From a4bfae1b551651e50f1484d6d488d3f598d274cf Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Wed, 30 Dec 2020 18:37:31 -0300 Subject: buffer_cache/buffer_base: Add a range tracking buffer container It keeps track of the modified CPU and GPU ranges on a CPU page granularity, notifying the given rasterizer about state changes in the tracking behavior of the buffer. Use a small vector optimization to store buffers smaller than 256 KiB locally instead of using free store memory allocations. --- src/video_core/CMakeLists.txt | 1 + src/video_core/buffer_cache/buffer_base.h | 495 ++++++++++++++++++++++++++++++ 2 files changed, 496 insertions(+) create mode 100644 src/video_core/buffer_cache/buffer_base.h (limited to 'src') diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt index f7b9d7f86..c3d0f4c31 100644 --- a/src/video_core/CMakeLists.txt +++ b/src/video_core/CMakeLists.txt @@ -1,6 +1,7 @@ add_subdirectory(host_shaders) add_library(video_core STATIC + buffer_cache/buffer_base.h buffer_cache/buffer_block.h buffer_cache/buffer_cache.h buffer_cache/map_interval.cpp diff --git a/src/video_core/buffer_cache/buffer_base.h b/src/video_core/buffer_cache/buffer_base.h new file mode 100644 index 000000000..fd740c2c1 --- /dev/null +++ b/src/video_core/buffer_cache/buffer_base.h @@ -0,0 +1,495 @@ +// Copyright 2020 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include +#include +#include +#include + +#include "common/alignment.h" +#include "common/common_funcs.h" +#include "common/common_types.h" +#include "common/div_ceil.h" +#include "core/memory.h" + +namespace VideoCommon { + +enum class BufferFlagBits { + Picked = 1 << 0, +}; +DECLARE_ENUM_FLAG_OPERATORS(BufferFlagBits) + +/// Tag for creating null buffers with no storage or size +struct NullBufferParams {}; + +/** + * Range tracking buffer container. + * + * It keeps track of the modified CPU and GPU ranges on a CPU page granularity, notifying the given + * rasterizer about state changes in the tracking behavior of the buffer. + * + * The buffer size and address is forcefully aligned to CPU page boundaries. + */ +template +class BufferBase { + static constexpr u64 PAGES_PER_WORD = 64; + static constexpr u64 BYTES_PER_PAGE = Core::Memory::PAGE_SIZE; + static constexpr u64 BYTES_PER_WORD = PAGES_PER_WORD * BYTES_PER_PAGE; + + /// Vector tracking modified pages tightly packed with small vector optimization + union WrittenWords { + /// Returns the pointer to the words state + [[nodiscard]] const u64* Pointer(bool is_short) const noexcept { + return is_short ? &stack : heap; + } + + /// Returns the pointer to the words state + [[nodiscard]] u64* Pointer(bool is_short) noexcept { + return is_short ? &stack : heap; + } + + u64 stack = 0; ///< Small buffers storage + u64* heap; ///< Not-small buffers pointer to the storage + }; + + struct GpuCpuWords { + explicit GpuCpuWords() = default; + explicit GpuCpuWords(u64 size_bytes_) : size_bytes{size_bytes_} { + if (IsShort()) { + cpu.stack = ~u64{0}; + gpu.stack = 0; + } else { + // Share allocation between CPU and GPU pages and set their default values + const size_t num_words = NumWords(); + u64* const alloc = new u64[num_words * 2]; + cpu.heap = alloc; + gpu.heap = alloc + num_words; + std::fill_n(cpu.heap, num_words, ~u64{0}); + std::fill_n(gpu.heap, num_words, 0); + } + // Clean up tailing bits + const u64 last_local_page = + Common::DivCeil(size_bytes % BYTES_PER_WORD, BYTES_PER_PAGE); + const u64 shift = (PAGES_PER_WORD - last_local_page) % PAGES_PER_WORD; + u64& last_word = cpu.Pointer(IsShort())[NumWords() - 1]; + last_word = (last_word << shift) >> shift; + } + + ~GpuCpuWords() { + Release(); + } + + GpuCpuWords& operator=(GpuCpuWords&& rhs) noexcept { + Release(); + size_bytes = rhs.size_bytes; + cpu = rhs.cpu; + gpu = rhs.gpu; + rhs.cpu.heap = nullptr; + return *this; + } + + GpuCpuWords(GpuCpuWords&& rhs) noexcept + : size_bytes{rhs.size_bytes}, cpu{rhs.cpu}, gpu{rhs.gpu} { + rhs.cpu.heap = nullptr; + } + + GpuCpuWords& operator=(const GpuCpuWords&) = delete; + GpuCpuWords(const GpuCpuWords&) = delete; + + /// Returns true when the buffer fits in the small vector optimization + [[nodiscard]] bool IsShort() const noexcept { + return size_bytes <= BYTES_PER_WORD; + } + + /// Returns the number of words of the buffer + [[nodiscard]] size_t NumWords() const noexcept { + return Common::DivCeil(size_bytes, BYTES_PER_WORD); + } + + /// Release buffer resources + void Release() { + if (!IsShort()) { + // CPU written words is the base for the heap allocation + delete[] cpu.heap; + } + } + + u64 size_bytes = 0; + WrittenWords cpu; + WrittenWords gpu; + }; + +public: + explicit BufferBase(RasterizerInterface& rasterizer_, VAddr cpu_addr_, u64 size_bytes) + : rasterizer{&rasterizer_}, cpu_addr{Common::AlignDown(cpu_addr_, BYTES_PER_PAGE)}, + words(Common::AlignUp(size_bytes + (cpu_addr_ - cpu_addr), BYTES_PER_PAGE)) {} + + explicit BufferBase(NullBufferParams) {} + + BufferBase& operator=(const BufferBase&) = delete; + BufferBase(const BufferBase&) = delete; + + /// Returns the inclusive CPU modified range in a begin end pair + [[nodiscard]] std::pair ModifiedCpuRegion(VAddr query_cpu_addr, + u64 query_size) const noexcept { + const u64 offset = query_cpu_addr - cpu_addr; + return ModifiedRegion(offset, query_size); + } + + /// Returns the inclusive GPU modified range in a begin end pair + [[nodiscard]] std::pair ModifiedGpuRegion(VAddr query_cpu_addr, + u64 query_size) const noexcept { + const u64 offset = query_cpu_addr - cpu_addr; + return ModifiedRegion(offset, query_size); + } + + /// Returns true if a region has been modified from the CPU + [[nodiscard]] bool IsRegionCpuModified(VAddr query_cpu_addr, u64 query_size) const noexcept { + const u64 offset = query_cpu_addr - cpu_addr; + return IsRegionModified(offset, query_size); + } + + /// Returns true if a region has been modified from the GPU + [[nodiscard]] bool IsRegionGpuModified(VAddr query_cpu_addr, u64 query_size) const noexcept { + const u64 offset = query_cpu_addr - cpu_addr; + return IsRegionModified(offset, query_size); + } + + /// Mark region as CPU modified, notifying the rasterizer about this change + void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 size) { + ChangeRegionState(words.cpu, dirty_cpu_addr, size); + } + + /// Unmark region as CPU modified, notifying the rasterizer about this change + void UnmarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 size) { + ChangeRegionState(words.cpu, dirty_cpu_addr, size); + } + + /// Mark region as modified from the host GPU + void MarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 size) noexcept { + ChangeRegionState(words.gpu, dirty_cpu_addr, size); + } + + /// Unmark region as modified from the host GPU + void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 size) noexcept { + ChangeRegionState(words.gpu, dirty_cpu_addr, size); + } + + /// Call 'func' for each CPU modified range and unmark those pages as CPU modified + template + void ForEachUploadRange(VAddr query_cpu_range, u64 size, Func&& func) { + ForEachModifiedRange(query_cpu_range, size, func); + } + + /// Call 'func' for each GPU modified range and unmark those pages as GPU modified + template + void ForEachDownloadRange(VAddr query_cpu_range, u64 size, Func&& func) { + ForEachModifiedRange(query_cpu_range, size, func); + } + + /// Call 'func' for each GPU modified range and unmark those pages as GPU modified + template + void ForEachDownloadRange(Func&& func) { + ForEachModifiedRange(cpu_addr, SizeBytes(), func); + } + + /// Mark buffer as picked + void Pick() noexcept { + flags |= BufferFlagBits::Picked; + } + + /// Unmark buffer as picked + void Unpick() noexcept { + flags &= ~BufferFlagBits::Picked; + } + + /// Returns true when vaddr -> vaddr+size is fully contained in the buffer + [[nodiscard]] bool IsInBounds(VAddr addr, u64 size) const noexcept { + return addr >= cpu_addr && addr + size <= cpu_addr + SizeBytes(); + } + + /// Returns true if the buffer has been marked as picked + [[nodiscard]] bool IsPicked() const noexcept { + return True(flags & BufferFlagBits::Picked); + } + + /// Returns the base CPU address of the buffer + [[nodiscard]] VAddr CpuAddr() const noexcept { + return cpu_addr; + } + + /// Returns the offset relative to the given CPU address + /// @pre IsInBounds returns true + [[nodiscard]] u32 Offset(VAddr other_cpu_addr) const noexcept { + return static_cast(other_cpu_addr - cpu_addr); + } + + /// Returns the size in bytes of the buffer + [[nodiscard]] u64 SizeBytes() const noexcept { + return words.size_bytes; + } + +private: + /** + * Change the state of a range of pages + * + * @param written_words Pages to be marked or unmarked as modified + * @param dirty_addr Base address to mark or unmark as modified + * @param size Size in bytes to mark or unmark as modified + * + * @tparam enable True when the bits will be set to one, false for zero + * @tparam notify_rasterizer True when the rasterizer has to be notified about the changes + */ + template + void ChangeRegionState(WrittenWords& written_words, u64 dirty_addr, + s64 size) noexcept(!notify_rasterizer) { + const s64 difference = dirty_addr - cpu_addr; + const u64 offset = std::max(difference, 0); + size += std::min(difference, 0); + if (offset >= SizeBytes() || size < 0) { + return; + } + u64* const state_words = written_words.Pointer(IsShort()); + const u64 offset_end = std::min(offset + size, SizeBytes()); + const u64 begin_page_index = offset / BYTES_PER_PAGE; + const u64 begin_word_index = begin_page_index / PAGES_PER_WORD; + const u64 end_page_index = Common::DivCeil(offset_end, BYTES_PER_PAGE); + const u64 end_word_index = Common::DivCeil(end_page_index, PAGES_PER_WORD); + u64 page_index = begin_page_index % PAGES_PER_WORD; + u64 word_index = begin_word_index; + while (word_index < end_word_index) { + const u64 next_word_first_page = (word_index + 1) * PAGES_PER_WORD; + const u64 left_offset = + std::min(next_word_first_page - end_page_index, PAGES_PER_WORD) % PAGES_PER_WORD; + const u64 right_offset = page_index; + u64 bits = ~u64{0}; + bits = (bits >> right_offset) << right_offset; + bits = (bits << left_offset) >> left_offset; + if constexpr (notify_rasterizer) { + NotifyRasterizer(word_index, state_words[word_index], bits); + } + if constexpr (enable) { + state_words[word_index] |= bits; + } else { + state_words[word_index] &= ~bits; + } + page_index = 0; + ++word_index; + } + } + + /** + * Notify rasterizer about changes in the CPU tracking state of a word in the buffer + * + * @param word_index Index to the word to notify to the rasterizer + * @param current_bits Current state of the word + * @param new_bits New state of the word + * + * @tparam add_to_rasterizer True when the rasterizer should start tracking the new pages + */ + template + void NotifyRasterizer(u64 word_index, u64 current_bits, u64 new_bits) { + u64 changed_bits = (add_to_rasterizer ? current_bits : ~current_bits) & new_bits; + VAddr addr = cpu_addr + word_index * BYTES_PER_WORD; + while (changed_bits != 0) { + const int empty_bits = std::countr_zero(changed_bits); + addr += empty_bits * BYTES_PER_PAGE; + changed_bits >>= empty_bits; + + const u32 continuous_bits = std::countr_one(changed_bits); + const u64 size = continuous_bits * BYTES_PER_PAGE; + const VAddr begin_addr = addr; + addr += size; + changed_bits = continuous_bits < PAGES_PER_WORD ? (changed_bits >> continuous_bits) : 0; + rasterizer->UpdatePagesCachedCount(begin_addr, size, add_to_rasterizer ? 1 : -1); + } + } + + /** + * Loop over each page in the given range, turn off those bits and notify the rasterizer if + * needed. Call the given function on each turned off range. + * + * @param query_cpu_range Base CPU address to loop over + * @param size Size in bytes of the CPU range to loop over + * @param func Function to call for each turned off region + * + * @tparam gpu True for host GPU pages, false for CPU pages + * @tparam notify_rasterizer True when the rasterizer should be notified about state changes + */ + template + void ForEachModifiedRange(VAddr query_cpu_range, s64 size, Func&& func) { + const s64 difference = query_cpu_range - cpu_addr; + const u64 query_begin = std::max(difference, 0); + size += std::min(difference, 0); + if (query_begin >= SizeBytes() || size < 0) { + return; + } + const u64* const cpu_words = words.cpu.Pointer(IsShort()); + const u64 query_end = query_begin + std::min(static_cast(size), SizeBytes()); + u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort()); + u64* const words_begin = state_words + query_begin / BYTES_PER_WORD; + u64* const words_end = state_words + Common::DivCeil(query_end, BYTES_PER_WORD); + + const auto modified = [](u64 word) { return word != 0; }; + const auto first_modified_word = std::find_if(words_begin, words_end, modified); + if (first_modified_word == words_end) { + // Exit early when the buffer is not modified + return; + } + const auto last_modified_word = std::find_if_not(first_modified_word, words_end, modified); + + const u64 word_index_begin = std::distance(state_words, first_modified_word); + const u64 word_index_end = std::distance(state_words, last_modified_word); + + const unsigned local_page_begin = std::countr_zero(*first_modified_word); + const unsigned local_page_end = PAGES_PER_WORD - std::countl_zero(last_modified_word[-1]); + const u64 word_page_begin = word_index_begin * PAGES_PER_WORD; + const u64 word_page_end = (word_index_end - 1) * PAGES_PER_WORD; + const u64 query_page_begin = query_begin / BYTES_PER_PAGE; + const u64 query_page_end = Common::DivCeil(query_end, BYTES_PER_PAGE); + const u64 page_index_begin = std::max(word_page_begin + local_page_begin, query_page_begin); + const u64 page_index_end = std::min(word_page_end + local_page_end, query_page_end); + const u64 first_word_page_begin = page_index_begin % PAGES_PER_WORD; + const u64 last_word_page_end = (page_index_end - 1) % PAGES_PER_WORD + 1; + + u64 page_begin = first_word_page_begin; + u64 current_base = 0; + u64 current_size = 0; + bool on_going = false; + for (u64 word_index = word_index_begin; word_index < word_index_end; ++word_index) { + const bool is_last_word = word_index + 1 == word_index_end; + const u64 page_end = is_last_word ? last_word_page_end : PAGES_PER_WORD; + const u64 right_offset = page_begin; + const u64 left_offset = PAGES_PER_WORD - page_end; + u64 bits = ~u64{0}; + bits = (bits >> right_offset) << right_offset; + bits = (bits << left_offset) >> left_offset; + + const u64 current_word = state_words[word_index] & bits; + state_words[word_index] &= ~bits; + + // Exclude CPU modified pages when visiting GPU pages + const u64 word = current_word & ~(gpu ? cpu_words[word_index] : 0); + if constexpr (notify_rasterizer) { + NotifyRasterizer(word_index, word, ~u64{0}); + } + u64 page = page_begin; + page_begin = 0; + + while (page < page_end) { + const int empty_bits = std::countr_zero(word >> page); + if (on_going && empty_bits != 0) { + InvokeModifiedRange(func, current_size, current_base); + current_size = 0; + on_going = false; + } + page += empty_bits; + + const int continuous_bits = std::countr_one(word >> page); + if (!on_going && continuous_bits != 0) { + current_base = word_index * PAGES_PER_WORD + page; + on_going = true; + } + current_size += continuous_bits; + page += continuous_bits; + } + } + if (on_going && current_size > 0) { + InvokeModifiedRange(func, current_size, current_base); + } + } + + template + void InvokeModifiedRange(Func&& func, u64 current_size, u64 current_base) { + const u64 current_size_bytes = current_size * BYTES_PER_PAGE; + const u64 offset_begin = current_base * BYTES_PER_PAGE; + const u64 offset_end = std::min(offset_begin + current_size_bytes, SizeBytes()); + func(offset_begin, offset_end - offset_begin); + } + + /** + * Returns true when a region has been modified + * + * @param offset Offset in bytes from the start of the buffer + * @param size Size in bytes of the region to query for modifications + */ + template + [[nodiscard]] bool IsRegionModified(u64 offset, u64 size) const noexcept { + const u64* const cpu_words = words.cpu.Pointer(IsShort()); + const u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort()); + const u64 num_query_words = size / BYTES_PER_WORD + 1; + const u64 word_begin = offset / BYTES_PER_WORD; + const u64 word_end = std::min(word_begin + num_query_words, NumWords()); + const u64 page_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE); + u64 page_index = (offset / BYTES_PER_PAGE) % PAGES_PER_WORD; + for (u64 word_index = word_begin; word_index < word_end; ++word_index, page_index = 0) { + const u64 word = state_words[word_index] & ~(gpu ? cpu_words[word_index] : 0); + if (word == 0) { + continue; + } + const u64 page_end = std::min((word_index + 1) * PAGES_PER_WORD, page_limit); + const u64 local_page_end = page_end % PAGES_PER_WORD; + const u64 page_end_shift = (PAGES_PER_WORD - local_page_end) % PAGES_PER_WORD; + if (((word >> page_index) << page_index) << page_end_shift != 0) { + return true; + } + } + return false; + } + + /** + * Returns a begin end pair with the inclusive modified region + * + * @param offset Offset in bytes from the start of the buffer + * @param size Size in bytes of the region to query for modifications + * + * @tparam True to query GPU modified pages, false for CPU pages + */ + template + [[nodiscard]] std::pair ModifiedRegion(u64 offset, u64 size) const noexcept { + const u64* const cpu_words = words.cpu.Pointer(IsShort()); + const u64* const state_words = (gpu ? words.gpu : words.cpu).Pointer(IsShort()); + const u64 num_query_words = size / BYTES_PER_WORD + 1; + const u64 word_begin = offset / BYTES_PER_WORD; + const u64 word_end = std::min(word_begin + num_query_words, NumWords()); + const u64 page_base = offset / BYTES_PER_PAGE; + const u64 page_limit = Common::DivCeil(offset + size, BYTES_PER_PAGE); + u64 begin = std::numeric_limits::max(); + u64 end = 0; + for (u64 word_index = word_begin; word_index < word_end; ++word_index) { + const u64 word = state_words[word_index] & ~(gpu ? cpu_words[word_index] : 0); + if (word == 0) { + continue; + } + const u64 local_page_begin = std::countr_zero(word); + const u64 local_page_end = PAGES_PER_WORD - std::countl_zero(word); + const u64 page_index = word_index * PAGES_PER_WORD; + const u64 page_begin = std::max(page_index + local_page_begin, page_base); + const u64 page_end = std::min(page_index + local_page_end, page_limit); + begin = std::min(begin, page_begin); + end = std::max(end, page_end); + } + static constexpr std::pair EMPTY{0, 0}; + return begin < end ? std::make_pair(begin * BYTES_PER_PAGE, end * BYTES_PER_PAGE) : EMPTY; + } + + /// Returns the number of words of the buffer + [[nodiscard]] size_t NumWords() const noexcept { + return words.NumWords(); + } + + /// Returns true when the buffer fits in the small vector optimization + [[nodiscard]] bool IsShort() const noexcept { + return words.IsShort(); + } + + RasterizerInterface* rasterizer = nullptr; + VAddr cpu_addr = 0; + GpuCpuWords words; + BufferFlagBits flags{}; +}; + +} // namespace VideoCommon -- cgit v1.2.3