nvdec: Defer texture cache invalidation until end of video stream

src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp

@@ -68,6 +68,7 @@ void nvhost_nvdec::OnOpen(DeviceFD fd) {}
 void nvhost_nvdec::OnClose(DeviceFD fd) {
     LOG_INFO(Service_NVDRV, "NVDEC video stream ended");
     system.GPU().ClearCdmaInstance();
+    system.GPU().MemoryManager().InvalidateQueuedCaches();
 }
 
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp

@@ -193,7 +193,13 @@ NvResult nvhost_nvdec_common::UnmapBuffer(const std::vector<u8>& input, std::vec
             return NvResult::InvalidState;
         }
         if (const auto size{RemoveBufferMap(object->dma_map_addr)}; size) {
-            gpu.MemoryManager().Unmap(object->dma_map_addr, *size);
+            if (vic_device) {
+                // UnmapVicFrame defers texture_cache invalidation of the frame address until
+                // the stream is over
+                gpu.MemoryManager().UnmapVicFrame(object->dma_map_addr, *size);
+            } else {
+                gpu.MemoryManager().Unmap(object->dma_map_addr, *size);
+            }
         } else {
             // This occurs quite frequently, however does not seem to impact functionality
             LOG_DEBUG(Service_NVDRV, "invalid offset=0x{:X} dma=0x{:X}", object->addr,

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h

@@ -160,6 +160,7 @@ protected:
 
     s32_le nvmap_fd{};
     u32_le submit_timeout{};
+    bool vic_device{};
     std::shared_ptr<nvmap> nvmap_dev;
     SyncpointManager& syncpoint_manager;
     std::array<u32, MaxSyncPoints> device_syncpoints{};

src/core/hle/service/nvdrv/devices/nvhost_vic.cpp

@@ -12,8 +12,9 @@
 namespace Service::Nvidia::Devices {
 nvhost_vic::nvhost_vic(Core::System& system_, std::shared_ptr<nvmap> nvmap_dev_,
                        SyncpointManager& syncpoint_manager_)
-    : nvhost_nvdec_common{system_, std::move(nvmap_dev_), syncpoint_manager_} {}
+    : nvhost_nvdec_common(system_, std::move(nvmap_dev_), syncpoint_manager_) {
-
+    vic_device = true;
+}
 nvhost_vic::~nvhost_vic() = default;
 
 NvResult nvhost_vic::Ioctl1(DeviceFD fd, Ioctl command, const std::vector<u8>& input,

src/video_core/memory_manager.cpp

@@ -119,6 +119,25 @@ void MemoryManager::TryUnlockPage(PageEntry page_entry, std::size_t size) {
                .IsSuccess());
 }
 
+void MemoryManager::UnmapVicFrame(GPUVAddr gpu_addr, std::size_t size) {
+    if (!size) {
+        return;
+    }
+
+    const std::optional<VAddr> cpu_addr = GpuToCpuAddress(gpu_addr);
+    ASSERT(cpu_addr);
+    rasterizer->InvalidateExceptTextureCache(*cpu_addr, size);
+    cache_invalidate_queue.push_back({*cpu_addr, size});
+
+    UpdateRange(gpu_addr, PageEntry::State::Unmapped, size);
+}
+
+void MemoryManager::InvalidateQueuedCaches() {
+    for (const auto& entry : cache_invalidate_queue) {
+        rasterizer->InvalidateTextureCache(entry.first, entry.second);
+    }
+    cache_invalidate_queue.clear();
+}
 PageEntry MemoryManager::GetPageEntry(GPUVAddr gpu_addr) const {
     return page_table[PageEntryIndex(gpu_addr)];
 }

src/video_core/memory_manager.h

@@ -143,6 +143,14 @@ public:
     [[nodiscard]] GPUVAddr Allocate(std::size_t size, std::size_t align);
     void Unmap(GPUVAddr gpu_addr, std::size_t size);
 
+    /**
+     * Some Decoded NVDEC frames require that texture cache does not get invalidated.
+     * UnmapVicFrame defers the texture cache invalidation until the stream ends
+     * by invoking InvalidateQueuedCaches to invalidate all frame texture caches.
+     */
+    void UnmapVicFrame(GPUVAddr gpu_addr, std::size_t size);
+    void InvalidateQueuedCaches();
+
 private:
     [[nodiscard]] PageEntry GetPageEntry(GPUVAddr gpu_addr) const;
     void SetPageEntry(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size = page_size);

src/video_core/rasterizer_interface.h

@@ -78,6 +78,12 @@ public:
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     virtual void FlushRegion(VAddr addr, u64 size) = 0;
 
+    /// Notify rasterizer to flush the texture cache to Switch memory
+    virtual void InvalidateExceptTextureCache(VAddr addr, u64 size) = 0;
+
+    /// Notify rasterizer to invalidate the texture cache
+    virtual void InvalidateTextureCache(VAddr addr, u64 size) = 0;
+
     /// Check if the the specified memory area requires flushing to CPU Memory.
     virtual bool MustFlushRegion(VAddr addr, u64 size) = 0;
 

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -548,6 +548,26 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
     query_cache.FlushRegion(addr, size);
 }
 
+void RasterizerOpenGL::InvalidateExceptTextureCache(VAddr addr, u64 size) {
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    shader_cache.InvalidateRegion(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
+    query_cache.InvalidateRegion(addr, size);
+}
+
+void RasterizerOpenGL::InvalidateTextureCache(VAddr addr, u64 size) {
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    std::scoped_lock lock{texture_cache.mutex};
+    texture_cache.UnmapMemory(addr, size);
+}
+
 bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size) {
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     if (!Settings::IsGPULevelHigh()) {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -88,6 +88,8 @@ public:
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
+    void InvalidateExceptTextureCache(VAddr addr, u64 size) override;
+    void InvalidateTextureCache(VAddr addr, u64 size) override;
     bool MustFlushRegion(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size) override;
     void OnCPUWrite(VAddr addr, u64 size) override;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -499,6 +499,26 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) {
     query_cache.FlushRegion(addr, size);
 }
 
+void Vulkan::RasterizerVulkan::InvalidateExceptTextureCache(VAddr addr, u64 size) {
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    pipeline_cache.InvalidateRegion(addr, size);
+    {
+        std::scoped_lock lock{buffer_cache.mutex};
+        buffer_cache.WriteMemory(addr, size);
+    }
+    query_cache.InvalidateRegion(addr, size);
+}
+
+void Vulkan::RasterizerVulkan::InvalidateTextureCache(VAddr addr, u64 size) {
+    if (addr == 0 || size == 0) {
+        return;
+    }
+    std::scoped_lock lock{texture_cache.mutex};
+    texture_cache.UnmapMemory(addr, size);
+}
+
 bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size) {
     std::scoped_lock lock{texture_cache.mutex, buffer_cache.mutex};
     if (!Settings::IsGPULevelHigh()) {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -80,6 +80,8 @@ public:
     void DisableGraphicsUniformBuffer(size_t stage, u32 index) override;
     void FlushAll() override;
     void FlushRegion(VAddr addr, u64 size) override;
+    void InvalidateExceptTextureCache(VAddr addr, u64 size) override;
+    void InvalidateTextureCache(VAddr addr, u64 size) override;
     bool MustFlushRegion(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size) override;
     void OnCPUWrite(VAddr addr, u64 size) override;