buffer_cache: Do not cache buffer views

src/video_core/buffer_cache/buffer.cpp

@@ -91,10 +91,10 @@ void UniqueBuffer::Create(const vk::BufferCreateInfo& buffer_ci, MemoryUsage usa
     buffer = vk::Buffer{unsafe_buffer};
 }
 
-Buffer::Buffer(const Vulkan::Instance& instance_, MemoryUsage usage_, VAddr cpu_addr_,
+Buffer::Buffer(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_, MemoryUsage usage_,
-               vk::BufferUsageFlags flags, u64 size_bytes_)
+               VAddr cpu_addr_, vk::BufferUsageFlags flags, u64 size_bytes_)
-    : cpu_addr{cpu_addr_}, size_bytes{size_bytes_}, instance{&instance_}, usage{usage_},
+    : cpu_addr{cpu_addr_}, size_bytes{size_bytes_}, instance{&instance_}, scheduler{&scheduler_},
-      buffer{instance->GetDevice(), instance->GetAllocator()} {
+      usage{usage_}, buffer{instance->GetDevice(), instance->GetAllocator()} {
     // Create buffer object.
     const vk::BufferCreateInfo buffer_ci = {
         .size = size_bytes,
@@ -117,13 +117,6 @@ Buffer::Buffer(const Vulkan::Instance& instance_, MemoryUsage usage_, VAddr cpu_
 
 vk::BufferView Buffer::View(u32 offset, u32 size, bool is_written, AmdGpu::DataFormat dfmt,
                             AmdGpu::NumberFormat nfmt) {
-    const auto it{std::ranges::find_if(views, [=](const BufferView& view) {
-        return offset == view.offset && size == view.size && is_written == view.is_written &&
-               dfmt == view.dfmt && nfmt == view.nfmt;
-    })};
-    if (it != views.end()) {
-        return *it->handle;
-    }
     const vk::BufferUsageFlags2CreateInfoKHR usage_flags = {
         .usage = is_written ? vk::BufferUsageFlagBits2KHR::eStorageTexelBuffer
                             : vk::BufferUsageFlagBits2KHR::eUniformTexelBuffer,
@@ -135,23 +128,18 @@ vk::BufferView Buffer::View(u32 offset, u32 size, bool is_written, AmdGpu::DataF
         .offset = offset,
         .range = size,
     };
-    views.push_back({
+    const auto view = instance->GetDevice().createBufferView(view_ci);
-        .offset = offset,
+    scheduler->DeferOperation(
-        .size = size,
+        [view, device = instance->GetDevice()] { device.destroyBufferView(view); });
-        .is_written = is_written,
+    return view;
-        .dfmt = dfmt,
-        .nfmt = nfmt,
-        .handle = instance->GetDevice().createBufferViewUnique(view_ci),
-    });
-    return *views.back().handle;
 }
 
 constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000;
 constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000;
 
-StreamBuffer::StreamBuffer(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler_,
+StreamBuffer::StreamBuffer(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
                            MemoryUsage usage, u64 size_bytes)
-    : Buffer{instance, usage, 0, AllFlags, size_bytes}, scheduler{scheduler_} {
+    : Buffer{instance, scheduler, usage, 0, AllFlags, size_bytes} {
     ReserveWatches(current_watches, WATCHES_INITIAL_RESERVE);
     ReserveWatches(previous_watches, WATCHES_INITIAL_RESERVE);
     const auto device = instance.GetDevice();
@@ -206,7 +194,7 @@ void StreamBuffer::Commit() {
 
     auto& watch = current_watches[current_watch_cursor++];
     watch.upper_bound = offset;
-    watch.tick = scheduler.CurrentTick();
+    watch.tick = scheduler->CurrentTick();
 }
 
 void StreamBuffer::ReserveWatches(std::vector<Watch>& watches, std::size_t grow_size) {
@@ -220,7 +208,7 @@ void StreamBuffer::WaitPendingOperations(u64 requested_upper_bound) {
     while (requested_upper_bound > wait_bound && wait_cursor < *invalidation_mark) {
         auto& watch = previous_watches[wait_cursor];
         wait_bound = watch.upper_bound;
-        scheduler.Wait(watch.tick);
+        scheduler->Wait(watch.tick);
         ++wait_cursor;
     }
 }

src/video_core/buffer_cache/buffer.h

@@ -73,8 +73,9 @@ struct UniqueBuffer {
 
 class Buffer {
 public:
-    explicit Buffer(const Vulkan::Instance& instance, MemoryUsage usage, VAddr cpu_addr_,
+    explicit Buffer(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
-                    vk::BufferUsageFlags flags, u64 size_bytes_);
+                    MemoryUsage usage, VAddr cpu_addr_, vk::BufferUsageFlags flags,
+                    u64 size_bytes_);
 
     Buffer& operator=(const Buffer&) = delete;
     Buffer(const Buffer&) = delete;
@@ -144,20 +145,12 @@ public:
     int stream_score = 0;
     size_t size_bytes = 0;
     std::span<u8> mapped_data;
-    const Vulkan::Instance* instance{};
+    const Vulkan::Instance* instance;
+    Vulkan::Scheduler* scheduler;
     MemoryUsage usage;
     UniqueBuffer buffer;
     vk::AccessFlagBits2 access_mask{vk::AccessFlagBits2::eNone};
     vk::PipelineStageFlagBits2 stage{vk::PipelineStageFlagBits2::eNone};
-    struct BufferView {
-        u32 offset;
-        u32 size;
-        bool is_written;
-        AmdGpu::DataFormat dfmt;
-        AmdGpu::NumberFormat nfmt;
-        vk::UniqueBufferView handle;
-    };
-    std::vector<BufferView> views;
 };
 
 class StreamBuffer : public Buffer {
@@ -196,7 +189,6 @@ private:
     void WaitPendingOperations(u64 requested_upper_bound);
 
 private:
-    Vulkan::Scheduler& scheduler;
     u64 offset{};
     u64 mapped_size{};
     std::vector<Watch> current_watches;

src/video_core/buffer_cache/buffer_cache.cpp

@@ -27,7 +27,7 @@ BufferCache::BufferCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& s
       stream_buffer{instance, scheduler, MemoryUsage::Stream, UboStreamBufferSize},
       memory_tracker{&tracker} {
     // Ensure the first slot is used for the null buffer
-    void(slot_buffers.insert(instance, MemoryUsage::DeviceLocal, 0, ReadFlags, 1));
+    void(slot_buffers.insert(instance, scheduler, MemoryUsage::DeviceLocal, 0, ReadFlags, 1));
 }
 
 BufferCache::~BufferCache() = default;
@@ -236,7 +236,7 @@ std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, b
                                                   bool is_texel_buffer) {
     static constexpr u64 StreamThreshold = CACHING_PAGESIZE;
     const bool is_gpu_dirty = memory_tracker.IsRegionGpuModified(device_addr, size);
-    if (!is_written && !is_texel_buffer && size <= StreamThreshold && !is_gpu_dirty) {
+    if (!is_written && size <= StreamThreshold && !is_gpu_dirty) {
         // For small uniform buffers that have not been modified by gpu
         // use device local stream buffer to reduce renderpass breaks.
         const u64 offset = stream_buffer.Copy(device_addr, size, instance.UniformMinAlignment());
@@ -424,8 +424,8 @@ BufferId BufferCache::CreateBuffer(VAddr device_addr, u32 wanted_size) {
     wanted_size = static_cast<u32>(device_addr_end - device_addr);
     const OverlapResult overlap = ResolveOverlaps(device_addr, wanted_size);
     const u32 size = static_cast<u32>(overlap.end - overlap.begin);
-    const BufferId new_buffer_id =
+    const BufferId new_buffer_id = slot_buffers.insert(
-        slot_buffers.insert(instance, MemoryUsage::DeviceLocal, overlap.begin, AllFlags, size);
+        instance, scheduler, MemoryUsage::DeviceLocal, overlap.begin, AllFlags, size);
     auto& new_buffer = slot_buffers[new_buffer_id];
     const size_t size_bytes = new_buffer.SizeBytes();
     const auto cmdbuf = scheduler.CommandBuffer();
@@ -505,7 +505,11 @@ void BufferCache::SynchronizeBuffer(Buffer& buffer, VAddr device_addr, u32 size,
     } else {
         // For large one time transfers use a temporary host buffer.
         // RenderDoc can lag quite a bit if the stream buffer is too large.
-        Buffer temp_buffer{instance, MemoryUsage::Upload, 0, vk::BufferUsageFlagBits::eTransferSrc,
+        Buffer temp_buffer{instance,
+                           scheduler,
+                           MemoryUsage::Upload,
+                           0,
+                           vk::BufferUsageFlagBits::eTransferSrc,
                            total_size_bytes};
         src_buffer = temp_buffer.Handle();
         u8* const staging = temp_buffer.mapped_data.data();