Memory Tracking: Optimize tracking to only use atomic writes when contested with the host GPU

src/core/core.cpp

@@ -55,7 +55,6 @@
 #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));
@@ -132,7 +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>();
@@ -236,6 +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();
 

src/core/gpu_dirty_memory_manager.h

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 #pragma once
 
 #include <atomic>
@@ -59,8 +62,7 @@ public:
                 mask = mask >> empty_bits;
 
                 const size_t continuous_bits = std::countr_one(mask);
-                callback((transform.address << Memory::YUZU_PAGEBITS) + offset,
-                         continuous_bits << align_bits);
+                callback((transform.address << page_bits) + offset, continuous_bits << align_bits);
                 mask = continuous_bits < align_size ? (mask >> continuous_bits) : 0;
                 offset += continuous_bits << align_bits;
             }
@@ -74,6 +76,10 @@ private:
         u64 mask;
     };
 
+    constexpr static size_t page_bits = Memory::YUZU_PAGEBITS;
+    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;
@@ -94,11 +100,11 @@ private:
     }
 
     TransformAddress BuildTransform(VAddr address, size_t size) {
-        const size_t minor_address = address & Memory::YUZU_PAGEMASK;
+        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 = address >> Memory::YUZU_PAGEBITS;
+        result.address = address >> page_bits;
         result.mask = CreateMask<u64>(top_bit, minor_bit);
         return result;
     }

src/core/memory.cpp

@@ -3,6 +3,7 @@
 
 #include <algorithm>
 #include <cstring>
+#include <span>
 
 #include "common/assert.h"
 #include "common/atomic_ops.h"
@@ -679,7 +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.CurrentGPUDirtyMemoryManager().Collect(GetInteger(vaddr), sizeof(T)); });
+            [&]() { HandleRasterizerWrite(GetInteger(vaddr), sizeof(T)); });
         if (ptr) {
             std::memcpy(ptr, &data, sizeof(T));
         }
@@ -693,7 +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.CurrentGPUDirtyMemoryManager().Collect(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);
@@ -708,7 +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.CurrentGPUDirtyMemoryManager().Collect(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);
@@ -718,7 +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]] {
@@ -727,9 +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_} {
@@ -877,6 +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);
 }

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;