gpu_thread: Move to bounded queue

3 files changed, 185 insertions, 4 deletions

diff --git a/src/common/bounded_threadsafe_queue.h b/src/common/bounded_threadsafe_queue.h
new file mode 100644
index 000000000..e83064c7f
--- /dev/null
+++ b/src/common/bounded_threadsafe_queue.h
@@ -0,0 +1,180 @@
+// SPDX-FileCopyrightText: Copyright (c) 2020 Erik Rigtorp <erik@rigtorp.se>
+// SPDX-License-Identifier: MIT
+#pragma once
+#ifdef _MSC_VER
+#pragma warning(push)
+#pragma warning(disable : 4324)
+#endif
+
+#include <atomic>
+#include <bit>
+#include <condition_variable>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <stdexcept>
+#include <stop_token>
+#include <type_traits>
+#include <utility>
+
+namespace Common {
+namespace mpsc {
+#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
+
+template <typename T>
+using AlignedAllocator = std::allocator<T>;
+
+template <typename T>
+struct Slot {
+    ~Slot() noexcept {
+        if (turn.test()) {
+            destroy();
+        }
+    }
+
+    template <typename... Args>
+    void construct(Args&&... args) noexcept {
+        static_assert(std::is_nothrow_constructible_v<T, Args&&...>,
+                      "T must be nothrow constructible with Args&&...");
+        std::construct_at(reinterpret_cast<T*>(&storage), std::forward<Args>(args)...);
+    }
+
+    void destroy() noexcept {
+        static_assert(std::is_nothrow_destructible_v<T>, "T must be nothrow destructible");
+        std::destroy_at(reinterpret_cast<T*>(&storage));
+    }
+
+    T&& move() noexcept {
+        return reinterpret_cast<T&&>(storage);
+    }
+
+    // Align to avoid false sharing between adjacent slots
+    alignas(hardware_interference_size) std::atomic_flag turn{};
+    struct aligned_store {
+        struct type {
+            alignas(T) unsigned char data[sizeof(T)];
+        };
+    };
+    typename aligned_store::type storage;
+};
+
+template <typename T, typename Allocator = AlignedAllocator<Slot<T>>>
+class Queue {
+public:
+    explicit Queue(const size_t capacity, const Allocator& allocator = Allocator())
+        : allocator_(allocator) {
+        if (capacity < 1) {
+            throw std::invalid_argument("capacity < 1");
+        }
+        // Ensure that the queue length is an integer power of 2
+        // This is so that idx(i) can be a simple i & mask_ insted of i % capacity
+        // https://github.com/rigtorp/MPMCQueue/pull/36
+        if (!std::has_single_bit(capacity)) {
+            throw std::invalid_argument("capacity must be an integer power of 2");
+        }
+
+        mask_ = capacity - 1;
+
+        // Allocate one extra slot to prevent false sharing on the last slot
+        slots_ = allocator_.allocate(mask_ + 2);
+        // 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<uintptr_t>(slots_) % alignof(Slot<T>) != 0) {
+            allocator_.deallocate(slots_, mask_ + 2);
+            throw std::bad_alloc();
+        }
+        for (size_t i = 0; i < mask_ + 1; ++i) {
+            std::construct_at(&slots_[i]);
+        }
+        static_assert(alignof(Slot<T>) == hardware_interference_size,
+                      "Slot must be aligned to cache line boundary to prevent false sharing");
+        static_assert(sizeof(Slot<T>) % hardware_interference_size == 0,
+                      "Slot size must be a multiple of cache line size to prevent "
+                      "false sharing between adjacent slots");
+        static_assert(sizeof(Queue) % hardware_interference_size == 0,
+                      "Queue size must be a multiple of cache line size to "
+                      "prevent false sharing between adjacent queues");
+    }
+
+    ~Queue() noexcept {
+        for (size_t i = 0; i < mask_ + 1; ++i) {
+            slots_[i].~Slot();
+        }
+        allocator_.deallocate(slots_, mask_ + 2);
+    }
+
+    // non-copyable and non-movable
+    Queue(const Queue&) = delete;
+    Queue& operator=(const Queue&) = delete;
+
+    void Push(const T& v) noexcept {
+        static_assert(std::is_nothrow_copy_constructible_v<T>,
+                      "T must be nothrow copy constructible");
+        emplace(v);
+    }
+
+    template <typename P, typename = std::enable_if_t<std::is_nothrow_constructible_v<T, P&&>>>
+    void Push(P&& v) noexcept {
+        emplace(std::forward<P>(v));
+    }
+
+    void Pop(T& v, std::stop_token stop) noexcept {
+        auto const tail = tail_.fetch_add(1);
+        auto& slot = slots_[idx(tail)];
+        if (false == slot.turn.test()) {
+            std::unique_lock lock{cv_mutex};
+            cv.wait(lock, stop, [&slot] { return slot.turn.test(); });
+        }
+        v = slot.move();
+        slot.destroy();
+        slot.turn.clear();
+        slot.turn.notify_one();
+    }
+
+private:
+    template <typename... Args>
+    void emplace(Args&&... args) noexcept {
+        static_assert(std::is_nothrow_constructible_v<T, Args&&...>,
+                      "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>(args)...);
+        slot.turn.test_and_set();
+        cv.notify_one();
+    }
+
+    constexpr size_t idx(size_t i) const noexcept {
+        return i & mask_;
+    }
+
+    std::conditional_t<true, std::condition_variable_any, std::condition_variable> cv;
+    std::mutex cv_mutex;
+    size_t mask_;
+    Slot<T>* slots_;
+    [[no_unique_address]] Allocator allocator_;
+
+    // Align to avoid false sharing between head_ and tail_
+    alignas(hardware_interference_size) std::atomic<size_t> head_{0};
+    alignas(hardware_interference_size) std::atomic<size_t> tail_{0};
+
+    static_assert(std::is_nothrow_copy_assignable_v<T> || std::is_nothrow_move_assignable_v<T>,
+                  "T must be nothrow copy or move assignable");
+
+    static_assert(std::is_nothrow_destructible_v<T>, "T must be nothrow destructible");
+};
+} // namespace mpsc
+
+template <typename T, typename Allocator = mpsc::AlignedAllocator<mpsc::Slot<T>>>
+using MPSCQueue = mpsc::Queue<T, Allocator>;
+
+} // namespace Common
+
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif
diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index b79a73132..8479dc6d2 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -31,7 +31,8 @@ static void RunThread(std::stop_token stop_token, Core::System& system,
     VideoCore::RasterizerInterface* const rasterizer = renderer.ReadRasterizer();
 
     while (!stop_token.stop_requested()) {
-        CommandDataContainer next = state.queue.PopWait(stop_token);
+        CommandDataContainer next;
+        state.queue.Pop(next, stop_token);
         if (stop_token.stop_requested()) {
             break;
         }
diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index 71cd35756..ad9fd5eff 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -10,7 +10,7 @@
 #include <thread>
 #include <variant>
 
-#include "common/threadsafe_queue.h"
+#include "common/bounded_threadsafe_queue.h"
 #include "video_core/framebuffer_config.h"
 
 namespace Tegra {
@@ -96,9 +96,9 @@ struct CommandDataContainer {
 
 /// Struct used to synchronize the GPU thread
 struct SynchState final {
-    using CommandQueue = Common::SPSCQueue<CommandDataContainer, true>;
+    using CommandQueue = Common::MPSCQueue<CommandDataContainer>;
     std::mutex write_lock;
-    CommandQueue queue;
+    CommandQueue queue{512}; // size must be 2^n
     u64 last_fence{};
     std::atomic<u64> signaled_fence{};
     std::condition_variable_any cv;