From 7290086e2d9d17436d3244d0e79859424154eb61 Mon Sep 17 00:00:00 2001 From: pineappleEA Date: Sun, 19 Mar 2023 11:35:25 +0100 Subject: [PATCH] early-access version 3464 --- README.md | 2 +- src/common/bounded_threadsafe_queue.h | 418 ++++++++++++++++++-------- src/common/logging/backend.cpp | 16 +- src/video_core/gpu_thread.cpp | 10 +- 4 files changed, 307 insertions(+), 139 deletions(-) diff --git a/README.md b/README.md index 6743716d2..9b3ffc452 100755 --- a/README.md +++ b/README.md @@ -1,7 +1,7 @@ yuzu emulator early access ============= -This is the source code for early-access 3463. +This is the source code for early-access 3464. ## Legal Notice diff --git a/src/common/bounded_threadsafe_queue.h b/src/common/bounded_threadsafe_queue.h index 8febc3002..b0a5a3bad 100755 --- a/src/common/bounded_threadsafe_queue.h +++ b/src/common/bounded_threadsafe_queue.h @@ -1,159 +1,323 @@ -// SPDX-FileCopyrightText: Copyright (c) 2020 Erik Rigtorp -// SPDX-License-Identifier: MIT +// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project +// SPDX-License-Identifier: GPL-2.0-or-later #pragma once #include -#include #include -#include +#include #include #include -#include -#include +#include #include "common/polyfill_thread.h" namespace Common { -#if defined(__cpp_lib_hardware_interference_size) -constexpr size_t hardware_interference_size = std::hardware_destructive_interference_size; -#else -constexpr size_t hardware_interference_size = 64; -#endif +namespace detail { +constexpr size_t DefaultCapacity = 0x1000; +} // namespace detail + +template +class SPSCQueue { + static_assert((Capacity & (Capacity - 1)) == 0, "Capacity must be a power of two."); -template -class MPSCQueue { public: - explicit MPSCQueue() : allocator{std::allocator>()} { - // Allocate one extra slot to prevent false sharing on the last slot - slots = allocator.allocate(capacity + 1); - // Allocators are not required to honor alignment for over-aligned types - // (see http://eel.is/c++draft/allocator.requirements#10) so we verify - // alignment here - if (reinterpret_cast(slots) % alignof(Slot) != 0) { - allocator.deallocate(slots, capacity + 1); - throw std::bad_alloc(); + bool TryPush(T&& t) { + const size_t write_index = m_write_index.load(); + + // Check if we have free slots to write to. + if ((write_index - m_read_index.load()) == Capacity) { + return false; } - for (size_t i = 0; i < capacity; ++i) { - std::construct_at(&slots[i]); - } - static_assert(std::has_single_bit(capacity), "capacity must be an integer power of 2"); - static_assert(alignof(Slot) == hardware_interference_size, - "Slot must be aligned to cache line boundary to prevent false sharing"); - static_assert(sizeof(Slot) % hardware_interference_size == 0, - "Slot size must be a multiple of cache line size to prevent " - "false sharing between adjacent slots"); - static_assert(sizeof(MPSCQueue) % hardware_interference_size == 0, - "Queue size must be a multiple of cache line size to " - "prevent false sharing between adjacent queues"); + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Push into the queue. + m_data[pos] = std::move(t); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; + cv.notify_one(); + + return true; } - ~MPSCQueue() noexcept { - for (size_t i = 0; i < capacity; ++i) { - std::destroy_at(&slots[i]); - } - allocator.deallocate(slots, capacity + 1); - } - - // The queue must be both non-copyable and non-movable - MPSCQueue(const MPSCQueue&) = delete; - MPSCQueue& operator=(const MPSCQueue&) = delete; - - MPSCQueue(MPSCQueue&&) = delete; - MPSCQueue& operator=(MPSCQueue&&) = delete; - - void Push(const T& v) noexcept { - static_assert(std::is_nothrow_copy_constructible_v, - "T must be nothrow copy constructible"); - emplace(v); - } - - template >> - void Push(P&& v) noexcept { - emplace(std::forward

(v)); - } - - void Pop(T& v, std::stop_token stop) noexcept { - auto const tail = tail_.fetch_add(1); - auto& slot = slots[idx(tail)]; - if (!slot.turn.test()) { - std::unique_lock lock{cv_mutex}; - Common::CondvarWait(cv, lock, stop, [&slot] { return slot.turn.test(); }); - } - v = slot.move(); - slot.destroy(); - slot.turn.clear(); - slot.turn.notify_one(); - } - -private: - template - struct Slot { - ~Slot() noexcept { - if (turn.test()) { - destroy(); - } - } - - template - void construct(Args&&... args) noexcept { - static_assert(std::is_nothrow_constructible_v, - "T must be nothrow constructible with Args&&..."); - std::construct_at(reinterpret_cast(&storage), std::forward(args)...); - } - - void destroy() noexcept { - static_assert(std::is_nothrow_destructible_v, "T must be nothrow destructible"); - std::destroy_at(reinterpret_cast(&storage)); - } - - U&& move() noexcept { - return reinterpret_cast(storage); - } - - // Align to avoid false sharing between adjacent slots - alignas(hardware_interference_size) std::atomic_flag turn{}; - struct aligned_store { - struct type { - alignas(U) unsigned char data[sizeof(U)]; - }; - }; - typename aligned_store::type storage; - }; - template - void emplace(Args&&... args) noexcept { - static_assert(std::is_nothrow_constructible_v, - "T must be nothrow constructible with Args&&..."); - auto const head = head_.fetch_add(1); - auto& slot = slots[idx(head)]; - slot.turn.wait(true); - slot.construct(std::forward(args)...); - slot.turn.test_and_set(); + bool TryPush(Args&&... args) { + const size_t write_index = m_write_index.load(); + + // Check if we have free slots to write to. + if ((write_index - m_read_index.load()) == Capacity) { + return false; + } + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Emplace into the queue. + std::construct_at(std::addressof(m_data[pos]), std::forward(args)...); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; + cv.notify_one(); + + return true; + } + + void Push(T&& t) { + const size_t write_index = m_write_index.load(); + + // Wait until we have free slots to write to. + while ((write_index - m_read_index.load()) == Capacity) { + std::this_thread::yield(); + } + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Push into the queue. + m_data[pos] = std::move(t); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; cv.notify_one(); } - constexpr size_t idx(size_t i) const noexcept { - return i & mask; + template + void Push(Args&&... args) { + const size_t write_index = m_write_index.load(); + + // Wait until we have free slots to write to. + while ((write_index - m_read_index.load()) == Capacity) { + std::this_thread::yield(); + } + + // Determine the position to write to. + const size_t pos = write_index % Capacity; + + // Emplace into the queue. + std::construct_at(std::addressof(m_data[pos]), std::forward(args)...); + + // Increment the write index. + ++m_write_index; + + // Notify the consumer that we have pushed into the queue. + std::scoped_lock lock{cv_mutex}; + cv.notify_one(); } - static constexpr size_t mask = capacity - 1; + bool TryPop(T& t) { + return Pop(t); + } - // Align to avoid false sharing between head_ and tail_ - alignas(hardware_interference_size) std::atomic head_{0}; - alignas(hardware_interference_size) std::atomic tail_{0}; + void PopWait(T& t, std::stop_token stop_token) { + Wait(stop_token); + Pop(t); + } + + T PopWait(std::stop_token stop_token) { + Wait(stop_token); + T t; + Pop(t); + return t; + } + + void Clear() { + while (!Empty()) { + Pop(); + } + } + + bool Empty() const { + return m_read_index.load() == m_write_index.load(); + } + + size_t Size() const { + return m_write_index.load() - m_read_index.load(); + } + +private: + void Pop() { + const size_t read_index = m_read_index.load(); + + // Check if the queue is empty. + if (read_index == m_write_index.load()) { + return; + } + + // Determine the position to read from. + const size_t pos = read_index % Capacity; + + // Pop the data off the queue, deleting it. + std::destroy_at(std::addressof(m_data[pos])); + + // Increment the read index. + ++m_read_index; + } + + bool Pop(T& t) { + const size_t read_index = m_read_index.load(); + + // Check if the queue is empty. + if (read_index == m_write_index.load()) { + return false; + } + + // Determine the position to read from. + const size_t pos = read_index % Capacity; + + // Pop the data off the queue, moving it. + t = std::move(m_data[pos]); + + // Increment the read index. + ++m_read_index; + + return true; + } + + void Wait(std::stop_token stop_token) { + std::unique_lock lock{cv_mutex}; + Common::CondvarWait(cv, lock, stop_token, [this] { return !Empty(); }); + } + +#ifdef __cpp_lib_hardware_interference_size + alignas(std::hardware_destructive_interference_size) std::atomic_size_t m_read_index{0}; + alignas(std::hardware_destructive_interference_size) std::atomic_size_t m_write_index{0}; +#else + alignas(64) std::atomic_size_t m_read_index{0}; + alignas(64) std::atomic_size_t m_write_index{0}; +#endif + + std::array m_data; - std::mutex cv_mutex; std::condition_variable_any cv; + std::mutex cv_mutex; +}; - Slot* slots; - [[no_unique_address]] std::allocator> allocator; +template +class MPSCQueue { +public: + void TryPush(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.TryPush(std::move(t)); + } - static_assert(std::is_nothrow_copy_assignable_v || std::is_nothrow_move_assignable_v, - "T must be nothrow copy or move assignable"); + template + void TryPush(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.TryPush(std::forward(args)...); + } - static_assert(std::is_nothrow_destructible_v, "T must be nothrow destructible"); + void Push(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::move(t)); + } + + template + void Push(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::forward(args)...); + } + + bool TryPop(T& t) { + return spsc_queue.TryPop(t); + } + + void PopWait(T& t, std::stop_token stop_token) { + spsc_queue.PopWait(t, stop_token); + } + + T PopWait(std::stop_token stop_token) { + return spsc_queue.PopWait(stop_token); + } + + void Clear() { + spsc_queue.Clear(); + } + + bool Empty() { + return spsc_queue.Empty(); + } + + size_t Size() { + return spsc_queue.Size(); + } + +private: + SPSCQueue spsc_queue; + std::mutex write_mutex; +}; + +template +class MPMCQueue { +public: + void TryPush(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.TryPush(std::move(t)); + } + + template + void TryPush(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.TryPush(std::forward(args)...); + } + + void Push(T&& t) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::move(t)); + } + + template + void Push(Args&&... args) { + std::scoped_lock lock{write_mutex}; + spsc_queue.Push(std::forward(args)...); + } + + bool TryPop(T& t) { + std::scoped_lock lock{read_mutex}; + return spsc_queue.TryPop(t); + } + + void PopWait(T& t, std::stop_token stop_token) { + std::scoped_lock lock{read_mutex}; + spsc_queue.PopWait(t, stop_token); + } + + T PopWait(std::stop_token stop_token) { + std::scoped_lock lock{read_mutex}; + return spsc_queue.PopWait(stop_token); + } + + void Clear() { + std::scoped_lock lock{read_mutex}; + spsc_queue.Clear(); + } + + bool Empty() { + std::scoped_lock lock{read_mutex}; + return spsc_queue.Empty(); + } + + size_t Size() { + std::scoped_lock lock{read_mutex}; + return spsc_queue.Size(); + } + +private: + SPSCQueue spsc_queue; + std::mutex write_mutex; + std::mutex read_mutex; }; } // namespace Common diff --git a/src/common/logging/backend.cpp b/src/common/logging/backend.cpp index 863970e33..caa707f2f 100755 --- a/src/common/logging/backend.cpp +++ b/src/common/logging/backend.cpp @@ -28,7 +28,7 @@ #ifdef _WIN32 #include "common/string_util.h" #endif -#include "common/threadsafe_queue.h" +#include "common/bounded_threadsafe_queue.h" namespace Common::Log { @@ -204,11 +204,11 @@ public: void PushEntry(Class log_class, Level log_level, const char* filename, unsigned int line_num, const char* function, std::string&& message) { - if (!filter.CheckMessage(log_class, log_level)) + if (!filter.CheckMessage(log_class, log_level)) { return; - const Entry& entry = - CreateEntry(log_class, log_level, filename, line_num, function, std::move(message)); - message_queue.Push(entry); + } + message_queue.Push( + CreateEntry(log_class, log_level, filename, line_num, function, std::move(message))); } private: @@ -225,7 +225,7 @@ private: ForEachBackend([&entry](Backend& backend) { backend.Write(entry); }); }; while (!stop_token.stop_requested()) { - entry = message_queue.PopWait(stop_token); + message_queue.PopWait(entry, stop_token); if (entry.filename != nullptr) { write_logs(); } @@ -233,7 +233,7 @@ private: // Drain the logging queue. Only writes out up to MAX_LOGS_TO_WRITE to prevent a // case where a system is repeatedly spamming logs even on close. int max_logs_to_write = filter.IsDebug() ? INT_MAX : 100; - while (max_logs_to_write-- && message_queue.Pop(entry)) { + while (max_logs_to_write-- && message_queue.TryPop(entry)) { write_logs(); } }); @@ -273,7 +273,7 @@ private: ColorConsoleBackend color_console_backend{}; FileBackend file_backend; - MPSCQueue message_queue{}; + MPSCQueue message_queue{}; std::chrono::steady_clock::time_point time_origin{std::chrono::steady_clock::now()}; std::jthread backend_thread; }; diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp index 83ed85817..acba77ba6 100755 --- a/src/video_core/gpu_thread.cpp +++ b/src/video_core/gpu_thread.cpp @@ -31,9 +31,10 @@ static void RunThread(std::stop_token stop_token, Core::System& system, auto current_context = context.Acquire(); VideoCore::RasterizerInterface* const rasterizer = renderer.ReadRasterizer(); + CommandDataContainer next; + while (!stop_token.stop_requested()) { - CommandDataContainer next; - state.queue.Pop(next, stop_token); + state.queue.PopWait(next, stop_token); if (stop_token.stop_requested()) { break; } @@ -58,6 +59,9 @@ static void RunThread(std::stop_token stop_token, Core::System& system, state.cv.notify_all(); } } + + // Drain the queue. + state.queue.Clear(); } ThreadManager::ThreadManager(Core::System& system_, bool is_async_) @@ -117,7 +121,7 @@ u64 ThreadManager::PushCommand(CommandData&& command_data, bool block) { std::unique_lock lk(state.write_lock); const u64 fence{++state.last_fence}; - state.queue.Push(CommandDataContainer(std::move(command_data), fence, block)); + state.queue.Push(std::move(command_data), fence, block); if (block) { Common::CondvarWait(state.cv, lk, thread.get_stop_token(), [this, fence] {