early-access version 1786

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 1785.
+This is the source code for early-access 1786.
 
 ## Legal Notice
 

src/common/settings.cpp

@@ -55,9 +55,11 @@ void LogSettings() {
     log_setting("Renderer_UseAsynchronousGpuEmulation",
                 values.use_asynchronous_gpu_emulation.GetValue());
     log_setting("Renderer_UseNvdecEmulation", values.use_nvdec_emulation.GetValue());
+    log_setting("Renderer_AccelerateASTC", values.accelerate_astc.GetValue());
     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());
@@ -135,10 +137,12 @@ void RestoreGlobalState(bool is_powered_on) {
     values.gpu_accuracy.SetGlobal(true);
     values.use_asynchronous_gpu_emulation.SetGlobal(true);
     values.use_nvdec_emulation.SetGlobal(true);
+    values.accelerate_astc.SetGlobal(true);
     values.use_vsync.SetGlobal(true);
     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

@@ -147,10 +147,12 @@ struct Values {
     Setting<GPUAccuracy> gpu_accuracy;
     Setting<bool> use_asynchronous_gpu_emulation;
     Setting<bool> use_nvdec_emulation;
+    Setting<bool> accelerate_astc;
     Setting<bool> use_vsync;
     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/core/telemetry_session.cpp

@@ -230,6 +230,7 @@ void TelemetrySession::AddInitialInfo(Loader::AppLoader& app_loader,
              Settings::values.use_asynchronous_gpu_emulation.GetValue());
     AddField(field_type, "Renderer_UseNvdecEmulation",
              Settings::values.use_nvdec_emulation.GetValue());
+    AddField(field_type, "Renderer_AccelerateASTC", Settings::values.accelerate_astc.GetValue());
     AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync.GetValue());
     AddField(field_type, "Renderer_UseAssemblyShaders",
              Settings::values.use_assembly_shaders.GetValue());

src/video_core/CMakeLists.txt

@@ -237,6 +237,7 @@ add_library(video_core STATIC
     texture_cache/util.cpp
     texture_cache/util.h
     textures/astc.h
+    textures/astc.cpp
     textures/decoders.cpp
     textures/decoders.h
     textures/texture.cpp

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

@@ -65,6 +65,9 @@ class BufferCache {
 
     static constexpr BufferId NULL_BUFFER_ID{0};
 
+    static constexpr u64 expected_memory = 512ULL * 1024ULL * 1024ULL;
+    static constexpr u64 critical_memory = 1024ULL * 1024ULL * 1024ULL;
+
     using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
     using Runtime = typename P::Runtime;
@@ -243,6 +246,8 @@ 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 +260,10 @@ 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 +328,10 @@ 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;
 };
 
@@ -332,10 +345,16 @@ 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>::TickFrame() {
+    const bool enabled_gc = Settings::values.use_caches_gc.GetValue();
+    SCOPE_EXIT({
+        ++frame_tick;
+        delayed_destruction_ring.Tick();
+    });
     // Calculate hits and shots and move hit bits to the right
     const u32 hits = std::reduce(uniform_cache_hits.begin(), uniform_cache_hits.end());
     const u32 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end());
@@ -349,7 +368,27 @@ void BufferCache<P>::TickFrame() {
     const bool skip_preferred = hits * 256 < shots * 251;
     uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
 
-    delayed_destruction_ring.Tick();
+    const bool activate_gc = enabled_gc && total_used_memory >= expected_memory;
+    if (!activate_gc) {
+        return;
+    }
+    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>
@@ -371,50 +410,8 @@ 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) {
+    ForEachBufferInRange(cpu_addr, size,
-        boost::container::small_vector<BufferCopy, 1> copies;
+                         [&](BufferId, Buffer& buffer) { DownloadBufferMemory(buffer); });
-        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>
@@ -640,6 +637,7 @@ 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 +656,7 @@ 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 +692,7 @@ 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 +744,7 @@ 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 +767,7 @@ 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 +786,7 @@ 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 +806,7 @@ 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 +1105,7 @@ 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 +1127,14 @@ 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 +1146,11 @@ 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 +1227,57 @@ 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 +1303,7 @@ 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

@@ -9,6 +9,8 @@
 
 #include <glad/glad.h>
 
+#include "common/settings.h"
+
 #include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
 #include "video_core/renderer_opengl/gl_state_tracker.h"
@@ -307,7 +309,9 @@ void ApplySwizzle(GLuint handle, PixelFormat format, std::array<SwizzleSource, 4
 
 [[nodiscard]] bool CanBeAccelerated(const TextureCacheRuntime& runtime,
                                     const VideoCommon::ImageInfo& info) {
-    return !runtime.HasNativeASTC() && IsPixelFormatASTC(info.format);
+    if (IsPixelFormatASTC(info.format)) {
+        return !runtime.HasNativeASTC() && Settings::values.accelerate_astc.GetValue();
+    }
     // Disable other accelerated uploads for now as they don't implement swizzled uploads
     return false;
     switch (info.type) {

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -8,6 +8,7 @@
 #include <vector>
 
 #include "common/bit_cast.h"
+#include "common/settings.h"
 
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/renderer_vulkan/blit_image.h"
@@ -828,7 +829,11 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
         commit = runtime.memory_allocator.Commit(buffer, MemoryUsage::DeviceLocal);
     }
     if (IsPixelFormatASTC(info.format) && !runtime.device.IsOptimalAstcSupported()) {
-        flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
+        if (Settings::values.accelerate_astc.GetValue()) {
+            flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
+        } else {
+            flags |= VideoCommon::ImageFlagBits::Converted;
+        }
     }
     if (runtime.device.HasDebuggingToolAttached()) {
         if (image) {

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) 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

@@ -22,6 +22,7 @@
 #include "common/common_funcs.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"
@@ -75,6 +76,9 @@ class TextureCache {
     /// Sampler ID for bugged sampler ids
     static constexpr SamplerId NULL_SAMPLER_ID{0};
 
+    static constexpr u64 expected_memory = 1024ULL * 1024ULL * 1024ULL;
+    static constexpr u64 critical_memory = 2 * 1024ULL * 1024ULL * 1024ULL;
+
     using Runtime = typename P::Runtime;
     using Image = typename P::Image;
     using ImageAlloc = typename P::ImageAlloc;
@@ -333,6 +337,7 @@ private:
     std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> page_table;
 
     bool has_deleted_images = false;
+    u64 total_used_memory = 0;
 
     SlotVector<Image> slot_images;
     SlotVector<ImageView> slot_image_views;
@@ -353,6 +358,7 @@ private:
 
     u64 modification_tick = 0;
     u64 frame_tick = 0;
+    typename SlotVector<Image>::Iterator deletion_iterator;
 };
 
 template <class P>
@@ -373,11 +379,82 @@ 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();
 }
 
 template <class P>
 void TextureCache<P>::TickFrame() {
-    // Tick sentenced resources in this order to ensure they are destroyed in the right order
+    const bool enabled_gc = Settings::values.use_caches_gc.GetValue();
+    if (!enabled_gc) {
+        // @Note(Blinkhawk): compile error with SCOPE_EXIT on msvc.
+        sentenced_images.Tick();
+        sentenced_framebuffers.Tick();
+        sentenced_image_view.Tick();
+        ++frame_tick;
+        return;
+    }
+    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;
+    }
     sentenced_images.Tick();
     sentenced_framebuffers.Tick();
     sentenced_image_view.Tick();
@@ -568,17 +645,7 @@ 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 (!image.IsSafeDownload()) {
-        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");
             return;
         }
         image.flags &= ~ImageFlagBits::GpuModified;
@@ -967,6 +1034,7 @@ 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 +1060,14 @@ 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 +1095,18 @@ 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 +1276,8 @@ 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); });
+    total_used_memory +=
+        Common::AlignUp(std::max(image.guest_size_bytes, image.unswizzled_size_bytes), 1024);
 }
 
 template <class P>
@@ -1203,6 +1286,9 @@ 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;
+    total_used_memory -=
+        Common::AlignUp(std::max(image.guest_size_bytes, image.unswizzled_size_bytes), 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 +1362,19 @@ 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

@@ -47,6 +47,7 @@
 #include "video_core/texture_cache/formatter.h"
 #include "video_core/texture_cache/samples_helper.h"
 #include "video_core/texture_cache/util.h"
+#include "video_core/textures/astc.h"
 #include "video_core/textures/decoders.h"
 
 namespace VideoCommon {
@@ -580,6 +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);
 
@@ -884,8 +887,16 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
         ASSERT(copy.image_extent == mip_size);
         ASSERT(copy.buffer_row_length == Common::AlignUp(mip_size.width, tile_size.width));
         ASSERT(copy.buffer_image_height == Common::AlignUp(mip_size.height, tile_size.height));
-        DecompressBC4(input.subspan(copy.buffer_offset), copy.image_extent,
+        if (IsPixelFormatASTC(info.format)) {
-                      output.subspan(output_offset));
+            ASSERT(copy.image_extent.depth == 1);
+            Tegra::Texture::ASTC::Decompress(input.subspan(copy.buffer_offset),
+                                             copy.image_extent.width, copy.image_extent.height,
+                                             copy.image_subresource.num_layers, tile_size.width,
+                                             tile_size.height, output.subspan(output_offset));
+        } else {
+            DecompressBC4(input.subspan(copy.buffer_offset), copy.image_extent,
+                          output.subspan(output_offset));
+        }
         copy.buffer_offset = output_offset;
         copy.buffer_row_length = mip_size.width;
         copy.buffer_image_height = mip_size.height;

src/video_core/textures/astc.cpp

@@ -24,22 +24,8 @@
 #include <boost/container/static_vector.hpp>
 
 #include "common/common_types.h"
-
 #include "video_core/textures/astc.h"
 
-namespace {
-
-/// Count the number of bits set in a number.
-constexpr u32 Popcnt(u32 n) {
-    u32 c = 0;
-    for (; n; c++) {
-        n &= n - 1;
-    }
-    return c;
-}
-
-} // Anonymous namespace
-
 class InputBitStream {
 public:
     constexpr explicit InputBitStream(std::span<const u8> data, size_t start_offset = 0)
@@ -165,37 +151,7 @@ private:
     const IntType& m_Bits;
 };
 
-enum class IntegerEncoding { JustBits, Qus32, Trit };
+namespace Tegra::Texture::ASTC {
-
-struct IntegerEncodedValue {
-    constexpr IntegerEncodedValue() = default;
-
-    constexpr IntegerEncodedValue(IntegerEncoding encoding_, u32 num_bits_)
-        : encoding{encoding_}, num_bits{num_bits_} {}
-
-    constexpr bool MatchesEncoding(const IntegerEncodedValue& other) const {
-        return encoding == other.encoding && num_bits == other.num_bits;
-    }
-
-    // Returns the number of bits required to encode nVals values.
-    u32 GetBitLength(u32 nVals) const {
-        u32 totalBits = num_bits * nVals;
-        if (encoding == IntegerEncoding::Trit) {
-            totalBits += (nVals * 8 + 4) / 5;
-        } else if (encoding == IntegerEncoding::Qus32) {
-            totalBits += (nVals * 7 + 2) / 3;
-        }
-        return totalBits;
-    }
-
-    IntegerEncoding encoding{};
-    u32 num_bits = 0;
-    u32 bit_value = 0;
-    union {
-        u32 qus32_value = 0;
-        u32 trit_value;
-    };
-};
 using IntegerEncodedVector = boost::container::static_vector<
     IntegerEncodedValue, 256,
     boost::container::static_vector_options<
@@ -260,7 +216,7 @@ static void DecodeTritBlock(InputBitStream& bits, IntegerEncodedVector& result,
     }
 }
 
-static void DecodeQus32Block(InputBitStream& bits, IntegerEncodedVector& result,
+static void DecodeQuintBlock(InputBitStream& bits, IntegerEncodedVector& result,
                              u32 nBitsPerValue) {
     // Implement the algorithm in section C.2.12
     u32 m[3];
@@ -301,50 +257,12 @@ static void DecodeQus32Block(InputBitStream& bits, IntegerEncodedVector& result,
     }
 
     for (std::size_t i = 0; i < 3; ++i) {
-        IntegerEncodedValue& val = result.emplace_back(IntegerEncoding::Qus32, nBitsPerValue);
+        IntegerEncodedValue& val = result.emplace_back(IntegerEncoding::Quint, nBitsPerValue);
         val.bit_value = m[i];
-        val.qus32_value = q[i];
+        val.quint_value = q[i];
     }
 }
 
-// Returns a new instance of this struct that corresponds to the
-// can take no more than maxval values
-static constexpr IntegerEncodedValue CreateEncoding(u32 maxVal) {
-    while (maxVal > 0) {
-        u32 check = maxVal + 1;
-
-        // Is maxVal a power of two?
-        if (!(check & (check - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::JustBits, Popcnt(maxVal));
-        }
-
-        // Is maxVal of the type 3*2^n - 1?
-        if ((check % 3 == 0) && !((check / 3) & ((check / 3) - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::Trit, Popcnt(check / 3 - 1));
-        }
-
-        // Is maxVal of the type 5*2^n - 1?
-        if ((check % 5 == 0) && !((check / 5) & ((check / 5) - 1))) {
-            return IntegerEncodedValue(IntegerEncoding::Qus32, Popcnt(check / 5 - 1));
-        }
-
-        // Apparently it can't be represented with a bounded integer sequence...
-        // just iterate.
-        maxVal--;
-    }
-    return IntegerEncodedValue(IntegerEncoding::JustBits, 0);
-}
-
-static constexpr std::array<IntegerEncodedValue, 256> MakeEncodedValues() {
-    std::array<IntegerEncodedValue, 256> encodings{};
-    for (std::size_t i = 0; i < encodings.size(); ++i) {
-        encodings[i] = CreateEncoding(static_cast<u32>(i));
-    }
-    return encodings;
-}
-
-static constexpr std::array EncodingsValues = MakeEncodedValues();
-
 // Fills result with the values that are encoded in the given
 // bitstream. We must know beforehand what the maximum possible
 // value is, and how many values we're decoding.
@@ -357,8 +275,8 @@ static void DecodeIntegerSequence(IntegerEncodedVector& result, InputBitStream&
     u32 nValsDecoded = 0;
     while (nValsDecoded < nValues) {
         switch (val.encoding) {
-        case IntegerEncoding::Qus32:
+        case IntegerEncoding::Quint:
-            DecodeQus32Block(bits, result, val.num_bits);
+            DecodeQuintBlock(bits, result, val.num_bits);
             nValsDecoded += 3;
             break;
 
@@ -376,8 +294,6 @@ static void DecodeIntegerSequence(IntegerEncodedVector& result, InputBitStream&
     }
 }
 
-namespace ASTCC {
-
 struct TexelWeightParams {
     u32 m_Width = 0;
     u32 m_Height = 0;
@@ -635,48 +551,6 @@ static void FillError(std::span<u32> outBuf, u32 blockWidth, u32 blockHeight) {
         }
     }
 }
-
-// Replicates low numBits such that [(toBit - 1):(toBit - 1 - fromBit)]
-// is the same as [(numBits - 1):0] and repeats all the way down.
-template <typename IntType>
-static constexpr IntType Replicate(IntType val, u32 numBits, u32 toBit) {
-    if (numBits == 0) {
-        return 0;
-    }
-    if (toBit == 0) {
-        return 0;
-    }
-    const IntType v = val & static_cast<IntType>((1 << numBits) - 1);
-    IntType res = v;
-    u32 reslen = numBits;
-    while (reslen < toBit) {
-        u32 comp = 0;
-        if (numBits > toBit - reslen) {
-            u32 newshift = toBit - reslen;
-            comp = numBits - newshift;
-            numBits = newshift;
-        }
-        res = static_cast<IntType>(res << numBits);
-        res = static_cast<IntType>(res | (v >> comp));
-        reslen += numBits;
-    }
-    return res;
-}
-
-static constexpr std::size_t NumReplicateEntries(u32 num_bits) {
-    return std::size_t(1) << num_bits;
-}
-
-template <typename IntType, u32 num_bits, u32 to_bit>
-static constexpr auto MakeReplicateTable() {
-    std::array<IntType, NumReplicateEntries(num_bits)> table{};
-    for (IntType value = 0; value < static_cast<IntType>(std::size(table)); ++value) {
-        table[value] = Replicate(value, num_bits, to_bit);
-    }
-    return table;
-}
-
-static constexpr auto REPLICATE_BYTE_TO_16_TABLE = MakeReplicateTable<u32, 8, 16>();
 static constexpr u32 ReplicateByteTo16(std::size_t value) {
     return REPLICATE_BYTE_TO_16_TABLE[value];
 }
@@ -696,9 +570,6 @@ static constexpr auto REPLICATE_2_BIT_TO_8_TABLE = MakeReplicateTable<u32, 2, 8>
 static constexpr auto REPLICATE_3_BIT_TO_8_TABLE = MakeReplicateTable<u32, 3, 8>();
 static constexpr auto REPLICATE_4_BIT_TO_8_TABLE = MakeReplicateTable<u32, 4, 8>();
 static constexpr auto REPLICATE_5_BIT_TO_8_TABLE = MakeReplicateTable<u32, 5, 8>();
-static constexpr auto REPLICATE_6_BIT_TO_8_TABLE = MakeReplicateTable<u32, 6, 8>();
-static constexpr auto REPLICATE_7_BIT_TO_8_TABLE = MakeReplicateTable<u32, 7, 8>();
-static constexpr auto REPLICATE_8_BIT_TO_8_TABLE = MakeReplicateTable<u32, 8, 8>();
 /// Use a precompiled table with the most common usages, if it's not in the expected range, fallback
 /// to the runtime implementation
 static constexpr u32 FastReplicateTo8(u32 value, u32 num_bits) {
@@ -982,9 +853,9 @@ static void DecodeColorValues(u32* out, std::span<u8> data, const u32* modes, co
         }     // case IntegerEncoding::Trit
         break;
 
-        case IntegerEncoding::Qus32: {
+        case IntegerEncoding::Quint: {
 
-            D = val.qus32_value;
+            D = val.quint_value;
 
             switch (bitlen) {
             case 1: {
@@ -1023,7 +894,7 @@ static void DecodeColorValues(u32* out, std::span<u8> data, const u32* modes, co
                 assert(false && "Unsupported quint encoding for color values!");
                 break;
             } // switch(bitlen)
-        }     // case IntegerEncoding::Qus32
+        }     // case IntegerEncoding::Quint
         break;
         } // switch(val.encoding)
 
@@ -1086,8 +957,8 @@ static u32 UnquantizeTexelWeight(const IntegerEncodedValue& val) {
         }
     } break;
 
-    case IntegerEncoding::Qus32: {
+    case IntegerEncoding::Quint: {
-        D = val.qus32_value;
+        D = val.quint_value;
         assert(D < 5);
 
         switch (bitlen) {
@@ -1675,10 +1546,6 @@ static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
         }
 }
 
-} // namespace ASTCC
-
-namespace Tegra::Texture::ASTC {
-
 void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
                 uint32_t block_width, uint32_t block_height, std::span<uint8_t> output) {
     u32 block_index = 0;
@@ -1690,7 +1557,7 @@ void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height,
 
                 // Blocks can be at most 12x12
                 std::array<u32, 12 * 12> uncompData;
-                ASTCC::DecompressBlock(blockPtr, block_width, block_height, uncompData);
+                DecompressBlock(blockPtr, block_width, block_height, uncompData);
 
                 u32 decompWidth = std::min(block_width, width - x);
                 u32 decompHeight = std::min(block_height, height - y);

src/video_core/textures/astc.h

@@ -129,4 +129,7 @@ struct AstcBufferData {
     decltype(REPLICATE_BYTE_TO_16_TABLE) replicate_byte_to_16 = REPLICATE_BYTE_TO_16_TABLE;
 } constexpr ASTC_BUFFER_DATA;
 
+void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
+                uint32_t block_width, uint32_t block_height, std::span<uint8_t> output);
+
 } // namespace Tegra::Texture::ASTC

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,
+    explicit MemoryAllocation(MemoryAllocator* const allocator_, vk::DeviceMemory memory_,
-                              u64 allocation_size_, u32 type)
+                              VkMemoryPropertyFlags properties, u64 allocation_size_, u32 type)
-        : memory{std::move(memory_)}, allocation_size{allocation_size_}, property_flags{properties},
+        : allocator{allocator_}, memory{std::move(memory_)}, allocation_size{allocation_size_},
-          shifted_memory_type{1U << type} {}
+          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

@@ -809,6 +809,7 @@ void Config::ReadRendererValues() {
                       QStringLiteral("use_asynchronous_gpu_emulation"), true);
     ReadSettingGlobal(Settings::values.use_nvdec_emulation, QStringLiteral("use_nvdec_emulation"),
                       true);
+    ReadSettingGlobal(Settings::values.accelerate_astc, QStringLiteral("accelerate_astc"), true);
     ReadSettingGlobal(Settings::values.use_vsync, QStringLiteral("use_vsync"), true);
     ReadSettingGlobal(Settings::values.use_assembly_shaders, QStringLiteral("use_assembly_shaders"),
                       false);
@@ -816,6 +817,7 @@ 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);
@@ -1392,6 +1394,7 @@ void Config::SaveRendererValues() {
                        Settings::values.use_asynchronous_gpu_emulation, true);
     WriteSettingGlobal(QStringLiteral("use_nvdec_emulation"), Settings::values.use_nvdec_emulation,
                        true);
+    WriteSettingGlobal(QStringLiteral("accelerate_astc"), Settings::values.accelerate_astc, true);
     WriteSettingGlobal(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
     WriteSettingGlobal(QStringLiteral("use_assembly_shaders"),
                        Settings::values.use_assembly_shaders, false);
@@ -1399,6 +1402,7 @@ 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.cpp

@@ -70,10 +70,12 @@ void ConfigureGraphics::SetConfiguration() {
     ui->use_asynchronous_gpu_emulation->setEnabled(runtime_lock);
     ui->use_disk_shader_cache->setEnabled(runtime_lock);
     ui->use_nvdec_emulation->setEnabled(runtime_lock);
+    ui->accelerate_astc->setEnabled(runtime_lock);
     ui->use_disk_shader_cache->setChecked(Settings::values.use_disk_shader_cache.GetValue());
     ui->use_asynchronous_gpu_emulation->setChecked(
         Settings::values.use_asynchronous_gpu_emulation.GetValue());
     ui->use_nvdec_emulation->setChecked(Settings::values.use_nvdec_emulation.GetValue());
+    ui->accelerate_astc->setChecked(Settings::values.accelerate_astc.GetValue());
 
     if (Settings::IsConfiguringGlobal()) {
         ui->api->setCurrentIndex(static_cast<int>(Settings::values.renderer_backend.GetValue()));
@@ -118,6 +120,8 @@ void ConfigureGraphics::ApplyConfiguration() {
                                              use_asynchronous_gpu_emulation);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_nvdec_emulation,
                                              ui->use_nvdec_emulation, use_nvdec_emulation);
+    ConfigurationShared::ApplyPerGameSetting(&Settings::values.accelerate_astc, ui->accelerate_astc,
+                                             accelerate_astc);
 
     if (Settings::IsConfiguringGlobal()) {
         // Guard if during game and set to game-specific value
@@ -254,6 +258,7 @@ void ConfigureGraphics::SetupPerGameUI() {
         ui->use_asynchronous_gpu_emulation->setEnabled(
             Settings::values.use_asynchronous_gpu_emulation.UsingGlobal());
         ui->use_nvdec_emulation->setEnabled(Settings::values.use_nvdec_emulation.UsingGlobal());
+        ui->accelerate_astc->setEnabled(Settings::values.accelerate_astc.UsingGlobal());
         ui->use_disk_shader_cache->setEnabled(Settings::values.use_disk_shader_cache.UsingGlobal());
         ui->bg_button->setEnabled(Settings::values.bg_red.UsingGlobal());
 
@@ -269,6 +274,8 @@ void ConfigureGraphics::SetupPerGameUI() {
         ui->use_disk_shader_cache, Settings::values.use_disk_shader_cache, use_disk_shader_cache);
     ConfigurationShared::SetColoredTristate(
         ui->use_nvdec_emulation, Settings::values.use_nvdec_emulation, use_nvdec_emulation);
+    ConfigurationShared::SetColoredTristate(ui->accelerate_astc, Settings::values.accelerate_astc,
+                                            accelerate_astc);
     ConfigurationShared::SetColoredTristate(ui->use_asynchronous_gpu_emulation,
                                             Settings::values.use_asynchronous_gpu_emulation,
                                             use_asynchronous_gpu_emulation);

src/yuzu/configuration/configure_graphics.h

@@ -47,6 +47,7 @@ private:
     QColor bg_color;
 
     ConfigurationShared::CheckState use_nvdec_emulation;
+    ConfigurationShared::CheckState accelerate_astc;
     ConfigurationShared::CheckState use_disk_shader_cache;
     ConfigurationShared::CheckState use_asynchronous_gpu_emulation;
 

src/yuzu/configuration/configure_graphics.ui

@@ -104,6 +104,13 @@
           </property>
          </widget>
         </item>
+        <item>
+          <widget class="QCheckBox" name="accelerate_astc">
+            <property name="text">
+              <string>Accelerate ASTC texture decoding</string>
+            </property>
+          </widget>
+        </item>
         <item>
          <widget class="QWidget" name="fullscreen_mode_layout" native="true">
           <layout class="QHBoxLayout" name="horizontalLayout_1">

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -30,6 +30,7 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->use_vsync->setChecked(Settings::values.use_vsync.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()) {
@@ -62,6 +63,8 @@ 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);
 
@@ -101,6 +104,7 @@ 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());
 
@@ -115,6 +119,8 @@ 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

@@ -38,4 +38,5 @@ 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

@@ -103,6 +103,16 @@
           </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 caches garbage collection (unsafe)</string>
+          </property>
+         </widget>
+        </item>
         <item>
          <widget class="QWidget" name="af_layout" native="true">
           <layout class="QHBoxLayout" name="horizontalLayout_1">

src/yuzu_cmd/config.cpp

@@ -447,8 +447,10 @@ void Config::ReadValues() {
         sdl2_config->GetBoolean("Renderer", "use_assembly_shaders", true));
     Settings::values.use_asynchronous_shaders.SetValue(
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_shaders", false));
-    Settings::values.use_asynchronous_shaders.SetValue(
+    Settings::values.use_nvdec_emulation.SetValue(
-        sdl2_config->GetBoolean("Renderer", "use_asynchronous_shaders", false));
+        sdl2_config->GetBoolean("Renderer", "use_nvdec_emulation", true));
+    Settings::values.accelerate_astc.SetValue(
+        sdl2_config->GetBoolean("Renderer", "accelerate_astc", true));
     Settings::values.use_fast_gpu_time.SetValue(
         sdl2_config->GetBoolean("Renderer", "use_fast_gpu_time", true));
 

src/yuzu_cmd/default_ini.h

@@ -194,6 +194,14 @@ use_assembly_shaders =
 # 0 (default): Off, 1: On
 use_asynchronous_shaders =
 
+# Enable NVDEC emulation.
+# 0: Off, 1 (default): On
+use_nvdec_emulation =
+
+# Accelerate ASTC texture decoding.
+# 0: Off, 1 (default): On
+accelerate_astc =
+
 # Turns on the frame limiter, which will limit frames output to the target game speed
 # 0: Off, 1: On (default)
 use_frame_limit =
@@ -219,6 +227,10 @@ use_asynchronous_gpu_emulation =
 # 0: Off, 1 (default): On
 use_vsync =
 
+# Whether to use garbage collection or not.
+# 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 =