buffer: Don't leak texel buffers, blacklist stream buffer with texel buffers

src/video_core/buffer_cache/buffer.cpp

@@ -119,7 +119,7 @@ vk::BufferView Buffer::View(u32 offset, u32 size, bool is_written, AmdGpu::DataF
                dfmt == view.dfmt && nfmt == view.nfmt;
     })};
     if (it != views.end()) {
-        return it->handle;
+        return *it->handle;
     }
     const vk::BufferUsageFlags2CreateInfoKHR usage_flags = {
         .usage = is_written ? vk::BufferUsageFlagBits2KHR::eStorageTexelBuffer
@@ -138,9 +138,9 @@ vk::BufferView Buffer::View(u32 offset, u32 size, bool is_written, AmdGpu::DataF
         .is_written = is_written,
         .dfmt = dfmt,
         .nfmt = nfmt,
-        .handle = instance->GetDevice().createBufferView(view_ci),
+        .handle = instance->GetDevice().createBufferViewUnique(view_ci),
     });
-    return views.back().handle;
+    return *views.back().handle;
 }
 
 constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000;

src/video_core/buffer_cache/buffer.h

@@ -134,7 +134,7 @@ public:
         bool is_written;
         AmdGpu::DataFormat dfmt;
         AmdGpu::NumberFormat nfmt;
-        vk::BufferView handle;
+        vk::UniqueBufferView handle;
     };
     std::vector<BufferView> views;
 };

src/video_core/buffer_cache/buffer_cache.cpp

@@ -228,11 +228,12 @@ u32 BufferCache::BindIndexBuffer(bool& is_indexed, u32 index_offset) {
     return regs.num_indices;
 }
 
-std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, bool is_written) {
+std::pair<Buffer*, u32> BufferCache::ObtainBuffer(VAddr device_addr, u32 size, bool is_written,
+                                                  bool is_texel_buffer) {
     std::scoped_lock lk{mutex};
     static constexpr u64 StreamThreshold = CACHING_PAGESIZE;
     const bool is_gpu_dirty = memory_tracker.IsRegionGpuModified(device_addr, size);
-    if (!is_written && size < StreamThreshold && !is_gpu_dirty) {
+    if (!is_written && !is_texel_buffer && 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());

src/video_core/buffer_cache/buffer_cache.h

@@ -66,7 +66,8 @@ public:
     u32 BindIndexBuffer(bool& is_indexed, u32 index_offset);
 
     /// Obtains a buffer for the specified region.
-    [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(VAddr gpu_addr, u32 size, bool is_written);
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(VAddr gpu_addr, u32 size, bool is_written,
+                                                       bool is_texel_buffer = false);
 
     /// Obtains a temporary buffer for usage in texture cache.
     [[nodiscard]] std::pair<const Buffer*, u32> ObtainTempBuffer(VAddr gpu_addr, u32 size);

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -171,7 +171,7 @@ bool ComputePipeline::BindResources(VideoCore::BufferCache& buffer_cache,
             }
             const u32 alignment = instance.TexelBufferMinAlignment();
             const auto [vk_buffer, offset] =
-                buffer_cache.ObtainBuffer(address, size, desc.is_written);
+                buffer_cache.ObtainBuffer(address, size, desc.is_written, true);
             const u32 fmt_stride = AmdGpu::NumBits(vsharp.GetDataFmt()) >> 3;
             ASSERT_MSG(fmt_stride == vsharp.GetStride(),
                        "Texel buffer stride must match format stride");

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -412,7 +412,7 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
                 const u32 size = vsharp.GetSize();
                 const u32 alignment = instance.TexelBufferMinAlignment();
                 const auto [vk_buffer, offset] =
-                    buffer_cache.ObtainBuffer(address, size, tex_buffer.is_written);
+                    buffer_cache.ObtainBuffer(address, size, tex_buffer.is_written, true);
                 const u32 fmt_stride = AmdGpu::NumBits(vsharp.GetDataFmt()) >> 3;
                 ASSERT_MSG(fmt_stride == vsharp.GetStride(),
                            "Texel buffer stride must match format stride");