[TEST] Caching for texture + pages on NCE

src/core/device_memory_manager.h

@@ -126,6 +126,10 @@ public:
     // New batch API to update multiple ranges with a single lock acquisition.
     void UpdatePagesCachedBatch(std::span<const std::pair<DAddr, size_t>> ranges, s32 delta);
 
+    void UpdateTexturePagesCount(DAddr addr, size_t size, s32 delta);
+
+    [[nodiscard]] bool IsRegionTextureCached(DAddr addr, size_t size) const noexcept;
+
 private:
     struct TranslationEntry {
         DAddr guest_page{};
@@ -234,6 +238,7 @@ private:
         (1ULL << (device_virtual_bits - page_bits)) / subentries;
     using CachedPages = std::array<CounterEntry, num_counter_entries>;
     std::unique_ptr<CachedPages> cached_pages;
+    std::unique_ptr<CachedPages> texture_cached_pages;
     Common::RangeMutex counter_guard;
     std::mutex mapping_guard;
 

src/core/device_memory_manager.inc

@@ -177,6 +177,7 @@ DeviceMemoryManager<Traits>::DeviceMemoryManager(const DeviceMemory& device_memo
 {
     impl = std::make_unique<DeviceMemoryManagerAllocator<Traits>>();
     cached_pages = std::make_unique<CachedPages>();
+    texture_cached_pages = std::make_unique<CachedPages>();
 
     const size_t total_virtual = device_as_size >> Memory::YUZU_PAGEBITS;
     for (size_t i = 0; i < total_virtual; i++) {
@@ -625,6 +626,28 @@ void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size
     UpdatePagesCachedCountNoLock(addr, size, delta);
 }
 
+template <typename Traits>
+void DeviceMemoryManager<Traits>::UpdateTexturePagesCount(DAddr addr, size_t size, s32 delta) {
+    Common::ScopedRangeLock lk(counter_guard, addr, size);
+    const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
+    for (size_t page = addr >> Memory::YUZU_PAGEBITS; page != page_end; ++page) {
+        CounterAtomicType& count = texture_cached_pages->at(page >> subentries_shift).Count(page);
+        count.fetch_add(static_cast<CounterType>(delta), std::memory_order_release);
+    }
+}
+
+template <typename Traits>
+bool DeviceMemoryManager<Traits>::IsRegionTextureCached(DAddr addr, size_t size) const noexcept {
+    const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
+    for (size_t page = addr >> Memory::YUZU_PAGEBITS; page != page_end; ++page) {
+        if (texture_cached_pages->at(page >> subentries_shift).Count(page).load(
+                std::memory_order_acquire) != 0) {
+            return true;
+        }
+    }
+    return false;
+}
+
 template <typename Traits>
 void DeviceMemoryManager<Traits>::UpdatePagesCachedBatch(std::span<const std::pair<DAddr, size_t>> ranges, s32 delta) {
     if (ranges.empty()) {

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -772,12 +772,15 @@ bool RasterizerVulkan::OnCPUWrite(DAddr addr, u64 size) {
     return false;
 }
 
+static constexpr bool ENABLE_TEXTURE_CACHE_INVALIDATION_SKIP = true;
+
 void RasterizerVulkan::OnCacheInvalidation(DAddr addr, u64 size) {
     if (addr == 0 || size == 0) {
         return;
     }
 
-    {
+    if (!ENABLE_TEXTURE_CACHE_INVALIDATION_SKIP ||
+        device_memory.IsRegionTextureCached(addr, size)) {
         std::scoped_lock lock{texture_cache.mutex};
         texture_cache.WriteMemory(addr, size);
     }

src/video_core/texture_cache/texture_cache.h

@@ -2308,6 +2308,7 @@ void TextureCache<P>::TrackImage(ImageBase& image, ImageId image_id) {
     if (False(image.flags & ImageFlagBits::Sparse)) {
         if (image.cpu_addr < ~(1ULL << 40)) {
             device_memory.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, 1);
+            device_memory.UpdateTexturePagesCount(image.cpu_addr, image.guest_size_bytes, 1);
         }
         return;
     }
@@ -2320,12 +2321,14 @@ void TextureCache<P>::TrackImage(ImageBase& image, ImageId image_id) {
             const DAddr cpu_addr = map.cpu_addr;
             const std::size_t size = map.size;
             device_memory.UpdatePagesCachedCount(cpu_addr, size, 1);
+            device_memory.UpdateTexturePagesCount(cpu_addr, size, 1);
         }
         return;
     }
     ForEachSparseSegment(image,
                          [this]([[maybe_unused]] GPUVAddr gpu_addr, DAddr cpu_addr, size_t size) {
                              device_memory.UpdatePagesCachedCount(cpu_addr, size, 1);
+                             device_memory.UpdateTexturePagesCount(cpu_addr, size, 1);
                          });
 }
 
@@ -2336,6 +2339,7 @@ void TextureCache<P>::UntrackImage(ImageBase& image, ImageId image_id) {
     if (False(image.flags & ImageFlagBits::Sparse)) {
         if (image.cpu_addr < ~(1ULL << 40)) {
             device_memory.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, -1);
+            device_memory.UpdateTexturePagesCount(image.cpu_addr, image.guest_size_bytes, -1);
         }
         return;
     }
@@ -2348,6 +2352,7 @@ void TextureCache<P>::UntrackImage(ImageBase& image, ImageId image_id) {
         const DAddr cpu_addr = map.cpu_addr;
         const std::size_t size = map.size;
         device_memory.UpdatePagesCachedCount(cpu_addr, size, -1);
+        device_memory.UpdateTexturePagesCount(cpu_addr, size, -1);
     }
 }