8 files changed, 183 insertions, 14 deletions

diff --git a/src/core/core.cpp b/src/core/core.cpp
index b74fd0a58..deefeb301 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"
@@ -54,6 +55,7 @@
 #include "video_core/renderer_base.h"
 #include "video_core/video_core.h"
 
+
 MICROPROFILE_DEFINE(ARM_CPU0, "ARM", "CPU 0", MP_RGB(255, 64, 64));
 MICROPROFILE_DEFINE(ARM_CPU1, "ARM", "CPU 1", MP_RGB(255, 64, 64));
 MICROPROFILE_DEFINE(ARM_CPU2, "ARM", "CPU 2", MP_RGB(255, 64, 64));
@@ -540,6 +542,9 @@ 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 +634,31 @@ 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..9c3d41d11
--- /dev/null
+++ b/src/core/gpu_dirty_memory_manager.h
@@ -0,0 +1,112 @@
+#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} {}
+
+    ~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((transform.address << Memory::YUZU_PAGEBITS) + 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(16) TransformAddress {
+        VAddr address;
+        u64 mask;
+    };
+
+    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 = ~0ULL, .mask = 0ULL};
+
+    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 & Memory::YUZU_PAGEMASK;
+        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 = address >> Memory::YUZU_PAGEBITS;
+        result.mask = CreateMask<u64>(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..60b246bdd 100644
--- a/src/core/memory.cpp
+++ b/src/core/memory.cpp
@@ -13,6 +13,7 @@
 #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 +679,7 @@ 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)); });
+            [&]() { system.CurrentGPUDirtyMemoryManager().Collect(GetInteger(vaddr), sizeof(T)); });
         if (ptr) {
             std::memcpy(ptr, &data, sizeof(T));
         }
@@ -692,7 +693,7 @@ 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)); });
+            [&]() { system.CurrentGPUDirtyMemoryManager().Collect(GetInteger(vaddr), sizeof(T)); });
         if (ptr) {
             const auto volatile_pointer = reinterpret_cast<volatile T*>(ptr);
             return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
@@ -707,7 +708,7 @@ 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)); });
+            [&]() { system.CurrentGPUDirtyMemoryManager().Collect(GetInteger(vaddr), sizeof(u128)); });
         if (ptr) {
             const auto volatile_pointer = reinterpret_cast<volatile u64*>(ptr);
             return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 58a45ab67..9239ad862 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -115,7 +115,21 @@ void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
 
 template <class P>
 void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
-    memory_tracker.CachedCpuWrite(cpu_addr, size);
+    const bool is_dirty = IsRegionRegistered(cpu_addr, size);
+    if (!is_dirty) {
+        return;
+    }
+    VAddr aligned_start = Common::AlignDown(cpu_addr, YUZU_PAGESIZE);
+    VAddr aligned_end = Common::AlignUp(cpu_addr + size, YUZU_PAGESIZE);
+    if (!IsRegionGpuModified(aligned_start, aligned_end - aligned_start)) {
+        WriteMemory(cpu_addr, size);
+        return;
+    }
+
+    tmp_buffer.resize_destructive(size);
+    cpu_memory.ReadBlockUnsafe(cpu_addr, tmp_buffer.data(), size);
+
+    InlineMemoryImplementation(cpu_addr, size, tmp_buffer);
 }
 
 template <class P>
@@ -1553,6 +1567,14 @@ bool BufferCache<P>::InlineMemory(VAddr dest_address, size_t copy_size,
         return false;
     }
 
+    InlineMemoryImplementation(dest_address, copy_size, inlined_buffer);
+
+    return true;
+}
+
+template <class P>
+void BufferCache<P>::InlineMemoryImplementation(VAddr dest_address, size_t copy_size,
+                                  std::span<const u8> inlined_buffer) {
     const IntervalType subtract_interval{dest_address, dest_address + copy_size};
     ClearDownload(subtract_interval);
     common_ranges.subtract(subtract_interval);
@@ -1574,8 +1596,6 @@ bool BufferCache<P>::InlineMemory(VAddr dest_address, size_t copy_size,
     } else {
         buffer.ImmediateUpload(buffer.Offset(dest_address), inlined_buffer.first(copy_size));
     }
-
-    return true;
 }
 
 template <class P>
diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h
index fe6068cfe..4d9bab7f7 100644
--- a/src/video_core/buffer_cache/buffer_cache_base.h
+++ b/src/video_core/buffer_cache/buffer_cache_base.h
@@ -543,6 +543,8 @@ private:
 
     void ClearDownload(IntervalType subtract_interval);
 
+    void InlineMemoryImplementation(VAddr dest_address, size_t copy_size, std::span<const u8> inlined_buffer);
+
     VideoCore::RasterizerInterface& rasterizer;
     Core::Memory::Memory& cpu_memory;
 
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index db385076d..f823a1e2b 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -95,7 +95,9 @@ struct GPU::Impl {
 
     /// Synchronizes CPU writes with Host GPU memory.
     void InvalidateGPUCache() {
-        rasterizer->InvalidateGPUCache();
+        std::function<void(VAddr, size_t)> callback_writes(
+            [this](VAddr address, size_t size) { rasterizer->OnCPUWrite(address, size); });
+        system.GatherGPUDirtyMemory(callback_writes);
     }
 
     /// Signal the ending of command list.
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index f7c0d939a..a63a29e61 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -570,7 +570,7 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
     if (addr == 0 || size == 0) {
         return;
     }
-    pipeline_cache.OnCPUWrite(addr, size);
+
     {
         std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
@@ -579,14 +579,11 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
         std::scoped_lock lock{buffer_cache.mutex};
         buffer_cache.CachedWriteMemory(addr, size);
     }
+    pipeline_cache.InvalidateRegion(addr, size);
 }
 
 void RasterizerVulkan::InvalidateGPUCache() {
-    pipeline_cache.SyncGuestHost();
-    {
-        std::scoped_lock lock{buffer_cache.mutex};
-        buffer_cache.FlushCachedWrites();
-    }
+    gpu.InvalidateGPUCache();
 }
 
 void RasterizerVulkan::UnmapMemory(VAddr addr, u64 size) {