[buffer_cache] Add batching support for memory tracker updates (#3288)

I added a batching/ coalescing of ranges in WordManager to reduce calls per pages in UpdatePagesCachedCount, also a test to verify if FlushCachedWrites coalesced (reduces callings to UpdatePagesCachedCount) callings and register each of them to inspect them.

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/3288
Reviewed-by: Maufeat <sahyno1996@gmail.com>
Reviewed-by: DraVee <dravee@eden-emu.dev>
Co-authored-by: CamilleLaVey <camillelavey99@gmail.com>
Co-committed-by: CamilleLaVey <camillelavey99@gmail.com>

src/core/device_memory_manager.h

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -9,6 +12,7 @@
 #include <deque>
 #include <memory>
 #include <mutex>
+#include <vector>
 
 #include "common/common_types.h"
 #include "common/range_mutex.h"
@@ -44,6 +48,7 @@ public:
     ~DeviceMemoryManager();
 
     static constexpr bool HAS_FLUSH_INVALIDATION = true;
+    static constexpr size_t AS_BITS = Traits::device_virtual_bits;
 
     void BindInterface(DeviceInterface* device_inter);
 
@@ -117,7 +122,12 @@ public:
 
     void UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta);
 
-    static constexpr size_t AS_BITS = Traits::device_virtual_bits;
+    // New batch API to update multiple ranges with a single lock acquisition.
+    void UpdatePagesCachedBatch(const std::vector<std::pair<DAddr, size_t>>& ranges, s32 delta);
+
+private:
+    // Internal helper that performs the update assuming the caller already holds the necessary lock.
+    void UpdatePagesCachedCountNoLock(DAddr addr, size_t size, s32 delta);
 
 private:
     static constexpr size_t device_virtual_bits = Traits::device_virtual_bits;
@@ -214,6 +224,8 @@ private:
     std::unique_ptr<CachedPages> cached_pages;
     Common::RangeMutex counter_guard;
     std::mutex mapping_guard;
+
+
 };
 
 } // namespace Core

src/core/device_memory_manager.inc

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -5,6 +8,8 @@
 #include <limits>
 #include <memory>
 #include <type_traits>
+#include <algorithm>
+#include <vector>
 
 #include "common/address_space.h"
 #include "common/address_space.inc"
@@ -507,8 +512,7 @@ void DeviceMemoryManager<Traits>::UnregisterProcess(Asid asid) {
 }
 
 template <typename Traits>
-void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
-    Common::ScopedRangeLock lk(counter_guard, addr, size);
+void DeviceMemoryManager<Traits>::UpdatePagesCachedCountNoLock(DAddr addr, size_t size, s32 delta) {
     u64 uncache_begin = 0;
     u64 cache_begin = 0;
     u64 uncache_bytes = 0;
@@ -586,4 +590,47 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
     release_pending();
 }
 
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
+    Common::ScopedRangeLock lk(counter_guard, addr, size);
+    UpdatePagesCachedCountNoLock(addr, size, delta);
+} 
+
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdatePagesCachedBatch(const std::vector<std::pair<DAddr, size_t>>& ranges, s32 delta) {
+    if (ranges.empty()) {
+        return;
+    }
+    // Make a local copy and sort by address
+    std::vector<std::pair<DAddr, size_t>> tmp = ranges;
+    std::sort(tmp.begin(), tmp.end(), [](const auto& a, const auto& b) { return a.first < b.first; });
+
+    // Coalesce adjacent/overlapping ranges
+    std::vector<std::pair<DAddr, size_t>> coalesced;
+    DAddr cur_addr = tmp[0].first;
+    size_t cur_size = tmp[0].second;
+    for (size_t i = 1; i < tmp.size(); ++i) {
+        DAddr next_addr = tmp[i].first;
+        size_t next_size = tmp[i].second;
+        if (cur_addr + cur_size >= next_addr) {
+            // overlapping or contiguous
+            const DAddr end = std::max(cur_addr + cur_size, next_addr + next_size);
+            cur_size = end - cur_addr;
+        } else {
+            coalesced.emplace_back(cur_addr, cur_size);
+            cur_addr = next_addr;
+            cur_size = next_size;
+        }
+    }
+    coalesced.emplace_back(cur_addr, cur_size);
+
+    const DAddr lock_begin = coalesced.front().first;
+    const DAddr lock_end = coalesced.back().first + coalesced.back().second;
+    Common::ScopedRangeLock lk(counter_guard, lock_begin, static_cast<size_t>(lock_end - lock_begin));
+
+    for (const auto& [addr, size] : coalesced) {
+        UpdatePagesCachedCountNoLock(addr, size, delta);
+    }
+}
+
 } // namespace Core

src/tests/video_core/memory_tracker.cpp

@@ -4,11 +4,15 @@
 #include <memory>
 #include <stdexcept>
 #include <unordered_map>
+#include <tuple>
+#include <vector>
 
 #include <catch2/catch_test_macros.hpp>
 
 #include "common/common_types.h"
 #include "video_core/buffer_cache/memory_tracker_base.h"
+#include "core/device_memory.h"
+#include "video_core/host1x/gpu_device_memory_manager.h"
 
 namespace {
 using Range = std::pair<u64, u64>;
@@ -23,6 +27,8 @@ constexpr VAddr c = 16 * HIGH_PAGE_SIZE;
 class RasterizerInterface {
 public:
     void UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
+        ++update_calls;
+        calls.emplace_back(addr, size, delta);
         const u64 page_start{addr >> Core::DEVICE_PAGEBITS};
         const u64 page_end{(addr + size + Core::DEVICE_PAGESIZE - 1) >> Core::DEVICE_PAGEBITS};
         for (u64 page = page_start; page < page_end; ++page) {
@@ -36,6 +42,9 @@ public:
         }
     }
 
+    [[nodiscard]] size_t UpdateCalls() const noexcept { return update_calls; }
+    [[nodiscard]] const std::vector<std::tuple<VAddr, u64, int>>& UpdateCallsList() const noexcept { return calls; }
+
     [[nodiscard]] int Count(VAddr addr) const noexcept {
         const auto it = page_table.find(addr >> Core::DEVICE_PAGEBITS);
         return it == page_table.end() ? 0 : it->second;
@@ -51,7 +60,10 @@ public:
 
 private:
     std::unordered_map<u64, int> page_table;
+    std::vector<std::tuple<VAddr, u64, int>> calls;
+    size_t update_calls = 0;
 };
+
 } // Anonymous namespace
 
 using MemoryTracker = VideoCommon::MemoryTrackerBase<RasterizerInterface>;
@@ -544,3 +556,34 @@ TEST_CASE("MemoryTracker: Cached write downloads") {
     memory_track->MarkRegionAsCpuModified(c, WORD);
     REQUIRE(rasterizer.Count() == 0);
 }
+
+TEST_CASE("MemoryTracker: FlushCachedWrites batching") {
+    RasterizerInterface rasterizer;
+    std::unique_ptr<MemoryTracker> memory_track(std::make_unique<MemoryTracker>(rasterizer));
+    memory_track->UnmarkRegionAsCpuModified(c, WORD * 2);
+    memory_track->CachedCpuWrite(c + PAGE, PAGE);
+    memory_track->CachedCpuWrite(c + PAGE * 2, PAGE);
+    memory_track->CachedCpuWrite(c + PAGE * 4, PAGE);
+    REQUIRE(rasterizer.UpdateCalls() == 0);
+    memory_track->FlushCachedWrites();
+    // Now we expect a single batch call (coalesced ranges) to the device memory manager
+    REQUIRE(rasterizer.UpdateCalls() == 1);
+    const auto& calls = rasterizer.UpdateCallsList();
+    REQUIRE(std::get<0>(calls[0]) == c + PAGE);
+    REQUIRE(std::get<1>(calls[0]) == PAGE * 3);
+}
+
+TEST_CASE("DeviceMemoryManager: UpdatePagesCachedBatch basic") {
+    Core::DeviceMemory device_memory;
+    Tegra::MaxwellDeviceMemoryManager manager(device_memory);
+    // empty should be a no-op
+    std::vector<std::pair<Core::DAddr, size_t>> empty;
+    manager.UpdatePagesCachedBatch(empty, 1);
+
+    // small ranges should be accepted and not crash
+    std::vector<std::pair<Core::DAddr, size_t>> ranges;
+    ranges.emplace_back(0, Core::Memory::YUZU_PAGESIZE);
+    ranges.emplace_back(Core::Memory::YUZU_PAGESIZE, Core::Memory::YUZU_PAGESIZE);
+    manager.UpdatePagesCachedBatch(ranges, 1);
+    SUCCEED("UpdatePagesCachedBatch executed without error");
+}

src/video_core/buffer_cache/word_manager.h

@@ -11,6 +11,7 @@
 #include <limits>
 #include <span>
 #include <utility>
+#include <vector>
 
 #include "common/alignment.h"
 #include "common/common_funcs.h"
@@ -256,9 +257,10 @@ public:
         std::span<u64> state_words = words.template Span<type>();
         [[maybe_unused]] std::span<u64> untracked_words = words.template Span<Type::Untracked>();
         [[maybe_unused]] std::span<u64> cached_words = words.template Span<Type::CachedCPU>();
+        std::vector<std::pair<VAddr, u64>> ranges;
         IterateWords(dirty_addr - cpu_addr, size, [&](size_t index, u64 mask) {
             if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-                NotifyRasterizer<!enable>(index, untracked_words[index], mask);
+                CollectChangedRanges<(!enable)>(index, untracked_words[index], mask, ranges);
             }
             if constexpr (enable) {
                 state_words[index] |= mask;
@@ -279,6 +281,9 @@ public:
                 }
             }
         });
+        if (!ranges.empty()) {
+            ApplyCollectedRanges(ranges, (!enable) ? 1 : -1);
+        }
     }
 
     /**
@@ -304,6 +309,7 @@ public:
             func(cpu_addr + pending_offset * BYTES_PER_PAGE,
                  (pending_pointer - pending_offset) * BYTES_PER_PAGE);
         };
+        std::vector<std::pair<VAddr, u64>> ranges;
         IterateWords(offset, size, [&](size_t index, u64 mask) {
             if constexpr (type == Type::GPU) {
                 mask &= ~untracked_words[index];
@@ -311,7 +317,7 @@ public:
             const u64 word = state_words[index] & mask;
             if constexpr (clear) {
                 if constexpr (type == Type::CPU || type == Type::CachedCPU) {
-                    NotifyRasterizer<true>(index, untracked_words[index], mask);
+                    CollectChangedRanges<true>(index, untracked_words[index], mask, ranges);
                 }
                 state_words[index] &= ~mask;
                 if constexpr (type == Type::CPU || type == Type::CachedCPU) {
@@ -343,6 +349,9 @@ public:
         if (pending) {
             release();
         }
+        if (!ranges.empty()) {
+            ApplyCollectedRanges(ranges, 1);
+        }
     }
 
     /**
@@ -425,13 +434,17 @@ public:
         u64* const cached_words = Array<Type::CachedCPU>();
         u64* const untracked_words = Array<Type::Untracked>();
         u64* const cpu_words = Array<Type::CPU>();
+        std::vector<std::pair<VAddr, u64>> ranges;
         for (u64 word_index = 0; word_index < num_words; ++word_index) {
             const u64 cached_bits = cached_words[word_index];
-            NotifyRasterizer<false>(word_index, untracked_words[word_index], cached_bits);
+            CollectChangedRanges<false>(word_index, untracked_words[word_index], cached_bits, ranges);
             untracked_words[word_index] |= cached_bits;
             cpu_words[word_index] |= cached_bits;
             cached_words[word_index] = 0;
         }
+        if (!ranges.empty()) {
+            ApplyCollectedRanges(ranges, -1);
+        }
     }
 
 private:
@@ -470,6 +483,40 @@ private:
      *
      * @tparam add_to_tracker True when the tracker should start tracking the new pages
      */
+    template <bool add_to_tracker>
+    void CollectChangedRanges(u64 word_index, u64 current_bits, u64 new_bits,
+                              std::vector<std::pair<VAddr, u64>>& out_ranges) const {
+        u64 changed_bits = (add_to_tracker ? current_bits : ~current_bits) & new_bits;
+        VAddr addr = cpu_addr + word_index * BYTES_PER_WORD;
+        IteratePages(changed_bits, [&](size_t offset, size_t size) {
+            out_ranges.emplace_back(addr + offset * BYTES_PER_PAGE, size * BYTES_PER_PAGE);
+        });
+    }
+
+    void ApplyCollectedRanges(std::vector<std::pair<VAddr, u64>>& ranges, int delta) const {
+        if (ranges.empty()) return;
+        std::sort(ranges.begin(), ranges.end(),
+                  [](const auto& a, const auto& b) { return a.first < b.first; });
+        // Coalesce adjacent/contiguous ranges
+        std::vector<std::pair<VAddr, size_t>> coalesced;
+        coalesced.reserve(ranges.size());
+        VAddr cur_addr = ranges[0].first;
+        size_t cur_size = static_cast<size_t>(ranges[0].second);
+        for (size_t i = 1; i < ranges.size(); ++i) {
+            if (cur_addr + cur_size == ranges[i].first) {
+                cur_size += static_cast<size_t>(ranges[i].second);
+            } else {
+                coalesced.emplace_back(cur_addr, cur_size);
+                cur_addr = ranges[i].first;
+                cur_size = static_cast<size_t>(ranges[i].second);
+            }
+        }
+        coalesced.emplace_back(cur_addr, cur_size);
+        // Use batch API to reduce lock acquisitions and contention.
+        tracker->UpdatePagesCachedBatch(coalesced, delta);
+        ranges.clear();
+    }
+
     template <bool add_to_tracker>
     void NotifyRasterizer(u64 word_index, u64 current_bits, u64 new_bits) const {
         u64 changed_bits = (add_to_tracker ? current_bits : ~current_bits) & new_bits;