From 72c1ee1bf96be45e2794f26ad83dbd13e892c22d Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Tue, 23 May 2023 01:34:46 -0400 Subject: [PATCH] texture_cache: process aliases and overlaps in the correct order --- src/video_core/texture_cache/image_base.cpp | 7 +- src/video_core/texture_cache/image_base.h | 2 +- src/video_core/texture_cache/texture_cache.h | 141 +++++++++++------- .../texture_cache/texture_cache_base.h | 16 ++ 4 files changed, 105 insertions(+), 61 deletions(-) diff --git a/src/video_core/texture_cache/image_base.cpp b/src/video_core/texture_cache/image_base.cpp index 91512022f..d79594ce5 100644 --- a/src/video_core/texture_cache/image_base.cpp +++ b/src/video_core/texture_cache/image_base.cpp @@ -155,7 +155,7 @@ void ImageBase::CheckAliasState() { flags &= ~ImageFlagBits::Alias; } -void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id) { +bool AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id) { static constexpr auto OPTIONS = RelaxedOptions::Size | RelaxedOptions::Format; ASSERT(lhs.info.type == rhs.info.type); std::optional base; @@ -169,7 +169,7 @@ void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_i } if (!base) { LOG_ERROR(HW_GPU, "Image alias should have been flipped"); - return; + return false; } const PixelFormat lhs_format = lhs.info.format; const PixelFormat rhs_format = rhs.info.format; @@ -248,12 +248,13 @@ void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_i } ASSERT(lhs_alias.copies.empty() == rhs_alias.copies.empty()); if (lhs_alias.copies.empty()) { - return; + return false; } lhs.aliased_images.push_back(std::move(lhs_alias)); rhs.aliased_images.push_back(std::move(rhs_alias)); lhs.flags &= ~ImageFlagBits::IsRescalable; rhs.flags &= ~ImageFlagBits::IsRescalable; + return true; } } // namespace VideoCommon diff --git a/src/video_core/texture_cache/image_base.h b/src/video_core/texture_cache/image_base.h index 329396bb6..1b8a17ee8 100644 --- a/src/video_core/texture_cache/image_base.h +++ b/src/video_core/texture_cache/image_base.h @@ -142,6 +142,6 @@ struct ImageAllocBase { std::vector images; }; -void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id); +bool AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id); } // namespace VideoCommon diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index b24086fce..f5c12d992 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -1274,17 +1274,18 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA const size_t size_bytes = CalculateGuestSizeInBytes(new_info); const bool broken_views = runtime.HasBrokenTextureViewFormats(); const bool native_bgr = runtime.HasNativeBgr(); - boost::container::small_vector overlap_ids; - std::unordered_set overlaps_found; - boost::container::small_vector left_aliased_ids; - boost::container::small_vector right_aliased_ids; - std::unordered_set ignore_textures; - boost::container::small_vector bad_overlap_ids; - boost::container::small_vector all_siblings; + join_overlap_ids.clear(); + join_overlaps_found.clear(); + join_left_aliased_ids.clear(); + join_right_aliased_ids.clear(); + join_ignore_textures.clear(); + join_bad_overlap_ids.clear(); + join_copies_to_do.clear(); + join_alias_indices.clear(); const bool this_is_linear = info.type == ImageType::Linear; const auto region_check = [&](ImageId overlap_id, ImageBase& overlap) { if (True(overlap.flags & ImageFlagBits::Remapped)) { - ignore_textures.insert(overlap_id); + join_ignore_textures.insert(overlap_id); return; } const bool overlap_is_linear = overlap.info.type == ImageType::Linear; @@ -1294,11 +1295,11 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA if (this_is_linear && overlap_is_linear) { if (info.pitch == overlap.info.pitch && gpu_addr == overlap.gpu_addr) { // Alias linear images with the same pitch - left_aliased_ids.push_back(overlap_id); + join_left_aliased_ids.push_back(overlap_id); } return; } - overlaps_found.insert(overlap_id); + join_overlaps_found.insert(overlap_id); static constexpr bool strict_size = true; const std::optional solution = ResolveOverlap( new_info, gpu_addr, cpu_addr, overlap, strict_size, broken_views, native_bgr); @@ -1306,33 +1307,33 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA gpu_addr = solution->gpu_addr; cpu_addr = solution->cpu_addr; new_info.resources = solution->resources; - overlap_ids.push_back(overlap_id); - all_siblings.push_back(overlap_id); + join_overlap_ids.push_back(overlap_id); + join_copies_to_do.emplace_back(JoinCopy{false, overlap_id}); return; } static constexpr auto options = RelaxedOptions::Size | RelaxedOptions::Format; const ImageBase new_image_base(new_info, gpu_addr, cpu_addr); if (IsSubresource(new_info, overlap, gpu_addr, options, broken_views, native_bgr)) { - left_aliased_ids.push_back(overlap_id); + join_left_aliased_ids.push_back(overlap_id); overlap.flags |= ImageFlagBits::Alias; - all_siblings.push_back(overlap_id); + join_copies_to_do.emplace_back(JoinCopy{true, overlap_id}); } else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options, broken_views, native_bgr)) { - right_aliased_ids.push_back(overlap_id); + join_right_aliased_ids.push_back(overlap_id); overlap.flags |= ImageFlagBits::Alias; - all_siblings.push_back(overlap_id); + join_copies_to_do.emplace_back(JoinCopy{true, overlap_id}); } else { - bad_overlap_ids.push_back(overlap_id); + join_bad_overlap_ids.push_back(overlap_id); } }; ForEachImageInRegion(cpu_addr, size_bytes, region_check); const auto region_check_gpu = [&](ImageId overlap_id, ImageBase& overlap) { - if (!overlaps_found.contains(overlap_id)) { + if (!join_overlaps_found.contains(overlap_id)) { if (True(overlap.flags & ImageFlagBits::Remapped)) { - ignore_textures.insert(overlap_id); + join_ignore_textures.insert(overlap_id); } if (overlap.gpu_addr == gpu_addr && overlap.guest_size_bytes == size_bytes) { - ignore_textures.insert(overlap_id); + join_ignore_textures.insert(overlap_id); } } }; @@ -1340,11 +1341,11 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA bool can_rescale = info.rescaleable; bool any_rescaled = false; - for (const ImageId sibling_id : all_siblings) { + for (const auto& copy : join_copies_to_do) { if (!can_rescale) { break; } - Image& sibling = slot_images[sibling_id]; + Image& sibling = slot_images[copy.id]; can_rescale &= ImageCanRescale(sibling); any_rescaled |= True(sibling.flags & ImageFlagBits::Rescaled); } @@ -1352,13 +1353,13 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA can_rescale &= any_rescaled; if (can_rescale) { - for (const ImageId sibling_id : all_siblings) { - Image& sibling = slot_images[sibling_id]; + for (const auto& copy : join_copies_to_do) { + Image& sibling = slot_images[copy.id]; ScaleUp(sibling); } } else { - for (const ImageId sibling_id : all_siblings) { - Image& sibling = slot_images[sibling_id]; + for (const auto& copy : join_copies_to_do) { + Image& sibling = slot_images[copy.id]; ScaleDown(sibling); } } @@ -1370,7 +1371,7 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA new_image.flags |= ImageFlagBits::Sparse; } - for (const ImageId overlap_id : ignore_textures) { + for (const ImageId overlap_id : join_ignore_textures) { Image& overlap = slot_images[overlap_id]; if (True(overlap.flags & ImageFlagBits::GpuModified)) { UNIMPLEMENTED(); @@ -1391,14 +1392,60 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA ScaleDown(new_image); } - std::ranges::sort(overlap_ids, [this](const ImageId lhs, const ImageId rhs) { - const ImageBase& lhs_image = slot_images[lhs]; - const ImageBase& rhs_image = slot_images[rhs]; + std::ranges::sort(join_copies_to_do, [this](const JoinCopy& lhs, const JoinCopy& rhs) { + const ImageBase& lhs_image = slot_images[lhs.id]; + const ImageBase& rhs_image = slot_images[rhs.id]; return lhs_image.modification_tick < rhs_image.modification_tick; }); - for (const ImageId overlap_id : overlap_ids) { - Image& overlap = slot_images[overlap_id]; + ImageBase& new_image_base = new_image; + for (const ImageId aliased_id : join_right_aliased_ids) { + ImageBase& aliased = slot_images[aliased_id]; + size_t alias_index = new_image_base.aliased_images.size(); + if (!AddImageAlias(new_image_base, aliased, new_image_id, aliased_id)) { + continue; + } + join_alias_indices.emplace(aliased_id, alias_index); + new_image.flags |= ImageFlagBits::Alias; + } + for (const ImageId aliased_id : join_left_aliased_ids) { + ImageBase& aliased = slot_images[aliased_id]; + size_t alias_index = new_image_base.aliased_images.size(); + if (!AddImageAlias(aliased, new_image_base, aliased_id, new_image_id)) { + continue; + } + join_alias_indices.emplace(aliased_id, alias_index); + new_image.flags |= ImageFlagBits::Alias; + } + for (const ImageId aliased_id : join_bad_overlap_ids) { + ImageBase& aliased = slot_images[aliased_id]; + aliased.overlapping_images.push_back(new_image_id); + new_image.overlapping_images.push_back(aliased_id); + if (aliased.info.resources.levels == 1 && aliased.info.block.depth == 0 && + aliased.overlapping_images.size() > 1) { + aliased.flags |= ImageFlagBits::BadOverlap; + } + if (new_image.info.resources.levels == 1 && new_image.info.block.depth == 0 && + new_image.overlapping_images.size() > 1) { + new_image.flags |= ImageFlagBits::BadOverlap; + } + } + + for (const auto& copy_object : join_copies_to_do) { + Image& overlap = slot_images[copy_object.id]; + if (copy_object.is_alias) { + if (!overlap.IsSafeDownload()) { + continue; + } + const auto alias_pointer = join_alias_indices.find(copy_object.id); + if (alias_pointer == join_alias_indices.end()) { + continue; + } + const AliasedImage& aliased = new_image.aliased_images[alias_pointer->second]; + CopyImage(new_image_id, aliased.id, aliased.copies); + new_image.modification_tick = overlap.modification_tick; + continue; + } if (True(overlap.flags & ImageFlagBits::GpuModified)) { new_image.flags |= ImageFlagBits::GpuModified; const auto& resolution = Settings::values.resolution_info; @@ -1411,35 +1458,15 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA } else { runtime.CopyImage(new_image, overlap, std::move(copies)); } + new_image.modification_tick = overlap.modification_tick; } if (True(overlap.flags & ImageFlagBits::Tracked)) { - UntrackImage(overlap, overlap_id); - } - UnregisterImage(overlap_id); - DeleteImage(overlap_id); - } - ImageBase& new_image_base = new_image; - for (const ImageId aliased_id : right_aliased_ids) { - ImageBase& aliased = slot_images[aliased_id]; - AddImageAlias(new_image_base, aliased, new_image_id, aliased_id); - new_image.flags |= ImageFlagBits::Alias; - } - for (const ImageId aliased_id : left_aliased_ids) { - ImageBase& aliased = slot_images[aliased_id]; - AddImageAlias(aliased, new_image_base, aliased_id, new_image_id); - new_image.flags |= ImageFlagBits::Alias; - } - for (const ImageId aliased_id : bad_overlap_ids) { - ImageBase& aliased = slot_images[aliased_id]; - aliased.overlapping_images.push_back(new_image_id); - new_image.overlapping_images.push_back(aliased_id); - if (aliased.info.resources.levels == 1 && aliased.overlapping_images.size() > 1) { - aliased.flags |= ImageFlagBits::BadOverlap; - } - if (new_image.info.resources.levels == 1 && new_image.overlapping_images.size() > 1) { - new_image.flags |= ImageFlagBits::BadOverlap; + UntrackImage(overlap, copy_object.id); } + UnregisterImage(copy_object.id); + DeleteImage(copy_object.id); } + RegisterImage(new_image_id); return new_image_id; } diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h index 0720494e5..9e5289977 100644 --- a/src/video_core/texture_cache/texture_cache_base.h +++ b/src/video_core/texture_cache/texture_cache_base.h @@ -10,7 +10,9 @@ #include #include #include +#include #include +#include #include #include "common/common_types.h" @@ -474,6 +476,20 @@ private: Common::ThreadWorker texture_decode_worker{1, "TextureDecoder"}; std::vector> async_decodes; + + // Join caching + boost::container::small_vector join_overlap_ids; + std::unordered_set join_overlaps_found; + boost::container::small_vector join_left_aliased_ids; + boost::container::small_vector join_right_aliased_ids; + std::unordered_set join_ignore_textures; + boost::container::small_vector join_bad_overlap_ids; + struct JoinCopy { + bool is_alias; + ImageId id; + }; + boost::container::small_vector join_copies_to_do; + std::unordered_map join_alias_indices; }; } // namespace VideoCommon