diff options
author | Zach Hilman <DarkLordZach@users.noreply.github.com> | 2019-07-05 19:39:13 +0200 |
---|---|---|
committer | GitHub <noreply@github.com> | 2019-07-05 19:39:13 +0200 |
commit | 772c86a260eb446b0fe4232b0a50666511bef25c (patch) | |
tree | 013d92268c06454c93565c83eff2b79b56a00839 /src/video_core/texture_cache | |
parent | Merge pull request #2669 from FearlessTobi/move-cpujit-setting (diff) | |
parent | texture_cache: Address Feedback (diff) | |
download | yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar.gz yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar.bz2 yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar.lz yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar.xz yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.tar.zst yuzu-772c86a260eb446b0fe4232b0a50666511bef25c.zip |
Diffstat (limited to 'src/video_core/texture_cache')
-rw-r--r-- | src/video_core/texture_cache/copy_params.h | 36 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_base.cpp | 300 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_base.h | 317 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_params.cpp | 334 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_params.h | 286 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_view.cpp | 23 | ||||
-rw-r--r-- | src/video_core/texture_cache/surface_view.h | 67 | ||||
-rw-r--r-- | src/video_core/texture_cache/texture_cache.h | 814 |
8 files changed, 2177 insertions, 0 deletions
diff --git a/src/video_core/texture_cache/copy_params.h b/src/video_core/texture_cache/copy_params.h new file mode 100644 index 000000000..9c21a0649 --- /dev/null +++ b/src/video_core/texture_cache/copy_params.h @@ -0,0 +1,36 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include "common/common_types.h" + +namespace VideoCommon { + +struct CopyParams { + constexpr CopyParams(u32 source_x, u32 source_y, u32 source_z, u32 dest_x, u32 dest_y, + u32 dest_z, u32 source_level, u32 dest_level, u32 width, u32 height, + u32 depth) + : source_x{source_x}, source_y{source_y}, source_z{source_z}, dest_x{dest_x}, + dest_y{dest_y}, dest_z{dest_z}, source_level{source_level}, + dest_level{dest_level}, width{width}, height{height}, depth{depth} {} + + constexpr CopyParams(u32 width, u32 height, u32 depth, u32 level) + : source_x{}, source_y{}, source_z{}, dest_x{}, dest_y{}, dest_z{}, source_level{level}, + dest_level{level}, width{width}, height{height}, depth{depth} {} + + u32 source_x; + u32 source_y; + u32 source_z; + u32 dest_x; + u32 dest_y; + u32 dest_z; + u32 source_level; + u32 dest_level; + u32 width; + u32 height; + u32 depth; +}; + +} // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_base.cpp b/src/video_core/texture_cache/surface_base.cpp new file mode 100644 index 000000000..7a0fdb19b --- /dev/null +++ b/src/video_core/texture_cache/surface_base.cpp @@ -0,0 +1,300 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include "common/assert.h" +#include "common/common_types.h" +#include "common/microprofile.h" +#include "video_core/memory_manager.h" +#include "video_core/texture_cache/surface_base.h" +#include "video_core/texture_cache/surface_params.h" +#include "video_core/textures/convert.h" + +namespace VideoCommon { + +MICROPROFILE_DEFINE(GPU_Load_Texture, "GPU", "Texture Load", MP_RGB(128, 192, 128)); +MICROPROFILE_DEFINE(GPU_Flush_Texture, "GPU", "Texture Flush", MP_RGB(128, 192, 128)); + +using Tegra::Texture::ConvertFromGuestToHost; +using VideoCore::MortonSwizzleMode; +using VideoCore::Surface::SurfaceCompression; + +StagingCache::StagingCache() = default; + +StagingCache::~StagingCache() = default; + +SurfaceBaseImpl::SurfaceBaseImpl(GPUVAddr gpu_addr, const SurfaceParams& params) + : params{params}, mipmap_sizes(params.num_levels), + mipmap_offsets(params.num_levels), gpu_addr{gpu_addr}, host_memory_size{ + params.GetHostSizeInBytes()} { + std::size_t offset = 0; + for (u32 level = 0; level < params.num_levels; ++level) { + const std::size_t mipmap_size{params.GetGuestMipmapSize(level)}; + mipmap_sizes[level] = mipmap_size; + mipmap_offsets[level] = offset; + offset += mipmap_size; + } + layer_size = offset; + if (params.is_layered) { + if (params.is_tiled) { + layer_size = + SurfaceParams::AlignLayered(layer_size, params.block_height, params.block_depth); + } + guest_memory_size = layer_size * params.depth; + } else { + guest_memory_size = layer_size; + } +} + +MatchTopologyResult SurfaceBaseImpl::MatchesTopology(const SurfaceParams& rhs) const { + const u32 src_bpp{params.GetBytesPerPixel()}; + const u32 dst_bpp{rhs.GetBytesPerPixel()}; + const bool ib1 = params.IsBuffer(); + const bool ib2 = rhs.IsBuffer(); + if (std::tie(src_bpp, params.is_tiled, ib1) == std::tie(dst_bpp, rhs.is_tiled, ib2)) { + const bool cb1 = params.IsCompressed(); + const bool cb2 = rhs.IsCompressed(); + if (cb1 == cb2) { + return MatchTopologyResult::FullMatch; + } + return MatchTopologyResult::CompressUnmatch; + } + return MatchTopologyResult::None; +} + +MatchStructureResult SurfaceBaseImpl::MatchesStructure(const SurfaceParams& rhs) const { + // Buffer surface Check + if (params.IsBuffer()) { + const std::size_t wd1 = params.width * params.GetBytesPerPixel(); + const std::size_t wd2 = rhs.width * rhs.GetBytesPerPixel(); + if (wd1 == wd2) { + return MatchStructureResult::FullMatch; + } + return MatchStructureResult::None; + } + + // Linear Surface check + if (!params.is_tiled) { + if (std::tie(params.width, params.height, params.pitch) == + std::tie(rhs.width, rhs.height, rhs.pitch)) { + return MatchStructureResult::FullMatch; + } + return MatchStructureResult::None; + } + + // Tiled Surface check + if (std::tie(params.depth, params.block_width, params.block_height, params.block_depth, + params.tile_width_spacing, params.num_levels) == + std::tie(rhs.depth, rhs.block_width, rhs.block_height, rhs.block_depth, + rhs.tile_width_spacing, rhs.num_levels)) { + if (std::tie(params.width, params.height) == std::tie(rhs.width, rhs.height)) { + return MatchStructureResult::FullMatch; + } + const u32 ws = SurfaceParams::ConvertWidth(rhs.GetBlockAlignedWidth(), params.pixel_format, + rhs.pixel_format); + const u32 hs = + SurfaceParams::ConvertHeight(rhs.height, params.pixel_format, rhs.pixel_format); + const u32 w1 = params.GetBlockAlignedWidth(); + if (std::tie(w1, params.height) == std::tie(ws, hs)) { + return MatchStructureResult::SemiMatch; + } + } + return MatchStructureResult::None; +} + +std::optional<std::pair<u32, u32>> SurfaceBaseImpl::GetLayerMipmap( + const GPUVAddr candidate_gpu_addr) const { + if (gpu_addr == candidate_gpu_addr) { + return {{0, 0}}; + } + if (candidate_gpu_addr < gpu_addr) { + return {}; + } + const auto relative_address{static_cast<GPUVAddr>(candidate_gpu_addr - gpu_addr)}; + const auto layer{static_cast<u32>(relative_address / layer_size)}; + const GPUVAddr mipmap_address = relative_address - layer_size * layer; + const auto mipmap_it = + Common::BinaryFind(mipmap_offsets.begin(), mipmap_offsets.end(), mipmap_address); + if (mipmap_it == mipmap_offsets.end()) { + return {}; + } + const auto level{static_cast<u32>(std::distance(mipmap_offsets.begin(), mipmap_it))}; + return std::make_pair(layer, level); +} + +std::vector<CopyParams> SurfaceBaseImpl::BreakDownLayered(const SurfaceParams& in_params) const { + const u32 layers{params.depth}; + const u32 mipmaps{params.num_levels}; + std::vector<CopyParams> result; + result.reserve(static_cast<std::size_t>(layers) * static_cast<std::size_t>(mipmaps)); + + for (u32 layer = 0; layer < layers; layer++) { + for (u32 level = 0; level < mipmaps; level++) { + const u32 width = SurfaceParams::IntersectWidth(params, in_params, level, level); + const u32 height = SurfaceParams::IntersectHeight(params, in_params, level, level); + result.emplace_back(width, height, layer, level); + } + } + return result; +} + +std::vector<CopyParams> SurfaceBaseImpl::BreakDownNonLayered(const SurfaceParams& in_params) const { + const u32 mipmaps{params.num_levels}; + std::vector<CopyParams> result; + result.reserve(mipmaps); + + for (u32 level = 0; level < mipmaps; level++) { + const u32 width = SurfaceParams::IntersectWidth(params, in_params, level, level); + const u32 height = SurfaceParams::IntersectHeight(params, in_params, level, level); + const u32 depth{std::min(params.GetMipDepth(level), in_params.GetMipDepth(level))}; + result.emplace_back(width, height, depth, level); + } + return result; +} + +void SurfaceBaseImpl::SwizzleFunc(MortonSwizzleMode mode, u8* memory, const SurfaceParams& params, + u8* buffer, u32 level) { + const u32 width{params.GetMipWidth(level)}; + const u32 height{params.GetMipHeight(level)}; + const u32 block_height{params.GetMipBlockHeight(level)}; + const u32 block_depth{params.GetMipBlockDepth(level)}; + + std::size_t guest_offset{mipmap_offsets[level]}; + if (params.is_layered) { + std::size_t host_offset{0}; + const std::size_t guest_stride = layer_size; + const std::size_t host_stride = params.GetHostLayerSize(level); + for (u32 layer = 0; layer < params.depth; ++layer) { + MortonSwizzle(mode, params.pixel_format, width, block_height, height, block_depth, 1, + params.tile_width_spacing, buffer + host_offset, memory + guest_offset); + guest_offset += guest_stride; + host_offset += host_stride; + } + } else { + MortonSwizzle(mode, params.pixel_format, width, block_height, height, block_depth, + params.GetMipDepth(level), params.tile_width_spacing, buffer, + memory + guest_offset); + } +} + +void SurfaceBaseImpl::LoadBuffer(Tegra::MemoryManager& memory_manager, + StagingCache& staging_cache) { + MICROPROFILE_SCOPE(GPU_Load_Texture); + auto& staging_buffer = staging_cache.GetBuffer(0); + u8* host_ptr; + is_continuous = memory_manager.IsBlockContinuous(gpu_addr, guest_memory_size); + + // Handle continuouty + if (is_continuous) { + // Use physical memory directly + host_ptr = memory_manager.GetPointer(gpu_addr); + if (!host_ptr) { + return; + } + } else { + // Use an extra temporal buffer + auto& tmp_buffer = staging_cache.GetBuffer(1); + tmp_buffer.resize(guest_memory_size); + host_ptr = tmp_buffer.data(); + memory_manager.ReadBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); + } + + if (params.is_tiled) { + ASSERT_MSG(params.block_width == 0, "Block width is defined as {} on texture target {}", + params.block_width, static_cast<u32>(params.target)); + for (u32 level = 0; level < params.num_levels; ++level) { + const std::size_t host_offset{params.GetHostMipmapLevelOffset(level)}; + SwizzleFunc(MortonSwizzleMode::MortonToLinear, host_ptr, params, + staging_buffer.data() + host_offset, level); + } + } else { + ASSERT_MSG(params.num_levels == 1, "Linear mipmap loading is not implemented"); + const u32 bpp{params.GetBytesPerPixel()}; + const u32 block_width{params.GetDefaultBlockWidth()}; + const u32 block_height{params.GetDefaultBlockHeight()}; + const u32 width{(params.width + block_width - 1) / block_width}; + const u32 height{(params.height + block_height - 1) / block_height}; + const u32 copy_size{width * bpp}; + if (params.pitch == copy_size) { + std::memcpy(staging_buffer.data(), host_ptr, params.GetHostSizeInBytes()); + } else { + const u8* start{host_ptr}; + u8* write_to{staging_buffer.data()}; + for (u32 h = height; h > 0; --h) { + std::memcpy(write_to, start, copy_size); + start += params.pitch; + write_to += copy_size; + } + } + } + + auto compression_type = params.GetCompressionType(); + if (compression_type == SurfaceCompression::None || + compression_type == SurfaceCompression::Compressed) + return; + + for (u32 level_up = params.num_levels; level_up > 0; --level_up) { + const u32 level = level_up - 1; + const std::size_t in_host_offset{params.GetHostMipmapLevelOffset(level)}; + const std::size_t out_host_offset = compression_type == SurfaceCompression::Rearranged + ? in_host_offset + : params.GetConvertedMipmapOffset(level); + u8* in_buffer = staging_buffer.data() + in_host_offset; + u8* out_buffer = staging_buffer.data() + out_host_offset; + ConvertFromGuestToHost(in_buffer, out_buffer, params.pixel_format, + params.GetMipWidth(level), params.GetMipHeight(level), + params.GetMipDepth(level), true, true); + } +} + +void SurfaceBaseImpl::FlushBuffer(Tegra::MemoryManager& memory_manager, + StagingCache& staging_cache) { + MICROPROFILE_SCOPE(GPU_Flush_Texture); + auto& staging_buffer = staging_cache.GetBuffer(0); + u8* host_ptr; + + // Handle continuouty + if (is_continuous) { + // Use physical memory directly + host_ptr = memory_manager.GetPointer(gpu_addr); + if (!host_ptr) { + return; + } + } else { + // Use an extra temporal buffer + auto& tmp_buffer = staging_cache.GetBuffer(1); + tmp_buffer.resize(guest_memory_size); + host_ptr = tmp_buffer.data(); + } + + if (params.is_tiled) { + ASSERT_MSG(params.block_width == 0, "Block width is defined as {}", params.block_width); + for (u32 level = 0; level < params.num_levels; ++level) { + const std::size_t host_offset{params.GetHostMipmapLevelOffset(level)}; + SwizzleFunc(MortonSwizzleMode::LinearToMorton, host_ptr, params, + staging_buffer.data() + host_offset, level); + } + } else { + ASSERT(params.target == SurfaceTarget::Texture2D); + ASSERT(params.num_levels == 1); + + const u32 bpp{params.GetBytesPerPixel()}; + const u32 copy_size{params.width * bpp}; + if (params.pitch == copy_size) { + std::memcpy(host_ptr, staging_buffer.data(), guest_memory_size); + } else { + u8* start{host_ptr}; + const u8* read_to{staging_buffer.data()}; + for (u32 h = params.height; h > 0; --h) { + std::memcpy(start, read_to, copy_size); + start += params.pitch; + read_to += copy_size; + } + } + } + if (!is_continuous) { + memory_manager.WriteBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); + } +} + +} // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_base.h b/src/video_core/texture_cache/surface_base.h new file mode 100644 index 000000000..8ba386a8a --- /dev/null +++ b/src/video_core/texture_cache/surface_base.h @@ -0,0 +1,317 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include <algorithm> +#include <unordered_map> +#include <vector> + +#include "common/assert.h" +#include "common/binary_find.h" +#include "common/common_types.h" +#include "video_core/gpu.h" +#include "video_core/morton.h" +#include "video_core/texture_cache/copy_params.h" +#include "video_core/texture_cache/surface_params.h" +#include "video_core/texture_cache/surface_view.h" + +namespace Tegra { +class MemoryManager; +} + +namespace VideoCommon { + +using VideoCore::MortonSwizzleMode; +using VideoCore::Surface::SurfaceTarget; + +enum class MatchStructureResult : u32 { + FullMatch = 0, + SemiMatch = 1, + None = 2, +}; + +enum class MatchTopologyResult : u32 { + FullMatch = 0, + CompressUnmatch = 1, + None = 2, +}; + +class StagingCache { +public: + explicit StagingCache(); + ~StagingCache(); + + std::vector<u8>& GetBuffer(std::size_t index) { + return staging_buffer[index]; + } + + const std::vector<u8>& GetBuffer(std::size_t index) const { + return staging_buffer[index]; + } + + void SetSize(std::size_t size) { + staging_buffer.resize(size); + } + +private: + std::vector<std::vector<u8>> staging_buffer; +}; + +class SurfaceBaseImpl { +public: + void LoadBuffer(Tegra::MemoryManager& memory_manager, StagingCache& staging_cache); + + void FlushBuffer(Tegra::MemoryManager& memory_manager, StagingCache& staging_cache); + + GPUVAddr GetGpuAddr() const { + return gpu_addr; + } + + bool Overlaps(const CacheAddr start, const CacheAddr end) const { + return (cache_addr < end) && (cache_addr_end > start); + } + + bool IsInside(const GPUVAddr other_start, const GPUVAddr other_end) { + const GPUVAddr gpu_addr_end = gpu_addr + guest_memory_size; + return (gpu_addr <= other_start && other_end <= gpu_addr_end); + } + + // Use only when recycling a surface + void SetGpuAddr(const GPUVAddr new_addr) { + gpu_addr = new_addr; + } + + VAddr GetCpuAddr() const { + return cpu_addr; + } + + void SetCpuAddr(const VAddr new_addr) { + cpu_addr = new_addr; + } + + CacheAddr GetCacheAddr() const { + return cache_addr; + } + + CacheAddr GetCacheAddrEnd() const { + return cache_addr_end; + } + + void SetCacheAddr(const CacheAddr new_addr) { + cache_addr = new_addr; + cache_addr_end = new_addr + guest_memory_size; + } + + const SurfaceParams& GetSurfaceParams() const { + return params; + } + + std::size_t GetSizeInBytes() const { + return guest_memory_size; + } + + std::size_t GetHostSizeInBytes() const { + return host_memory_size; + } + + std::size_t GetMipmapSize(const u32 level) const { + return mipmap_sizes[level]; + } + + void MarkAsContinuous(const bool is_continuous) { + this->is_continuous = is_continuous; + } + + bool IsContinuous() const { + return is_continuous; + } + + bool IsLinear() const { + return !params.is_tiled; + } + + bool MatchFormat(VideoCore::Surface::PixelFormat pixel_format) const { + return params.pixel_format == pixel_format; + } + + VideoCore::Surface::PixelFormat GetFormat() const { + return params.pixel_format; + } + + bool MatchTarget(VideoCore::Surface::SurfaceTarget target) const { + return params.target == target; + } + + MatchTopologyResult MatchesTopology(const SurfaceParams& rhs) const; + + MatchStructureResult MatchesStructure(const SurfaceParams& rhs) const; + + bool MatchesSubTexture(const SurfaceParams& rhs, const GPUVAddr other_gpu_addr) const { + return std::tie(gpu_addr, params.target, params.num_levels) == + std::tie(other_gpu_addr, rhs.target, rhs.num_levels) && + params.target == SurfaceTarget::Texture2D && params.num_levels == 1; + } + + std::optional<std::pair<u32, u32>> GetLayerMipmap(const GPUVAddr candidate_gpu_addr) const; + + std::vector<CopyParams> BreakDown(const SurfaceParams& in_params) const { + return params.is_layered ? BreakDownLayered(in_params) : BreakDownNonLayered(in_params); + } + +protected: + explicit SurfaceBaseImpl(GPUVAddr gpu_addr, const SurfaceParams& params); + ~SurfaceBaseImpl() = default; + + virtual void DecorateSurfaceName() = 0; + + const SurfaceParams params; + std::size_t layer_size; + std::size_t guest_memory_size; + const std::size_t host_memory_size; + GPUVAddr gpu_addr{}; + CacheAddr cache_addr{}; + CacheAddr cache_addr_end{}; + VAddr cpu_addr{}; + bool is_continuous{}; + + std::vector<std::size_t> mipmap_sizes; + std::vector<std::size_t> mipmap_offsets; + +private: + void SwizzleFunc(MortonSwizzleMode mode, u8* memory, const SurfaceParams& params, u8* buffer, + u32 level); + + std::vector<CopyParams> BreakDownLayered(const SurfaceParams& in_params) const; + + std::vector<CopyParams> BreakDownNonLayered(const SurfaceParams& in_params) const; +}; + +template <typename TView> +class SurfaceBase : public SurfaceBaseImpl { +public: + virtual void UploadTexture(const std::vector<u8>& staging_buffer) = 0; + + virtual void DownloadTexture(std::vector<u8>& staging_buffer) = 0; + + void MarkAsModified(const bool is_modified_, const u64 tick) { + is_modified = is_modified_ || is_target; + modification_tick = tick; + } + + void MarkAsRenderTarget(const bool is_target) { + this->is_target = is_target; + } + + void MarkAsPicked(const bool is_picked) { + this->is_picked = is_picked; + } + + bool IsModified() const { + return is_modified; + } + + bool IsProtected() const { + // Only 3D Slices are to be protected + return is_target && params.block_depth > 0; + } + + bool IsRenderTarget() const { + return is_target; + } + + bool IsRegistered() const { + return is_registered; + } + + bool IsPicked() const { + return is_picked; + } + + void MarkAsRegistered(bool is_reg) { + is_registered = is_reg; + } + + u64 GetModificationTick() const { + return modification_tick; + } + + TView EmplaceOverview(const SurfaceParams& overview_params) { + const u32 num_layers{(params.is_layered && !overview_params.is_layered) ? 1 : params.depth}; + return GetView(ViewParams(overview_params.target, 0, num_layers, 0, params.num_levels)); + } + + std::optional<TView> EmplaceIrregularView(const SurfaceParams& view_params, + const GPUVAddr view_addr, + const std::size_t candidate_size, const u32 mipmap, + const u32 layer) { + const auto layer_mipmap{GetLayerMipmap(view_addr + candidate_size)}; + if (!layer_mipmap) { + return {}; + } + const u32 end_layer{layer_mipmap->first}; + const u32 end_mipmap{layer_mipmap->second}; + if (layer != end_layer) { + if (mipmap == 0 && end_mipmap == 0) { + return GetView(ViewParams(view_params.target, layer, end_layer - layer + 1, 0, 1)); + } + return {}; + } else { + return GetView( + ViewParams(view_params.target, layer, 1, mipmap, end_mipmap - mipmap + 1)); + } + } + + std::optional<TView> EmplaceView(const SurfaceParams& view_params, const GPUVAddr view_addr, + const std::size_t candidate_size) { + if (params.target == SurfaceTarget::Texture3D || + (params.num_levels == 1 && !params.is_layered) || + view_params.target == SurfaceTarget::Texture3D) { + return {}; + } + const auto layer_mipmap{GetLayerMipmap(view_addr)}; + if (!layer_mipmap) { + return {}; + } + const u32 layer{layer_mipmap->first}; + const u32 mipmap{layer_mipmap->second}; + if (GetMipmapSize(mipmap) != candidate_size) { + return EmplaceIrregularView(view_params, view_addr, candidate_size, mipmap, layer); + } + return GetView(ViewParams(view_params.target, layer, 1, mipmap, 1)); + } + + TView GetMainView() const { + return main_view; + } + +protected: + explicit SurfaceBase(const GPUVAddr gpu_addr, const SurfaceParams& params) + : SurfaceBaseImpl(gpu_addr, params) {} + + ~SurfaceBase() = default; + + virtual TView CreateView(const ViewParams& view_key) = 0; + + TView main_view; + std::unordered_map<ViewParams, TView> views; + +private: + TView GetView(const ViewParams& key) { + const auto [entry, is_cache_miss] = views.try_emplace(key); + auto& view{entry->second}; + if (is_cache_miss) { + view = CreateView(key); + } + return view; + } + + bool is_modified{}; + bool is_target{}; + bool is_registered{}; + bool is_picked{}; + u64 modification_tick{}; +}; + +} // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_params.cpp b/src/video_core/texture_cache/surface_params.cpp new file mode 100644 index 000000000..9c56e2b4f --- /dev/null +++ b/src/video_core/texture_cache/surface_params.cpp @@ -0,0 +1,334 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include <map> + +#include "common/alignment.h" +#include "common/bit_util.h" +#include "core/core.h" +#include "video_core/engines/shader_bytecode.h" +#include "video_core/surface.h" +#include "video_core/texture_cache/surface_params.h" + +namespace VideoCommon { + +using VideoCore::Surface::ComponentTypeFromDepthFormat; +using VideoCore::Surface::ComponentTypeFromRenderTarget; +using VideoCore::Surface::ComponentTypeFromTexture; +using VideoCore::Surface::PixelFormat; +using VideoCore::Surface::PixelFormatFromDepthFormat; +using VideoCore::Surface::PixelFormatFromRenderTargetFormat; +using VideoCore::Surface::PixelFormatFromTextureFormat; +using VideoCore::Surface::SurfaceTarget; +using VideoCore::Surface::SurfaceTargetFromTextureType; +using VideoCore::Surface::SurfaceType; + +SurfaceTarget TextureType2SurfaceTarget(Tegra::Shader::TextureType type, bool is_array) { + switch (type) { + case Tegra::Shader::TextureType::Texture1D: { + if (is_array) + return SurfaceTarget::Texture1DArray; + else + return SurfaceTarget::Texture1D; + } + case Tegra::Shader::TextureType::Texture2D: { + if (is_array) + return SurfaceTarget::Texture2DArray; + else + return SurfaceTarget::Texture2D; + } + case Tegra::Shader::TextureType::Texture3D: { + ASSERT(!is_array); + return SurfaceTarget::Texture3D; + } + case Tegra::Shader::TextureType::TextureCube: { + if (is_array) + return SurfaceTarget::TextureCubeArray; + else + return SurfaceTarget::TextureCubemap; + } + default: { + UNREACHABLE(); + return SurfaceTarget::Texture2D; + } + } +} + +namespace { +constexpr u32 GetMipmapSize(bool uncompressed, u32 mip_size, u32 tile) { + return uncompressed ? mip_size : std::max(1U, (mip_size + tile - 1) / tile); +} +} // Anonymous namespace + +SurfaceParams SurfaceParams::CreateForTexture(Core::System& system, + const Tegra::Texture::FullTextureInfo& config, + const VideoCommon::Shader::Sampler& entry) { + SurfaceParams params; + params.is_tiled = config.tic.IsTiled(); + params.srgb_conversion = config.tic.IsSrgbConversionEnabled(); + params.block_width = params.is_tiled ? config.tic.BlockWidth() : 0, + params.block_height = params.is_tiled ? config.tic.BlockHeight() : 0, + params.block_depth = params.is_tiled ? config.tic.BlockDepth() : 0, + params.tile_width_spacing = params.is_tiled ? (1 << config.tic.tile_width_spacing.Value()) : 1; + params.pixel_format = PixelFormatFromTextureFormat(config.tic.format, config.tic.r_type.Value(), + params.srgb_conversion); + params.type = GetFormatType(params.pixel_format); + if (entry.IsShadow() && params.type == SurfaceType::ColorTexture) { + switch (params.pixel_format) { + case PixelFormat::R16U: + case PixelFormat::R16F: { + params.pixel_format = PixelFormat::Z16; + break; + } + case PixelFormat::R32F: { + params.pixel_format = PixelFormat::Z32F; + break; + } + default: { + UNIMPLEMENTED_MSG("Unimplemented shadow convert format: {}", + static_cast<u32>(params.pixel_format)); + } + } + params.type = GetFormatType(params.pixel_format); + } + params.component_type = ComponentTypeFromTexture(config.tic.r_type.Value()); + params.type = GetFormatType(params.pixel_format); + // TODO: on 1DBuffer we should use the tic info. + if (!config.tic.IsBuffer()) { + params.target = TextureType2SurfaceTarget(entry.GetType(), entry.IsArray()); + params.width = config.tic.Width(); + params.height = config.tic.Height(); + params.depth = config.tic.Depth(); + params.pitch = params.is_tiled ? 0 : config.tic.Pitch(); + if (params.target == SurfaceTarget::TextureCubemap || + params.target == SurfaceTarget::TextureCubeArray) { + params.depth *= 6; + } + params.num_levels = config.tic.max_mip_level + 1; + params.emulated_levels = std::min(params.num_levels, params.MaxPossibleMipmap()); + params.is_layered = params.IsLayered(); + } else { + params.target = SurfaceTarget::TextureBuffer; + params.width = config.tic.Width(); + params.pitch = params.width * params.GetBytesPerPixel(); + params.height = 1; + params.depth = 1; + params.num_levels = 1; + params.emulated_levels = 1; + params.is_layered = false; + } + return params; +} + +SurfaceParams SurfaceParams::CreateForDepthBuffer( + Core::System& system, u32 zeta_width, u32 zeta_height, Tegra::DepthFormat format, + u32 block_width, u32 block_height, u32 block_depth, + Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type) { + SurfaceParams params; + params.is_tiled = type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear; + params.srgb_conversion = false; + params.block_width = std::min(block_width, 5U); + params.block_height = std::min(block_height, 5U); + params.block_depth = std::min(block_depth, 5U); + params.tile_width_spacing = 1; + params.pixel_format = PixelFormatFromDepthFormat(format); + params.component_type = ComponentTypeFromDepthFormat(format); + params.type = GetFormatType(params.pixel_format); + params.width = zeta_width; + params.height = zeta_height; + params.target = SurfaceTarget::Texture2D; + params.depth = 1; + params.pitch = 0; + params.num_levels = 1; + params.emulated_levels = 1; + params.is_layered = false; + return params; +} + +SurfaceParams SurfaceParams::CreateForFramebuffer(Core::System& system, std::size_t index) { + const auto& config{system.GPU().Maxwell3D().regs.rt[index]}; + SurfaceParams params; + params.is_tiled = + config.memory_layout.type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear; + params.srgb_conversion = config.format == Tegra::RenderTargetFormat::BGRA8_SRGB || + config.format == Tegra::RenderTargetFormat::RGBA8_SRGB; + params.block_width = config.memory_layout.block_width; + params.block_height = config.memory_layout.block_height; + params.block_depth = config.memory_layout.block_depth; + params.tile_width_spacing = 1; + params.pixel_format = PixelFormatFromRenderTargetFormat(config.format); + params.component_type = ComponentTypeFromRenderTarget(config.format); + params.type = GetFormatType(params.pixel_format); + if (params.is_tiled) { + params.pitch = 0; + params.width = config.width; + } else { + const u32 bpp = GetFormatBpp(params.pixel_format) / CHAR_BIT; + params.pitch = config.width; + params.width = params.pitch / bpp; + } + params.height = config.height; + params.depth = 1; + params.target = SurfaceTarget::Texture2D; + params.num_levels = 1; + params.emulated_levels = 1; + params.is_layered = false; + return params; +} + +SurfaceParams SurfaceParams::CreateForFermiCopySurface( + const Tegra::Engines::Fermi2D::Regs::Surface& config) { + SurfaceParams params{}; + params.is_tiled = !config.linear; + params.srgb_conversion = config.format == Tegra::RenderTargetFormat::BGRA8_SRGB || + config.format == Tegra::RenderTargetFormat::RGBA8_SRGB; + params.block_width = params.is_tiled ? std::min(config.BlockWidth(), 5U) : 0, + params.block_height = params.is_tiled ? std::min(config.BlockHeight(), 5U) : 0, + params.block_depth = params.is_tiled ? std::min(config.BlockDepth(), 5U) : 0, + params.tile_width_spacing = 1; + params.pixel_format = PixelFormatFromRenderTargetFormat(config.format); + params.component_type = ComponentTypeFromRenderTarget(config.format); + params.type = GetFormatType(params.pixel_format); + params.width = config.width; + params.height = config.height; + params.pitch = config.pitch; + // TODO(Rodrigo): Try to guess the surface target from depth and layer parameters + params.target = SurfaceTarget::Texture2D; + params.depth = 1; + params.num_levels = 1; + params.emulated_levels = 1; + params.is_layered = params.IsLayered(); + return params; +} + +bool SurfaceParams::IsLayered() const { + switch (target) { + case SurfaceTarget::Texture1DArray: + case SurfaceTarget::Texture2DArray: + case SurfaceTarget::TextureCubemap: + case SurfaceTarget::TextureCubeArray: + return true; + default: + return false; + } +} + +// Auto block resizing algorithm from: +// https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nv50/nv50_miptree.c +u32 SurfaceParams::GetMipBlockHeight(u32 level) const { + if (level == 0) { + return this->block_height; + } + + const u32 height_new{GetMipHeight(level)}; + const u32 default_block_height{GetDefaultBlockHeight()}; + const u32 blocks_in_y{(height_new + default_block_height - 1) / default_block_height}; + const u32 block_height_new = Common::Log2Ceil32(blocks_in_y); + return std::clamp(block_height_new, 3U, 7U) - 3U; +} + +u32 SurfaceParams::GetMipBlockDepth(u32 level) const { + if (level == 0) { + return this->block_depth; + } + if (is_layered) { + return 0; + } + + const u32 depth_new{GetMipDepth(level)}; + const u32 block_depth_new = Common::Log2Ceil32(depth_new); + if (block_depth_new > 4) { + return 5 - (GetMipBlockHeight(level) >= 2); + } + return block_depth_new; +} + +std::size_t SurfaceParams::GetGuestMipmapLevelOffset(u32 level) const { + std::size_t offset = 0; + for (u32 i = 0; i < level; i++) { + offset += GetInnerMipmapMemorySize(i, false, false); + } + return offset; +} + +std::size_t SurfaceParams::GetHostMipmapLevelOffset(u32 level) const { + std::size_t offset = 0; + for (u32 i = 0; i < level; i++) { + offset += GetInnerMipmapMemorySize(i, true, false) * GetNumLayers(); + } + return offset; +} + +std::size_t SurfaceParams::GetConvertedMipmapOffset(u32 level) const { + std::size_t offset = 0; + for (u32 i = 0; i < level; i++) { + offset += GetConvertedMipmapSize(i); + } + return offset; +} + +std::size_t SurfaceParams::GetConvertedMipmapSize(u32 level) const { + constexpr std::size_t rgba8_bpp = 4ULL; + const std::size_t width_t = GetMipWidth(level); + const std::size_t height_t = GetMipHeight(level); + const std::size_t depth_t = is_layered ? depth : GetMipDepth(level); + return width_t * height_t * depth_t * rgba8_bpp; +} + +std::size_t SurfaceParams::GetLayerSize(bool as_host_size, bool uncompressed) const { + std::size_t size = 0; + for (u32 level = 0; level < num_levels; ++level) { + size += GetInnerMipmapMemorySize(level, as_host_size, uncompressed); + } + if (is_tiled && is_layered) { + return Common::AlignBits(size, + Tegra::Texture::GetGOBSizeShift() + block_height + block_depth); + } + return size; +} + +std::size_t SurfaceParams::GetInnerMipmapMemorySize(u32 level, bool as_host_size, + bool uncompressed) const { + const bool tiled{as_host_size ? false : is_tiled}; + const u32 width{GetMipmapSize(uncompressed, GetMipWidth(level), GetDefaultBlockWidth())}; + const u32 height{GetMipmapSize(uncompressed, GetMipHeight(level), GetDefaultBlockHeight())}; + const u32 depth{is_layered ? 1U : GetMipDepth(level)}; + return Tegra::Texture::CalculateSize(tiled, GetBytesPerPixel(), width, height, depth, + GetMipBlockHeight(level), GetMipBlockDepth(level)); +} + +bool SurfaceParams::operator==(const SurfaceParams& rhs) const { + return std::tie(is_tiled, block_width, block_height, block_depth, tile_width_spacing, width, + height, depth, pitch, num_levels, pixel_format, component_type, type, target) == + std::tie(rhs.is_tiled, rhs.block_width, rhs.block_height, rhs.block_depth, + rhs.tile_width_spacing, rhs.width, rhs.height, rhs.depth, rhs.pitch, + rhs.num_levels, rhs.pixel_format, rhs.component_type, rhs.type, rhs.target); +} + +std::string SurfaceParams::TargetName() const { + switch (target) { + case SurfaceTarget::Texture1D: + return "1D"; + case SurfaceTarget::TextureBuffer: + return "TexBuffer"; + case SurfaceTarget::Texture2D: + return "2D"; + case SurfaceTarget::Texture3D: + return "3D"; + case SurfaceTarget::Texture1DArray: + return "1DArray"; + case SurfaceTarget::Texture2DArray: + return "2DArray"; + case SurfaceTarget::TextureCubemap: + return "Cube"; + case SurfaceTarget::TextureCubeArray: + return "CubeArray"; + default: + LOG_CRITICAL(HW_GPU, "Unimplemented surface_target={}", static_cast<u32>(target)); + UNREACHABLE(); + return fmt::format("TUK({})", static_cast<u32>(target)); + } +} + +} // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_params.h b/src/video_core/texture_cache/surface_params.h new file mode 100644 index 000000000..358d6757c --- /dev/null +++ b/src/video_core/texture_cache/surface_params.h @@ -0,0 +1,286 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include <map> + +#include "common/alignment.h" +#include "common/bit_util.h" +#include "common/cityhash.h" +#include "common/common_types.h" +#include "video_core/engines/fermi_2d.h" +#include "video_core/engines/maxwell_3d.h" +#include "video_core/shader/shader_ir.h" +#include "video_core/surface.h" +#include "video_core/textures/decoders.h" + +namespace VideoCommon { + +using VideoCore::Surface::SurfaceCompression; + +class SurfaceParams { +public: + /// Creates SurfaceCachedParams from a texture configuration. + static SurfaceParams CreateForTexture(Core::System& system, + const Tegra::Texture::FullTextureInfo& config, + const VideoCommon::Shader::Sampler& entry); + + /// Creates SurfaceCachedParams for a depth buffer configuration. + static SurfaceParams CreateForDepthBuffer( + Core::System& system, u32 zeta_width, u32 zeta_height, Tegra::DepthFormat format, + u32 block_width, u32 block_height, u32 block_depth, + Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type); + + /// Creates SurfaceCachedParams from a framebuffer configuration. + static SurfaceParams CreateForFramebuffer(Core::System& system, std::size_t index); + + /// Creates SurfaceCachedParams from a Fermi2D surface configuration. + static SurfaceParams CreateForFermiCopySurface( + const Tegra::Engines::Fermi2D::Regs::Surface& config); + + std::size_t Hash() const { + return static_cast<std::size_t>( + Common::CityHash64(reinterpret_cast<const char*>(this), sizeof(*this))); + } + + bool operator==(const SurfaceParams& rhs) const; + + bool operator!=(const SurfaceParams& rhs) const { + return !operator==(rhs); + } + + std::size_t GetGuestSizeInBytes() const { + return GetInnerMemorySize(false, false, false); + } + + std::size_t GetHostSizeInBytes() const { + std::size_t host_size_in_bytes; + if (GetCompressionType() == SurfaceCompression::Converted) { + constexpr std::size_t rgb8_bpp = 4ULL; + // ASTC is uncompressed in software, in emulated as RGBA8 + host_size_in_bytes = 0; + for (u32 level = 0; level < num_levels; ++level) { + host_size_in_bytes += GetConvertedMipmapSize(level); + } + } else { + host_size_in_bytes = GetInnerMemorySize(true, false, false); + } + return host_size_in_bytes; + } + + u32 GetBlockAlignedWidth() const { + return Common::AlignUp(width, 64 / GetBytesPerPixel()); + } + + /// Returns the width of a given mipmap level. + u32 GetMipWidth(u32 level) const { + return std::max(1U, width >> level); + } + + /// Returns the height of a given mipmap level. + u32 GetMipHeight(u32 level) const { + return std::max(1U, height >> level); + } + + /// Returns the depth of a given mipmap level. + u32 GetMipDepth(u32 level) const { + return is_layered ? depth : std::max(1U, depth >> level); + } + + /// Returns the block height of a given mipmap level. + u32 GetMipBlockHeight(u32 level) const; + + /// Returns the block depth of a given mipmap level. + u32 GetMipBlockDepth(u32 level) const; + + /// Returns the best possible row/pitch alignment for the surface. + u32 GetRowAlignment(u32 level) const { + const u32 bpp = + GetCompressionType() == SurfaceCompression::Converted ? 4 : GetBytesPerPixel(); + return 1U << Common::CountTrailingZeroes32(GetMipWidth(level) * bpp); + } + + /// Returns the offset in bytes in guest memory of a given mipmap level. + std::size_t GetGuestMipmapLevelOffset(u32 level) const; + + /// Returns the offset in bytes in host memory (linear) of a given mipmap level. + std::size_t GetHostMipmapLevelOffset(u32 level) const; + + /// Returns the offset in bytes in host memory (linear) of a given mipmap level + /// for a texture that is converted in host gpu. + std::size_t GetConvertedMipmapOffset(u32 level) const; + + /// Returns the size in bytes in guest memory of a given mipmap level. + std::size_t GetGuestMipmapSize(u32 level) const { + return GetInnerMipmapMemorySize(level, false, false); + } + + /// Returns the size in bytes in host memory (linear) of a given mipmap level. + std::size_t GetHostMipmapSize(u32 level) const { + return GetInnerMipmapMemorySize(level, true, false) * GetNumLayers(); + } + + std::size_t GetConvertedMipmapSize(u32 level) const; + + /// Returns the size of a layer in bytes in guest memory. + std::size_t GetGuestLayerSize() const { + return GetLayerSize(false, false); + } + + /// Returns the size of a layer in bytes in host memory for a given mipmap level. + std::size_t GetHostLayerSize(u32 level) const { + ASSERT(target != VideoCore::Surface::SurfaceTarget::Texture3D); + return GetInnerMipmapMemorySize(level, true, false); + } + + /// Returns the max possible mipmap that the texture can have in host gpu + u32 MaxPossibleMipmap() const { + const u32 max_mipmap_w = Common::Log2Ceil32(width) + 1U; + const u32 max_mipmap_h = Common::Log2Ceil32(height) + 1U; + const u32 max_mipmap = std::max(max_mipmap_w, max_mipmap_h); + if (target != VideoCore::Surface::SurfaceTarget::Texture3D) + return max_mipmap; + return std::max(max_mipmap, Common::Log2Ceil32(depth) + 1U); + } + + /// Returns if the guest surface is a compressed surface. + bool IsCompressed() const { + return GetDefaultBlockHeight() > 1 || GetDefaultBlockWidth() > 1; + } + + /// Returns the default block width. + u32 GetDefaultBlockWidth() const { + return VideoCore::Surface::GetDefaultBlockWidth(pixel_format); + } + + /// Returns the default block height. + u32 GetDefaultBlockHeight() const { + return VideoCore::Surface::GetDefaultBlockHeight(pixel_format); + } + + /// Returns the bits per pixel. + u32 GetBitsPerPixel() const { + return VideoCore::Surface::GetFormatBpp(pixel_format); + } + + /// Returns the bytes per pixel. + u32 GetBytesPerPixel() const { + return VideoCore::Surface::GetBytesPerPixel(pixel_format); + } + + /// Returns true if the pixel format is a depth and/or stencil format. + bool IsPixelFormatZeta() const { + return pixel_format >= VideoCore::Surface::PixelFormat::MaxColorFormat && + pixel_format < VideoCore::Surface::PixelFormat::MaxDepthStencilFormat; + } + + /// Returns how the compression should be handled for this texture. + SurfaceCompression GetCompressionType() const { + return VideoCore::Surface::GetFormatCompressionType(pixel_format); + } + + /// Returns is the surface is a TextureBuffer type of surface. + bool IsBuffer() const { + return target == VideoCore::Surface::SurfaceTarget::TextureBuffer; + } + + /// Returns the debug name of the texture for use in graphic debuggers. + std::string TargetName() const; + + // Helper used for out of class size calculations + static std::size_t AlignLayered(const std::size_t out_size, const u32 block_height, + const u32 block_depth) { + return Common::AlignBits(out_size, + Tegra::Texture::GetGOBSizeShift() + block_height + block_depth); + } + + /// Converts a width from a type of surface into another. This helps represent the + /// equivalent value between compressed/non-compressed textures. + static u32 ConvertWidth(u32 width, VideoCore::Surface::PixelFormat pixel_format_from, + VideoCore::Surface::PixelFormat pixel_format_to) { + const u32 bw1 = VideoCore::Surface::GetDefaultBlockWidth(pixel_format_from); + const u32 bw2 = VideoCore::Surface::GetDefaultBlockWidth(pixel_format_to); + return (width * bw2 + bw1 - 1) / bw1; + } + + /// Converts a height from a type of surface into another. This helps represent the + /// equivalent value between compressed/non-compressed textures. + static u32 ConvertHeight(u32 height, VideoCore::Surface::PixelFormat pixel_format_from, + VideoCore::Surface::PixelFormat pixel_format_to) { + const u32 bh1 = VideoCore::Surface::GetDefaultBlockHeight(pixel_format_from); + const u32 bh2 = VideoCore::Surface::GetDefaultBlockHeight(pixel_format_to); + return (height * bh2 + bh1 - 1) / bh1; + } + + // Finds the maximun possible width between 2 2D layers of different formats + static u32 IntersectWidth(const SurfaceParams& src_params, const SurfaceParams& dst_params, + const u32 src_level, const u32 dst_level) { + const u32 bw1 = src_params.GetDefaultBlockWidth(); + const u32 bw2 = dst_params.GetDefaultBlockWidth(); + const u32 t_src_width = (src_params.GetMipWidth(src_level) * bw2 + bw1 - 1) / bw1; + const u32 t_dst_width = (dst_params.GetMipWidth(dst_level) * bw1 + bw2 - 1) / bw2; + return std::min(t_src_width, t_dst_width); + } + + // Finds the maximun possible height between 2 2D layers of different formats + static u32 IntersectHeight(const SurfaceParams& src_params, const SurfaceParams& dst_params, + const u32 src_level, const u32 dst_level) { + const u32 bh1 = src_params.GetDefaultBlockHeight(); + const u32 bh2 = dst_params.GetDefaultBlockHeight(); + const u32 t_src_height = (src_params.GetMipHeight(src_level) * bh2 + bh1 - 1) / bh1; + const u32 t_dst_height = (dst_params.GetMipHeight(dst_level) * bh1 + bh2 - 1) / bh2; + return std::min(t_src_height, t_dst_height); + } + + bool is_tiled; + bool srgb_conversion; + bool is_layered; + u32 block_width; + u32 block_height; + u32 block_depth; + u32 tile_width_spacing; + u32 width; + u32 height; + u32 depth; + u32 pitch; + u32 num_levels; + u32 emulated_levels; + VideoCore::Surface::PixelFormat pixel_format; + VideoCore::Surface::ComponentType component_type; + VideoCore::Surface::SurfaceType type; + VideoCore::Surface::SurfaceTarget target; + +private: + /// Returns the size of a given mipmap level inside a layer. + std::size_t GetInnerMipmapMemorySize(u32 level, bool as_host_size, bool uncompressed) const; + + /// Returns the size of all mipmap levels and aligns as needed. + std::size_t GetInnerMemorySize(bool as_host_size, bool layer_only, bool uncompressed) const { + return GetLayerSize(as_host_size, uncompressed) * (layer_only ? 1U : depth); + } + + /// Returns the size of a layer + std::size_t GetLayerSize(bool as_host_size, bool uncompressed) const; + + std::size_t GetNumLayers() const { + return is_layered ? depth : 1; + } + + /// Returns true if these parameters are from a layered surface. + bool IsLayered() const; +}; + +} // namespace VideoCommon + +namespace std { + +template <> +struct hash<VideoCommon::SurfaceParams> { + std::size_t operator()(const VideoCommon::SurfaceParams& k) const noexcept { + return k.Hash(); + } +}; + +} // namespace std diff --git a/src/video_core/texture_cache/surface_view.cpp b/src/video_core/texture_cache/surface_view.cpp new file mode 100644 index 000000000..467696a4c --- /dev/null +++ b/src/video_core/texture_cache/surface_view.cpp @@ -0,0 +1,23 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include <tuple> + +#include "common/common_types.h" +#include "video_core/texture_cache/surface_view.h" + +namespace VideoCommon { + +std::size_t ViewParams::Hash() const { + return static_cast<std::size_t>(base_layer) ^ static_cast<std::size_t>(num_layers << 16) ^ + (static_cast<std::size_t>(base_level) << 24) ^ + (static_cast<std::size_t>(num_levels) << 32) ^ (static_cast<std::size_t>(target) << 36); +} + +bool ViewParams::operator==(const ViewParams& rhs) const { + return std::tie(base_layer, num_layers, base_level, num_levels, target) == + std::tie(rhs.base_layer, rhs.num_layers, rhs.base_level, rhs.num_levels, rhs.target); +} + +} // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_view.h b/src/video_core/texture_cache/surface_view.h new file mode 100644 index 000000000..04ca5639b --- /dev/null +++ b/src/video_core/texture_cache/surface_view.h @@ -0,0 +1,67 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include <functional> + +#include "common/common_types.h" +#include "video_core/surface.h" +#include "video_core/texture_cache/surface_params.h" + +namespace VideoCommon { + +struct ViewParams { + ViewParams(VideoCore::Surface::SurfaceTarget target, u32 base_layer, u32 num_layers, + u32 base_level, u32 num_levels) + : target{target}, base_layer{base_layer}, num_layers{num_layers}, base_level{base_level}, + num_levels{num_levels} {} + + std::size_t Hash() const; + + bool operator==(const ViewParams& rhs) const; + + VideoCore::Surface::SurfaceTarget target{}; + u32 base_layer{}; + u32 num_layers{}; + u32 base_level{}; + u32 num_levels{}; + + bool IsLayered() const { + switch (target) { + case VideoCore::Surface::SurfaceTarget::Texture1DArray: + case VideoCore::Surface::SurfaceTarget::Texture2DArray: + case VideoCore::Surface::SurfaceTarget::TextureCubemap: + case VideoCore::Surface::SurfaceTarget::TextureCubeArray: + return true; + default: + return false; + } + } +}; + +class ViewBase { +public: + ViewBase(const ViewParams& params) : params{params} {} + + const ViewParams& GetViewParams() const { + return params; + } + +protected: + ViewParams params; +}; + +} // namespace VideoCommon + +namespace std { + +template <> +struct hash<VideoCommon::ViewParams> { + std::size_t operator()(const VideoCommon::ViewParams& k) const noexcept { + return k.Hash(); + } +}; + +} // namespace std diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h new file mode 100644 index 000000000..c9e72531a --- /dev/null +++ b/src/video_core/texture_cache/texture_cache.h @@ -0,0 +1,814 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include <algorithm> +#include <array> +#include <memory> +#include <mutex> +#include <set> +#include <tuple> +#include <unordered_map> +#include <vector> + +#include <boost/icl/interval_map.hpp> +#include <boost/range/iterator_range.hpp> + +#include "common/assert.h" +#include "common/common_types.h" +#include "common/math_util.h" +#include "core/core.h" +#include "core/memory.h" +#include "core/settings.h" +#include "video_core/engines/fermi_2d.h" +#include "video_core/engines/maxwell_3d.h" +#include "video_core/gpu.h" +#include "video_core/memory_manager.h" +#include "video_core/rasterizer_interface.h" +#include "video_core/surface.h" +#include "video_core/texture_cache/copy_params.h" +#include "video_core/texture_cache/surface_base.h" +#include "video_core/texture_cache/surface_params.h" +#include "video_core/texture_cache/surface_view.h" + +namespace Tegra::Texture { +struct FullTextureInfo; +} + +namespace VideoCore { +class RasterizerInterface; +} + +namespace VideoCommon { + +using VideoCore::Surface::PixelFormat; + +using VideoCore::Surface::SurfaceTarget; +using RenderTargetConfig = Tegra::Engines::Maxwell3D::Regs::RenderTargetConfig; + +template <typename TSurface, typename TView> +class TextureCache { + using IntervalMap = boost::icl::interval_map<CacheAddr, std::set<TSurface>>; + using IntervalType = typename IntervalMap::interval_type; + +public: + void InvalidateRegion(CacheAddr addr, std::size_t size) { + std::lock_guard lock{mutex}; + + for (const auto& surface : GetSurfacesInRegion(addr, size)) { + Unregister(surface); + } + } + + /*** + * `Guard` guarantees that rendertargets don't unregister themselves if the + * collide. Protection is currently only done on 3D slices. + ***/ + void GuardRenderTargets(bool new_guard) { + guard_render_targets = new_guard; + } + + void GuardSamplers(bool new_guard) { + guard_samplers = new_guard; + } + + void FlushRegion(CacheAddr addr, std::size_t size) { + std::lock_guard lock{mutex}; + + auto surfaces = GetSurfacesInRegion(addr, size); + if (surfaces.empty()) { + return; + } + std::sort(surfaces.begin(), surfaces.end(), [](const TSurface& a, const TSurface& b) { + return a->GetModificationTick() < b->GetModificationTick(); + }); + for (const auto& surface : surfaces) { + FlushSurface(surface); + } + } + + TView GetTextureSurface(const Tegra::Texture::FullTextureInfo& config, + const VideoCommon::Shader::Sampler& entry) { + std::lock_guard lock{mutex}; + const auto gpu_addr{config.tic.Address()}; + if (!gpu_addr) { + return {}; + } + const auto params{SurfaceParams::CreateForTexture(system, config, entry)}; + const auto [surface, view] = GetSurface(gpu_addr, params, true, false); + if (guard_samplers) { + sampled_textures.push_back(surface); + } + return view; + } + + bool TextureBarrier() { + const bool any_rt = + std::any_of(sampled_textures.begin(), sampled_textures.end(), + [](const auto& surface) { return surface->IsRenderTarget(); }); + sampled_textures.clear(); + return any_rt; + } + + TView GetDepthBufferSurface(bool preserve_contents) { + std::lock_guard lock{mutex}; + auto& maxwell3d = system.GPU().Maxwell3D(); + + if (!maxwell3d.dirty_flags.zeta_buffer) { + return depth_buffer.view; + } + maxwell3d.dirty_flags.zeta_buffer = false; + + const auto& regs{maxwell3d.regs}; + const auto gpu_addr{regs.zeta.Address()}; + if (!gpu_addr || !regs.zeta_enable) { + SetEmptyDepthBuffer(); + return {}; + } + const auto depth_params{SurfaceParams::CreateForDepthBuffer( + system, regs.zeta_width, regs.zeta_height, regs.zeta.format, + regs.zeta.memory_layout.block_width, regs.zeta.memory_layout.block_height, + regs.zeta.memory_layout.block_depth, regs.zeta.memory_layout.type)}; + auto surface_view = GetSurface(gpu_addr, depth_params, preserve_contents, true); + if (depth_buffer.target) + depth_buffer.target->MarkAsRenderTarget(false); + depth_buffer.target = surface_view.first; + depth_buffer.view = surface_view.second; + if (depth_buffer.target) + depth_buffer.target->MarkAsRenderTarget(true); + return surface_view.second; + } + + TView GetColorBufferSurface(std::size_t index, bool preserve_contents) { + std::lock_guard lock{mutex}; + ASSERT(index < Tegra::Engines::Maxwell3D::Regs::NumRenderTargets); + auto& maxwell3d = system.GPU().Maxwell3D(); + if (!maxwell3d.dirty_flags.color_buffer[index]) { + return render_targets[index].view; + } + maxwell3d.dirty_flags.color_buffer.reset(index); + + const auto& regs{maxwell3d.regs}; + if (index >= regs.rt_control.count || regs.rt[index].Address() == 0 || + regs.rt[index].format == Tegra::RenderTargetFormat::NONE) { + SetEmptyColorBuffer(index); + return {}; + } + + const auto& config{regs.rt[index]}; + const auto gpu_addr{config.Address()}; + if (!gpu_addr) { + SetEmptyColorBuffer(index); + return {}; + } + + auto surface_view = GetSurface(gpu_addr, SurfaceParams::CreateForFramebuffer(system, index), + preserve_contents, true); + if (render_targets[index].target) + render_targets[index].target->MarkAsRenderTarget(false); + render_targets[index].target = surface_view.first; + render_targets[index].view = surface_view.second; + if (render_targets[index].target) + render_targets[index].target->MarkAsRenderTarget(true); + return surface_view.second; + } + + void MarkColorBufferInUse(std::size_t index) { + if (auto& render_target = render_targets[index].target) { + render_target->MarkAsModified(true, Tick()); + } + } + + void MarkDepthBufferInUse() { + if (depth_buffer.target) { + depth_buffer.target->MarkAsModified(true, Tick()); + } + } + + void SetEmptyDepthBuffer() { + if (depth_buffer.target == nullptr) { + return; + } + depth_buffer.target->MarkAsRenderTarget(false); + depth_buffer.target = nullptr; + depth_buffer.view = nullptr; + } + + void SetEmptyColorBuffer(std::size_t index) { + if (render_targets[index].target == nullptr) { + return; + } + render_targets[index].target->MarkAsRenderTarget(false); + render_targets[index].target = nullptr; + render_targets[index].view = nullptr; + } + + void DoFermiCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src_config, + const Tegra::Engines::Fermi2D::Regs::Surface& dst_config, + const Tegra::Engines::Fermi2D::Config& copy_config) { + std::lock_guard lock{mutex}; + std::pair<TSurface, TView> dst_surface = GetFermiSurface(dst_config); + std::pair<TSurface, TView> src_surface = GetFermiSurface(src_config); + ImageBlit(src_surface.second, dst_surface.second, copy_config); + dst_surface.first->MarkAsModified(true, Tick()); + } + + TSurface TryFindFramebufferSurface(const u8* host_ptr) { + const CacheAddr cache_addr = ToCacheAddr(host_ptr); + if (!cache_addr) { + return nullptr; + } + const CacheAddr page = cache_addr >> registry_page_bits; + std::vector<TSurface>& list = registry[page]; + for (auto& surface : list) { + if (surface->GetCacheAddr() == cache_addr) { + return surface; + } + } + return nullptr; + } + + u64 Tick() { + return ++ticks; + } + +protected: + TextureCache(Core::System& system, VideoCore::RasterizerInterface& rasterizer) + : system{system}, rasterizer{rasterizer} { + for (std::size_t i = 0; i < Tegra::Engines::Maxwell3D::Regs::NumRenderTargets; i++) { + SetEmptyColorBuffer(i); + } + + SetEmptyDepthBuffer(); + staging_cache.SetSize(2); + + const auto make_siblings = [this](PixelFormat a, PixelFormat b) { + siblings_table[static_cast<std::size_t>(a)] = b; + siblings_table[static_cast<std::size_t>(b)] = a; + }; + std::fill(siblings_table.begin(), siblings_table.end(), PixelFormat::Invalid); + make_siblings(PixelFormat::Z16, PixelFormat::R16U); + make_siblings(PixelFormat::Z32F, PixelFormat::R32F); + make_siblings(PixelFormat::Z32FS8, PixelFormat::RG32F); + + sampled_textures.reserve(64); + } + + ~TextureCache() = default; + + virtual TSurface CreateSurface(GPUVAddr gpu_addr, const SurfaceParams& params) = 0; + + virtual void ImageCopy(TSurface& src_surface, TSurface& dst_surface, + const CopyParams& copy_params) = 0; + + virtual void ImageBlit(TView& src_view, TView& dst_view, + const Tegra::Engines::Fermi2D::Config& copy_config) = 0; + + // Depending on the backend, a buffer copy can be slow as it means deoptimizing the texture + // and reading it from a sepparate buffer. + virtual void BufferCopy(TSurface& src_surface, TSurface& dst_surface) = 0; + + void Register(TSurface surface) { + const GPUVAddr gpu_addr = surface->GetGpuAddr(); + const CacheAddr cache_ptr = ToCacheAddr(system.GPU().MemoryManager().GetPointer(gpu_addr)); + const std::size_t size = surface->GetSizeInBytes(); + const std::optional<VAddr> cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); + if (!cache_ptr || !cpu_addr) { + LOG_CRITICAL(HW_GPU, "Failed to register surface with unmapped gpu_address 0x{:016x}", + gpu_addr); + return; + } + const bool continuous = system.GPU().MemoryManager().IsBlockContinuous(gpu_addr, size); + surface->MarkAsContinuous(continuous); + surface->SetCacheAddr(cache_ptr); + surface->SetCpuAddr(*cpu_addr); + RegisterInnerCache(surface); + surface->MarkAsRegistered(true); + rasterizer.UpdatePagesCachedCount(*cpu_addr, size, 1); + } + + void Unregister(TSurface surface) { + if (guard_render_targets && surface->IsProtected()) { + return; + } + const GPUVAddr gpu_addr = surface->GetGpuAddr(); + const CacheAddr cache_ptr = surface->GetCacheAddr(); + const std::size_t size = surface->GetSizeInBytes(); + const VAddr cpu_addr = surface->GetCpuAddr(); + rasterizer.UpdatePagesCachedCount(cpu_addr, size, -1); + UnregisterInnerCache(surface); + surface->MarkAsRegistered(false); + ReserveSurface(surface->GetSurfaceParams(), surface); + } + + TSurface GetUncachedSurface(const GPUVAddr gpu_addr, const SurfaceParams& params) { + if (const auto surface = TryGetReservedSurface(params); surface) { + surface->SetGpuAddr(gpu_addr); + return surface; + } + // No reserved surface available, create a new one and reserve it + auto new_surface{CreateSurface(gpu_addr, params)}; + return new_surface; + } + + std::pair<TSurface, TView> GetFermiSurface( + const Tegra::Engines::Fermi2D::Regs::Surface& config) { + SurfaceParams params = SurfaceParams::CreateForFermiCopySurface(config); + const GPUVAddr gpu_addr = config.Address(); + return GetSurface(gpu_addr, params, true, false); + } + + Core::System& system; + +private: + enum class RecycleStrategy : u32 { + Ignore = 0, + Flush = 1, + BufferCopy = 3, + }; + + /** + * `PickStrategy` takes care of selecting a proper strategy to deal with a texture recycle. + * @param overlaps, the overlapping surfaces registered in the cache. + * @param params, the paremeters on the new surface. + * @param gpu_addr, the starting address of the new surface. + * @param untopological, tells the recycler that the texture has no way to match the overlaps + * due to topological reasons. + **/ + RecycleStrategy PickStrategy(std::vector<TSurface>& overlaps, const SurfaceParams& params, + const GPUVAddr gpu_addr, const MatchTopologyResult untopological) { + if (Settings::values.use_accurate_gpu_emulation) { + return RecycleStrategy::Flush; + } + // 3D Textures decision + if (params.block_depth > 1 || params.target == SurfaceTarget::Texture3D) { + return RecycleStrategy::Flush; + } + for (auto s : overlaps) { + const auto& s_params = s->GetSurfaceParams(); + if (s_params.block_depth > 1 || s_params.target == SurfaceTarget::Texture3D) { + return RecycleStrategy::Flush; + } + } + // Untopological decision + if (untopological == MatchTopologyResult::CompressUnmatch) { + return RecycleStrategy::Flush; + } + if (untopological == MatchTopologyResult::FullMatch && !params.is_tiled) { + return RecycleStrategy::Flush; + } + return RecycleStrategy::Ignore; + } + + /** + * `RecycleSurface` es a method we use to decide what to do with textures we can't resolve in + *the cache It has 2 implemented strategies: Ignore and Flush. Ignore just unregisters all the + *overlaps and loads the new texture. Flush, flushes all the overlaps into memory and loads the + *new surface from that data. + * @param overlaps, the overlapping surfaces registered in the cache. + * @param params, the paremeters on the new surface. + * @param gpu_addr, the starting address of the new surface. + * @param preserve_contents, tells if the new surface should be loaded from meory or left blank + * @param untopological, tells the recycler that the texture has no way to match the overlaps + * due to topological reasons. + **/ + std::pair<TSurface, TView> RecycleSurface(std::vector<TSurface>& overlaps, + const SurfaceParams& params, const GPUVAddr gpu_addr, + const bool preserve_contents, + const MatchTopologyResult untopological) { + const bool do_load = preserve_contents && Settings::values.use_accurate_gpu_emulation; + for (auto& surface : overlaps) { + Unregister(surface); + } + switch (PickStrategy(overlaps, params, gpu_addr, untopological)) { + case RecycleStrategy::Ignore: { + return InitializeSurface(gpu_addr, params, do_load); + } + case RecycleStrategy::Flush: { + std::sort(overlaps.begin(), overlaps.end(), + [](const TSurface& a, const TSurface& b) -> bool { + return a->GetModificationTick() < b->GetModificationTick(); + }); + for (auto& surface : overlaps) { + FlushSurface(surface); + } + return InitializeSurface(gpu_addr, params, preserve_contents); + } + case RecycleStrategy::BufferCopy: { + auto new_surface = GetUncachedSurface(gpu_addr, params); + BufferCopy(overlaps[0], new_surface); + return {new_surface, new_surface->GetMainView()}; + } + default: { + UNIMPLEMENTED_MSG("Unimplemented Texture Cache Recycling Strategy!"); + return InitializeSurface(gpu_addr, params, do_load); + } + } + } + + /** + * `RebuildSurface` this method takes a single surface and recreates into another that + * may differ in format, target or width alingment. + * @param current_surface, the registered surface in the cache which we want to convert. + * @param params, the new surface params which we'll use to recreate the surface. + **/ + std::pair<TSurface, TView> RebuildSurface(TSurface current_surface, const SurfaceParams& params, + bool is_render) { + const auto gpu_addr = current_surface->GetGpuAddr(); + const auto& cr_params = current_surface->GetSurfaceParams(); + TSurface new_surface; + if (cr_params.pixel_format != params.pixel_format && !is_render && + GetSiblingFormat(cr_params.pixel_format) == params.pixel_format) { + SurfaceParams new_params = params; + new_params.pixel_format = cr_params.pixel_format; + new_params.component_type = cr_params.component_type; + new_params.type = cr_params.type; + new_surface = GetUncachedSurface(gpu_addr, new_params); + } else { + new_surface = GetUncachedSurface(gpu_addr, params); + } + const auto& final_params = new_surface->GetSurfaceParams(); + if (cr_params.type != final_params.type || + (cr_params.component_type != final_params.component_type)) { + BufferCopy(current_surface, new_surface); + } else { + std::vector<CopyParams> bricks = current_surface->BreakDown(final_params); + for (auto& brick : bricks) { + ImageCopy(current_surface, new_surface, brick); + } + } + Unregister(current_surface); + Register(new_surface); + new_surface->MarkAsModified(current_surface->IsModified(), Tick()); + return {new_surface, new_surface->GetMainView()}; + } + + /** + * `ManageStructuralMatch` this method takes a single surface and checks with the new surface's + * params if it's an exact match, we return the main view of the registered surface. If it's + * formats don't match, we rebuild the surface. We call this last method a `Mirage`. If formats + * match but the targets don't, we create an overview View of the registered surface. + * @param current_surface, the registered surface in the cache which we want to convert. + * @param params, the new surface params which we want to check. + **/ + std::pair<TSurface, TView> ManageStructuralMatch(TSurface current_surface, + const SurfaceParams& params, bool is_render) { + const bool is_mirage = !current_surface->MatchFormat(params.pixel_format); + const bool matches_target = current_surface->MatchTarget(params.target); + const auto match_check = [&]() -> std::pair<TSurface, TView> { + if (matches_target) { + return {current_surface, current_surface->GetMainView()}; + } + return {current_surface, current_surface->EmplaceOverview(params)}; + }; + if (!is_mirage) { + return match_check(); + } + if (!is_render && GetSiblingFormat(current_surface->GetFormat()) == params.pixel_format) { + return match_check(); + } + return RebuildSurface(current_surface, params, is_render); + } + + /** + * `TryReconstructSurface` unlike `RebuildSurface` where we know the registered surface + * matches the candidate in some way, we got no guarantess here. We try to see if the overlaps + * are sublayers/mipmaps of the new surface, if they all match we end up recreating a surface + * for them, else we return nothing. + * @param overlaps, the overlapping surfaces registered in the cache. + * @param params, the paremeters on the new surface. + * @param gpu_addr, the starting address of the new surface. + **/ + std::optional<std::pair<TSurface, TView>> TryReconstructSurface(std::vector<TSurface>& overlaps, + const SurfaceParams& params, + const GPUVAddr gpu_addr) { + if (params.target == SurfaceTarget::Texture3D) { + return {}; + } + bool modified = false; + TSurface new_surface = GetUncachedSurface(gpu_addr, params); + u32 passed_tests = 0; + for (auto& surface : overlaps) { + const SurfaceParams& src_params = surface->GetSurfaceParams(); + if (src_params.is_layered || src_params.num_levels > 1) { + // We send this cases to recycle as they are more complex to handle + return {}; + } + const std::size_t candidate_size = surface->GetSizeInBytes(); + auto mipmap_layer{new_surface->GetLayerMipmap(surface->GetGpuAddr())}; + if (!mipmap_layer) { + continue; + } + const auto [layer, mipmap] = *mipmap_layer; + if (new_surface->GetMipmapSize(mipmap) != candidate_size) { + continue; + } + modified |= surface->IsModified(); + // Now we got all the data set up + const u32 width = SurfaceParams::IntersectWidth(src_params, params, 0, mipmap); + const u32 height = SurfaceParams::IntersectHeight(src_params, params, 0, mipmap); + const CopyParams copy_params(0, 0, 0, 0, 0, layer, 0, mipmap, width, height, 1); + passed_tests++; + ImageCopy(surface, new_surface, copy_params); + } + if (passed_tests == 0) { + return {}; + // In Accurate GPU all tests should pass, else we recycle + } else if (Settings::values.use_accurate_gpu_emulation && passed_tests != overlaps.size()) { + return {}; + } + for (auto surface : overlaps) { + Unregister(surface); + } + new_surface->MarkAsModified(modified, Tick()); + Register(new_surface); + return {{new_surface, new_surface->GetMainView()}}; + } + + /** + * `GetSurface` gets the starting address and parameters of a candidate surface and tries + * to find a matching surface within the cache. This is done in 3 big steps. The first is to + * check the 1st Level Cache in order to find an exact match, if we fail, we move to step 2. + * Step 2 is checking if there are any overlaps at all, if none, we just load the texture from + * memory else we move to step 3. Step 3 consists on figuring the relationship between the + * candidate texture and the overlaps. We divide the scenarios depending if there's 1 or many + * overlaps. If there's many, we just try to reconstruct a new surface out of them based on the + * candidate's parameters, if we fail, we recycle. When there's only 1 overlap then we have to + * check if the candidate is a view (layer/mipmap) of the overlap or if the registered surface + * is a mipmap/layer of the candidate. In this last case we reconstruct a new surface. + * @param gpu_addr, the starting address of the candidate surface. + * @param params, the paremeters on the candidate surface. + * @param preserve_contents, tells if the new surface should be loaded from meory or left blank. + **/ + std::pair<TSurface, TView> GetSurface(const GPUVAddr gpu_addr, const SurfaceParams& params, + bool preserve_contents, bool is_render) { + const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; + const auto cache_addr{ToCacheAddr(host_ptr)}; + + // Step 0: guarantee a valid surface + if (!cache_addr) { + // Return a null surface if it's invalid + SurfaceParams new_params = params; + new_params.width = 1; + new_params.height = 1; + new_params.depth = 1; + new_params.block_height = 0; + new_params.block_depth = 0; + return InitializeSurface(gpu_addr, new_params, false); + } + + // Step 1 + // Check Level 1 Cache for a fast structural match. If candidate surface + // matches at certain level we are pretty much done. + if (const auto iter = l1_cache.find(cache_addr); iter != l1_cache.end()) { + TSurface& current_surface = iter->second; + const auto topological_result = current_surface->MatchesTopology(params); + if (topological_result != MatchTopologyResult::FullMatch) { + std::vector<TSurface> overlaps{current_surface}; + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + topological_result); + } + const auto struct_result = current_surface->MatchesStructure(params); + if (struct_result != MatchStructureResult::None && + (params.target != SurfaceTarget::Texture3D || + current_surface->MatchTarget(params.target))) { + if (struct_result == MatchStructureResult::FullMatch) { + return ManageStructuralMatch(current_surface, params, is_render); + } else { + return RebuildSurface(current_surface, params, is_render); + } + } + } + + // Step 2 + // Obtain all possible overlaps in the memory region + const std::size_t candidate_size = params.GetGuestSizeInBytes(); + auto overlaps{GetSurfacesInRegion(cache_addr, candidate_size)}; + + // If none are found, we are done. we just load the surface and create it. + if (overlaps.empty()) { + return InitializeSurface(gpu_addr, params, preserve_contents); + } + + // Step 3 + // Now we need to figure the relationship between the texture and its overlaps + // we do a topological test to ensure we can find some relationship. If it fails + // inmediatly recycle the texture + for (const auto& surface : overlaps) { + const auto topological_result = surface->MatchesTopology(params); + if (topological_result != MatchTopologyResult::FullMatch) { + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + topological_result); + } + } + + // Split cases between 1 overlap or many. + if (overlaps.size() == 1) { + TSurface current_surface = overlaps[0]; + // First check if the surface is within the overlap. If not, it means + // two things either the candidate surface is a supertexture of the overlap + // or they don't match in any known way. + if (!current_surface->IsInside(gpu_addr, gpu_addr + candidate_size)) { + if (current_surface->GetGpuAddr() == gpu_addr) { + std::optional<std::pair<TSurface, TView>> view = + TryReconstructSurface(overlaps, params, gpu_addr); + if (view) { + return *view; + } + } + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + MatchTopologyResult::FullMatch); + } + // Now we check if the candidate is a mipmap/layer of the overlap + std::optional<TView> view = + current_surface->EmplaceView(params, gpu_addr, candidate_size); + if (view) { + const bool is_mirage = !current_surface->MatchFormat(params.pixel_format); + if (is_mirage) { + // On a mirage view, we need to recreate the surface under this new view + // and then obtain a view again. + SurfaceParams new_params = current_surface->GetSurfaceParams(); + const u32 wh = SurfaceParams::ConvertWidth( + new_params.width, new_params.pixel_format, params.pixel_format); + const u32 hh = SurfaceParams::ConvertHeight( + new_params.height, new_params.pixel_format, params.pixel_format); + new_params.width = wh; + new_params.height = hh; + new_params.pixel_format = params.pixel_format; + std::pair<TSurface, TView> pair = + RebuildSurface(current_surface, new_params, is_render); + std::optional<TView> mirage_view = + pair.first->EmplaceView(params, gpu_addr, candidate_size); + if (mirage_view) + return {pair.first, *mirage_view}; + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + MatchTopologyResult::FullMatch); + } + return {current_surface, *view}; + } + // The next case is unsafe, so if we r in accurate GPU, just skip it + if (Settings::values.use_accurate_gpu_emulation) { + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + MatchTopologyResult::FullMatch); + } + // This is the case the texture is a part of the parent. + if (current_surface->MatchesSubTexture(params, gpu_addr)) { + return RebuildSurface(current_surface, params, is_render); + } + } else { + // If there are many overlaps, odds are they are subtextures of the candidate + // surface. We try to construct a new surface based on the candidate parameters, + // using the overlaps. If a single overlap fails, this will fail. + std::optional<std::pair<TSurface, TView>> view = + TryReconstructSurface(overlaps, params, gpu_addr); + if (view) { + return *view; + } + } + // We failed all the tests, recycle the overlaps into a new texture. + return RecycleSurface(overlaps, params, gpu_addr, preserve_contents, + MatchTopologyResult::FullMatch); + } + + std::pair<TSurface, TView> InitializeSurface(GPUVAddr gpu_addr, const SurfaceParams& params, + bool preserve_contents) { + auto new_surface{GetUncachedSurface(gpu_addr, params)}; + Register(new_surface); + if (preserve_contents) { + LoadSurface(new_surface); + } + return {new_surface, new_surface->GetMainView()}; + } + + void LoadSurface(const TSurface& surface) { + staging_cache.GetBuffer(0).resize(surface->GetHostSizeInBytes()); + surface->LoadBuffer(system.GPU().MemoryManager(), staging_cache); + surface->UploadTexture(staging_cache.GetBuffer(0)); + surface->MarkAsModified(false, Tick()); + } + + void FlushSurface(const TSurface& surface) { + if (!surface->IsModified()) { + return; + } + staging_cache.GetBuffer(0).resize(surface->GetHostSizeInBytes()); + surface->DownloadTexture(staging_cache.GetBuffer(0)); + surface->FlushBuffer(system.GPU().MemoryManager(), staging_cache); + surface->MarkAsModified(false, Tick()); + } + + void RegisterInnerCache(TSurface& surface) { + const CacheAddr cache_addr = surface->GetCacheAddr(); + CacheAddr start = cache_addr >> registry_page_bits; + const CacheAddr end = (surface->GetCacheAddrEnd() - 1) >> registry_page_bits; + l1_cache[cache_addr] = surface; + while (start <= end) { + registry[start].push_back(surface); + start++; + } + } + + void UnregisterInnerCache(TSurface& surface) { + const CacheAddr cache_addr = surface->GetCacheAddr(); + CacheAddr start = cache_addr >> registry_page_bits; + const CacheAddr end = (surface->GetCacheAddrEnd() - 1) >> registry_page_bits; + l1_cache.erase(cache_addr); + while (start <= end) { + auto& reg{registry[start]}; + reg.erase(std::find(reg.begin(), reg.end(), surface)); + start++; + } + } + + std::vector<TSurface> GetSurfacesInRegion(const CacheAddr cache_addr, const std::size_t size) { + if (size == 0) { + return {}; + } + const CacheAddr cache_addr_end = cache_addr + size; + CacheAddr start = cache_addr >> registry_page_bits; + const CacheAddr end = (cache_addr_end - 1) >> registry_page_bits; + std::vector<TSurface> surfaces; + while (start <= end) { + std::vector<TSurface>& list = registry[start]; + for (auto& surface : list) { + if (!surface->IsPicked() && surface->Overlaps(cache_addr, cache_addr_end)) { + surface->MarkAsPicked(true); + surfaces.push_back(surface); + } + } + start++; + } + for (auto& surface : surfaces) { + surface->MarkAsPicked(false); + } + return surfaces; + } + + void ReserveSurface(const SurfaceParams& params, TSurface surface) { + surface_reserve[params].push_back(std::move(surface)); + } + + TSurface TryGetReservedSurface(const SurfaceParams& params) { + auto search{surface_reserve.find(params)}; + if (search == surface_reserve.end()) { + return {}; + } + for (auto& surface : search->second) { + if (!surface->IsRegistered()) { + return surface; + } + } + return {}; + } + + constexpr PixelFormat GetSiblingFormat(PixelFormat format) const { + return siblings_table[static_cast<std::size_t>(format)]; + } + + struct FramebufferTargetInfo { + TSurface target; + TView view; + }; + + VideoCore::RasterizerInterface& rasterizer; + + u64 ticks{}; + + // Guards the cache for protection conflicts. + bool guard_render_targets{}; + bool guard_samplers{}; + + // The siblings table is for formats that can inter exchange with one another + // without causing issues. This is only valid when a conflict occurs on a non + // rendering use. + std::array<PixelFormat, static_cast<std::size_t>(PixelFormat::Max)> siblings_table; + + // The internal Cache is different for the Texture Cache. It's based on buckets + // of 1MB. This fits better for the purpose of this cache as textures are normaly + // large in size. + static constexpr u64 registry_page_bits{20}; + static constexpr u64 registry_page_size{1 << registry_page_bits}; + std::unordered_map<CacheAddr, std::vector<TSurface>> registry; + + // The L1 Cache is used for fast texture lookup before checking the overlaps + // This avoids calculating size and other stuffs. + std::unordered_map<CacheAddr, TSurface> l1_cache; + + /// The surface reserve is a "backup" cache, this is where we put unique surfaces that have + /// previously been used. This is to prevent surfaces from being constantly created and + /// destroyed when used with different surface parameters. + std::unordered_map<SurfaceParams, std::vector<TSurface>> surface_reserve; + std::array<FramebufferTargetInfo, Tegra::Engines::Maxwell3D::Regs::NumRenderTargets> + render_targets; + FramebufferTargetInfo depth_buffer; + + std::vector<TSurface> sampled_textures; + + StagingCache staging_cache; + std::recursive_mutex mutex; +}; + +} // namespace VideoCommon |