[video_core] reduce SPSC/MPSC queue contention for commands

Signed-off-by: lizzie <lizzie@eden-emu.dev>

src/common/bounded_threadsafe_queue.h

@@ -1,3 +1,5 @@
+// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -23,34 +25,34 @@ class SPSCQueue {
 
 public:
     template <typename... Args>
-    bool TryEmplace(Args&&... args) {
+    bool TryEmplace(Args&&... args) noexcept {
         return Emplace<PushMode::Try>(std::forward<Args>(args)...);
     }
 
     template <typename... Args>
-    void EmplaceWait(Args&&... args) {
+    void EmplaceWait(Args&&... args) noexcept {
         Emplace<PushMode::Wait>(std::forward<Args>(args)...);
     }
 
-    bool TryPop(T& t) {
+    bool TryPop(T& t) noexcept {
         return Pop<PopMode::Try>(t);
     }
 
-    void PopWait(T& t) {
+    void PopWait(T& t) noexcept {
         Pop<PopMode::Wait>(t);
     }
 
-    void PopWait(T& t, std::stop_token stop_token) {
+    void PopWait(T& t, const std::stop_token stop_token) noexcept {
         Pop<PopMode::WaitWithStopToken>(t, stop_token);
     }
 
-    T PopWait() {
+    T PopWait() noexcept {
         T t{};
         Pop<PopMode::Wait>(t);
         return t;
     }
 
-    T PopWait(std::stop_token stop_token) {
+    T PopWait(const std::stop_token stop_token) noexcept {
         T t{};
         Pop<PopMode::WaitWithStopToken>(t, stop_token);
         return t;
@@ -71,60 +73,53 @@ private:
     };
 
     template <PushMode Mode, typename... Args>
-    bool Emplace(Args&&... args) {
+    bool Emplace(Args&&... args) noexcept {
-        const size_t write_index = m_write_index.load(std::memory_order::relaxed);
+        const std::size_t write_index = producer.index.load(std::memory_order::relaxed);
-
         if constexpr (Mode == PushMode::Try) {
             // Check if we have free slots to write to.
-            if ((write_index - m_read_index.load(std::memory_order::acquire)) == Capacity) {
+            if ((write_index - consumer.index.load(std::memory_order::acquire)) == Capacity) {
                 return false;
             }
         } else if constexpr (Mode == PushMode::Wait) {
             // Wait until we have free slots to write to.
-            std::unique_lock lock{producer_cv_mutex};
+            std::unique_lock lock{producer.cv_mutex};
-            producer_cv.wait(lock, [this, write_index] {
+            producer.cv.wait(lock, [this, write_index] {
-                return (write_index - m_read_index.load(std::memory_order::acquire)) < Capacity;
+                return (write_index - consumer.index.load(std::memory_order::acquire)) < Capacity;
             });
         } else {
             static_assert(Mode < PushMode::Count, "Invalid PushMode.");
         }
-
         // Determine the position to write to.
-        const size_t pos = write_index % Capacity;
+        const std::size_t pos = write_index % Capacity;
-
         // Emplace into the queue.
         std::construct_at(std::addressof(m_data[pos]), std::forward<Args>(args)...);
-
         // Increment the write index.
-        ++m_write_index;
+        ++producer.index;
-
         // Notify the consumer that we have pushed into the queue.
-        std::scoped_lock lock{consumer_cv_mutex};
+        std::scoped_lock lock{consumer.cv_mutex};
-        consumer_cv.notify_one();
+        consumer.cv.notify_one();
-
         return true;
     }
 
     template <PopMode Mode>
-    bool Pop(T& t, [[maybe_unused]] std::stop_token stop_token = {}) {
+    bool Pop(T& t, [[maybe_unused]] std::stop_token stop_token = {}) noexcept {
-        const size_t read_index = m_read_index.load(std::memory_order::relaxed);
+        const std::size_t read_index = consumer.index.load(std::memory_order::relaxed);
-
         if constexpr (Mode == PopMode::Try) {
             // Check if the queue is empty.
-            if (read_index == m_write_index.load(std::memory_order::acquire)) {
+            if (read_index == producer.index.load(std::memory_order::acquire)) {
                 return false;
             }
         } else if constexpr (Mode == PopMode::Wait) {
             // Wait until the queue is not empty.
-            std::unique_lock lock{consumer_cv_mutex};
+            std::unique_lock lock{consumer.cv_mutex};
-            consumer_cv.wait(lock, [this, read_index] {
+            consumer.cv.wait(lock, [this, read_index] {
-                return read_index != m_write_index.load(std::memory_order::acquire);
+                return read_index != producer.index.load(std::memory_order::acquire);
             });
         } else if constexpr (Mode == PopMode::WaitWithStopToken) {
             // Wait until the queue is not empty.
-            std::unique_lock lock{consumer_cv_mutex};
+            std::unique_lock lock{consumer.cv_mutex};
-            Common::CondvarWait(consumer_cv, lock, stop_token, [this, read_index] {
+            Common::CondvarWait(consumer.cv, lock, stop_token, [this, read_index] {
-                return read_index != m_write_index.load(std::memory_order::acquire);
+                return read_index != producer.index.load(std::memory_order::acquire);
             });
             if (stop_token.stop_requested()) {
                 return false;
@@ -132,32 +127,29 @@ private:
         } else {
             static_assert(Mode < PopMode::Count, "Invalid PopMode.");
         }
-
         // Determine the position to read from.
-        const size_t pos = read_index % Capacity;
+        const std::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;
+        ++consumer.index;
-
         // Notify the producer that we have popped off the queue.
-        std::scoped_lock lock{producer_cv_mutex};
+        std::scoped_lock lock{producer.cv_mutex};
-        producer_cv.notify_one();
+        producer.cv.notify_one();
-
         return true;
     }
 
-    alignas(128) std::atomic_size_t m_read_index{0};
-    alignas(128) std::atomic_size_t m_write_index{0};
-
     std::array<T, Capacity> m_data;
-
+    alignas(64) struct {
-    std::condition_variable_any producer_cv;
+        std::atomic_size_t index{0};
-    std::mutex producer_cv_mutex;
+        std::condition_variable_any cv;
-    std::condition_variable_any consumer_cv;
+        std::mutex cv_mutex;
-    std::mutex consumer_cv_mutex;
+    } producer;
+    alignas(64) struct {
+        std::atomic_size_t index{0};
+        std::condition_variable_any cv;
+        std::mutex cv_mutex;
+    } consumer;
 };
 
 template <typename T, size_t Capacity = detail::DefaultCapacity>
@@ -242,8 +234,8 @@ public:
 
 private:
     SPSCQueue<T, Capacity> spsc_queue;
-    std::mutex write_mutex;
+    alignas(64) std::mutex write_mutex;
-    std::mutex read_mutex;
+    alignas(64) std::mutex read_mutex;
 };
 
 } // namespace Common