diff options
Diffstat (limited to 'src/video_core/buffer_cache')
-rw-r--r-- | src/video_core/buffer_cache/buffer_base.h | 19 | ||||
-rw-r--r-- | src/video_core/buffer_cache/buffer_cache.h | 292 |
2 files changed, 237 insertions, 74 deletions
diff --git a/src/video_core/buffer_cache/buffer_base.h b/src/video_core/buffer_cache/buffer_base.h index b121d36a3..c3318095c 100644 --- a/src/video_core/buffer_cache/buffer_base.h +++ b/src/video_core/buffer_cache/buffer_base.h @@ -226,19 +226,24 @@ public: /// Call 'func' for each CPU modified range and unmark those pages as CPU modified template <typename Func> void ForEachUploadRange(VAddr query_cpu_range, u64 size, Func&& func) { - ForEachModifiedRange<Type::CPU>(query_cpu_range, size, func); + ForEachModifiedRange<Type::CPU>(query_cpu_range, size, true, func); } /// Call 'func' for each GPU modified range and unmark those pages as GPU modified template <typename Func> - void ForEachDownloadRange(VAddr query_cpu_range, u64 size, Func&& func) { - ForEachModifiedRange<Type::GPU>(query_cpu_range, size, func); + void ForEachDownloadRange(VAddr query_cpu_range, u64 size, bool clear, Func&& func) { + ForEachModifiedRange<Type::GPU>(query_cpu_range, size, clear, func); + } + + template <typename Func> + void ForEachDownloadRangeAndClear(VAddr query_cpu_range, u64 size, Func&& func) { + ForEachModifiedRange<Type::GPU>(query_cpu_range, size, true, func); } /// Call 'func' for each GPU modified range and unmark those pages as GPU modified template <typename Func> void ForEachDownloadRange(Func&& func) { - ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), func); + ForEachModifiedRange<Type::GPU>(cpu_addr, SizeBytes(), true, func); } /// Mark buffer as picked @@ -415,7 +420,7 @@ private: * @param func Function to call for each turned off region */ template <Type type, typename Func> - void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) { + void ForEachModifiedRange(VAddr query_cpu_range, s64 size, bool clear, Func&& func) { static_assert(type != Type::Untracked); const s64 difference = query_cpu_range - cpu_addr; @@ -467,7 +472,9 @@ private: bits = (bits << left_offset) >> left_offset; const u64 current_word = state_words[word_index] & bits; - state_words[word_index] &= ~bits; + if (clear) { + state_words[word_index] &= ~bits; + } if constexpr (type == Type::CPU) { const u64 current_bits = untracked_words[word_index] & bits; diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index cad7f902d..2871682f6 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" @@ -77,6 +78,9 @@ 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 { @@ -148,18 +152,26 @@ public: /// 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(); + [[nodiscard]] bool DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount); + /// Return true when a CPU region is modified from the GPU [[nodiscard]] bool IsRegionGpuModified(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; private: @@ -190,6 +202,36 @@ 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); @@ -272,8 +314,6 @@ 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; @@ -327,9 +367,9 @@ private: 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; @@ -352,6 +392,7 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_, // 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 +463,68 @@ void BufferCache<P>::DownloadMemory(VAddr cpu_addr, u64 size) { } 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 = IsRegionGpuModified(*cpu_src_address, amount); + const bool dest_dirty = IsRegionGpuModified(*cpu_dest_address, amount); + if (!source_dirty && !dest_dirty) { + return false; + } + + const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount}; + uncommitted_ranges.subtract(subtract_interval); + for (auto& interval_set : committed_ranges) { + interval_set.subtract(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); + for (const IntervalType add_interval : tmp_intervals) { + common_ranges.add(add_interval); + } + + runtime.CopyBuffer(dest_buffer, src_buffer, copies); + if (source_dirty) { + 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> 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); @@ -547,29 +650,30 @@ 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 +681,43 @@ 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 +750,19 @@ 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 +782,25 @@ bool BufferCache<P>::IsRegionGpuModified(VAddr addr, size_t size) { } 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 +808,9 @@ 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, @@ -863,7 +1024,7 @@ 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; @@ -1010,16 +1171,16 @@ 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) { - return; - } - if (std::ranges::find(uncommitted_downloads, buffer_id) != uncommitted_downloads.end()) { - // Already inserted + if (!is_async && !is_accuracy_high) { return; } - uncommitted_downloads.push_back(buffer_id); + uncommitted_ranges.add(base_interval); } template <class P> @@ -1103,7 +1264,6 @@ 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 +1404,28 @@ 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}; + common_ranges.subtract(subtract_interval); }); if (total_size_bytes == 0) { return; @@ -1316,18 +1490,6 @@ 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}); @@ -1349,15 +1511,9 @@ 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; |