diff options
Diffstat (limited to '')
| -rw-r--r-- | src/video_core/buffer_cache/buffer_cache.h | 619 |
1 files changed, 519 insertions, 100 deletions
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index cad7f902d..24c858104 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -15,6 +15,7 @@ #include <vector> #include <boost/container/small_vector.hpp> +#include <boost/icl/interval_set.hpp> #include "common/common_types.h" #include "common/div_ceil.h" @@ -30,6 +31,7 @@ #include "video_core/engines/maxwell_3d.h" #include "video_core/memory_manager.h" #include "video_core/rasterizer_interface.h" +#include "video_core/surface.h" #include "video_core/texture_cache/slot_vector.h" #include "video_core/texture_cache/types.h" @@ -41,14 +43,19 @@ MICROPROFILE_DECLARE(GPU_DownloadMemory); using BufferId = SlotId; +using VideoCore::Surface::PixelFormat; +using namespace Common::Literals; + constexpr u32 NUM_VERTEX_BUFFERS = 32; constexpr u32 NUM_TRANSFORM_FEEDBACK_BUFFERS = 4; constexpr u32 NUM_GRAPHICS_UNIFORM_BUFFERS = 18; constexpr u32 NUM_COMPUTE_UNIFORM_BUFFERS = 8; constexpr u32 NUM_STORAGE_BUFFERS = 16; +constexpr u32 NUM_TEXTURE_BUFFERS = 16; constexpr u32 NUM_STAGES = 5; -using namespace Common::Literals; +using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>; +using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>; template <typename P> class BufferCache { @@ -66,6 +73,7 @@ class BufferCache { static constexpr bool NEEDS_BIND_UNIFORM_INDEX = P::NEEDS_BIND_UNIFORM_INDEX; static constexpr bool NEEDS_BIND_STORAGE_INDEX = P::NEEDS_BIND_STORAGE_INDEX; static constexpr bool USE_MEMORY_MAPS = P::USE_MEMORY_MAPS; + static constexpr bool SEPARATE_IMAGE_BUFFERS_BINDINGS = P::SEPARATE_IMAGE_BUFFER_BINDINGS; static constexpr BufferId NULL_BUFFER_ID{0}; @@ -77,6 +85,9 @@ class BufferCache { using Runtime = typename P::Runtime; using Buffer = typename P::Buffer; + using IntervalSet = boost::icl::interval_set<VAddr>; + using IntervalType = typename IntervalSet::interval_type; + struct Empty {}; struct OverlapResult { @@ -92,6 +103,10 @@ class BufferCache { BufferId buffer_id; }; + struct TextureBufferBinding : Binding { + PixelFormat format; + }; + static constexpr Binding NULL_BINDING{ .cpu_addr = 0, .size = 0, @@ -129,38 +144,63 @@ public: void BindHostComputeBuffers(); - void SetEnabledUniformBuffers(size_t stage, u32 enabled); + void SetUniformBuffersState(const std::array<u32, NUM_STAGES>& mask, + const UniformBufferSizes* sizes); - void SetEnabledComputeUniformBuffers(u32 enabled); + void SetComputeUniformBufferState(u32 mask, const ComputeUniformBufferSizes* sizes); void UnbindGraphicsStorageBuffers(size_t stage); void BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset, bool is_written); + void UnbindGraphicsTextureBuffers(size_t stage); + + void BindGraphicsTextureBuffer(size_t stage, size_t tbo_index, GPUVAddr gpu_addr, u32 size, + PixelFormat format, bool is_written, bool is_image); + void UnbindComputeStorageBuffers(); void BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset, bool is_written); + void UnbindComputeTextureBuffers(); + + void BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size, PixelFormat format, + bool is_written, bool is_image); + void FlushCachedWrites(); /// Return true when there are uncommitted buffers to be downloaded [[nodiscard]] bool HasUncommittedFlushes() const noexcept; + void AccumulateFlushes(); + /// Return true when the caller should wait for async downloads [[nodiscard]] bool ShouldWaitAsyncFlushes() const noexcept; /// Commit asynchronous downloads void CommitAsyncFlushes(); + void CommitAsyncFlushesHigh(); /// Pop asynchronous downloads void PopAsyncFlushes(); + bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount); + + bool DMAClear(GPUVAddr src_address, u64 amount, u32 value); + /// Return true when a CPU region is modified from the GPU [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size); + /// Return true when a region is registered on the cache + [[nodiscard]] bool IsRegionRegistered(VAddr addr, size_t size); + + /// Return true when a CPU region is modified from the CPU + [[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size); + std::mutex mutex; + Runtime& runtime; private: template <typename Func> @@ -190,6 +230,36 @@ private: } } + template <typename Func> + void ForEachWrittenRange(VAddr cpu_addr, u64 size, Func&& func) { + const VAddr start_address = cpu_addr; + const VAddr end_address = start_address + size; + const VAddr search_base = + static_cast<VAddr>(std::min<s64>(0LL, static_cast<s64>(start_address - size))); + const IntervalType search_interval{search_base, search_base + 1}; + auto it = common_ranges.lower_bound(search_interval); + if (it == common_ranges.end()) { + it = common_ranges.begin(); + } + for (; it != common_ranges.end(); it++) { + VAddr inter_addr_end = it->upper(); + VAddr inter_addr = it->lower(); + if (inter_addr >= end_address) { + break; + } + if (inter_addr_end <= start_address) { + continue; + } + if (inter_addr_end > end_address) { + inter_addr_end = end_address; + } + if (inter_addr < start_address) { + inter_addr = start_address; + } + func(inter_addr, inter_addr_end); + } + } + static bool IsRangeGranular(VAddr cpu_addr, size_t size) { return (cpu_addr & ~Core::Memory::PAGE_MASK) == ((cpu_addr + size) & ~Core::Memory::PAGE_MASK); @@ -207,12 +277,16 @@ private: void BindHostGraphicsStorageBuffers(size_t stage); + void BindHostGraphicsTextureBuffers(size_t stage); + void BindHostTransformFeedbackBuffers(); void BindHostComputeUniformBuffers(); void BindHostComputeStorageBuffers(); + void BindHostComputeTextureBuffers(); + void DoUpdateGraphicsBuffers(bool is_indexed); void DoUpdateComputeBuffers(); @@ -227,6 +301,8 @@ private: void UpdateStorageBuffers(size_t stage); + void UpdateTextureBuffers(size_t stage); + void UpdateTransformFeedbackBuffers(); void UpdateTransformFeedbackBuffer(u32 index); @@ -235,6 +311,8 @@ private: void UpdateComputeStorageBuffers(); + void UpdateComputeTextureBuffers(); + void MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 size); [[nodiscard]] BufferId FindBuffer(VAddr cpu_addr, u32 size); @@ -272,24 +350,26 @@ private: void DeleteBuffer(BufferId buffer_id); - void ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id); - void NotifyBufferDeletion(); [[nodiscard]] Binding StorageBufferBinding(GPUVAddr ssbo_addr) const; + [[nodiscard]] TextureBufferBinding GetTextureBufferBinding(GPUVAddr gpu_addr, u32 size, + PixelFormat format); + [[nodiscard]] std::span<const u8> ImmediateBufferWithData(VAddr cpu_addr, size_t size); [[nodiscard]] std::span<u8> ImmediateBuffer(size_t wanted_capacity); [[nodiscard]] bool HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept; + void ClearDownload(IntervalType subtract_interval); + VideoCore::RasterizerInterface& rasterizer; Tegra::Engines::Maxwell3D& maxwell3d; Tegra::Engines::KeplerCompute& kepler_compute; Tegra::MemoryManager& gpu_memory; Core::Memory::Memory& cpu_memory; - Runtime& runtime; SlotVector<Buffer> slot_buffers; DelayedDestructionRing<Buffer, 8> delayed_destruction_ring; @@ -300,20 +380,30 @@ private: std::array<Binding, NUM_VERTEX_BUFFERS> vertex_buffers; std::array<std::array<Binding, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES> uniform_buffers; std::array<std::array<Binding, NUM_STORAGE_BUFFERS>, NUM_STAGES> storage_buffers; + std::array<std::array<TextureBufferBinding, NUM_TEXTURE_BUFFERS>, NUM_STAGES> texture_buffers; std::array<Binding, NUM_TRANSFORM_FEEDBACK_BUFFERS> transform_feedback_buffers; std::array<Binding, NUM_COMPUTE_UNIFORM_BUFFERS> compute_uniform_buffers; std::array<Binding, NUM_STORAGE_BUFFERS> compute_storage_buffers; + std::array<TextureBufferBinding, NUM_TEXTURE_BUFFERS> compute_texture_buffers; + + std::array<u32, NUM_STAGES> enabled_uniform_buffer_masks{}; + u32 enabled_compute_uniform_buffer_mask = 0; - std::array<u32, NUM_STAGES> enabled_uniform_buffers{}; - u32 enabled_compute_uniform_buffers = 0; + const UniformBufferSizes* uniform_buffer_sizes{}; + const ComputeUniformBufferSizes* compute_uniform_buffer_sizes{}; std::array<u32, NUM_STAGES> enabled_storage_buffers{}; std::array<u32, NUM_STAGES> written_storage_buffers{}; u32 enabled_compute_storage_buffers = 0; u32 written_compute_storage_buffers = 0; - std::array<u32, NUM_STAGES> fast_bound_uniform_buffers{}; + std::array<u32, NUM_STAGES> enabled_texture_buffers{}; + std::array<u32, NUM_STAGES> written_texture_buffers{}; + std::array<u32, NUM_STAGES> image_texture_buffers{}; + u32 enabled_compute_texture_buffers = 0; + u32 written_compute_texture_buffers = 0; + u32 image_compute_texture_buffers = 0; std::array<u32, 16> uniform_cache_hits{}; std::array<u32, 16> uniform_cache_shots{}; @@ -324,12 +414,16 @@ private: std::conditional_t<HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS, std::array<u32, NUM_STAGES>, Empty> dirty_uniform_buffers{}; + std::conditional_t<IS_OPENGL, std::array<u32, NUM_STAGES>, Empty> fast_bound_uniform_buffers{}; + std::conditional_t<HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS, + std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>, Empty> + uniform_buffer_binding_sizes{}; std::vector<BufferId> cached_write_buffer_ids; - // TODO: This data structure is not optimal and it should be reworked - std::vector<BufferId> uncommitted_downloads; - std::deque<std::vector<BufferId>> committed_downloads; + IntervalSet uncommitted_ranges; + IntervalSet common_ranges; + std::deque<IntervalSet> committed_ranges; size_t immediate_buffer_capacity = 0; std::unique_ptr<u8[]> immediate_buffer_alloc; @@ -347,11 +441,12 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_, Tegra::Engines::KeplerCompute& kepler_compute_, Tegra::MemoryManager& gpu_memory_, Core::Memory::Memory& cpu_memory_, Runtime& runtime_) - : rasterizer{rasterizer_}, maxwell3d{maxwell3d_}, kepler_compute{kepler_compute_}, - gpu_memory{gpu_memory_}, cpu_memory{cpu_memory_}, runtime{runtime_} { + : runtime{runtime_}, rasterizer{rasterizer_}, maxwell3d{maxwell3d_}, + kepler_compute{kepler_compute_}, gpu_memory{gpu_memory_}, cpu_memory{cpu_memory_} { // Ensure the first slot is used for the null buffer void(slot_buffers.insert(runtime, NullBufferParams{})); deletion_iterator = slot_buffers.end(); + common_ranges.clear(); } template <class P> @@ -422,6 +517,97 @@ void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) { } template <class P> +void BufferCache<P>::ClearDownload(IntervalType subtract_interval) { + uncommitted_ranges.subtract(subtract_interval); + for (auto& interval_set : committed_ranges) { + interval_set.subtract(subtract_interval); + } +} + +template <class P> +bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) { + const std::optional<VAddr> cpu_src_address = gpu_memory.GpuToCpuAddress(src_address); + const std::optional<VAddr> cpu_dest_address = gpu_memory.GpuToCpuAddress(dest_address); + if (!cpu_src_address || !cpu_dest_address) { + return false; + } + const bool source_dirty = IsRegionRegistered(*cpu_src_address, amount); + const bool dest_dirty = IsRegionRegistered(*cpu_dest_address, amount); + if (!source_dirty && !dest_dirty) { + return false; + } + + const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount}; + ClearDownload(subtract_interval); + + BufferId buffer_a; + BufferId buffer_b; + do { + has_deleted_buffers = false; + buffer_a = FindBuffer(*cpu_src_address, static_cast<u32>(amount)); + buffer_b = FindBuffer(*cpu_dest_address, static_cast<u32>(amount)); + } while (has_deleted_buffers); + auto& src_buffer = slot_buffers[buffer_a]; + auto& dest_buffer = slot_buffers[buffer_b]; + SynchronizeBuffer(src_buffer, *cpu_src_address, static_cast<u32>(amount)); + SynchronizeBuffer(dest_buffer, *cpu_dest_address, static_cast<u32>(amount)); + std::array copies{BufferCopy{ + .src_offset = src_buffer.Offset(*cpu_src_address), + .dst_offset = dest_buffer.Offset(*cpu_dest_address), + .size = amount, + }}; + + boost::container::small_vector<IntervalType, 4> tmp_intervals; + auto mirror = [&](VAddr base_address, VAddr base_address_end) { + const u64 size = base_address_end - base_address; + const VAddr diff = base_address - *cpu_src_address; + const VAddr new_base_address = *cpu_dest_address + diff; + const IntervalType add_interval{new_base_address, new_base_address + size}; + uncommitted_ranges.add(add_interval); + tmp_intervals.push_back(add_interval); + }; + ForEachWrittenRange(*cpu_src_address, amount, mirror); + // This subtraction in this order is important for overlapping copies. + common_ranges.subtract(subtract_interval); + bool atleast_1_download = tmp_intervals.size() != 0; + for (const IntervalType add_interval : tmp_intervals) { + common_ranges.add(add_interval); + } + + runtime.CopyBuffer(dest_buffer, src_buffer, copies); + if (atleast_1_download) { + dest_buffer.MarkRegionAsGpuModified(*cpu_dest_address, amount); + } + std::vector<u8> tmp_buffer(amount); + cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount); + cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount); + return true; +} + +template <class P> +bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) { + const std::optional<VAddr> cpu_dst_address = gpu_memory.GpuToCpuAddress(dst_address); + if (!cpu_dst_address) { + return false; + } + const bool dest_dirty = IsRegionRegistered(*cpu_dst_address, amount); + if (!dest_dirty) { + return false; + } + + const size_t size = amount * sizeof(u32); + const IntervalType subtract_interval{*cpu_dst_address, *cpu_dst_address + size}; + ClearDownload(subtract_interval); + common_ranges.subtract(subtract_interval); + + const BufferId buffer = FindBuffer(*cpu_dst_address, static_cast<u32>(size)); + auto& dest_buffer = slot_buffers[buffer]; + const u32 offset = dest_buffer.Offset(*cpu_dst_address); + runtime.ClearBuffer(dest_buffer, offset, size, value); + return true; +} + +template <class P> void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size) { const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr); @@ -476,6 +662,7 @@ void BufferCache<P>::BindHostStageBuffers(size_t stage) { MICROPROFILE_SCOPE(GPU_BindUploadBuffers); BindHostGraphicsUniformBuffers(stage); BindHostGraphicsStorageBuffers(stage); + BindHostGraphicsTextureBuffers(stage); } template <class P> @@ -483,21 +670,30 @@ void BufferCache<P>::BindHostComputeBuffers() { MICROPROFILE_SCOPE(GPU_BindUploadBuffers); BindHostComputeUniformBuffers(); BindHostComputeStorageBuffers(); + BindHostComputeTextureBuffers(); } template <class P> -void BufferCache<P>::SetEnabledUniformBuffers(size_t stage, u32 enabled) { +void BufferCache<P>::SetUniformBuffersState(const std::array<u32, NUM_STAGES>& mask, + const UniformBufferSizes* sizes) { if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) { - if (enabled_uniform_buffers[stage] != enabled) { - dirty_uniform_buffers[stage] = ~u32{0}; + if (enabled_uniform_buffer_masks != mask) { + if constexpr (IS_OPENGL) { + fast_bound_uniform_buffers.fill(0); + } + dirty_uniform_buffers.fill(~u32{0}); + uniform_buffer_binding_sizes.fill({}); } } - enabled_uniform_buffers[stage] = enabled; + enabled_uniform_buffer_masks = mask; + uniform_buffer_sizes = sizes; } template <class P> -void BufferCache<P>::SetEnabledComputeUniformBuffers(u32 enabled) { - enabled_compute_uniform_buffers = enabled; +void BufferCache<P>::SetComputeUniformBufferState(u32 mask, + const ComputeUniformBufferSizes* sizes) { + enabled_compute_uniform_buffer_mask = mask; + compute_uniform_buffer_sizes = sizes; } template <class P> @@ -518,9 +714,29 @@ void BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, } template <class P> +void BufferCache<P>::UnbindGraphicsTextureBuffers(size_t stage) { + enabled_texture_buffers[stage] = 0; + written_texture_buffers[stage] = 0; + image_texture_buffers[stage] = 0; +} + +template <class P> +void BufferCache<P>::BindGraphicsTextureBuffer(size_t stage, size_t tbo_index, GPUVAddr gpu_addr, + u32 size, PixelFormat format, bool is_written, + bool is_image) { + enabled_texture_buffers[stage] |= 1U << tbo_index; + written_texture_buffers[stage] |= (is_written ? 1U : 0U) << tbo_index; + if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) { + image_texture_buffers[stage] |= (is_image ? 1U : 0U) << tbo_index; + } + texture_buffers[stage][tbo_index] = GetTextureBufferBinding(gpu_addr, size, format); +} + +template <class P> void BufferCache<P>::UnbindComputeStorageBuffers() { enabled_compute_storage_buffers = 0; written_compute_storage_buffers = 0; + image_compute_texture_buffers = 0; } template <class P> @@ -538,6 +754,24 @@ void BufferCache<P>::BindComputeStorageBuffer(size_t ssbo_index, u32 cbuf_index, } template <class P> +void BufferCache<P>::UnbindComputeTextureBuffers() { + enabled_compute_texture_buffers = 0; + written_compute_texture_buffers = 0; + image_compute_texture_buffers = 0; +} + +template <class P> +void BufferCache<P>::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size, + PixelFormat format, bool is_written, bool is_image) { + enabled_compute_texture_buffers |= 1U << tbo_index; + written_compute_texture_buffers |= (is_written ? 1U : 0U) << tbo_index; + if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) { + image_compute_texture_buffers |= (is_image ? 1U : 0U) << tbo_index; + } + compute_texture_buffers[tbo_index] = GetTextureBufferBinding(gpu_addr, size, format); +} + +template <class P> void BufferCache<P>::FlushCachedWrites() { for (const BufferId buffer_id : cached_write_buffer_ids) { slot_buffers[buffer_id].FlushCachedWrites(); @@ -547,29 +781,30 @@ void BufferCache<P>::FlushCachedWrites() { template <class P> bool BufferCache<P>::HasUncommittedFlushes() const noexcept { - return !uncommitted_downloads.empty(); + return !uncommitted_ranges.empty() || !committed_ranges.empty(); } template <class P> -bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept { - return !committed_downloads.empty() && !committed_downloads.front().empty(); +void BufferCache<P>::AccumulateFlushes() { + if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) { + uncommitted_ranges.clear(); + return; + } + if (uncommitted_ranges.empty()) { + return; + } + committed_ranges.emplace_back(std::move(uncommitted_ranges)); } template <class P> -void BufferCache<P>::CommitAsyncFlushes() { - // This is intentionally passing the value by copy - committed_downloads.push_front(uncommitted_downloads); - uncommitted_downloads.clear(); +bool BufferCache<P>::ShouldWaitAsyncFlushes() const noexcept { + return false; } template <class P> -void BufferCache<P>::PopAsyncFlushes() { - if (committed_downloads.empty()) { - return; - } - auto scope_exit_pop_download = detail::ScopeExit([this] { committed_downloads.pop_back(); }); - const std::span<const BufferId> download_ids = committed_downloads.back(); - if (download_ids.empty()) { +void BufferCache<P>::CommitAsyncFlushesHigh() { + AccumulateFlushes(); + if (committed_ranges.empty()) { return; } MICROPROFILE_SCOPE(GPU_DownloadMemory); @@ -577,20 +812,43 @@ void BufferCache<P>::PopAsyncFlushes() { boost::container::small_vector<std::pair<BufferCopy, BufferId>, 1> downloads; u64 total_size_bytes = 0; u64 largest_copy = 0; - for (const BufferId buffer_id : download_ids) { - slot_buffers[buffer_id].ForEachDownloadRange([&](u64 range_offset, u64 range_size) { - downloads.push_back({ - BufferCopy{ - .src_offset = range_offset, - .dst_offset = total_size_bytes, - .size = range_size, - }, - buffer_id, + for (const IntervalSet& intervals : committed_ranges) { + for (auto& interval : intervals) { + const std::size_t size = interval.upper() - interval.lower(); + const VAddr cpu_addr = interval.lower(); + ForEachBufferInRange(cpu_addr, size, [&](BufferId buffer_id, Buffer& buffer) { + boost::container::small_vector<BufferCopy, 1> copies; + buffer.ForEachDownloadRangeAndClear( + cpu_addr, size, [&](u64 range_offset, u64 range_size) { + const VAddr buffer_addr = buffer.CpuAddr(); + const auto add_download = [&](VAddr start, VAddr end) { + const u64 new_offset = start - buffer_addr; + const u64 new_size = end - start; + downloads.push_back({ + BufferCopy{ + .src_offset = new_offset, + .dst_offset = total_size_bytes, + .size = new_size, + }, + buffer_id, + }); + // Align up to avoid cache conflicts + constexpr u64 align = 256ULL; + constexpr u64 mask = ~(align - 1ULL); + total_size_bytes += (new_size + align - 1) & mask; + largest_copy = std::max(largest_copy, new_size); + }; + + const VAddr start_address = buffer_addr + range_offset; + const VAddr end_address = start_address + range_size; + ForEachWrittenRange(start_address, range_size, add_download); + const IntervalType subtract_interval{start_address, end_address}; + common_ranges.subtract(subtract_interval); + }); }); - total_size_bytes += range_size; - largest_copy = std::max(largest_copy, range_size); - }); + } } + committed_ranges.clear(); if (downloads.empty()) { return; } @@ -623,6 +881,19 @@ void BufferCache<P>::PopAsyncFlushes() { } template <class P> +void BufferCache<P>::CommitAsyncFlushes() { + if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) { + CommitAsyncFlushesHigh(); + } else { + uncommitted_ranges.clear(); + committed_ranges.clear(); + } +} + +template <class P> +void BufferCache<P>::PopAsyncFlushes() {} + +template <class P> bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) { const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE); for (u64 page = addr >> PAGE_BITS; page < page_end;) { @@ -642,6 +913,46 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) { } template <class P> +bool BufferCache<P>::IsRegionRegistered(VAddr addr, size_t size) { + const VAddr end_addr = addr + size; + const u64 page_end = Common::DivCeil(end_addr, PAGE_SIZE); + for (u64 page = addr >> PAGE_BITS; page < page_end;) { + const BufferId buffer_id = page_table[page]; + if (!buffer_id) { + ++page; + continue; + } + Buffer& buffer = slot_buffers[buffer_id]; + const VAddr buf_start_addr = buffer.CpuAddr(); + const VAddr buf_end_addr = buf_start_addr + buffer.SizeBytes(); + if (buf_start_addr < end_addr && addr < buf_end_addr) { + return true; + } + page = Common::DivCeil(end_addr, PAGE_SIZE); + } + return false; +} + +template <class P> +bool BufferCache<P>::IsRegionCpuModified(VAddr addr, size_t size) { + const u64 page_end = Common::DivCeil(addr + size, PAGE_SIZE); + for (u64 page = addr >> PAGE_BITS; page < page_end;) { + const BufferId image_id = page_table[page]; + if (!image_id) { + ++page; + continue; + } + Buffer& buffer = slot_buffers[image_id]; + if (buffer.IsRegionCpuModified(addr, size)) { + return true; + } + const VAddr end_addr = buffer.CpuAddr() + buffer.SizeBytes(); + page = Common::DivCeil(end_addr, PAGE_SIZE); + } + return false; +} + +template <class P> void BufferCache<P>::BindHostIndexBuffer() { Buffer& buffer = slot_buffers[index_buffer.buffer_id]; TouchBuffer(buffer); @@ -649,7 +960,9 @@ void BufferCache<P>::BindHostIndexBuffer() { const u32 size = index_buffer.size; SynchronizeBuffer(buffer, index_buffer.cpu_addr, size); if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) { - runtime.BindIndexBuffer(buffer, offset, size); + const u32 new_offset = offset + maxwell3d.regs.index_array.first * + maxwell3d.regs.index_array.FormatSizeInBytes(); + runtime.BindIndexBuffer(buffer, new_offset, size); } else { runtime.BindIndexBuffer(maxwell3d.regs.draw.topology, maxwell3d.regs.index_array.format, maxwell3d.regs.index_array.first, maxwell3d.regs.index_array.count, @@ -683,7 +996,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffers(size_t stage) { dirty = std::exchange(dirty_uniform_buffers[stage], 0); } u32 binding_index = 0; - ForEachEnabledBit(enabled_uniform_buffers[stage], [&](u32 index) { + ForEachEnabledBit(enabled_uniform_buffer_masks[stage], [&](u32 index) { const bool needs_bind = ((dirty >> index) & 1) != 0; BindHostGraphicsUniformBuffer(stage, index, binding_index, needs_bind); if constexpr (NEEDS_BIND_UNIFORM_INDEX) { @@ -697,7 +1010,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 bool needs_bind) { const Binding& binding = uniform_buffers[stage][index]; const VAddr cpu_addr = binding.cpu_addr; - const u32 size = binding.size; + const u32 size = std::min(binding.size, (*uniform_buffer_sizes)[stage][index]); Buffer& buffer = slot_buffers[binding.buffer_id]; TouchBuffer(buffer); const bool use_fast_buffer = binding.buffer_id != NULL_BUFFER_ID && @@ -707,8 +1020,13 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 if constexpr (IS_OPENGL) { if (runtime.HasFastBufferSubData()) { // Fast path for Nvidia - if (!HasFastUniformBufferBound(stage, binding_index)) { + const bool should_fast_bind = + !HasFastUniformBufferBound(stage, binding_index) || + uniform_buffer_binding_sizes[stage][binding_index] != size; + if (should_fast_bind) { // We only have to bind when the currently bound buffer is not the fast version + fast_bound_uniform_buffers[stage] |= 1U << binding_index; + uniform_buffer_binding_sizes[stage][binding_index] = size; runtime.BindFastUniformBuffer(stage, binding_index, size); } const auto span = ImmediateBufferWithData(cpu_addr, size); @@ -716,8 +1034,10 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 return; } } - fast_bound_uniform_buffers[stage] |= 1U << binding_index; - + if constexpr (IS_OPENGL) { + fast_bound_uniform_buffers[stage] |= 1U << binding_index; + uniform_buffer_binding_sizes[stage][binding_index] = size; + } // Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size); cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size); @@ -730,14 +1050,27 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32 } ++uniform_cache_shots[0]; - if (!needs_bind && !HasFastUniformBufferBound(stage, binding_index)) { - // Skip binding if it's not needed and if the bound buffer is not the fast version - // This exists to avoid instances where the fast buffer is bound and a GPU write happens + // Skip binding if it's not needed and if the bound buffer is not the fast version + // This exists to avoid instances where the fast buffer is bound and a GPU write happens + needs_bind |= HasFastUniformBufferBound(stage, binding_index); + if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) { + needs_bind |= uniform_buffer_binding_sizes[stage][binding_index] != size; + } + if (!needs_bind) { return; } - fast_bound_uniform_buffers[stage] &= ~(1U << binding_index); - const u32 offset = buffer.Offset(cpu_addr); + if constexpr (IS_OPENGL) { + // Fast buffer will be unbound + fast_bound_uniform_buffers[stage] &= ~(1U << binding_index); + + // Mark the index as dirty if offset doesn't match + const bool is_copy_bind = offset != 0 && !runtime.SupportsNonZeroUniformOffset(); + dirty_uniform_buffers[stage] |= (is_copy_bind ? 1U : 0U) << index; + } + if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) { + uniform_buffer_binding_sizes[stage][binding_index] = size; + } if constexpr (NEEDS_BIND_UNIFORM_INDEX) { runtime.BindUniformBuffer(stage, binding_index, buffer, offset, size); } else { @@ -767,6 +1100,28 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) { } template <class P> +void BufferCache<P>::BindHostGraphicsTextureBuffers(size_t stage) { + ForEachEnabledBit(enabled_texture_buffers[stage], [&](u32 index) { + const TextureBufferBinding& binding = texture_buffers[stage][index]; + Buffer& buffer = slot_buffers[binding.buffer_id]; + const u32 size = binding.size; + SynchronizeBuffer(buffer, binding.cpu_addr, size); + + const u32 offset = buffer.Offset(binding.cpu_addr); + const PixelFormat format = binding.format; + if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) { + if (((image_texture_buffers[stage] >> index) & 1) != 0) { + runtime.BindImageBuffer(buffer, offset, size, format); + } else { + runtime.BindTextureBuffer(buffer, offset, size, format); + } + } else { + runtime.BindTextureBuffer(buffer, offset, size, format); + } + }); +} + +template <class P> void BufferCache<P>::BindHostTransformFeedbackBuffers() { if (maxwell3d.regs.tfb_enabled == 0) { return; @@ -788,13 +1143,14 @@ void BufferCache<P>::BindHostComputeUniformBuffers() { if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) { // Mark all uniform buffers as dirty dirty_uniform_buffers.fill(~u32{0}); + fast_bound_uniform_buffers.fill(0); } u32 binding_index = 0; - ForEachEnabledBit(enabled_compute_uniform_buffers, [&](u32 index) { + ForEachEnabledBit(enabled_compute_uniform_buffer_mask, [&](u32 index) { const Binding& binding = compute_uniform_buffers[index]; Buffer& buffer = slot_buffers[binding.buffer_id]; TouchBuffer(buffer); - const u32 size = binding.size; + const u32 size = std::min(binding.size, (*compute_uniform_buffer_sizes)[index]); SynchronizeBuffer(buffer, binding.cpu_addr, size); const u32 offset = buffer.Offset(binding.cpu_addr); @@ -829,6 +1185,28 @@ void BufferCache<P>::BindHostComputeStorageBuffers() { } template <class P> +void BufferCache<P>::BindHostComputeTextureBuffers() { + ForEachEnabledBit(enabled_compute_texture_buffers, [&](u32 index) { + const TextureBufferBinding& binding = compute_texture_buffers[index]; + Buffer& buffer = slot_buffers[binding.buffer_id]; + const u32 size = binding.size; + SynchronizeBuffer(buffer, binding.cpu_addr, size); + + const u32 offset = buffer.Offset(binding.cpu_addr); + const PixelFormat format = binding.format; + if constexpr (SEPARATE_IMAGE_BUFFERS_BINDINGS) { + if (((image_compute_texture_buffers >> index) & 1) != 0) { + runtime.BindImageBuffer(buffer, offset, size, format); + } else { + runtime.BindTextureBuffer(buffer, offset, size, format); + } + } else { + runtime.BindTextureBuffer(buffer, offset, size, format); + } + }); +} + +template <class P> void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) { if (is_indexed) { UpdateIndexBuffer(); @@ -838,6 +1216,7 @@ void BufferCache<P>::DoUpdateGraphicsBuffers(bool is_indexed) { for (size_t stage = 0; stage < NUM_STAGES; ++stage) { UpdateUniformBuffers(stage); UpdateStorageBuffers(stage); + UpdateTextureBuffers(stage); } } @@ -845,6 +1224,7 @@ template <class P> void BufferCache<P>::DoUpdateComputeBuffers() { UpdateComputeUniformBuffers(); UpdateComputeStorageBuffers(); + UpdateComputeTextureBuffers(); } template <class P> @@ -863,7 +1243,7 @@ void BufferCache<P>::UpdateIndexBuffer() { const GPUVAddr gpu_addr_end = index_array.EndAddress(); const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr_begin); const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin); - const u32 draw_size = index_array.count * index_array.FormatSizeInBytes(); + const u32 draw_size = (index_array.count + index_array.first) * index_array.FormatSizeInBytes(); const u32 size = std::min(address_size, draw_size); if (size == 0 || !cpu_addr) { index_buffer = NULL_BINDING; @@ -914,7 +1294,7 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) { template <class P> void BufferCache<P>::UpdateUniformBuffers(size_t stage) { - ForEachEnabledBit(enabled_uniform_buffers[stage], [&](u32 index) { + ForEachEnabledBit(enabled_uniform_buffer_masks[stage], [&](u32 index) { Binding& binding = uniform_buffers[stage][index]; if (binding.buffer_id) { // Already updated @@ -945,6 +1325,18 @@ void BufferCache<P>::UpdateStorageBuffers(size_t stage) { } template <class P> +void BufferCache<P>::UpdateTextureBuffers(size_t stage) { + ForEachEnabledBit(enabled_texture_buffers[stage], [&](u32 index) { + Binding& binding = texture_buffers[stage][index]; + binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size); + // Mark buffer as written if needed + if (((written_texture_buffers[stage] >> index) & 1) != 0) { + MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, binding.size); + } + }); +} + +template <class P> void BufferCache<P>::UpdateTransformFeedbackBuffers() { if (maxwell3d.regs.tfb_enabled == 0) { return; @@ -975,7 +1367,7 @@ void BufferCache<P>::UpdateTransformFeedbackBuffer(u32 index) { template <class P> void BufferCache<P>::UpdateComputeUniformBuffers() { - ForEachEnabledBit(enabled_compute_uniform_buffers, [&](u32 index) { + ForEachEnabledBit(enabled_compute_uniform_buffer_mask, [&](u32 index) { Binding& binding = compute_uniform_buffers[index]; binding = NULL_BINDING; const auto& launch_desc = kepler_compute.launch_description; @@ -996,11 +1388,22 @@ void BufferCache<P>::UpdateComputeStorageBuffers() { ForEachEnabledBit(enabled_compute_storage_buffers, [&](u32 index) { // Resolve buffer Binding& binding = compute_storage_buffers[index]; - const BufferId buffer_id = FindBuffer(binding.cpu_addr, binding.size); - binding.buffer_id = buffer_id; + binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size); // Mark as written if needed if (((written_compute_storage_buffers >> index) & 1) != 0) { - MarkWrittenBuffer(buffer_id, binding.cpu_addr, binding.size); + MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, binding.size); + } + }); +} + +template <class P> +void BufferCache<P>::UpdateComputeTextureBuffers() { + ForEachEnabledBit(enabled_compute_texture_buffers, [&](u32 index) { + Binding& binding = compute_texture_buffers[index]; + binding.buffer_id = FindBuffer(binding.cpu_addr, binding.size); + // Mark as written if needed + if (((written_compute_texture_buffers >> index) & 1) != 0) { + MarkWrittenBuffer(binding.buffer_id, binding.cpu_addr, binding.size); } }); } @@ -1010,16 +1413,16 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s Buffer& buffer = slot_buffers[buffer_id]; buffer.MarkRegionAsGpuModified(cpu_addr, size); - const bool is_accuracy_high = Settings::IsGPULevelHigh(); + const IntervalType base_interval{cpu_addr, cpu_addr + size}; + common_ranges.add(base_interval); + + const bool is_accuracy_high = + Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High; const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue(); - if (!is_accuracy_high || !is_async) { + if (!is_async && !is_accuracy_high) { return; } - if (std::ranges::find(uncommitted_downloads, buffer_id) != uncommitted_downloads.end()) { - // Already inserted - return; - } - uncommitted_downloads.push_back(buffer_id); + uncommitted_ranges.add(base_interval); } template <class P> @@ -1103,7 +1506,6 @@ void BufferCache<P>::JoinOverlap(BufferId new_buffer_id, BufferId overlap_id, if (!copies.empty()) { runtime.CopyBuffer(slot_buffers[new_buffer_id], overlap, copies); } - ReplaceBufferDownloads(overlap_id, new_buffer_id); DeleteBuffer(overlap_id); } @@ -1244,14 +1646,29 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si boost::container::small_vector<BufferCopy, 1> copies; u64 total_size_bytes = 0; u64 largest_copy = 0; - buffer.ForEachDownloadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) { - copies.push_back(BufferCopy{ - .src_offset = range_offset, - .dst_offset = total_size_bytes, - .size = range_size, - }); - total_size_bytes += range_size; - largest_copy = std::max(largest_copy, range_size); + buffer.ForEachDownloadRangeAndClear(cpu_addr, size, [&](u64 range_offset, u64 range_size) { + const VAddr buffer_addr = buffer.CpuAddr(); + const auto add_download = [&](VAddr start, VAddr end) { + const u64 new_offset = start - buffer_addr; + const u64 new_size = end - start; + copies.push_back(BufferCopy{ + .src_offset = new_offset, + .dst_offset = total_size_bytes, + .size = new_size, + }); + // Align up to avoid cache conflicts + constexpr u64 align = 256ULL; + constexpr u64 mask = ~(align - 1ULL); + total_size_bytes += (new_size + align - 1) & mask; + largest_copy = std::max(largest_copy, new_size); + }; + + const VAddr start_address = buffer_addr + range_offset; + const VAddr end_address = start_address + range_size; + ForEachWrittenRange(start_address, range_size, add_download); + const IntervalType subtract_interval{start_address, end_address}; + ClearDownload(subtract_interval); + common_ranges.subtract(subtract_interval); }); if (total_size_bytes == 0) { return; @@ -1316,21 +1733,10 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id) { } template <class P> -void BufferCache<P>::ReplaceBufferDownloads(BufferId old_buffer_id, BufferId new_buffer_id) { - const auto replace = [old_buffer_id, new_buffer_id](std::vector<BufferId>& buffers) { - std::ranges::replace(buffers, old_buffer_id, new_buffer_id); - if (auto it = std::ranges::find(buffers, new_buffer_id); it != buffers.end()) { - buffers.erase(std::remove(it + 1, buffers.end(), new_buffer_id), buffers.end()); - } - }; - replace(uncommitted_downloads); - std::ranges::for_each(committed_downloads, replace); -} - -template <class P> void BufferCache<P>::NotifyBufferDeletion() { if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) { dirty_uniform_buffers.fill(~u32{0}); + uniform_buffer_binding_sizes.fill({}); } auto& flags = maxwell3d.dirty.flags; flags[Dirty::IndexBuffer] = true; @@ -1349,21 +1755,34 @@ typename BufferCache<P>::Binding BufferCache<P>::StorageBufferBinding(GPUVAddr s if (!cpu_addr || size == 0) { return NULL_BINDING; } - // HACK(Rodrigo): This is the number of bytes bound in host beyond the guest API's range. - // It exists due to some games like Astral Chain operate out of bounds. - // Binding the whole map range would be technically correct, but games have large maps that make - // this approach unaffordable for now. - static constexpr u32 arbitrary_extra_bytes = 0xc000; - const u32 bytes_to_map_end = static_cast<u32>(gpu_memory.BytesToMapEnd(gpu_addr)); const Binding binding{ .cpu_addr = *cpu_addr, - .size = std::min(size + arbitrary_extra_bytes, bytes_to_map_end), + .size = size, .buffer_id = BufferId{}, }; return binding; } template <class P> +typename BufferCache<P>::TextureBufferBinding BufferCache<P>::GetTextureBufferBinding( + GPUVAddr gpu_addr, u32 size, PixelFormat format) { + const std::optional<VAddr> cpu_addr = gpu_memory.GpuToCpuAddress(gpu_addr); + TextureBufferBinding binding; + if (!cpu_addr || size == 0) { + binding.cpu_addr = 0; + binding.size = 0; + binding.buffer_id = NULL_BUFFER_ID; + binding.format = PixelFormat::Invalid; + } else { + binding.cpu_addr = *cpu_addr; + binding.size = size; + binding.buffer_id = BufferId{}; + binding.format = format; + } + return binding; +} + +template <class P> std::span<const u8> BufferCache<P>::ImmediateBufferWithData(VAddr cpu_addr, size_t size) { u8* const base_pointer = cpu_memory.GetPointer(cpu_addr); if (IsRangeGranular(cpu_addr, size) || |