[device_memory, buffer_cache] Batching API

src/core/device_memory_manager.h

@@ -12,6 +12,7 @@
 #include <deque>
 #include <memory>
 #include <mutex>
+#include <vector>
 
 #include "common/common_types.h"
 #include "common/range_mutex.h"
@@ -120,19 +121,12 @@ public:
 
     void UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta);
 
-#if defined(YUZU_TESTS)
-    // Instrumentation getters for testing
-    [[nodiscard]] size_t UpdatePagesCachedCalls() const noexcept { return update_pages_cached_calls.load(std::memory_order_relaxed); }
-    [[nodiscard]] uint64_t UpdatePagesCachedTotalNs() const noexcept { return update_pages_cached_total_ns.load(std::memory_order_relaxed); }
-    [[nodiscard]] uint64_t UpdatePagesCachedMaxNs() const noexcept { return update_pages_cached_max_ns.load(std::memory_order_relaxed); }
-    [[nodiscard]] size_t UpdatePagesCachedTotalBytes() const noexcept { return update_pages_cached_total_bytes.load(std::memory_order_relaxed); }
-    void ResetUpdatePagesCachedMetrics() noexcept {
-        update_pages_cached_calls.store(0, std::memory_order_relaxed);
-        update_pages_cached_total_ns.store(0, std::memory_order_relaxed);
-        update_pages_cached_max_ns.store(0, std::memory_order_relaxed);
-        update_pages_cached_total_bytes.store(0, std::memory_order_relaxed);
-    }
-#endif
+    // 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);
 
     static constexpr size_t AS_BITS = Traits::device_virtual_bits;
 
@@ -232,13 +226,7 @@ private:
     Common::RangeMutex counter_guard;
     std::mutex mapping_guard;
 
-#if defined(YUZU_TESTS)
-    // Instrumentation counters for UpdatePagesCachedCount
-    mutable std::atomic_size_t update_pages_cached_calls{0};
-    mutable std::atomic<uint64_t> update_pages_cached_total_ns{0};
-    mutable std::atomic<uint64_t> update_pages_cached_max_ns{0};
-    mutable std::atomic_size_t update_pages_cached_total_bytes{0};
-#endif
+
 };
 
 } // namespace Core

src/core/device_memory_manager.inc

@@ -8,7 +8,8 @@
 #include <limits>
 #include <memory>
 #include <type_traits>
-#include <chrono>
+#include <algorithm>
+#include <vector>
 
 #include "common/address_space.h"
 #include "common/address_space.inc"
@@ -167,9 +168,6 @@ template <typename Traits>
 DeviceMemoryManager<Traits>::DeviceMemoryManager(const DeviceMemory& device_memory_)
     : physical_base{reinterpret_cast<const uintptr_t>(device_memory_.buffer.BackingBasePointer())},
       device_inter{nullptr}, compressed_physical_ptr(device_as_size >> Memory::YUZU_PAGEBITS),
-#if defined(YUZU_TESTS)
-      update_pages_cached_calls{0}, update_pages_cached_total_ns{0}, update_pages_cached_max_ns{0}, update_pages_cached_total_bytes{0},
-#endif
       compressed_device_addr(1ULL << ((Settings::values.memory_layout_mode.GetValue() ==
                                                Settings::MemoryLayout::Memory_4Gb
                                            ? physical_min_bits
@@ -514,11 +512,7 @@ void DeviceMemoryManager<Traits>::UnregisterProcess(Asid asid) {
 }
 
 template <typename Traits>
-void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
-#if defined(YUZU_TESTS)
-    const auto start_time = std::chrono::steady_clock::now();
-#endif
-    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;
@@ -594,19 +588,49 @@ 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));
 
-#if defined(YUZU_TESTS)
-    const auto end_time = std::chrono::steady_clock::now();
-    const uint64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count();
-    update_pages_cached_calls.fetch_add(1, std::memory_order_relaxed);
-    update_pages_cached_total_ns.fetch_add(ns, std::memory_order_relaxed);
-    update_pages_cached_total_bytes.fetch_add(size, std::memory_order_relaxed);
-    // Update max
-    uint64_t old_max = update_pages_cached_max_ns.load(std::memory_order_relaxed);
-    while (old_max < ns && !update_pages_cached_max_ns.compare_exchange_weak(old_max, ns, std::memory_order_relaxed)) {
-        // loop until updated
+    for (const auto& [addr, size] : coalesced) {
+        UpdatePagesCachedCountNoLock(addr, size, delta);
     }
-#endif
 }
 
 } // namespace Core

src/tests/video_core/memory_tracker.cpp

@@ -566,26 +566,24 @@ TEST_CASE("MemoryTracker: FlushCachedWrites batching") {
     memory_track->CachedCpuWrite(c + PAGE * 4, PAGE);
     REQUIRE(rasterizer.UpdateCalls() == 0);
     memory_track->FlushCachedWrites();
-    REQUIRE(rasterizer.UpdateCalls() == 2);
+    // 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 * 2);
-    REQUIRE(std::get<0>(calls[1]) == c + PAGE * 4);
-    REQUIRE(std::get<1>(calls[1]) == PAGE);
+    REQUIRE(std::get<1>(calls[0]) == PAGE * 3);
 }
 
-TEST_CASE("DeviceMemoryManager: UpdatePagesCachedCount instrumentation") {
+TEST_CASE("DeviceMemoryManager: UpdatePagesCachedBatch basic") {
     Core::DeviceMemory device_memory;
     Tegra::MaxwellDeviceMemoryManager manager(device_memory);
-#if defined(YUZU_TESTS)
-    manager.ResetUpdatePagesCachedMetrics();
-    REQUIRE(manager.UpdatePagesCachedCalls() == 0);
-    manager.UpdatePagesCachedCount(0, Core::Memory::YUZU_PAGESIZE, 1);
-    REQUIRE(manager.UpdatePagesCachedCalls() == 1);
-    REQUIRE(manager.UpdatePagesCachedTotalBytes() >= Core::Memory::YUZU_PAGESIZE);
-    REQUIRE(manager.UpdatePagesCachedTotalNs() > 0);
-    REQUIRE(manager.UpdatePagesCachedMaxNs() > 0);
-#else
-    SUCCEED("Instrumentation only available in test builds");
-#endif
+    // 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

@@ -503,12 +503,14 @@ private:
             if (cur_addr + cur_size == ranges[i].first) {
                 cur_size += ranges[i].second;
             } else {
-                tracker->UpdatePagesCachedCount(cur_addr, cur_size, delta);
+                coalesced.emplace_back(cur_addr, cur_size);
                 cur_addr = ranges[i].first;
                 cur_size = ranges[i].second;
             }
         }
-        tracker->UpdatePagesCachedCount(cur_addr, cur_size, delta);
+        coalesced.emplace_back(cur_addr, cur_size);
+        // Use batch API to reduce lock acquisitions and contention.
+        tracker->UpdatePagesCachedBatch(coalesced, delta);
         ranges.clear();
     }