Merge pull request #10457 from Kelebek1/optimise

Remove memory allocations in some hot paths

5 files changed, 55 insertions, 47 deletions

diff --git a/src/video_core/texture_cache/image_base.h b/src/video_core/texture_cache/image_base.h
index 1b8a17ee8..55d49d017 100644
--- a/src/video_core/texture_cache/image_base.h
+++ b/src/video_core/texture_cache/image_base.h
@@ -6,6 +6,7 @@
 #include <array>
 #include <optional>
 #include <vector>
+#include <boost/container/small_vector.hpp>
 
 #include "common/common_funcs.h"
 #include "common/common_types.h"
@@ -108,8 +109,8 @@ struct ImageBase {
     std::vector<ImageViewInfo> image_view_infos;
     std::vector<ImageViewId> image_view_ids;
 
-    std::vector<u32> slice_offsets;
-    std::vector<SubresourceBase> slice_subresources;
+    boost::container::small_vector<u32, 16> slice_offsets;
+    boost::container::small_vector<SubresourceBase, 16> slice_subresources;
 
     std::vector<AliasedImage> aliased_images;
     std::vector<ImageId> overlapping_images;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index d58bb69ff..d3f03a995 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -526,7 +526,7 @@ void TextureCache<P>::WriteMemory(VAddr cpu_addr, size_t size) {
 
 template <class P>
 void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
-    std::vector<ImageId> images;
+    boost::container::small_vector<ImageId, 16> images;
     ForEachImageInRegion(cpu_addr, size, [&images](ImageId image_id, ImageBase& image) {
         if (!image.IsSafeDownload()) {
             return;
@@ -579,7 +579,7 @@ std::optional<VideoCore::RasterizerDownloadArea> TextureCache<P>::GetFlushArea(V
 
 template <class P>
 void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
-    std::vector<ImageId> deleted_images;
+    boost::container::small_vector<ImageId, 16> deleted_images;
     ForEachImageInRegion(cpu_addr, size, [&](ImageId id, Image&) { deleted_images.push_back(id); });
     for (const ImageId id : deleted_images) {
         Image& image = slot_images[id];
@@ -593,7 +593,7 @@ void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
 
 template <class P>
 void TextureCache<P>::UnmapGPUMemory(size_t as_id, GPUVAddr gpu_addr, size_t size) {
-    std::vector<ImageId> deleted_images;
+    boost::container::small_vector<ImageId, 16> deleted_images;
     ForEachImageInRegionGPU(as_id, gpu_addr, size,
                             [&](ImageId id, Image&) { deleted_images.push_back(id); });
     for (const ImageId id : deleted_images) {
@@ -1101,7 +1101,7 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
     const bool native_bgr = runtime.HasNativeBgr();
     const bool flexible_formats = True(options & RelaxedOptions::Format);
     ImageId image_id{};
-    boost::container::small_vector<ImageId, 1> image_ids;
+    boost::container::small_vector<ImageId, 8> image_ids;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
         if (True(existing_image.flags & ImageFlagBits::Remapped)) {
             return false;
@@ -1622,7 +1622,7 @@ ImageId TextureCache<P>::FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr)
         }
     }
     ImageId image_id{};
-    boost::container::small_vector<ImageId, 1> image_ids;
+    boost::container::small_vector<ImageId, 8> image_ids;
     const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
         if (True(existing_image.flags & ImageFlagBits::Remapped)) {
             return false;
@@ -1942,7 +1942,7 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
         image.map_view_id = map_id;
         return;
     }
-    std::vector<ImageViewId> sparse_maps{};
+    boost::container::small_vector<ImageViewId, 16> sparse_maps;
     ForEachSparseSegment(
         image, [this, image_id, &sparse_maps](GPUVAddr gpu_addr, VAddr cpu_addr, size_t size) {
             auto map_id = slot_map_views.insert(gpu_addr, cpu_addr, size, image_id);
@@ -2217,7 +2217,7 @@ void TextureCache<P>::MarkModification(ImageBase& image) noexcept {
 
 template <class P>
 void TextureCache<P>::SynchronizeAliases(ImageId image_id) {
-    boost::container::small_vector<const AliasedImage*, 1> aliased_images;
+    boost::container::small_vector<const AliasedImage*, 8> aliased_images;
     Image& image = slot_images[image_id];
     bool any_rescaled = True(image.flags & ImageFlagBits::Rescaled);
     bool any_modified = True(image.flags & ImageFlagBits::GpuModified);
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 44232b961..e9ec91265 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -56,7 +56,7 @@ struct ImageViewInOut {
 struct AsyncDecodeContext {
     ImageId image_id;
     Common::ScratchBuffer<u8> decoded_data;
-    std::vector<BufferImageCopy> copies;
+    boost::container::small_vector<BufferImageCopy, 16> copies;
     std::mutex mutex;
     std::atomic_bool complete;
 };
@@ -429,7 +429,7 @@ private:
 
     std::unordered_map<u64, std::vector<ImageMapId>, Common::IdentityHash<u64>> page_table;
     std::unordered_map<u64, std::vector<ImageId>, Common::IdentityHash<u64>> sparse_page_table;
-    std::unordered_map<ImageId, std::vector<ImageViewId>> sparse_views;
+    std::unordered_map<ImageId, boost::container::small_vector<ImageViewId, 16>> sparse_views;
 
     VAddr virtual_invalid_space{};
 
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index 95a5b47d8..f781cb7a0 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -329,13 +329,13 @@ template <u32 GOB_EXTENT>
 
 [[nodiscard]] std::optional<SubresourceExtent> ResolveOverlapRightAddress3D(
     const ImageInfo& new_info, GPUVAddr gpu_addr, const ImageBase& overlap, bool strict_size) {
-    const std::vector<u32> slice_offsets = CalculateSliceOffsets(new_info);
+    const auto slice_offsets = CalculateSliceOffsets(new_info);
     const u32 diff = static_cast<u32>(overlap.gpu_addr - gpu_addr);
     const auto it = std::ranges::find(slice_offsets, diff);
     if (it == slice_offsets.end()) {
         return std::nullopt;
     }
-    const std::vector subresources = CalculateSliceSubresources(new_info);
+    const auto subresources = CalculateSliceSubresources(new_info);
     const SubresourceBase base = subresources[std::distance(slice_offsets.begin(), it)];
     const ImageInfo& info = overlap.info;
     if (!IsBlockLinearSizeCompatible(new_info, info, base.level, 0, strict_size)) {
@@ -655,9 +655,9 @@ LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept {
     return sizes;
 }
 
-std::vector<u32> CalculateSliceOffsets(const ImageInfo& info) {
+boost::container::small_vector<u32, 16> CalculateSliceOffsets(const ImageInfo& info) {
     ASSERT(info.type == ImageType::e3D);
-    std::vector<u32> offsets;
+    boost::container::small_vector<u32, 16> offsets;
     offsets.reserve(NumSlices(info));
 
     const LevelInfo level_info = MakeLevelInfo(info);
@@ -679,9 +679,10 @@ std::vector<u32> CalculateSliceOffsets(const ImageInfo& info) {
     return offsets;
 }
 
-std::vector<SubresourceBase> CalculateSliceSubresources(const ImageInfo& info) {
+boost::container::small_vector<SubresourceBase, 16> CalculateSliceSubresources(
+    const ImageInfo& info) {
     ASSERT(info.type == ImageType::e3D);
-    std::vector<SubresourceBase> subresources;
+    boost::container::small_vector<SubresourceBase, 16> subresources;
     subresources.reserve(NumSlices(info));
     for (s32 level = 0; level < info.resources.levels; ++level) {
         const s32 depth = AdjustMipSize(info.size.depth, level);
@@ -723,8 +724,10 @@ ImageViewType RenderTargetImageViewType(const ImageInfo& info) noexcept {
     }
 }
 
-std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageInfo& src,
-                                             SubresourceBase base, u32 up_scale, u32 down_shift) {
+boost::container::small_vector<ImageCopy, 16> MakeShrinkImageCopies(const ImageInfo& dst,
+                                                                    const ImageInfo& src,
+                                                                    SubresourceBase base,
+                                                                    u32 up_scale, u32 down_shift) {
     ASSERT(dst.resources.levels >= src.resources.levels);
 
     const bool is_dst_3d = dst.type == ImageType::e3D;
@@ -733,7 +736,7 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
         ASSERT(src.resources.levels == 1);
     }
     const bool both_2d{src.type == ImageType::e2D && dst.type == ImageType::e2D};
-    std::vector<ImageCopy> copies;
+    boost::container::small_vector<ImageCopy, 16> copies;
     copies.reserve(src.resources.levels);
     for (s32 level = 0; level < src.resources.levels; ++level) {
         ImageCopy& copy = copies.emplace_back();
@@ -770,9 +773,10 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
     return copies;
 }
 
-std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src, u32 up_scale,
-                                                  u32 down_shift) {
-    std::vector<ImageCopy> copies;
+boost::container::small_vector<ImageCopy, 16> MakeReinterpretImageCopies(const ImageInfo& src,
+                                                                         u32 up_scale,
+                                                                         u32 down_shift) {
+    boost::container::small_vector<ImageCopy, 16> copies;
     copies.reserve(src.resources.levels);
     const bool is_3d = src.type == ImageType::e3D;
     for (s32 level = 0; level < src.resources.levels; ++level) {
@@ -824,9 +828,11 @@ bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config
     return gpu_memory.GpuToCpuAddress(address, guest_size_bytes).has_value();
 }
 
-std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                                            const ImageInfo& info, std::span<const u8> input,
-                                            std::span<u8> output) {
+boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
+                                                                   GPUVAddr gpu_addr,
+                                                                   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;
@@ -861,7 +867,7 @@ std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory, GP
                                             info.tile_width_spacing);
     size_t guest_offset = 0;
     u32 host_offset = 0;
-    std::vector<BufferImageCopy> copies(num_levels);
+    boost::container::small_vector<BufferImageCopy, 16> copies(num_levels);
 
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
@@ -978,7 +984,7 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
     }
 }
 
-std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info) {
+boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(const ImageInfo& info) {
     const Extent3D size = info.size;
     const u32 bytes_per_block = BytesPerBlock(info.format);
     if (info.type == ImageType::Linear) {
@@ -1006,7 +1012,7 @@ std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info) {
 
     u32 host_offset = 0;
 
-    std::vector<BufferImageCopy> copies(num_levels);
+    boost::container::small_vector<BufferImageCopy, 16> copies(num_levels);
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
         const u32 num_blocks_per_layer = NumBlocks(level_size, tile_size);
@@ -1042,10 +1048,10 @@ Extent3D MipBlockSize(const ImageInfo& info, u32 level) {
     return AdjustMipBlockSize(num_tiles, level_info.block, level);
 }
 
-std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info) {
+boost::container::small_vector<SwizzleParameters, 16> FullUploadSwizzles(const ImageInfo& info) {
     const Extent2D tile_size = DefaultBlockSize(info.format);
     if (info.type == ImageType::Linear) {
-        return std::vector{SwizzleParameters{
+        return {SwizzleParameters{
             .num_tiles = AdjustTileSize(info.size, tile_size),
             .block = {},
             .buffer_offset = 0,
@@ -1057,7 +1063,7 @@ std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info) {
     const s32 num_levels = info.resources.levels;
 
     u32 guest_offset = 0;
-    std::vector<SwizzleParameters> params(num_levels);
+    boost::container::small_vector<SwizzleParameters, 16> params(num_levels);
     for (s32 level = 0; level < num_levels; ++level) {
         const Extent3D level_size = AdjustMipSize(size, level);
         const Extent3D num_tiles = AdjustTileSize(level_size, tile_size);
diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h
index 84aa6880d..ab45a43c4 100644
--- a/src/video_core/texture_cache/util.h
+++ b/src/video_core/texture_cache/util.h
@@ -5,6 +5,7 @@
 
 #include <optional>
 #include <span>
+#include <boost/container/small_vector.hpp>
 
 #include "common/common_types.h"
 #include "common/scratch_buffer.h"
@@ -40,9 +41,10 @@ struct OverlapResult {
 
 [[nodiscard]] LevelArray CalculateMipLevelSizes(const ImageInfo& info) noexcept;
 
-[[nodiscard]] std::vector<u32> CalculateSliceOffsets(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<u32, 16> CalculateSliceOffsets(const ImageInfo& info);
 
-[[nodiscard]] std::vector<SubresourceBase> CalculateSliceSubresources(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<SubresourceBase, 16> CalculateSliceSubresources(
+    const ImageInfo& info);
 
 [[nodiscard]] u32 CalculateLevelStrideAlignment(const ImageInfo& info, u32 level);
 
@@ -51,21 +53,18 @@ struct OverlapResult {
 
 [[nodiscard]] ImageViewType RenderTargetImageViewType(const ImageInfo& info) noexcept;
 
-[[nodiscard]] std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst,
-                                                           const ImageInfo& src,
-                                                           SubresourceBase base, u32 up_scale = 1,
-                                                           u32 down_shift = 0);
+[[nodiscard]] boost::container::small_vector<ImageCopy, 16> MakeShrinkImageCopies(
+    const ImageInfo& dst, const ImageInfo& src, SubresourceBase base, u32 up_scale = 1,
+    u32 down_shift = 0);
 
-[[nodiscard]] std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src,
-                                                                u32 up_scale = 1,
-                                                                u32 down_shift = 0);
+[[nodiscard]] boost::container::small_vector<ImageCopy, 16> MakeReinterpretImageCopies(
+    const ImageInfo& src, u32 up_scale = 1, u32 down_shift = 0);
 
 [[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
 
-[[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]] boost::container::small_vector<BufferImageCopy, 16> 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,
                                           const ImageBase& image, std::span<u8> output);
@@ -73,13 +72,15 @@ struct OverlapResult {
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies);
 
-[[nodiscard]] std::vector<BufferImageCopy> FullDownloadCopies(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(
+    const ImageInfo& info);
 
 [[nodiscard]] Extent3D MipSize(Extent3D size, u32 level);
 
 [[nodiscard]] Extent3D MipBlockSize(const ImageInfo& info, u32 level);
 
-[[nodiscard]] std::vector<SwizzleParameters> FullUploadSwizzles(const ImageInfo& info);
+[[nodiscard]] boost::container::small_vector<SwizzleParameters, 16> 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,