texture_cache: Use pre-allocated buffer for texture uploads

src/video_core/texture_cache/texture_cache.h

@@ -39,6 +39,12 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
     sampler_descriptor.mipmap_filter.Assign(Tegra::Texture::TextureMipmapFilter::Linear);
     sampler_descriptor.cubemap_anisotropy.Assign(1);
 
+    // These values were chosen based on typical peak swizzle data sizes seen in some titles
+    static constexpr size_t SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 8_MiB;
+    static constexpr size_t UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 1_MiB;
+    swizzle_data_buffer.reserve(SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY);
+    unswizzle_data_buffer.reserve(UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY);
+
     // Make sure the first index is reserved for the null resources
     // This way the null resource becomes a compile time constant
     void(slot_images.insert(NullImageParams{}));
@@ -734,13 +740,21 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
         gpu_memory->ReadBlockUnsafe(gpu_addr, mapped_span.data(), mapped_span.size_bytes());
         const auto uploads = FullUploadSwizzles(image.info);
         runtime.AccelerateImageUpload(image, staging, uploads);
-    } else if (True(image.flags & ImageFlagBits::Converted)) {
-        std::vector<u8> unswizzled_data(image.unswizzled_size_bytes);
-        auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, unswizzled_data);
-        ConvertImage(unswizzled_data, image.info, mapped_span, copies);
+        return;
+    }
+    const size_t guest_size_bytes = image.guest_size_bytes;
+    swizzle_data_buffer.resize(guest_size_bytes);
+    gpu_memory->ReadBlockUnsafe(gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
+
+    if (True(image.flags & ImageFlagBits::Converted)) {
+        unswizzle_data_buffer.resize(image.unswizzled_size_bytes);
+        auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer,
+                                     unswizzle_data_buffer);
+        ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies);
         image.UploadMemory(staging, copies);
     } else {
-        const auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, mapped_span);
+        const auto copies =
+            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, mapped_span);
         image.UploadMemory(staging, copies);
     }
 }
@@ -910,7 +924,7 @@ void TextureCache<P>::InvalidateScale(Image& image) {
 }
 
 template <class P>
-u64 TextureCache<P>::GetScaledImageSizeBytes(ImageBase& image) {
+u64 TextureCache<P>::GetScaledImageSizeBytes(const ImageBase& image) {
     const u64 scale_up = static_cast<u64>(Settings::values.resolution_info.up_scale *
                                           Settings::values.resolution_info.up_scale);
     const u64 down_shift = static_cast<u64>(Settings::values.resolution_info.down_shift +

src/video_core/texture_cache/texture_cache_base.h

@@ -368,7 +368,7 @@ private:
     void InvalidateScale(Image& image);
     bool ScaleUp(Image& image);
     bool ScaleDown(Image& image);
-    u64 GetScaledImageSizeBytes(ImageBase& image);
+    u64 GetScaledImageSizeBytes(const ImageBase& image);
 
     Runtime& runtime;
 
@@ -417,6 +417,9 @@ private:
 
     std::unordered_map<GPUVAddr, ImageAllocId> image_allocs_table;
 
+    std::vector<u8> swizzle_data_buffer;
+    std::vector<u8> unswizzle_data_buffer;
+
     u64 modification_tick = 0;
     u64 frame_tick = 0;
 };

src/video_core/texture_cache/util.cpp

@@ -765,8 +765,9 @@ bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config
 }
 
 std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                                            const ImageInfo& info, std::span<u8> output) {
-    const size_t guest_size_bytes = CalculateGuestSizeInBytes(info);
+                                            const ImageInfo& info, std::span<const u8> input,
+                                            std::span<u8> output) {
+    const size_t guest_size_bytes = input.size_bytes();
     const u32 bpp_log2 = BytesPerBlockLog2(info.format);
     const Extent3D size = info.size;
 
@@ -789,10 +790,6 @@ std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GP
             .image_extent = size,
         }};
     }
-    const auto input_data = std::make_unique<u8[]>(guest_size_bytes);
-    gpu_memory.ReadBlockUnsafe(gpu_addr, input_data.get(), guest_size_bytes);
-    const std::span<const u8> input(input_data.get(), guest_size_bytes);
-
     const LevelInfo level_info = MakeLevelInfo(info);
     const s32 num_layers = info.resources.layers;
     const s32 num_levels = info.resources.levels;

src/video_core/texture_cache/util.h

@@ -59,6 +59,7 @@ struct OverlapResult {
 
 [[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
                                                           GPUVAddr gpu_addr, const ImageInfo& info,
+                                                          std::span<const u8> input,
                                                           std::span<u8> output);
 
 [[nodiscard]] BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,