Merge pull request #9490 from ameerj/texture-cache-prealloc

texture_cache: Use pre-allocated heap buffer for texture swizzles

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.resize_destructive(SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY);
+    unswizzle_data_buffer.resize_destructive(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{}));
@@ -90,7 +96,8 @@ void TextureCache<P>::RunGarbageCollector() {
             const auto copies = FullDownloadCopies(image.info);
             image.DownloadMemory(map, copies);
             runtime.Finish();
-            SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span);
+            SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span,
+                         swizzle_data_buffer);
         }
         if (True(image.flags & ImageFlagBits::Tracked)) {
             UntrackImage(image, image_id);
@@ -461,7 +468,8 @@ void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
         const auto copies = FullDownloadCopies(image.info);
         image.DownloadMemory(map, copies);
         runtime.Finish();
-        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span);
+        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span,
+                     swizzle_data_buffer);
     }
 }
 
@@ -672,7 +680,8 @@ void TextureCache<P>::PopAsyncFlushes() {
     for (const ImageId image_id : download_ids) {
         const ImageBase& image = slot_images[image_id];
         const auto copies = FullDownloadCopies(image.info);
-        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span);
+        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
+                     swizzle_data_buffer);
         download_map.offset += image.unswizzled_size_bytes;
         download_span = download_span.subspan(image.unswizzled_size_bytes);
     }
@@ -734,13 +743,21 @@ 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_destructive(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_destructive(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 +927,7 @@ 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

@@ -17,6 +17,7 @@
 #include "common/literals.h"
 #include "common/lru_cache.h"
 #include "common/polyfill_ranges.h"
+#include "common/scratch_buffer.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/control/channel_state_cache.h"
 #include "video_core/delayed_destruction_ring.h"
@@ -368,7 +369,7 @@ 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 +418,9 @@ private:
 
     std::unordered_map<GPUVAddr, ImageAllocId> image_allocs_table;
 
+    Common::ScratchBuffer<u8> swizzle_data_buffer;
+    Common::ScratchBuffer<u8> unswizzle_data_buffer;
+
     u64 modification_tick = 0;
     u64 frame_tick = 0;
 };

src/video_core/texture_cache/util.cpp

@@ -505,7 +505,7 @@ void SwizzlePitchLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
 
 void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
                              const ImageInfo& info, const BufferImageCopy& copy,
-                             std::span<const u8> input) {
+                             std::span<const u8> input, Common::ScratchBuffer<u8>& tmp_buffer) {
     const Extent3D size = info.size;
     const LevelInfo level_info = MakeLevelInfo(info);
     const Extent2D tile_size = DefaultBlockSize(info.format);
@@ -534,8 +534,8 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
                        tile_size.height, info.tile_width_spacing);
     const size_t subresource_size = sizes[level];
 
-    const auto dst_data = std::make_unique<u8[]>(subresource_size);
-    const std::span<u8> dst(dst_data.get(), subresource_size);
+    tmp_buffer.resize_destructive(subresource_size);
+    const std::span<u8> dst(tmp_buffer);
 
     for (s32 layer = 0; layer < info.resources.layers; ++layer) {
         const std::span<const u8> src = input.subspan(host_offset);
@@ -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;
@@ -980,13 +977,14 @@ std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info) {
 }
 
 void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
-                  std::span<const BufferImageCopy> copies, std::span<const u8> memory) {
+                  std::span<const BufferImageCopy> copies, std::span<const u8> memory,
+                  Common::ScratchBuffer<u8>& tmp_buffer) {
     const bool is_pitch_linear = info.type == ImageType::Linear;
     for (const BufferImageCopy& copy : copies) {
         if (is_pitch_linear) {
             SwizzlePitchLinearImage(gpu_memory, gpu_addr, info, copy, memory);
         } else {
-            SwizzleBlockLinearImage(gpu_memory, gpu_addr, info, copy, memory);
+            SwizzleBlockLinearImage(gpu_memory, gpu_addr, info, copy, memory, tmp_buffer);
         }
     }
 }

src/video_core/texture_cache/util.h

@@ -7,6 +7,7 @@
 #include <span>
 
 #include "common/common_types.h"
+#include "common/scratch_buffer.h"
 
 #include "video_core/surface.h"
 #include "video_core/texture_cache/image_base.h"
@@ -59,6 +60,7 @@ 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,
@@ -76,7 +78,8 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
 [[nodiscard]] std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info);
 
 void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
-                  std::span<const BufferImageCopy> copies, std::span<const u8> memory);
+                  std::span<const BufferImageCopy> copies, std::span<const u8> memory,
+                  Common::ScratchBuffer<u8>& tmp_buffer);
 
 [[nodiscard]] bool IsBlockLinearSizeCompatible(const ImageInfo& new_info,
                                                const ImageInfo& overlap_info, u32 new_level,