Initial Reaper Setup

WIP

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

@@ -243,6 +243,8 @@ 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 +257,10 @@ 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 +325,9 @@ private:
     size_t immediate_buffer_capacity = 0;
     std::unique_ptr<u8[]> immediate_buffer_alloc;
 
+    typename SlotVector<Buffer>::Iterator deletion_iterator;
+    u64 frame_tick = 0;
+
     std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
 };
 
@@ -332,6 +341,7 @@ 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>
@@ -349,7 +359,24 @@ void BufferCache<P>::TickFrame() {
     const bool skip_preferred = hits * 256 < shots * 251;
     uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
 
+    static constexpr u64 ticks_to_destroy = 120;
+    int num_iterations = 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);
+        }
+    }
     delayed_destruction_ring.Tick();
+    ++frame_tick;
 }
 
 template <class P>
@@ -371,50 +398,8 @@ 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 +625,7 @@ 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 +644,7 @@ 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 +680,7 @@ 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 +732,7 @@ 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 +755,7 @@ 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 +774,7 @@ 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 +794,7 @@ 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 +1093,7 @@ 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);
     }
@@ -1135,6 +1128,11 @@ 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 +1209,57 @@ 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 +1285,7 @@ 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/texture_cache/image_base.cpp

@@ -113,6 +113,23 @@ 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 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

@@ -44,6 +44,8 @@ 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;

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

@@ -353,6 +353,7 @@ private:
 
     u64 modification_tick = 0;
     u64 frame_tick = 0;
+    typename SlotVector<Image>::Iterator deletion_iterator;
 };
 
 template <class P>
@@ -373,10 +374,41 @@ 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() {
+    static constexpr u64 ticks_to_destroy = 120;
+    int num_iterations = 32;
+    for (; num_iterations > 0; --num_iterations) {
+        if (deletion_iterator == slot_images.end()) {
+            deletion_iterator = slot_images.begin();
+            if (deletion_iterator == slot_images.end()) {
+                break;
+            }
+        }
+        const auto [image_id, image] = *deletion_iterator;
+        if (image->frame_tick + ticks_to_destroy < frame_tick) {
+            if (image->IsSafeDownload() &&
+                std::ranges::none_of(image->aliased_images, [&](const AliasedImage& alias) {
+                    return slot_images[alias.id].modification_tick > image->modification_tick;
+                })) {
+                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);
+        }
+        ++deletion_iterator;
+    }
     // Tick sentenced resources in this order to ensure they are destroyed in the right order
     sentenced_images.Tick();
     sentenced_framebuffers.Tick();
@@ -568,17 +600,7 @@ 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;