Merge pull request #6465 from FernandoS27/sex-on-the-beach

GPU: Implement a garbage collector for GPU Caches (project Reaper+)

src/common/common_sizes.h

@@ -24,6 +24,7 @@ enum : u64 {
     Size_128_MB = 128ULL * Size_1_MB,
     Size_448_MB = 448ULL * Size_1_MB,
     Size_507_MB = 507ULL * Size_1_MB,
+    Size_512_MB = 512ULL * Size_1_MB,
     Size_562_MB = 562ULL * Size_1_MB,
     Size_1554_MB = 1554ULL * Size_1_MB,
     Size_2048_MB = 2048ULL * Size_1_MB,

src/common/settings.cpp

@@ -59,6 +59,7 @@ void LogSettings() {
     log_setting("Renderer_UseVsync", values.use_vsync.GetValue());
     log_setting("Renderer_UseAssemblyShaders", values.use_assembly_shaders.GetValue());
     log_setting("Renderer_UseAsynchronousShaders", values.use_asynchronous_shaders.GetValue());
+    log_setting("Renderer_UseGarbageCollection", values.use_caches_gc.GetValue());
     log_setting("Renderer_AnisotropicFilteringLevel", values.max_anisotropy.GetValue());
     log_setting("Audio_OutputEngine", values.sink_id);
     log_setting("Audio_EnableAudioStretching", values.enable_audio_stretching.GetValue());
@@ -142,6 +143,7 @@ void RestoreGlobalState(bool is_powered_on) {
     values.use_assembly_shaders.SetGlobal(true);
     values.use_asynchronous_shaders.SetGlobal(true);
     values.use_fast_gpu_time.SetGlobal(true);
+    values.use_caches_gc.SetGlobal(true);
     values.bg_red.SetGlobal(true);
     values.bg_green.SetGlobal(true);
     values.bg_blue.SetGlobal(true);

src/common/settings.h

@@ -154,6 +154,7 @@ struct Values {
     Setting<bool> use_assembly_shaders;
     Setting<bool> use_asynchronous_shaders;
     Setting<bool> use_fast_gpu_time;
+    Setting<bool> use_caches_gc;
 
     Setting<float> bg_red;
     Setting<float> bg_green;

src/video_core/buffer_cache/buffer_base.h

@@ -256,6 +256,16 @@ public:
         stream_score += score;
     }
 
+    /// Sets the new frame tick
+    void SetFrameTick(u64 new_frame_tick) noexcept {
+        frame_tick = new_frame_tick;
+    }
+
+    /// Returns the new frame tick
+    [[nodiscard]] u64 FrameTick() const noexcept {
+        return frame_tick;
+    }
+
     /// Returns the likeliness of this being a stream buffer
     [[nodiscard]] int StreamScore() const noexcept {
         return stream_score;
@@ -586,6 +596,7 @@ private:
     RasterizerInterface* rasterizer = nullptr;
     VAddr cpu_addr = 0;
     Words words;
+    u64 frame_tick = 0;
     BufferFlagBits flags{};
     int stream_score = 0;
 };

src/video_core/buffer_cache/buffer_cache.h

@@ -16,6 +16,7 @@
 
 #include <boost/container/small_vector.hpp>
 
+#include "common/common_sizes.h"
 #include "common/common_types.h"
 #include "common/div_ceil.h"
 #include "common/microprofile.h"
@@ -65,6 +66,9 @@ class BufferCache {
 
     static constexpr BufferId NULL_BUFFER_ID{0};
 
+    static constexpr u64 EXPECTED_MEMORY = Common::Size_512_MB;
+    static constexpr u64 CRITICAL_MEMORY = Common::Size_1_GB;
+
     using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
     using Runtime = typename P::Runtime;
@@ -102,6 +106,8 @@ public:
 
     void TickFrame();
 
+    void RunGarbageCollector();
+
     void WriteMemory(VAddr cpu_addr, u64 size);
 
     void CachedWriteMemory(VAddr cpu_addr, u64 size);
@@ -243,6 +249,8 @@ private:
     template <bool insert>
     void ChangeRegister(BufferId buffer_id);
 
+    void TouchBuffer(Buffer& buffer) const noexcept;
+
     bool SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size);
 
     bool SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size);
@@ -255,6 +263,10 @@ private:
 
     void MappedUploadMemory(Buffer& buffer, u64 total_size_bytes, std::span<BufferCopy> copies);
 
+    void DownloadBufferMemory(Buffer& buffer_id);
+
+    void DownloadBufferMemory(Buffer& buffer_id, VAddr cpu_addr, u64 size);
+
     void DeleteBuffer(BufferId buffer_id);
 
     void ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id);
@@ -319,6 +331,10 @@ private:
     size_t immediate_buffer_capacity = 0;
     std::unique_ptr<u8[]> immediate_buffer_alloc;
 
+    typename SlotVector<Buffer>::Iterator deletion_iterator;
+    u64 frame_tick = 0;
+    u64 total_used_memory = 0;
+
     std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
 };
 
@@ -332,6 +348,28 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
       gpu_memory{gpu_memory_}, cpu_memory{cpu_memory_}, runtime{runtime_} {
     // Ensure the first slot is used for the null buffer
     void(slot_buffers.insert(runtime, NullBufferParams{}));
+    deletion_iterator = slot_buffers.end();
+}
+
+template <class P>
+void BufferCache<P>::RunGarbageCollector() {
+    const bool aggressive_gc = total_used_memory >= CRITICAL_MEMORY;
+    const u64 ticks_to_destroy = aggressive_gc ? 60 : 120;
+    int num_iterations = aggressive_gc ? 64 : 32;
+    for (; num_iterations > 0; --num_iterations) {
+        if (deletion_iterator == slot_buffers.end()) {
+            deletion_iterator = slot_buffers.begin();
+        }
+        ++deletion_iterator;
+        if (deletion_iterator == slot_buffers.end()) {
+            break;
+        }
+        const auto [buffer_id, buffer] = *deletion_iterator;
+        if (buffer->FrameTick() + ticks_to_destroy < frame_tick) {
+            DownloadBufferMemory(*buffer);
+            DeleteBuffer(buffer_id);
+        }
+    }
 }
 
 template <class P>
@@ -349,6 +387,10 @@ void BufferCache<P>::TickFrame() {
     const bool skip_preferred = hits * 256 < shots * 251;
     uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
 
+    if (Settings::values.use_caches_gc.GetValue() && total_used_memory >= EXPECTED_MEMORY) {
+        RunGarbageCollector();
+    }
+    ++frame_tick;
     delayed_destruction_ring.Tick();
 }
 
@@ -371,50 +413,8 @@ void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
 
 template <class P>
 void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) {
-    ForEachBufferInRange(cpu_addr, size, [&](BufferId, Buffer& buffer) {
-        boost::container::small_vector<BufferCopy, 1> copies;
-        u64 total_size_bytes = 0;
-        u64 largest_copy = 0;
-        buffer.ForEachDownloadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
-            copies.push_back(BufferCopy{
-                .src_offset = range_offset,
-                .dst_offset = total_size_bytes,
-                .size = range_size,
-            });
-            total_size_bytes += range_size;
-            largest_copy = std::max(largest_copy, range_size);
-        });
-        if (total_size_bytes == 0) {
-            return;
-        }
-        MICROPROFILE_SCOPE(GPU_DownloadMemory);
-
-        if constexpr (USE_MEMORY_MAPS) {
-            auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
-            const u8* const mapped_memory = download_staging.mapped_span.data();
-            const std::span<BufferCopy> copies_span(copies.data(), copies.data() + copies.size());
-            for (BufferCopy& copy : copies) {
-                // Modify copies to have the staging offset in mind
-                copy.dst_offset += download_staging.offset;
-            }
-            runtime.CopyBuffer(download_staging.buffer, buffer, copies_span);
-            runtime.Finish();
-            for (const BufferCopy& copy : copies) {
-                const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
-                // Undo the modified offset
-                const u64 dst_offset = copy.dst_offset - download_staging.offset;
-                const u8* copy_mapped_memory = mapped_memory + dst_offset;
-                cpu_memory.WriteBlockUnsafe(copy_cpu_addr, copy_mapped_memory, copy.size);
-            }
-        } else {
-            const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
-            for (const BufferCopy& copy : copies) {
-                buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
-                const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
-                cpu_memory.WriteBlockUnsafe(copy_cpu_addr, immediate_buffer.data(), copy.size);
-            }
-        }
-    });
+    ForEachBufferInRange(cpu_addr, size,
+                         [&](BufferId, Buffer& buffer) { DownloadBufferMemory(buffer); });
 }
 
 template <class P>
@@ -640,6 +640,7 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) {
 template <class P>
 void BufferCache<P>::BindHostIndexBuffer() {
     Buffer& buffer = slot_buffers[index_buffer.buffer_id];
+    TouchBuffer(buffer);
     const u32 offset = buffer.Offset(index_buffer.cpu_addr);
     const u32 size = index_buffer.size;
     SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
@@ -658,6 +659,7 @@ void BufferCache<P>::BindHostVertexBuffers() {
     for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
         const Binding& binding = vertex_buffers[index];
         Buffer& buffer = slot_buffers[binding.buffer_id];
+        TouchBuffer(buffer);
         SynchronizeBuffer(buffer, binding.cpu_addr, binding.size);
         if (!flags[Dirty::VertexBuffer0 + index]) {
             continue;
@@ -693,6 +695,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     const VAddr cpu_addr = binding.cpu_addr;
     const u32 size = binding.size;
     Buffer& buffer = slot_buffers[binding.buffer_id];
+    TouchBuffer(buffer);
     const bool use_fast_buffer = binding.buffer_id != NULL_BUFFER_ID &&
                                  size <= uniform_buffer_skip_cache_size &&
                                  !buffer.IsRegionGpuModified(cpu_addr, size);
@@ -744,6 +747,7 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
     ForEachEnabledBit(enabled_storage_buffers[stage], [&](u32 index) {
         const Binding& binding = storage_buffers[stage][index];
         Buffer& buffer = slot_buffers[binding.buffer_id];
+        TouchBuffer(buffer);
         const u32 size = binding.size;
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
@@ -766,6 +770,7 @@ void BufferCache<P>::BindHostTransformFeedbackBuffers() {
     for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
         const Binding& binding = transform_feedback_buffers[index];
         Buffer& buffer = slot_buffers[binding.buffer_id];
+        TouchBuffer(buffer);
         const u32 size = binding.size;
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
@@ -784,6 +789,7 @@ void BufferCache<P>::BindHostComputeUniformBuffers() {
     ForEachEnabledBit(enabled_compute_uniform_buffers, [&](u32 index) {
         const Binding& binding = compute_uniform_buffers[index];
         Buffer& buffer = slot_buffers[binding.buffer_id];
+        TouchBuffer(buffer);
         const u32 size = binding.size;
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
@@ -803,6 +809,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
     ForEachEnabledBit(enabled_compute_storage_buffers, [&](u32 index) {
         const Binding& binding = compute_storage_buffers[index];
         Buffer& buffer = slot_buffers[binding.buffer_id];
+        TouchBuffer(buffer);
         const u32 size = binding.size;
         SynchronizeBuffer(buffer, binding.cpu_addr, size);
 
@@ -1101,6 +1108,7 @@ BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
     const OverlapResult overlap = ResolveOverlaps(cpu_addr, wanted_size);
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
     const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);
+    TouchBuffer(slot_buffers[new_buffer_id]);
     for (const BufferId overlap_id : overlap.ids) {
         JoinOverlap(new_buffer_id, overlap_id, !overlap.has_stream_leap);
     }
@@ -1122,8 +1130,14 @@ template <class P>
 template <bool insert>
 void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
     const Buffer& buffer = slot_buffers[buffer_id];
+    const auto size = buffer.SizeBytes();
+    if (insert) {
+        total_used_memory += Common::AlignUp(size, 1024);
+    } else {
+        total_used_memory -= Common::AlignUp(size, 1024);
+    }
     const VAddr cpu_addr_begin = buffer.CpuAddr();
-    const VAddr cpu_addr_end = cpu_addr_begin + buffer.SizeBytes();
+    const VAddr cpu_addr_end = cpu_addr_begin + size;
     const u64 page_begin = cpu_addr_begin / PAGE_SIZE;
     const u64 page_end = Common::DivCeil(cpu_addr_end, PAGE_SIZE);
     for (u64 page = page_begin; page != page_end; ++page) {
@@ -1135,6 +1149,11 @@ void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
     }
 }
 
+template <class P>
+void BufferCache<P>::TouchBuffer(Buffer& buffer) const noexcept {
+    buffer.SetFrameTick(frame_tick);
+}
+
 template <class P>
 bool BufferCache<P>::SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size) {
     if (buffer.CpuAddr() == 0) {
@@ -1211,6 +1230,57 @@ void BufferCache<P>::MappedUploadMemory(Buffer& buffer, u64 total_size_bytes,
     runtime.CopyBuffer(buffer, upload_staging.buffer, copies);
 }
 
+template <class P>
+void BufferCache<P>::DownloadBufferMemory(Buffer& buffer) {
+    DownloadBufferMemory(buffer, buffer.CpuAddr(), buffer.SizeBytes());
+}
+
+template <class P>
+void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 size) {
+    boost::container::small_vector<BufferCopy, 1> copies;
+    u64 total_size_bytes = 0;
+    u64 largest_copy = 0;
+    buffer.ForEachDownloadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
+        copies.push_back(BufferCopy{
+            .src_offset = range_offset,
+            .dst_offset = total_size_bytes,
+            .size = range_size,
+        });
+        total_size_bytes += range_size;
+        largest_copy = std::max(largest_copy, range_size);
+    });
+    if (total_size_bytes == 0) {
+        return;
+    }
+    MICROPROFILE_SCOPE(GPU_DownloadMemory);
+
+    if constexpr (USE_MEMORY_MAPS) {
+        auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
+        const u8* const mapped_memory = download_staging.mapped_span.data();
+        const std::span<BufferCopy> copies_span(copies.data(), copies.data() + copies.size());
+        for (BufferCopy& copy : copies) {
+            // Modify copies to have the staging offset in mind
+            copy.dst_offset += download_staging.offset;
+        }
+        runtime.CopyBuffer(download_staging.buffer, buffer, copies_span);
+        runtime.Finish();
+        for (const BufferCopy& copy : copies) {
+            const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
+            // Undo the modified offset
+            const u64 dst_offset = copy.dst_offset - download_staging.offset;
+            const u8* copy_mapped_memory = mapped_memory + dst_offset;
+            cpu_memory.WriteBlockUnsafe(copy_cpu_addr, copy_mapped_memory, copy.size);
+        }
+    } else {
+        const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
+        for (const BufferCopy& copy : copies) {
+            buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
+            const VAddr copy_cpu_addr = buffer.CpuAddr() + copy.src_offset;
+            cpu_memory.WriteBlockUnsafe(copy_cpu_addr, immediate_buffer.data(), copy.size);
+        }
+    }
+}
+
 template <class P>
 void BufferCache<P>::DeleteBuffer(BufferId buffer_id) {
     const auto scalar_replace = [buffer_id](Binding& binding) {
@@ -1236,6 +1306,7 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id) {
 
     Unregister(buffer_id);
     delayed_destruction_ring.Push(std::move(slot_buffers[buffer_id]));
+    slot_buffers.erase(buffer_id);
 
     NotifyBufferDeletion();
 }

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -737,6 +737,8 @@ Image::Image(TextureCacheRuntime& runtime, const VideoCommon::ImageInfo& info_,
     }
 }
 
+Image::~Image() = default;
+
 void Image::UploadMemory(const ImageBufferMap& map,
                          std::span<const VideoCommon::BufferImageCopy> copies) {
     glBindBuffer(GL_PIXEL_UNPACK_BUFFER, map.buffer);

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -143,6 +143,14 @@ public:
     explicit Image(TextureCacheRuntime&, const VideoCommon::ImageInfo& info, GPUVAddr gpu_addr,
                    VAddr cpu_addr);
 
+    ~Image();
+
+    Image(const Image&) = delete;
+    Image& operator=(const Image&) = delete;
+
+    Image(Image&&) = default;
+    Image& operator=(Image&&) = default;
+
     void UploadMemory(const ImageBufferMap& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
@@ -235,6 +243,7 @@ struct TextureCacheParams {
     static constexpr bool ENABLE_VALIDATION = true;
     static constexpr bool FRAMEBUFFER_BLITS = true;
     static constexpr bool HAS_EMULATED_COPIES = true;
+    static constexpr bool HAS_DEVICE_MEMORY_INFO = false;
 
     using Runtime = OpenGL::TextureCacheRuntime;
     using Image = OpenGL::Image;

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -818,6 +818,10 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
     });
 }
 
+u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
+    return device.GetDeviceLocalMemory();
+}
+
 Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_addr_,
              VAddr cpu_addr_)
     : VideoCommon::ImageBase(info_, gpu_addr_, cpu_addr_), scheduler{&runtime.scheduler},
@@ -876,6 +880,8 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
     }
 }
 
+Image::~Image() = default;
+
 void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
     // TODO: Move this to another API
     scheduler->RequestOutsideRenderPassOperationContext();

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -97,6 +97,8 @@ struct TextureCacheRuntime {
         // All known Vulkan drivers can natively handle BGR textures
         return true;
     }
+
+    u64 GetDeviceLocalMemory() const;
 };
 
 class Image : public VideoCommon::ImageBase {
@@ -104,6 +106,14 @@ public:
     explicit Image(TextureCacheRuntime&, const VideoCommon::ImageInfo& info, GPUVAddr gpu_addr,
                    VAddr cpu_addr);
 
+    ~Image();
+
+    Image(const Image&) = delete;
+    Image& operator=(const Image&) = delete;
+
+    Image(Image&&) = default;
+    Image& operator=(Image&&) = default;
+
     void UploadMemory(const StagingBufferRef& map,
                       std::span<const VideoCommon::BufferImageCopy> copies);
 
@@ -257,6 +267,7 @@ struct TextureCacheParams {
     static constexpr bool ENABLE_VALIDATION = true;
     static constexpr bool FRAMEBUFFER_BLITS = false;
     static constexpr bool HAS_EMULATED_COPIES = false;
+    static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
 
     using Runtime = Vulkan::TextureCacheRuntime;
     using Image = Vulkan::Image;

src/video_core/surface.cpp

@@ -283,4 +283,11 @@ std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) {
     return {DefaultBlockWidth(format), DefaultBlockHeight(format)};
 }
 
+u64 EstimatedDecompressedSize(u64 base_size, PixelFormat format) {
+    constexpr u64 RGBA8_PIXEL_SIZE = 4;
+    const u64 base_block_size = static_cast<u64>(DefaultBlockWidth(format)) *
+                                static_cast<u64>(DefaultBlockHeight(format)) * RGBA8_PIXEL_SIZE;
+    return (base_size * base_block_size) / BytesPerBlock(format);
+}
+
 } // namespace VideoCore::Surface

src/video_core/surface.h

@@ -462,4 +462,6 @@ bool IsPixelFormatSRGB(PixelFormat format);
 
 std::pair<u32, u32> GetASTCBlockSize(PixelFormat format);
 
+u64 EstimatedDecompressedSize(u64 base_size, PixelFormat format);
+
 } // namespace VideoCore::Surface

src/video_core/texture_cache/image_base.cpp

@@ -113,6 +113,43 @@ void ImageBase::InsertView(const ImageViewInfo& view_info, ImageViewId image_vie
     image_view_ids.push_back(image_view_id);
 }
 
+bool ImageBase::IsSafeDownload() const noexcept {
+    // Skip images that were not modified from the GPU
+    if (False(flags & ImageFlagBits::GpuModified)) {
+        return false;
+    }
+    // Skip images that .are. modified from the CPU
+    // We don't want to write sensitive data from the guest
+    if (True(flags & ImageFlagBits::CpuModified)) {
+        return false;
+    }
+    if (info.num_samples > 1) {
+        LOG_WARNING(HW_GPU, "MSAA image downloads are not implemented");
+        return false;
+    }
+    return true;
+}
+
+void ImageBase::CheckBadOverlapState() {
+    if (False(flags & ImageFlagBits::BadOverlap)) {
+        return;
+    }
+    if (!overlapping_images.empty()) {
+        return;
+    }
+    flags &= ~ImageFlagBits::BadOverlap;
+}
+
+void ImageBase::CheckAliasState() {
+    if (False(flags & ImageFlagBits::Alias)) {
+        return;
+    }
+    if (!aliased_images.empty()) {
+        return;
+    }
+    flags &= ~ImageFlagBits::Alias;
+}
+
 void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id) {
     static constexpr auto OPTIONS = RelaxedOptions::Size | RelaxedOptions::Format;
     ASSERT(lhs.info.type == rhs.info.type);

src/video_core/texture_cache/image_base.h

@@ -25,6 +25,12 @@ enum class ImageFlagBits : u32 {
     Strong = 1 << 5,      ///< Exists in the image table, the dimensions are can be trusted
     Registered = 1 << 6,  ///< True when the image is registered
     Picked = 1 << 7,      ///< Temporary flag to mark the image as picked
+
+    // Garbage Collection Flags
+    BadOverlap = 1 << 8, ///< This image overlaps other but doesn't fit, has higher
+                         ///< garbage collection priority
+    Alias = 1 << 9,      ///< This image has aliases and has priority on garbage
+                         ///< collection
 };
 DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)
 
@@ -44,11 +50,16 @@ struct ImageBase {
 
     void InsertView(const ImageViewInfo& view_info, ImageViewId image_view_id);
 
+    [[nodiscard]] bool IsSafeDownload() const noexcept;
+
     [[nodiscard]] bool Overlaps(VAddr overlap_cpu_addr, size_t overlap_size) const noexcept {
         const VAddr overlap_end = overlap_cpu_addr + overlap_size;
         return cpu_addr < overlap_end && overlap_cpu_addr < cpu_addr_end;
     }
 
+    void CheckBadOverlapState();
+    void CheckAliasState();
+
     ImageInfo info;
 
     u32 guest_size_bytes = 0;
@@ -72,6 +83,7 @@ struct ImageBase {
     std::vector<SubresourceBase> slice_subresources;
 
     std::vector<AliasedImage> aliased_images;
+    std::vector<ImageId> overlapping_images;
 };
 
 struct ImageAllocBase {

src/video_core/texture_cache/slot_vector.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include <array>
+#include <bit>
 #include <concepts>
 #include <numeric>
 #include <type_traits>
@@ -32,6 +33,60 @@ template <class T>
 requires std::is_nothrow_move_assignable_v<T>&&
     std::is_nothrow_move_constructible_v<T> class SlotVector {
 public:
+    class Iterator {
+        friend SlotVector<T>;
+
+    public:
+        constexpr Iterator() = default;
+
+        Iterator& operator++() noexcept {
+            const u64* const bitset = slot_vector->stored_bitset.data();
+            const u32 size = static_cast<u32>(slot_vector->stored_bitset.size()) * 64;
+            if (id.index < size) {
+                do {
+                    ++id.index;
+                } while (id.index < size && !IsValid(bitset));
+                if (id.index == size) {
+                    id.index = SlotId::INVALID_INDEX;
+                }
+            }
+            return *this;
+        }
+
+        Iterator operator++(int) noexcept {
+            const Iterator copy{*this};
+            ++*this;
+            return copy;
+        }
+
+        bool operator==(const Iterator& other) const noexcept {
+            return id.index == other.id.index;
+        }
+
+        bool operator!=(const Iterator& other) const noexcept {
+            return id.index != other.id.index;
+        }
+
+        std::pair<SlotId, T*> operator*() const noexcept {
+            return {id, std::addressof((*slot_vector)[id])};
+        }
+
+        T* operator->() const noexcept {
+            return std::addressof((*slot_vector)[id]);
+        }
+
+    private:
+        Iterator(SlotVector<T>* slot_vector_, SlotId id_) noexcept
+            : slot_vector{slot_vector_}, id{id_} {}
+
+        bool IsValid(const u64* bitset) const noexcept {
+            return ((bitset[id.index / 64] >> (id.index % 64)) & 1) != 0;
+        }
+
+        SlotVector<T>* slot_vector;
+        SlotId id;
+    };
+
     ~SlotVector() noexcept {
         size_t index = 0;
         for (u64 bits : stored_bitset) {
@@ -70,6 +125,20 @@ public:
         ResetStorageBit(id.index);
     }
 
+    [[nodiscard]] Iterator begin() noexcept {
+        const auto it = std::ranges::find_if(stored_bitset, [](u64 value) { return value != 0; });
+        if (it == stored_bitset.end()) {
+            return end();
+        }
+        const u32 word_index = static_cast<u32>(std::distance(it, stored_bitset.begin()));
+        const SlotId first_id{word_index * 64 + static_cast<u32>(std::countr_zero(*it))};
+        return Iterator(this, first_id);
+    }
+
+    [[nodiscard]] Iterator end() noexcept {
+        return Iterator(this, SlotId{SlotId::INVALID_INDEX});
+    }
+
 private:
     struct NonTrivialDummy {
         NonTrivialDummy() noexcept {}
@@ -140,7 +209,6 @@ private:
 
     Entry* values = nullptr;
     size_t values_capacity = 0;
-    size_t values_size = 0;
 
     std::vector<u64> stored_bitset;
     std::vector<u32> free_list;

src/video_core/texture_cache/texture_cache.h

@@ -20,8 +20,10 @@
 
 #include "common/alignment.h"
 #include "common/common_funcs.h"
+#include "common/common_sizes.h"
 #include "common/common_types.h"
 #include "common/logging/log.h"
+#include "common/settings.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/delayed_destruction_ring.h"
 #include "video_core/dirty_flags.h"
@@ -69,12 +71,17 @@ class TextureCache {
     static constexpr bool FRAMEBUFFER_BLITS = P::FRAMEBUFFER_BLITS;
     /// True when some copies have to be emulated
     static constexpr bool HAS_EMULATED_COPIES = P::HAS_EMULATED_COPIES;
+    /// True when the API can provide info about the memory of the device.
+    static constexpr bool HAS_DEVICE_MEMORY_INFO = P::HAS_DEVICE_MEMORY_INFO;
 
     /// Image view ID for null descriptors
     static constexpr ImageViewId NULL_IMAGE_VIEW_ID{0};
     /// Sampler ID for bugged sampler ids
     static constexpr SamplerId NULL_SAMPLER_ID{0};
 
+    static constexpr u64 DEFAULT_EXPECTED_MEMORY = Common::Size_1_GB;
+    static constexpr u64 DEFAULT_CRITICAL_MEMORY = Common::Size_2_GB;
+
     using Runtime = typename P::Runtime;
     using Image = typename P::Image;
     using ImageAlloc = typename P::ImageAlloc;
@@ -103,6 +110,9 @@ public:
     /// Notify the cache that a new frame has been queued
     void TickFrame();
 
+    /// Runs the Garbage Collector.
+    void RunGarbageCollector();
+
     /// Return a constant reference to the given image view id
     [[nodiscard]] const ImageView& GetImageView(ImageViewId id) const noexcept;
 
@@ -333,6 +343,10 @@ private:
     std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> page_table;
 
     bool has_deleted_images = false;
+    u64 total_used_memory = 0;
+    u64 minimum_memory;
+    u64 expected_memory;
+    u64 critical_memory;
 
     SlotVector<Image> slot_images;
     SlotVector<ImageView> slot_image_views;
@@ -353,6 +367,7 @@ private:
 
     u64 modification_tick = 0;
     u64 frame_tick = 0;
+    typename SlotVector<Image>::Iterator deletion_iterator;
 };
 
 template <class P>
@@ -373,11 +388,94 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
     // This way the null resource becomes a compile time constant
     void(slot_image_views.insert(runtime, NullImageParams{}));
     void(slot_samplers.insert(runtime, sampler_descriptor));
+
+    deletion_iterator = slot_images.begin();
+
+    if constexpr (HAS_DEVICE_MEMORY_INFO) {
+        const auto device_memory = runtime.GetDeviceLocalMemory();
+        const u64 possible_expected_memory = (device_memory * 3) / 10;
+        const u64 possible_critical_memory = (device_memory * 6) / 10;
+        expected_memory = std::max(possible_expected_memory, DEFAULT_EXPECTED_MEMORY);
+        critical_memory = std::max(possible_critical_memory, DEFAULT_CRITICAL_MEMORY);
+        minimum_memory = 0;
+    } else {
+        // on OGL we can be more conservatives as the driver takes care.
+        expected_memory = DEFAULT_EXPECTED_MEMORY + Common::Size_512_MB;
+        critical_memory = DEFAULT_CRITICAL_MEMORY + Common::Size_1_GB;
+        minimum_memory = expected_memory;
+    }
+}
+
+template <class P>
+void TextureCache<P>::RunGarbageCollector() {
+    const bool high_priority_mode = total_used_memory >= expected_memory;
+    const bool aggressive_mode = total_used_memory >= critical_memory;
+    const u64 ticks_to_destroy = high_priority_mode ? 60 : 100;
+    int num_iterations = aggressive_mode ? 256 : (high_priority_mode ? 128 : 64);
+    for (; num_iterations > 0; --num_iterations) {
+        if (deletion_iterator == slot_images.end()) {
+            deletion_iterator = slot_images.begin();
+            if (deletion_iterator == slot_images.end()) {
+                break;
+            }
+        }
+        auto [image_id, image_tmp] = *deletion_iterator;
+        Image* image = image_tmp; // fix clang error.
+        const bool is_alias = True(image->flags & ImageFlagBits::Alias);
+        const bool is_bad_overlap = True(image->flags & ImageFlagBits::BadOverlap);
+        const bool must_download = image->IsSafeDownload();
+        bool should_care = is_bad_overlap || is_alias || (high_priority_mode && !must_download);
+        const u64 ticks_needed =
+            is_bad_overlap
+                ? ticks_to_destroy >> 4
+                : ((should_care && aggressive_mode) ? ticks_to_destroy >> 1 : ticks_to_destroy);
+        should_care |= aggressive_mode;
+        if (should_care && image->frame_tick + ticks_needed < frame_tick) {
+            if (is_bad_overlap) {
+                const bool overlap_check = std::ranges::all_of(
+                    image->overlapping_images, [&, image](const ImageId& overlap_id) {
+                        auto& overlap = slot_images[overlap_id];
+                        return overlap.frame_tick >= image->frame_tick;
+                    });
+                if (!overlap_check) {
+                    ++deletion_iterator;
+                    continue;
+                }
+            }
+            if (!is_bad_overlap && must_download) {
+                const bool alias_check = std::ranges::none_of(
+                    image->aliased_images, [&, image](const AliasedImage& alias) {
+                        auto& alias_image = slot_images[alias.id];
+                        return (alias_image.frame_tick < image->frame_tick) ||
+                               (alias_image.modification_tick < image->modification_tick);
+                    });
+
+                if (alias_check) {
+                    auto map = runtime.DownloadStagingBuffer(image->unswizzled_size_bytes);
+                    const auto copies = FullDownloadCopies(image->info);
+                    image->DownloadMemory(map, copies);
+                    runtime.Finish();
+                    SwizzleImage(gpu_memory, image->gpu_addr, image->info, copies, map.mapped_span);
+                }
+            }
+            if (True(image->flags & ImageFlagBits::Tracked)) {
+                UntrackImage(*image);
+            }
+            UnregisterImage(image_id);
+            DeleteImage(image_id);
+            if (is_bad_overlap) {
+                ++num_iterations;
+            }
+        }
+        ++deletion_iterator;
+    }
 }
 
 template <class P>
 void TextureCache<P>::TickFrame() {
-    // Tick sentenced resources in this order to ensure they are destroyed in the right order
+    if (Settings::values.use_caches_gc.GetValue() && total_used_memory > minimum_memory) {
+        RunGarbageCollector();
+    }
     sentenced_images.Tick();
     sentenced_framebuffers.Tick();
     sentenced_image_view.Tick();
@@ -568,17 +666,7 @@ template <class P>
 void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
     std::vector<ImageId> images;
     ForEachImageInRegion(cpu_addr, size, [this, &images](ImageId image_id, ImageBase& image) {
-        // Skip images that were not modified from the GPU
-        if (False(image.flags & ImageFlagBits::GpuModified)) {
-            return;
-        }
-        // Skip images that .are. modified from the CPU
-        // We don't want to write sensitive data from the guest
-        if (True(image.flags & ImageFlagBits::CpuModified)) {
-            return;
-        }
-        if (image.info.num_samples > 1) {
-            LOG_WARNING(HW_GPU, "MSAA image downloads are not implemented");
+        if (!image.IsSafeDownload()) {
             return;
         }
         image.flags &= ~ImageFlagBits::GpuModified;
@@ -967,6 +1055,7 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
     std::vector<ImageId> overlap_ids;
     std::vector<ImageId> left_aliased_ids;
     std::vector<ImageId> right_aliased_ids;
+    std::vector<ImageId> bad_overlap_ids;
     ForEachImageInRegion(cpu_addr, size_bytes, [&](ImageId overlap_id, ImageBase& overlap) {
         if (info.type != overlap.info.type) {
             return;
@@ -992,9 +1081,14 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
         const ImageBase new_image_base(new_info, gpu_addr, cpu_addr);
         if (IsSubresource(new_info, overlap, gpu_addr, options, broken_views, native_bgr)) {
             left_aliased_ids.push_back(overlap_id);
+            overlap.flags |= ImageFlagBits::Alias;
         } else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options,
                                  broken_views, native_bgr)) {
             right_aliased_ids.push_back(overlap_id);
+            overlap.flags |= ImageFlagBits::Alias;
+        } else {
+            bad_overlap_ids.push_back(overlap_id);
+            overlap.flags |= ImageFlagBits::BadOverlap;
         }
     });
     const ImageId new_image_id = slot_images.insert(runtime, new_info, gpu_addr, cpu_addr);
@@ -1022,10 +1116,18 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
     for (const ImageId aliased_id : right_aliased_ids) {
         ImageBase& aliased = slot_images[aliased_id];
         AddImageAlias(new_image_base, aliased, new_image_id, aliased_id);
+        new_image.flags |= ImageFlagBits::Alias;
     }
     for (const ImageId aliased_id : left_aliased_ids) {
         ImageBase& aliased = slot_images[aliased_id];
         AddImageAlias(aliased, new_image_base, aliased_id, new_image_id);
+        new_image.flags |= ImageFlagBits::Alias;
+    }
+    for (const ImageId aliased_id : bad_overlap_ids) {
+        ImageBase& aliased = slot_images[aliased_id];
+        aliased.overlapping_images.push_back(new_image_id);
+        new_image.overlapping_images.push_back(aliased_id);
+        new_image.flags |= ImageFlagBits::BadOverlap;
     }
     RegisterImage(new_image_id);
     return new_image_id;
@@ -1195,6 +1297,13 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
     image.flags |= ImageFlagBits::Registered;
     ForEachPage(image.cpu_addr, image.guest_size_bytes,
                 [this, image_id](u64 page) { page_table[page].push_back(image_id); });
+    u64 tentative_size = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
+    if ((IsPixelFormatASTC(image.info.format) &&
+         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
+        True(image.flags & ImageFlagBits::Converted)) {
+        tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
+    }
+    total_used_memory += Common::AlignUp(tentative_size, 1024);
 }
 
 template <class P>
@@ -1203,6 +1312,14 @@ void TextureCache<P>::UnregisterImage(ImageId image_id) {
     ASSERT_MSG(True(image.flags & ImageFlagBits::Registered),
                "Trying to unregister an already registered image");
     image.flags &= ~ImageFlagBits::Registered;
+    image.flags &= ~ImageFlagBits::BadOverlap;
+    u64 tentative_size = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
+    if ((IsPixelFormatASTC(image.info.format) &&
+         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
+        True(image.flags & ImageFlagBits::Converted)) {
+        tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
+    }
+    total_used_memory -= Common::AlignUp(tentative_size, 1024);
     ForEachPage(image.cpu_addr, image.guest_size_bytes, [this, image_id](u64 page) {
         const auto page_it = page_table.find(page);
         if (page_it == page_table.end()) {
@@ -1276,9 +1393,19 @@ void TextureCache<P>::DeleteImage(ImageId image_id) {
             std::erase_if(other_image.aliased_images, [image_id](const AliasedImage& other_alias) {
                 return other_alias.id == image_id;
             });
+        other_image.CheckAliasState();
         ASSERT_MSG(num_removed_aliases == 1, "Invalid number of removed aliases: {}",
                    num_removed_aliases);
     }
+    for (const ImageId overlap_id : image.overlapping_images) {
+        ImageBase& other_image = slot_images[overlap_id];
+        [[maybe_unused]] const size_t num_removed_overlaps = std::erase_if(
+            other_image.overlapping_images,
+            [image_id](const ImageId other_overlap_id) { return other_overlap_id == image_id; });
+        other_image.CheckBadOverlapState();
+        ASSERT_MSG(num_removed_overlaps == 1, "Invalid number of removed overlapps: {}",
+                   num_removed_overlaps);
+    }
     for (const ImageViewId image_view_id : image_view_ids) {
         sentenced_image_view.Push(std::move(slot_image_views[image_view_id]));
         slot_image_views.erase(image_view_id);

src/video_core/texture_cache/util.cpp

@@ -581,6 +581,8 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
 
     for (s32 layer = 0; layer < info.resources.layers; ++layer) {
         const std::span<const u8> src = input.subspan(host_offset);
+        gpu_memory.ReadBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
+
         SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
                        num_tiles.depth, block.height, block.depth);
 

src/video_core/vulkan_common/vulkan_device.cpp

@@ -408,6 +408,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
     }
     logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
 
+    CollectPhysicalMemoryInfo();
     CollectTelemetryParameters();
     CollectToolingInfo();
 
@@ -818,6 +819,17 @@ void Device::CollectTelemetryParameters() {
     }
 }
 
+void Device::CollectPhysicalMemoryInfo() {
+    const auto mem_properties = physical.GetMemoryProperties();
+    const std::size_t num_properties = mem_properties.memoryHeapCount;
+    device_access_memory = 0;
+    for (std::size_t element = 0; element < num_properties; element++) {
+        if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
+            device_access_memory += mem_properties.memoryHeaps[element].size;
+        }
+    }
+}
+
 void Device::CollectToolingInfo() {
     if (!ext_tooling_info) {
         return;

src/video_core/vulkan_common/vulkan_device.h

@@ -225,6 +225,10 @@ public:
         return use_asynchronous_shaders;
     }
 
+    u64 GetDeviceLocalMemory() const {
+        return device_access_memory;
+    }
+
 private:
     /// Checks if the physical device is suitable.
     void CheckSuitability(bool requires_swapchain) const;
@@ -244,6 +248,9 @@ private:
     /// Collects information about attached tools.
     void CollectToolingInfo();
 
+    /// Collects information about the device's local memory.
+    void CollectPhysicalMemoryInfo();
+
     /// Returns a list of queue initialization descriptors.
     std::vector<VkDeviceQueueCreateInfo> GetDeviceQueueCreateInfos() const;
 
@@ -302,6 +309,8 @@ private:
 
     /// Nsight Aftermath GPU crash tracker
     std::unique_ptr<NsightAftermathTracker> nsight_aftermath_tracker;
+
+    u64 device_access_memory;
 };
 
 } // namespace Vulkan

src/video_core/vulkan_common/vulkan_memory_allocator.cpp

@@ -69,10 +69,10 @@ constexpr VkExportMemoryAllocateInfo EXPORT_ALLOCATE_INFO{
 
 class MemoryAllocation {
 public:
-    explicit MemoryAllocation(vk::DeviceMemory memory_, VkMemoryPropertyFlags properties,
-                              u64 allocation_size_, u32 type)
-        : memory{std::move(memory_)}, allocation_size{allocation_size_}, property_flags{properties},
-          shifted_memory_type{1U << type} {}
+    explicit MemoryAllocation(MemoryAllocator* const allocator_, vk::DeviceMemory memory_,
+                              VkMemoryPropertyFlags properties, u64 allocation_size_, u32 type)
+        : allocator{allocator_}, memory{std::move(memory_)}, allocation_size{allocation_size_},
+          property_flags{properties}, shifted_memory_type{1U << type} {}
 
 #if defined(_WIN32) || defined(__unix__)
     ~MemoryAllocation() {
@@ -106,6 +106,10 @@ public:
         const auto it = std::ranges::find(commits, begin, &Range::begin);
         ASSERT_MSG(it != commits.end(), "Invalid commit");
         commits.erase(it);
+        if (commits.empty()) {
+            // Do not call any code involving 'this' after this call, the object will be destroyed
+            allocator->ReleaseMemory(this);
+        }
     }
 
     [[nodiscard]] std::span<u8> Map() {
@@ -171,6 +175,7 @@ private:
         return candidate;
     }
 
+    MemoryAllocator* const allocator;           ///< Parent memory allocation.
     const vk::DeviceMemory memory;              ///< Vulkan memory allocation handler.
     const u64 allocation_size;                  ///< Size of this allocation.
     const VkMemoryPropertyFlags property_flags; ///< Vulkan memory property flags.
@@ -275,10 +280,17 @@ bool MemoryAllocator::TryAllocMemory(VkMemoryPropertyFlags flags, u32 type_mask,
             return false;
         }
     }
-    allocations.push_back(std::make_unique<MemoryAllocation>(std::move(memory), flags, size, type));
+    allocations.push_back(
+        std::make_unique<MemoryAllocation>(this, std::move(memory), flags, size, type));
     return true;
 }
 
+void MemoryAllocator::ReleaseMemory(MemoryAllocation* alloc) {
+    const auto it = std::ranges::find(allocations, alloc, &std::unique_ptr<MemoryAllocation>::get);
+    ASSERT(it != allocations.end());
+    allocations.erase(it);
+}
+
 std::optional<MemoryCommit> MemoryAllocator::TryCommit(const VkMemoryRequirements& requirements,
                                                        VkMemoryPropertyFlags flags) {
     for (auto& allocation : allocations) {

src/video_core/vulkan_common/vulkan_memory_allocator.h

@@ -69,6 +69,8 @@ private:
 /// Memory allocator container.
 /// Allocates and releases memory allocations on demand.
 class MemoryAllocator {
+    friend MemoryAllocation;
+
 public:
     /**
      * Construct memory allocator
@@ -104,6 +106,9 @@ private:
     /// Tries to allocate a chunk of memory.
     bool TryAllocMemory(VkMemoryPropertyFlags flags, u32 type_mask, u64 size);
 
+    /// Releases a chunk of memory.
+    void ReleaseMemory(MemoryAllocation* alloc);
+
     /// Tries to allocate a memory commit.
     std::optional<MemoryCommit> TryCommit(const VkMemoryRequirements& requirements,
                                           VkMemoryPropertyFlags flags);

src/yuzu/configuration/config.cpp

@@ -822,6 +822,7 @@ void Config::ReadRendererValues() {
                       QStringLiteral("use_asynchronous_shaders"), false);
     ReadSettingGlobal(Settings::values.use_fast_gpu_time, QStringLiteral("use_fast_gpu_time"),
                       true);
+    ReadSettingGlobal(Settings::values.use_caches_gc, QStringLiteral("use_caches_gc"), false);
     ReadSettingGlobal(Settings::values.bg_red, QStringLiteral("bg_red"), 0.0);
     ReadSettingGlobal(Settings::values.bg_green, QStringLiteral("bg_green"), 0.0);
     ReadSettingGlobal(Settings::values.bg_blue, QStringLiteral("bg_blue"), 0.0);
@@ -1410,6 +1411,7 @@ void Config::SaveRendererValues() {
                        Settings::values.use_asynchronous_shaders, false);
     WriteSettingGlobal(QStringLiteral("use_fast_gpu_time"), Settings::values.use_fast_gpu_time,
                        true);
+    WriteSettingGlobal(QStringLiteral("use_caches_gc"), Settings::values.use_caches_gc, false);
     // Cast to double because Qt's written float values are not human-readable
     WriteSettingGlobal(QStringLiteral("bg_red"), Settings::values.bg_red, 0.0);
     WriteSettingGlobal(QStringLiteral("bg_green"), Settings::values.bg_green, 0.0);

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -31,6 +31,7 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->disable_fps_limit->setChecked(Settings::values.disable_fps_limit.GetValue());
     ui->use_assembly_shaders->setChecked(Settings::values.use_assembly_shaders.GetValue());
     ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue());
+    ui->use_caches_gc->setChecked(Settings::values.use_caches_gc.GetValue());
     ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time.GetValue());
 
     if (Settings::IsConfiguringGlobal()) {
@@ -65,6 +66,8 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_asynchronous_shaders,
                                              ui->use_asynchronous_shaders,
                                              use_asynchronous_shaders);
+    ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_caches_gc, ui->use_caches_gc,
+                                             use_caches_gc);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_fast_gpu_time,
                                              ui->use_fast_gpu_time, use_fast_gpu_time);
 
@@ -105,6 +108,7 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
         ui->use_asynchronous_shaders->setEnabled(
             Settings::values.use_asynchronous_shaders.UsingGlobal());
         ui->use_fast_gpu_time->setEnabled(Settings::values.use_fast_gpu_time.UsingGlobal());
+        ui->use_caches_gc->setEnabled(Settings::values.use_caches_gc.UsingGlobal());
         ui->anisotropic_filtering_combobox->setEnabled(
             Settings::values.max_anisotropy.UsingGlobal());
 
@@ -121,6 +125,8 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
                                             use_asynchronous_shaders);
     ConfigurationShared::SetColoredTristate(ui->use_fast_gpu_time,
                                             Settings::values.use_fast_gpu_time, use_fast_gpu_time);
+    ConfigurationShared::SetColoredTristate(ui->use_caches_gc, Settings::values.use_caches_gc,
+                                            use_caches_gc);
     ConfigurationShared::SetColoredComboBox(
         ui->gpu_accuracy, ui->label_gpu_accuracy,
         static_cast<int>(Settings::values.gpu_accuracy.GetValue(true)));

src/yuzu/configuration/configure_graphics_advanced.h

@@ -39,4 +39,5 @@ private:
     ConfigurationShared::CheckState use_assembly_shaders;
     ConfigurationShared::CheckState use_asynchronous_shaders;
     ConfigurationShared::CheckState use_fast_gpu_time;
+    ConfigurationShared::CheckState use_caches_gc;
 };

src/yuzu/configuration/configure_graphics_advanced.ui

@@ -121,6 +121,16 @@
           </property>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="use_caches_gc">
+          <property name="toolTip">
+           <string>Enables garbage collection for the GPU caches, this will try to keep VRAM within 3-4 GB by flushing the least used textures/buffers. May cause issues in a few games.</string>
+          </property>
+          <property name="text">
+           <string>Enable GPU cache garbage collection (experimental)</string>
+          </property>
+         </widget>
+        </item>
         <item>
          <widget class="QWidget" name="af_layout" native="true">
           <layout class="QHBoxLayout" name="horizontalLayout_1">

src/yuzu_cmd/default_ini.h

@@ -227,6 +227,10 @@ use_asynchronous_gpu_emulation =
 # 0: Off, 1 (default): On
 use_vsync =
 
+# Whether to use garbage collection or not for GPU caches.
+# 0 (default): Off, 1: On
+use_caches_gc =
+
 # The clear color for the renderer. What shows up on the sides of the bottom screen.
 # Must be in range of 0.0-1.0. Defaults to 1.0 for all.
 bg_red =