() const {
     return impl->kernel.GetCurrentHostThreadID();
 }
 
+void System::GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback) {
+    for (auto& manager : impl->gpu_dirty_memory_write_manager) {
+        manager.Gather(callback);
+    }
+}
+
 PerfStatsResults System::GetAndResetPerfStats() {
     return impl->GetAndResetPerfStats();
 }
diff --git a/src/core/core.h b/src/core/core.h
index 93afc9303..14b2f7785 100644
--- a/src/core/core.h
+++ b/src/core/core.h
@@ -108,9 +108,10 @@ class CpuManager;
 class Debugger;
 class DeviceMemory;
 class ExclusiveMonitor;
-class SpeedLimiter;
+class GPUDirtyMemoryManager;
 class PerfStats;
 class Reporter;
+class SpeedLimiter;
 class TelemetrySession;
 
 struct PerfStatsResults;
@@ -225,6 +226,14 @@ public:
     /// Prepare the core emulation for a reschedule
     void PrepareReschedule(u32 core_index);
 
+    /// Provides a reference to the gou dirty memory manager.
+    [[nodiscard]] Core::GPUDirtyMemoryManager& CurrentGPUDirtyMemoryManager();
+
+    /// Provides a constant reference to the current gou dirty memory manager.
+    [[nodiscard]] const Core::GPUDirtyMemoryManager& CurrentGPUDirtyMemoryManager() const;
+
+    void GatherGPUDirtyMemory(std::function<void(VAddr, size_t)>& callback);
+
     [[nodiscard]] size_t GetCurrentHostThreadID() const;
 
     /// Gets and resets core performance statistics
diff --git a/src/core/gpu_dirty_memory_manager.h b/src/core/gpu_dirty_memory_manager.h
new file mode 100644
index 000000000..9687531e8
--- /dev/null
+++ b/src/core/gpu_dirty_memory_manager.h
@@ -0,0 +1,122 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <atomic>
+#include <bit>
+#include <functional>
+#include <mutex>
+#include <utility>
+#include <vector>
+
+#include "core/memory.h"
+
+namespace Core {
+
+class GPUDirtyMemoryManager {
+public:
+    GPUDirtyMemoryManager() : current{default_transform} {
+        back_buffer.reserve(256);
+        front_buffer.reserve(256);
+    }
+
+    ~GPUDirtyMemoryManager() = default;
+
+    void Collect(VAddr address, size_t size) {
+        TransformAddress t = BuildTransform(address, size);
+        TransformAddress tmp, original;
+        do {
+            tmp = current.load(std::memory_order_acquire);
+            original = tmp;
+            if (tmp.address != t.address) {
+                if (IsValid(tmp.address)) {
+                    std::scoped_lock lk(guard);
+                    back_buffer.emplace_back(tmp);
+                    current.exchange(t, std::memory_order_relaxed);
+                    return;
+                }
+                tmp.address = t.address;
+                tmp.mask = 0;
+            }
+            if ((tmp.mask | t.mask) == tmp.mask) {
+                return;
+            }
+            tmp.mask |= t.mask;
+        } while (!current.compare_exchange_weak(original, tmp, std::memory_order_release,
+                                                std::memory_order_relaxed));
+    }
+
+    void Gather(std::function<void(VAddr, size_t)>& callback) {
+        {
+            std::scoped_lock lk(guard);
+            TransformAddress t = current.exchange(default_transform, std::memory_order_relaxed);
+            front_buffer.swap(back_buffer);
+            if (IsValid(t.address)) {
+                front_buffer.emplace_back(t);
+            }
+        }
+        for (auto& transform : front_buffer) {
+            size_t offset = 0;
+            u64 mask = transform.mask;
+            while (mask != 0) {
+                const size_t empty_bits = std::countr_zero(mask);
+                offset += empty_bits << align_bits;
+                mask = mask >> empty_bits;
+
+                const size_t continuous_bits = std::countr_one(mask);
+                callback((static_cast<VAddr>(transform.address) << page_bits) + offset,
+                         continuous_bits << align_bits);
+                mask = continuous_bits < align_size ? (mask >> continuous_bits) : 0;
+                offset += continuous_bits << align_bits;
+            }
+        }
+        front_buffer.clear();
+    }
+
+private:
+    struct alignas(8) TransformAddress {
+        u32 address;
+        u32 mask;
+    };
+
+    constexpr static size_t page_bits = Memory::YUZU_PAGEBITS - 1;
+    constexpr static size_t page_size = 1ULL << page_bits;
+    constexpr static size_t page_mask = page_size - 1;
+
+    constexpr static size_t align_bits = 6U;
+    constexpr static size_t align_size = 1U << align_bits;
+    constexpr static size_t align_mask = align_size - 1;
+    constexpr static TransformAddress default_transform = {.address = ~0U, .mask = 0U};
+
+    bool IsValid(VAddr address) {
+        return address < (1ULL << 39);
+    }
+
+    template <typename T>
+    T CreateMask(size_t top_bit, size_t minor_bit) {
+        T mask = ~T(0);
+        mask <<= (sizeof(T) * 8 - top_bit);
+        mask >>= (sizeof(T) * 8 - top_bit);
+        mask >>= minor_bit;
+        mask <<= minor_bit;
+        return mask;
+    }
+
+    TransformAddress BuildTransform(VAddr address, size_t size) {
+        const size_t minor_address = address & page_mask;
+        const size_t minor_bit = minor_address >> align_bits;
+        const size_t top_bit = (minor_address + size + align_mask) >> align_bits;
+        TransformAddress result{};
+        result.address = static_cast<u32>(address >> page_bits);
+        result.mask = CreateMask<u32>(top_bit, minor_bit);
+        return result;
+    }
+
+    std::atomic<TransformAddress> current{};
+    std::mutex guard;
+    std::vector<TransformAddress> back_buffer;
+    std::vector<TransformAddress> front_buffer;
+};
+
+} // namespace Core
diff --git a/src/core/memory.cpp b/src/core/memory.cpp
index 514ba0d66..257406f09 100644
--- a/src/core/memory.cpp
+++ b/src/core/memory.cpp
@@ -3,6 +3,7 @@
 
 #include <algorithm>
 #include <cstring>
+#include <span>
 
 #include "common/assert.h"
 #include "common/atomic_ops.h"
@@ -13,6 +14,7 @@
 #include "common/swap.h"
 #include "core/core.h"
 #include "core/device_memory.h"
+#include "core/gpu_dirty_memory_manager.h"
 #include "core/hardware_properties.h"
 #include "core/hle/kernel/k_page_table.h"
 #include "core/hle/kernel/k_process.h"
@@ -678,7 +680,7 @@ struct Memory::Impl {
                 LOG_ERROR(HW_Memory, "Unmapped Write{} @ 0x{:016X} = 0x{:016X}", sizeof(T) * 8,
                           GetInteger(vaddr), static_cast<u64>(data));
             },
-            [&]() { system.GPU().InvalidateRegion(GetInteger(vaddr), sizeof(T)); });
+            [&]() { HandleRasterizerWrite(GetInteger(vaddr), sizeof(T)); });
         if (ptr) {
             std::memcpy(ptr, &data, sizeof(T));
         }
@@ -692,7 +694,7 @@ struct Memory::Impl {
                 LOG_ERROR(HW_Memory, "Unmapped WriteExclusive{} @ 0x{:016X} = 0x{:016X}",
                           sizeof(T) * 8, GetInteger(vaddr), static_cast<u64>(data));
             },
-            [&]() { system.GPU().InvalidateRegion(GetInteger(vaddr), sizeof(T)); });
+            [&]() { HandleRasterizerWrite(GetInteger(vaddr), sizeof(T)); });
         if (ptr) {
             const auto volatile_pointer = reinterpret_cast<volatile T*>(ptr);
             return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
@@ -707,7 +709,7 @@ struct Memory::Impl {
                 LOG_ERROR(HW_Memory, "Unmapped WriteExclusive128 @ 0x{:016X} = 0x{:016X}{:016X}",
                           GetInteger(vaddr), static_cast<u64>(data[1]), static_cast<u64>(data[0]));
             },
-            [&]() { system.GPU().InvalidateRegion(GetInteger(vaddr), sizeof(u128)); });
+            [&]() { HandleRasterizerWrite(GetInteger(vaddr), sizeof(u128)); });
         if (ptr) {
             const auto volatile_pointer = reinterpret_cast<volatile u64*>(ptr);
             return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
@@ -717,7 +719,7 @@ struct Memory::Impl {
 
     void HandleRasterizerDownload(VAddr address, size_t size) {
         const size_t core = system.GetCurrentHostThreadID();
-        auto& current_area = rasterizer_areas[core];
+        auto& current_area = rasterizer_read_areas[core];
         const VAddr end_address = address + size;
         if (current_area.start_address <= address && end_address <= current_area.end_address)
             [[likely]] {
@@ -726,9 +728,31 @@ struct Memory::Impl {
         current_area = system.GPU().OnCPURead(address, size);
     }
 
-    Common::PageTable* current_page_table = nullptr;
-    std::array<VideoCore::RasterizerDownloadArea, Core::Hardware::NUM_CPU_CORES> rasterizer_areas{};
+    void HandleRasterizerWrite(VAddr address, size_t size) {
+        const size_t core = system.GetCurrentHostThreadID();
+        auto& current_area = rasterizer_write_areas[core];
+        VAddr subaddress = address >> YUZU_PAGEBITS;
+        bool do_collection = current_area.last_address == subaddress;
+        if (!do_collection) [[unlikely]] {
+            do_collection = system.GPU().OnCPUWrite(address, size);
+            if (!do_collection) {
+                return;
+            }
+            current_area.last_address = subaddress;
+        }
+        gpu_dirty_managers[core].Collect(address, size);
+    }
+
+    struct GPUDirtyState {
+        VAddr last_address;
+    };
+
     Core::System& system;
+    Common::PageTable* current_page_table = nullptr;
+    std::array<VideoCore::RasterizerDownloadArea, Core::Hardware::NUM_CPU_CORES>
+        rasterizer_read_areas{};
+    std::array<GPUDirtyState, Core::Hardware::NUM_CPU_CORES> rasterizer_write_areas{};
+    std::span<Core::GPUDirtyMemoryManager> gpu_dirty_managers;
 };
 
 Memory::Memory(Core::System& system_) : system{system_} {
@@ -876,6 +900,10 @@ void Memory::ZeroBlock(Common::ProcessAddress dest_addr, const std::size_t size)
     impl->ZeroBlock(*system.ApplicationProcess(), dest_addr, size);
 }
 
+void Memory::SetGPUDirtyManagers(std::span<Core::GPUDirtyMemoryManager> managers) {
+    impl->gpu_dirty_managers = managers;
+}
+
 Result Memory::InvalidateDataCache(Common::ProcessAddress dest_addr, const std::size_t size) {
     return impl->InvalidateDataCache(*system.ApplicationProcess(), dest_addr, size);
 }
diff --git a/src/core/memory.h b/src/core/memory.h
index 72a0be813..ea01824f8 100644
--- a/src/core/memory.h
+++ b/src/core/memory.h
@@ -5,6 +5,7 @@
 
 #include <cstddef>
 #include <memory>
+#include <span>
 #include <string>
 #include "common/typed_address.h"
 #include "core/hle/result.h"
@@ -15,7 +16,8 @@ struct PageTable;
 
 namespace Core {
 class System;
-}
+class GPUDirtyMemoryManager;
+} // namespace Core
 
 namespace Kernel {
 class PhysicalMemory;
@@ -458,6 +460,8 @@ public:
      */
     void MarkRegionDebug(Common::ProcessAddress vaddr, u64 size, bool debug);
 
+    void SetGPUDirtyManagers(std::span<Core::GPUDirtyMemoryManager> managers);
+
 private:
     Core::System& system;