GPU: Move backend work off CPU thread

src/common/threading.cpp

@@ -1,10 +1,17 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
+// SPDX-FileCopyrightText: 2019-2023 Connor McLaughlin <stenzek@gmail.com>, 2002-2023 PCSX2 Dev Team
+// SPDX-License-Identifier: LGPL-3.0
 
 #include "threading.h"
 #include "assert.h"
+#include "log.h"
+#include "timer.h"
+
 #include <memory>
 
+#if defined(CPU_ARCH_X86) || defined(CPU_ARCH_X64)
+#include <emmintrin.h>
+#endif
+
 #if !defined(_WIN32) && !defined(__APPLE__)
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
@@ -14,6 +21,10 @@
 #if defined(_WIN32)
 #include "windows_headers.h"
 #include <process.h>
+
+#if defined(CPU_ARCH_ARM64) && defined(_MSC_VER)
+#include <arm64intr.h>
+#endif
 #else
 #include <pthread.h>
 #include <unistd.h>
@@ -38,6 +49,8 @@
 #endif
 #endif
 
+Log_SetChannel(Threading);
+
 #ifdef _WIN32
 union FileTimeU64Union
 {
@@ -101,6 +114,138 @@ void Threading::Timeslice()
 #endif
 }
 
+static void MultiPause()
+{
+#if defined(CPU_ARCH_X86) || defined(CPU_ARCH_X64)
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+  _mm_pause();
+#elif defined(CPU_ARCH_ARM64) && defined(_MSC_VER)
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+  __isb(_ARM64_BARRIER_SY);
+#elif defined(CPU_ARCH_ARM64) || defined(CPU_ARCH_ARM32)
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+  __asm__ __volatile__("isb");
+#elif defined(CPU_ARCH_RISCV64)
+  // Probably wrong... pause is optional :/
+  asm volatile("fence" ::: "memory");
+#else
+#pragma warning("Missing implementation")
+#endif
+}
+
+// Apple uses a lower tick frequency, so we can't use the dynamic loop below.
+#if !defined(_M_ARM64) || defined(__APPLE__) || defined(_WIN32)
+
+static u32 PAUSE_TIME = 0;
+
+static u32 MeasurePauseTime()
+{
+  // GetCPUTicks may have resolution as low as 1us
+  // One call to MultiPause could take anywhere from 20ns (fast Haswell) to 400ns (slow Skylake)
+  // We want a measurement of reasonable resolution, but don't want to take too long
+  // So start at a fairly small number and increase it if it's too fast
+  for (int testcnt = 64; true; testcnt *= 2)
+  {
+    Common::Timer::Value start = Common::Timer::GetCurrentValue();
+    for (int i = 0; i < testcnt; i++)
+    {
+      MultiPause();
+    }
+    Common::Timer::Value time = Common::Timer::GetCurrentValue() - start;
+    if (time > 100)
+    {
+      const double nanos = Common::Timer::ConvertValueToNanoseconds(time);
+      return static_cast<u32>((nanos / testcnt) + 1);
+    }
+  }
+}
+
+NEVER_INLINE static void UpdatePauseTime()
+{
+  Common::Timer::BusyWait(10000000);
+  u32 pause = MeasurePauseTime();
+  // Take a few measurements in case something weird happens during one
+  // (e.g. OS interrupt)
+  for (int i = 0; i < 4; i++)
+    pause = std::min(pause, MeasurePauseTime());
+  PAUSE_TIME = pause;
+  VERBOSE_LOG("MultiPause time: {}ns", pause);
+}
+
+u32 Threading::ShortSpin()
+{
+  u32 inc = PAUSE_TIME;
+  if (inc == 0) [[unlikely]]
+  {
+    UpdatePauseTime();
+    inc = PAUSE_TIME;
+  }
+
+  u32 time = 0;
+  // Sleep for approximately 500ns
+  for (; time < 500; time += inc)
+    MultiPause();
+
+  return time;
+}
+
+#else
+
+// On ARM, we have big/little cores, and who knows which one we'll measure/run on..
+// TODO: Actually verify this code.
+const u32 SHORT_SPIN_TIME_TICKS = static_cast<u32>((Common::Timer::GetFrequency() * 500) / 1000000000);
+
+u32 Threading::ShortSpin()
+{
+  const Common::Timer::Value start = Common::Timer::GetCurrentValue();
+  Common::Timer::Value now = start;
+  while ((now - start) < SHORT_SPIN_TIME_TICKS)
+  {
+    MultiPause();
+    now = Common::Timer::GetCurrentValue();
+  }
+
+  return static_cast<u32>((Common::Timer::GetCurrentValue() * (now - start)) / 1000000000);
+}
+
+#endif
+
+static u32 GetSpinTime()
+{
+  if (char* req = std::getenv("WAIT_SPIN_MICROSECONDS"))
+  {
+    return 1000 * atoi(req);
+  }
+  else
+  {
+#ifndef _M_ARM64
+    return 50 * 1000; // 50us
+#else
+    return 200 * 1000; // 200us
+#endif
+  }
+}
+
+const u32 Threading::SPIN_TIME_NS = GetSpinTime();
+
 Threading::ThreadHandle::ThreadHandle() = default;
 
 #ifdef _WIN32
@@ -617,3 +762,130 @@ bool Threading::KernelSemaphore::TryWait()
   return sem_trywait(&m_sema) == 0;
 #endif
 }
+
+bool Threading::WorkSema::CheckForWork()
+{
+  s32 value = m_state.load(std::memory_order_relaxed);
+  DebugAssert(!IsDead(value));
+
+  // we want to switch to the running state, but preserve the waiting empty bit for RUNNING_N -> RUNNING_0
+  // otherwise, we clear the waiting flag (since we're notifying the waiter that we're empty below)
+  while (!m_state.compare_exchange_weak(value,
+                                        IsReadyForSleep(value) ? STATE_RUNNING_0 : (value & STATE_FLAG_WAITING_EMPTY),
+                                        std::memory_order_acq_rel, std::memory_order_relaxed))
+  {
+  }
+
+  // if we're not empty, we have work to do
+  if (!IsReadyForSleep(value))
+    return true;
+
+  // this means we're empty, so notify any waiters
+  if (value & STATE_FLAG_WAITING_EMPTY)
+    m_empty_sema.Post();
+
+  // no work to do
+  return false;
+}
+
+void Threading::WorkSema::WaitForWork()
+{
+  // State change:
+  // SLEEPING, SPINNING: This is the worker thread and it's clearly not asleep or spinning, so these states should be
+  // impossible RUNNING_0: Change state to SLEEPING, wake up thread if WAITING_EMPTY RUNNING_N: Change state to
+  // RUNNING_0 (and preserve WAITING_EMPTY flag)
+  s32 value = m_state.load(std::memory_order_relaxed);
+  DebugAssert(!IsDead(value));
+  while (!m_state.compare_exchange_weak(value, NextStateWaitForWork(value), std::memory_order_acq_rel,
+                                        std::memory_order_relaxed))
+    ;
+  if (IsReadyForSleep(value))
+  {
+    if (value & STATE_FLAG_WAITING_EMPTY)
+      m_empty_sema.Post();
+    m_sema.Wait();
+    // Acknowledge any additional work added between wake up request and getting here
+    m_state.fetch_and(STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire);
+  }
+}
+
+void Threading::WorkSema::WaitForWorkWithSpin()
+{
+  s32 value = m_state.load(std::memory_order_relaxed);
+  DebugAssert(!IsDead(value));
+  while (IsReadyForSleep(value))
+  {
+    if (m_state.compare_exchange_weak(value, STATE_SPINNING, std::memory_order_release, std::memory_order_relaxed))
+    {
+      if (value & STATE_FLAG_WAITING_EMPTY)
+        m_empty_sema.Post();
+      value = STATE_SPINNING;
+      break;
+    }
+  }
+  u32 waited = 0;
+  while (value < 0)
+  {
+    if (waited > SPIN_TIME_NS)
+    {
+      if (!m_state.compare_exchange_weak(value, STATE_SLEEPING, std::memory_order_relaxed))
+        continue;
+      m_sema.Wait();
+      break;
+    }
+    waited += ShortSpin();
+    value = m_state.load(std::memory_order_relaxed);
+  }
+  // Clear back to STATE_RUNNING_0 (but preserve waiting empty flag)
+  m_state.fetch_and(STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire);
+}
+
+bool Threading::WorkSema::WaitForEmpty()
+{
+  s32 value = m_state.load(std::memory_order_acquire);
+  while (true)
+  {
+    if (value < 0)
+      return !IsDead(value); // STATE_SLEEPING or STATE_SPINNING, queue is empty!
+    // Note: We technically only need memory_order_acquire on *failure* (because that's when we could leave without
+    // sleeping), but libstdc++ still asserts on failure < success
+    if (m_state.compare_exchange_weak(value, value | STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire))
+      break;
+  }
+  DebugAssertMsg(!(value & STATE_FLAG_WAITING_EMPTY),
+                 "Multiple threads attempted to wait for empty (not currently supported)");
+  m_empty_sema.Wait();
+  return !IsDead(m_state.load(std::memory_order_relaxed));
+}
+
+bool Threading::WorkSema::WaitForEmptyWithSpin()
+{
+  s32 value = m_state.load(std::memory_order_acquire);
+  u32 waited = 0;
+  while (true)
+  {
+    if (value < 0)
+      return !IsDead(value); // STATE_SLEEPING or STATE_SPINNING, queue is empty!
+    if (waited > SPIN_TIME_NS &&
+        m_state.compare_exchange_weak(value, value | STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire))
+      break;
+    waited += ShortSpin();
+    value = m_state.load(std::memory_order_acquire);
+  }
+  DebugAssertMsg(!(value & STATE_FLAG_WAITING_EMPTY),
+                 "Multiple threads attempted to wait for empty (not currently supported)");
+  m_empty_sema.Wait();
+  return !IsDead(m_state.load(std::memory_order_relaxed));
+}
+
+void Threading::WorkSema::Kill()
+{
+  s32 value = m_state.exchange(std::numeric_limits<s32>::min(), std::memory_order_release);
+  if (value & STATE_FLAG_WAITING_EMPTY)
+    m_empty_sema.Post();
+}
+
+void Threading::WorkSema::Reset()
+{
+  m_state = STATE_RUNNING_0;
+}

src/common/threading.h

@@ -1,5 +1,5 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
+// SPDX-FileCopyrightText: 2019-2023 Connor McLaughlin <stenzek@gmail.com>, 2002-2023 PCSX2 Dev Team
+// SPDX-License-Identifier: LGPL-3.0
 
 #pragma once
 #include "types.h"
@@ -20,9 +20,16 @@ extern u64 GetThreadTicksPerSecond();
 /// Set the name of the current thread
 extern void SetNameOfCurrentThread(const char* name);
 
-// Releases a timeslice to other threads.
+/// Releases a timeslice to other threads.
 extern void Timeslice();
 
+/// Spin for a short period of time (call while spinning waiting for a lock)
+/// Returns the approximate number of ns that passed
+extern u32 ShortSpin();
+
+/// Number of ns to spin for before sleeping a thread
+extern const u32 SPIN_TIME_NS;
+
 // --------------------------------------------------------------------------------------
 //  ThreadHandle
 // --------------------------------------------------------------------------------------
@@ -121,4 +128,90 @@ public:
   bool TryWait();
 };
 
+/// A semaphore for notifying a work-processing thread of new work in a (separate) queue
+///
+/// Usage:
+/// - Processing thread loops on `WaitForWork()` followed by processing all work in the queue
+/// - Threads adding work first add their work to the queue, then call `NotifyOfWork()`
+class WorkSema
+{
+  /// Semaphore for sleeping the worker thread
+  KernelSemaphore m_sema;
+
+  /// Semaphore for sleeping thread waiting on worker queue empty
+  KernelSemaphore m_empty_sema;
+
+  /// Current state (see enum below)
+  std::atomic<s32> m_state{0};
+
+  // Expected call frequency is NotifyOfWork > WaitForWork > WaitForEmpty
+  // So optimize states for fast NotifyOfWork
+  enum
+  {
+    /* Any <-2 state: STATE_DEAD: Thread has crashed and is awaiting revival */
+    STATE_SPINNING = -2, ///< Worker thread is spinning waiting for work
+    STATE_SLEEPING = -1, ///< Worker thread is sleeping on m_sema
+    STATE_RUNNING_0 =
+      0, ///< Worker thread is processing work, but no work has been added since it last checked for new work
+    /* Any >0 state: STATE_RUNNING_N: Worker thread is processing work, and work has been added since it last checked
+       for new work */
+    STATE_FLAG_WAITING_EMPTY =
+      1 << 30, ///< Flag to indicate that a thread is sleeping on m_empty_sema (can be applied to any STATE_RUNNING)
+  };
+
+  bool IsDead(s32 state) { return state < STATE_SPINNING; }
+
+  bool IsReadyForSleep(s32 state)
+  {
+    s32 waiting_empty_cleared = state & (STATE_FLAG_WAITING_EMPTY - 1);
+    return waiting_empty_cleared == STATE_RUNNING_0;
+  }
+
+  s32 NextStateWaitForWork(s32 current)
+  {
+    s32 new_state = IsReadyForSleep(current) ? STATE_SLEEPING : STATE_RUNNING_0;
+    return new_state | (current & STATE_FLAG_WAITING_EMPTY); // Preserve waiting empty flag for RUNNING_N -> RUNNING_0
+  }
+
+public:
+  /// Notify the worker thread that you've added new work to its queue
+  void NotifyOfWork()
+  {
+    // State change:
+    // DEAD: Stay in DEAD (starting DEAD state is INT_MIN so we can assume we won't flip over to anything else)
+    // SPINNING: Change state to RUNNING.  Thread will notice and process the new data
+    // SLEEPING: Change state to RUNNING and wake worker.  Thread will wake up and process the new data.
+    // RUNNING_0: Change state to RUNNING_N.
+    // RUNNING_N: Stay in RUNNING_N
+    s32 old = m_state.fetch_add(2, std::memory_order_release);
+    if (old == STATE_SLEEPING)
+      m_sema.Post();
+  }
+
+  /// Checks if there's any work in the queue
+  bool CheckForWork();
+
+  /// Wait for work to be added to the queue
+  void WaitForWork();
+
+  /// Wait for work to be added to the queue, spinning for a bit before sleeping the thread
+  void WaitForWorkWithSpin();
+
+  /// Wait for the worker thread to finish processing all entries in the queue or die
+  /// Returns false if the thread is dead
+  bool WaitForEmpty();
+
+  /// Wait for the worker thread to finish processing all entries in the queue or die, spinning a bit before sleeping
+  /// the thread Returns false if the thread is dead
+  bool WaitForEmptyWithSpin();
+
+  /// Called by the worker thread to notify others of its death
+  /// Dead threads don't process work, and WaitForEmpty will return instantly even though there may be work in the queue
+  void Kill();
+
+  /// Reset the semaphore to the initial state
+  /// Should be called by the worker thread if it restarts after dying
+  void Reset();
+};
+
 } // namespace Threading

src/common/types.h

@@ -26,6 +26,18 @@
 #define ALWAYS_INLINE_RELEASE ALWAYS_INLINE
 #endif
 
+// Avoid inline helper
+#ifndef NEVER_INLINE
+#if defined(_MSC_VER)
+#define NEVER_INLINE __declspec(noinline)
+#elif defined(__GNUC__) || defined(__clang__)
+#define NEVER_INLINE __attribute__((noinline))
+#else
+#define NEVER_INLINE
+#endif
+#endif
+
+
 // unreferenced parameter macro
 #ifndef UNREFERENCED_VARIABLE
 #if defined(__GNUC__) || defined(__clang__) || defined(__EMSCRIPTEN__)

src/core/CMakeLists.txt

@@ -55,10 +55,10 @@ add_library(core
   gpu_shadergen.h
   gpu_sw.cpp
   gpu_sw.h
-  gpu_sw_backend.cpp
-  gpu_sw_backend.h
   gpu_sw_rasterizer.cpp
   gpu_sw_rasterizer.h
+  gpu_thread.cpp
+  gpu_thread.h
   gpu_types.h
   guncon.cpp
   guncon.h

src/core/achievements.cpp

@@ -11,6 +11,7 @@
 #include "bus.h"
 #include "cpu_core.h"
 #include "fullscreen_ui.h"
+#include "gpu_thread.h"
 #include "host.h"
 #include "system.h"
 
@@ -1001,7 +1002,7 @@ void Achievements::ClientLoadGameCallback(int result, const char* error_message,
 
   // ensure fullscreen UI is ready for notifications
   if (display_summary)
-    FullscreenUI::Initialize();
+    GPUThread::RunOnThread(&FullscreenUI::Initialize);
 
   if (const std::string_view badge_name = info->badge_name; !badge_name.empty())
   {
@@ -1062,7 +1063,7 @@ void Achievements::ClearGameHash()
 
 void Achievements::DisplayAchievementSummary()
 {
-  if (g_settings.achievements_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_notifications)
   {
     std::string title;
     if (IsHardcoreModeActive())
@@ -1087,8 +1088,13 @@ void Achievements::DisplayAchievementSummary()
       summary = TRANSLATE_STR("Achievements", "This game has no achievements.");
     }
 
+    GPUThread::RunOnThread([title = std::move(title), summary = std::move(summary)]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
       ImGuiFullscreen::AddNotification("achievement_summary", ACHIEVEMENT_SUMMARY_NOTIFICATION_TIME, std::move(title),
                                        std::move(summary), s_game_icon);
+    });
   }
 
   // Technically not going through the resource API, but since we're passing this to something else, we can't.
@@ -1098,11 +1104,16 @@ void Achievements::DisplayAchievementSummary()
 
 void Achievements::DisplayHardcoreDeferredMessage()
 {
-  if (g_settings.achievements_hardcore_mode && !s_hardcore_mode && System::IsValid() && FullscreenUI::Initialize())
+  if (g_settings.achievements_hardcore_mode && !s_hardcore_mode && System::IsValid())
   {
+    GPUThread::RunOnThread([]() {
+      if (!FullscreenUI::Initialize())
+        return;
+
       ImGuiFullscreen::ShowToast(std::string(),
                                  TRANSLATE_STR("Achievements", "Hardcore mode will be enabled on system reset."),
                                  Host::OSD_WARNING_DURATION);
+    });
   }
 }
 
@@ -1124,7 +1135,7 @@ void Achievements::HandleUnlockEvent(const rc_client_event_t* event)
   INFO_LOG("Achievement {} ({}) for game {} unlocked", cheevo->title, cheevo->id, s_game_id);
   UpdateGameSummary();
 
-  if (g_settings.achievements_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_notifications)
   {
     std::string title;
     if (cheevo->category == RC_CLIENT_ACHIEVEMENT_CATEGORY_UNOFFICIAL)
@@ -1134,9 +1145,15 @@ void Achievements::HandleUnlockEvent(const rc_client_event_t* event)
 
     std::string badge_path = GetAchievementBadgePath(cheevo, cheevo->state);
 
-    ImGuiFullscreen::AddNotification(fmt::format("achievement_unlock_{}", cheevo->id),
-                                     static_cast<float>(g_settings.achievements_notification_duration),
-                                     std::move(title), cheevo->description, std::move(badge_path));
+    GPUThread::RunOnThread([id = cheevo->id, duration = g_settings.achievements_notification_duration,
+                            title = std::move(title), description = std::string(cheevo->description),
+                            badge_path = std::move(badge_path)]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("achievement_unlock_{}", id), static_cast<float>(duration),
+                                       std::move(title), std::move(description), std::move(badge_path));
+    });
   }
 
   if (g_settings.achievements_sound_effects)
@@ -1148,7 +1165,7 @@ void Achievements::HandleGameCompleteEvent(const rc_client_event_t* event)
   INFO_LOG("Game {} complete", s_game_id);
   UpdateGameSummary();
 
-  if (g_settings.achievements_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_notifications)
   {
     std::string title = fmt::format(TRANSLATE_FS("Achievements", "Mastered {}"), s_game_title);
     std::string message =
@@ -1157,8 +1174,13 @@ void Achievements::HandleGameCompleteEvent(const rc_client_event_t* event)
                                        s_game_summary.num_unlocked_achievements),
                   TRANSLATE_PLURAL_STR("Achievements", "%n points", "Mastery popup", s_game_summary.points_unlocked));
 
+    GPUThread::RunOnThread([title = std::move(title), message = std::move(message), icon = s_game_icon]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
       ImGuiFullscreen::AddNotification("achievement_mastery", GAME_COMPLETE_NOTIFICATION_TIME, std::move(title),
-                                     std::move(message), s_game_icon);
+                                       std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1166,14 +1188,19 @@ void Achievements::HandleLeaderboardStartedEvent(const rc_client_event_t* event)
 {
   DEV_LOG("Leaderboard {} ({}) started", event->leaderboard->id, event->leaderboard->title);
 
-  if (g_settings.achievements_leaderboard_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_leaderboard_notifications)
   {
     std::string title = event->leaderboard->title;
     std::string message = TRANSLATE_STR("Achievements", "Leaderboard attempt started.");
 
-    ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", event->leaderboard->id),
-                                     LEADERBOARD_STARTED_NOTIFICATION_TIME, std::move(title), std::move(message),
-                                     s_game_icon);
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
+                            icon = s_game_icon]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id), LEADERBOARD_STARTED_NOTIFICATION_TIME,
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1181,14 +1208,19 @@ void Achievements::HandleLeaderboardFailedEvent(const rc_client_event_t* event)
 {
   DEV_LOG("Leaderboard {} ({}) failed", event->leaderboard->id, event->leaderboard->title);
 
-  if (g_settings.achievements_leaderboard_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_leaderboard_notifications)
   {
     std::string title = event->leaderboard->title;
     std::string message = TRANSLATE_STR("Achievements", "Leaderboard attempt failed.");
 
-    ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", event->leaderboard->id),
-                                     LEADERBOARD_FAILED_NOTIFICATION_TIME, std::move(title), std::move(message),
-                                     s_game_icon);
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
+                            icon = s_game_icon]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id), LEADERBOARD_FAILED_NOTIFICATION_TIME,
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1196,7 +1228,7 @@ void Achievements::HandleLeaderboardSubmittedEvent(const rc_client_event_t* even
 {
   DEV_LOG("Leaderboard {} ({}) submitted", event->leaderboard->id, event->leaderboard->title);
 
-  if (g_settings.achievements_leaderboard_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_leaderboard_notifications)
   {
     static const char* value_strings[NUM_RC_CLIENT_LEADERBOARD_FORMATS] = {
       TRANSLATE_NOOP("Achievements", "Your Time: {}{}"),
@@ -1212,9 +1244,14 @@ void Achievements::HandleLeaderboardSubmittedEvent(const rc_client_event_t* even
       event->leaderboard->tracker_value ? event->leaderboard->tracker_value : "Unknown",
       g_settings.achievements_spectator_mode ? std::string_view() : TRANSLATE_SV("Achievements", " (Submitting)"));
 
-    ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", event->leaderboard->id),
-                                     static_cast<float>(g_settings.achievements_leaderboard_duration), std::move(title),
-                                     std::move(message), s_game_icon);
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
+                            icon = s_game_icon]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id),
+                                       static_cast<float>(g_settings.achievements_leaderboard_duration),
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 
   if (g_settings.achievements_sound_effects)
@@ -1226,7 +1263,7 @@ void Achievements::HandleLeaderboardScoreboardEvent(const rc_client_event_t* eve
   DEV_LOG("Leaderboard {} scoreboard rank {} of {}", event->leaderboard_scoreboard->leaderboard_id,
           event->leaderboard_scoreboard->new_rank, event->leaderboard_scoreboard->num_entries);
 
-  if (g_settings.achievements_leaderboard_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_leaderboard_notifications)
   {
     static const char* value_strings[NUM_RC_CLIENT_LEADERBOARD_FORMATS] = {
       TRANSLATE_NOOP("Achievements", "Your Time: {} (Best: {})"),
@@ -1243,9 +1280,15 @@ void Achievements::HandleLeaderboardScoreboardEvent(const rc_client_event_t* eve
                   event->leaderboard_scoreboard->submitted_score, event->leaderboard_scoreboard->best_score),
       event->leaderboard_scoreboard->new_rank, event->leaderboard_scoreboard->num_entries);
 
-    ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", event->leaderboard->id),
-                                     static_cast<float>(g_settings.achievements_leaderboard_duration), std::move(title),
-                                     std::move(message), s_game_icon);
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
+                            icon = s_game_icon]() mutable {
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id),
+                                       static_cast<float>(g_settings.achievements_leaderboard_duration),
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1375,26 +1418,30 @@ void Achievements::HandleServerDisconnectedEvent(const rc_client_event_t* event)
 {
   WARNING_LOG("Server disconnected.");
 
-  if (FullscreenUI::Initialize())
-  {
+  GPUThread::RunOnThread([]() {
+    if (!FullscreenUI::Initialize())
+      return;
+
     ImGuiFullscreen::ShowToast(
       TRANSLATE_STR("Achievements", "Achievements Disconnected"),
       TRANSLATE_STR("Achievements",
                     "An unlock request could not be completed. We will keep retrying to submit this request."),
       Host::OSD_ERROR_DURATION);
-  }
+  });
 }
 
 void Achievements::HandleServerReconnectedEvent(const rc_client_event_t* event)
 {
   WARNING_LOG("Server reconnected.");
 
-  if (FullscreenUI::Initialize())
-  {
+  GPUThread::RunOnThread([]() {
+    if (!FullscreenUI::Initialize())
+      return;
+
     ImGuiFullscreen::ShowToast(TRANSLATE_STR("Achievements", "Achievements Reconnected"),
                                TRANSLATE_STR("Achievements", "All pending unlock requests have completed."),
                                Host::OSD_INFO_DURATION);
-  }
+  });
 }
 
 void Achievements::ResetClient()
@@ -1472,12 +1519,17 @@ void Achievements::SetHardcoreMode(bool enabled, bool force_display_message)
   // new mode
   s_hardcore_mode = enabled;
 
-  if (System::IsValid() && (HasActiveGame() || force_display_message) && FullscreenUI::Initialize())
+  if (System::IsValid() && (HasActiveGame() || force_display_message))
   {
+    GPUThread::RunOnThread([enabled]() {
+      if (!FullscreenUI::Initialize())
+        return;
+
       ImGuiFullscreen::ShowToast(std::string(),
                                  enabled ? TRANSLATE_STR("Achievements", "Hardcore mode is now enabled.") :
                                            TRANSLATE_STR("Achievements", "Hardcore mode is now disabled."),
                                  Host::OSD_INFO_DURATION);
+    });
   }
 
   rc_client_set_hardcore_enabled(s_client, enabled);
@@ -1806,7 +1858,7 @@ void Achievements::ShowLoginNotification()
   if (!user)
     return;
 
-  if (g_settings.achievements_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_notifications)
   {
     std::string badge_path = GetLoggedInUserBadgePath();
     std::string title = user->display_name;
@@ -1815,8 +1867,14 @@ void Achievements::ShowLoginNotification()
     std::string summary = fmt::format(TRANSLATE_FS("Achievements", "Score: {} ({} softcore)\nUnread messages: {}"),
                                       user->score, user->score_softcore, user->num_unread_messages);
 
+    GPUThread::RunOnThread(
+      [title = std::move(title), summary = std::move(summary), badge_path = std::move(badge_path)]() mutable {
+        if (!FullscreenUI::Initialize())
+          return;
+
         ImGuiFullscreen::AddNotification("achievements_login", LOGIN_NOTIFICATION_TIME, std::move(title),
                                          std::move(summary), std::move(badge_path));
+      });
   }
 }
 
@@ -1913,14 +1971,6 @@ void Achievements::ConfirmHardcoreModeDisableAsync(const char* trigger, std::fun
   }
 #endif
 
-  if (!FullscreenUI::Initialize())
-  {
-    Host::AddOSDMessage(fmt::format(TRANSLATE_FS("Achievements", "Cannot {} while hardcode mode is active."), trigger),
-                        Host::OSD_WARNING_DURATION);
-    callback(false);
-    return;
-  }
-
   auto real_callback = [callback = std::move(callback)](bool res) mutable {
     // don't run the callback in the middle of rendering the UI
     Host::RunOnCPUThread([callback = std::move(callback), res]() {
@@ -1930,13 +1980,25 @@ void Achievements::ConfirmHardcoreModeDisableAsync(const char* trigger, std::fun
     });
   };
 
+  GPUThread::RunOnThread([trigger = std::string(trigger), real_callback = std::move(real_callback)]() mutable {
+    if (!FullscreenUI::Initialize())
+    {
+      Host::AddOSDMessage(
+        fmt::format(TRANSLATE_FS("Achievements", "Cannot {} while hardcode mode is active."), trigger),
+        Host::OSD_WARNING_DURATION);
+      real_callback(false);
+      return;
+    }
+
     ImGuiFullscreen::OpenConfirmMessageDialog(
       TRANSLATE_STR("Achievements", "Confirm Hardcore Mode"),
-    fmt::format(TRANSLATE_FS("Achievements", "{0} cannot be performed while hardcore mode is active. Do you "
+      fmt::format(TRANSLATE_FS("Achievements",
+                               "{0} cannot be performed while hardcore mode is active. Do you "
                                "want to disable hardcore mode? {0} will be cancelled if you select No."),
                   trigger),
       std::move(real_callback), fmt::format(ICON_FA_CHECK " {}", TRANSLATE_SV("Achievements", "Yes")),
       fmt::format(ICON_FA_TIMES " {}", TRANSLATE_SV("Achievements", "No")));
+  });
 #else
   Host::AddOSDMessage(fmt::format(TRANSLATE_FS("Achievements", "Cannot {} while hardcode mode is active."), trigger),
                       Host::OSD_WARNING_DURATION);

src/core/core.vcxproj

@@ -49,7 +49,6 @@
     <ClCompile Include="gpu_hw_shadergen.cpp" />
     <ClCompile Include="gpu_shadergen.cpp" />
     <ClCompile Include="gpu_sw.cpp" />
-    <ClCompile Include="gpu_sw_backend.cpp" />
     <ClCompile Include="gpu_sw_rasterizer.cpp" />
     <ClCompile Include="gpu_sw_rasterizer_avx2.cpp">
       <EnableEnhancedInstructionSet>AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
@@ -57,6 +56,7 @@
       <ExcludedFromBuild Condition="'$(Platform)'!='x64'">true</ExcludedFromBuild>
       <PrecompiledHeader>NotUsing</PrecompiledHeader>
     </ClCompile>
+    <ClCompile Include="gpu_thread.cpp" />
     <ClCompile Include="gte.cpp" />
     <ClCompile Include="dma.cpp" />
     <ClCompile Include="gpu.cpp" />
@@ -134,8 +134,8 @@
     <ClInclude Include="gpu_hw_shadergen.h" />
     <ClInclude Include="gpu_shadergen.h" />
     <ClInclude Include="gpu_sw.h" />
-    <ClInclude Include="gpu_sw_backend.h" />
     <ClInclude Include="gpu_sw_rasterizer.h" />
+    <ClInclude Include="gpu_thread.h" />
     <ClInclude Include="gpu_types.h" />
     <ClInclude Include="gte.h" />
     <ClInclude Include="cpu_types.h" />

src/core/core.vcxproj.filters

@@ -46,7 +46,6 @@
     <ClCompile Include="analog_joystick.cpp" />
     <ClCompile Include="cpu_recompiler_code_generator_aarch32.cpp" />
     <ClCompile Include="gpu_backend.cpp" />
-    <ClCompile Include="gpu_sw_backend.cpp" />
     <ClCompile Include="texture_replacements.cpp" />
     <ClCompile Include="multitap.cpp" />
     <ClCompile Include="host.cpp" />
@@ -70,6 +69,7 @@
     <ClCompile Include="gdb_server.cpp" />
     <ClCompile Include="gpu_sw_rasterizer.cpp" />
     <ClCompile Include="gpu_sw_rasterizer_avx2.cpp" />
+    <ClCompile Include="gpu_thread.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="types.h" />
@@ -119,7 +119,6 @@
     <ClInclude Include="analog_joystick.h" />
     <ClInclude Include="gpu_types.h" />
     <ClInclude Include="gpu_backend.h" />
-    <ClInclude Include="gpu_sw_backend.h" />
     <ClInclude Include="texture_replacements.h" />
     <ClInclude Include="multitap.h" />
     <ClInclude Include="host.h" />
@@ -145,6 +144,7 @@
     <ClInclude Include="pine_server.h" />
     <ClInclude Include="gdb_server.h" />
     <ClInclude Include="gpu_sw_rasterizer.h" />
+    <ClInclude Include="gpu_thread.h" />
   </ItemGroup>
   <ItemGroup>
     <None Include="gpu_sw_rasterizer.inl" />

src/core/fullscreen_ui.cpp

@@ -12,6 +12,7 @@
 #include "cpu_core.h"
 #include "game_list.h"
 #include "gpu.h"
+#include "gpu_thread.h"
 #include "host.h"
 #include "resources.h"
 #include "settings.h"
@@ -594,6 +595,7 @@ bool FullscreenUI::Initialize()
   s_was_paused_on_quick_menu_open = false;
   s_about_window_open = false;
   s_hotkey_list_cache = InputManager::GetHotkeyList();
+  GPUThread::SetRunIdle(true);
 
   if (!System::IsValid())
     SwitchToLanding();
@@ -624,6 +626,7 @@ bool FullscreenUI::AreAnyDialogsOpen()
 
 void FullscreenUI::CheckForConfigChanges(const Settings& old_settings)
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
@@ -631,54 +634,114 @@ void FullscreenUI::CheckForConfigChanges(const Settings& old_settings)
   // That means we're going to be reading achievement state.
   if (old_settings.achievements_enabled && !g_settings.achievements_enabled)
   {
-    if (s_current_main_window == MainWindowType::Achievements || s_current_main_window == MainWindowType::Leaderboards)
+    if (!IsInitialized())
+      return;
+
+    GPUThread::RunOnThread([]() {
+      if (s_current_main_window == MainWindowType::Achievements ||
+          s_current_main_window == MainWindowType::Leaderboards)
+      {
         ReturnToPreviousWindow();
       }
+    });
+  }
 }
 
 void FullscreenUI::OnSystemStarted()
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
+      return;
+
+    GPUThread::SetRunIdle(false);
+
     s_current_main_window = MainWindowType::None;
     QueueResetFocus();
+  });
 }
 
 void FullscreenUI::OnSystemPaused()
 {
-  // noop
+  // NOTE: Called on CPU thread.
+  if (!IsInitialized())
+    return;
+
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
+      return;
+
+    GPUThread::SetRunIdle(true);
+  });
 }
 
 void FullscreenUI::OnSystemResumed()
 {
+  // NOTE: Called on CPU thread.
+  if (!IsInitialized())
+    return;
+
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
+      return;
+
+    GPUThread::SetRunIdle(false);
+
     // get rid of pause menu if we unpaused another way
     if (s_current_main_window == MainWindowType::PauseMenu)
       ClosePauseMenu();
+  });
 }
 
 void FullscreenUI::OnSystemDestroyed()
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
+      return;
+
+    // If we didn't start big picture before the system, shut ourselves down.
+    if (!GPUThread::WasFullscreenUIRequested())
+    {
+      Shutdown();
+      return;
+    }
+
+    GPUThread::SetRunIdle(true);
     s_pause_menu_was_open = false;
     s_was_paused_on_quick_menu_open = false;
     s_current_pause_submenu = PauseSubMenu::None;
     SwitchToLanding();
+  });
 }
 
 void FullscreenUI::OnRunningGameChanged()
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
   const std::string& path = System::GetDiscPath();
   const std::string& serial = System::GetGameSerial();
+
+  std::string subtitle;
   if (!serial.empty())
-    s_current_game_subtitle = fmt::format("{0} - {1}", serial, Path::GetFileName(path));
+    subtitle = fmt::format("{0} - {1}", serial, Path::GetFileName(path));
   else
-    s_current_game_subtitle = {};
+    subtitle = {};
+
+  GPUThread::RunOnThread([subtitle = std::move(subtitle)]() mutable {
+    if (!IsInitialized())
+      return;
+
+    s_current_game_subtitle = std::move(subtitle);
+  });
 }
 
 void FullscreenUI::PauseForMenuOpen(bool set_pause_menu_open)
@@ -695,6 +758,7 @@ void FullscreenUI::OpenPauseMenu()
   if (!System::IsValid())
     return;
 
+  GPUThread::RunOnThread([]() {
     if (!Initialize() || s_current_main_window != MainWindowType::None)
       return;
 
@@ -704,6 +768,7 @@ void FullscreenUI::OpenPauseMenu()
     QueueResetFocus();
     ForceKeyNavEnabled();
     FixStateIfPaused();
+  });
 }
 
 void FullscreenUI::FixStateIfPaused()
@@ -713,21 +778,15 @@ void FullscreenUI::FixStateIfPaused()
 
   // When we're paused, we won't have trickled the key up event for escape yet. Do it now.
   ImGui::UpdateInputEvents(false);
-
-  Host::OnIdleStateChanged();
-  Host::RunOnCPUThread([]() {
-    if (System::IsValid())
-    {
-      // Why twice? To clear the "wants keyboard input" flag.
-      System::InvalidateDisplay();
-      System::InvalidateDisplay();
-    }
-  });
 }
 
 void FullscreenUI::ClosePauseMenu()
 {
-  if (!IsInitialized() || !System::IsValid())
+  if (!System::IsValid())
+    return;
+
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
       return;
 
     if (System::GetState() == System::State::Paused && !s_was_paused_on_quick_menu_open)
@@ -738,6 +797,7 @@ void FullscreenUI::ClosePauseMenu()
     s_pause_menu_was_open = false;
     QueueResetFocus();
     FixStateIfPaused();
+  });
 }
 
 void FullscreenUI::OpenPauseSubMenu(PauseSubMenu submenu)
@@ -749,6 +809,8 @@ void FullscreenUI::OpenPauseSubMenu(PauseSubMenu submenu)
 
 void FullscreenUI::Shutdown()
 {
+  GPUThread::SetRunIdle(false);
+
   Achievements::ClearUIState();
   CloseSaveStateSelector();
   s_cover_image_map.clear();
@@ -1135,6 +1197,7 @@ void FullscreenUI::DoChangeDiscFromFile()
 
 void FullscreenUI::DoChangeDisc()
 {
+  Host::RunOnCPUThread([]() {
     ImGuiFullscreen::ChoiceDialogOptions options;
 
     if (System::HasMediaSubImages())
@@ -1147,6 +1210,7 @@ void FullscreenUI::DoChangeDisc()
       for (u32 i = 0; i < count; i++)
         options.emplace_back(System::GetMediaSubImageTitle(i), i == current_index);
 
+      GPUThread::RunOnThread([options = std::move(options)]() mutable {
         auto callback = [](s32 index, const std::string& title, bool checked) {
           if (index == 0)
           {
@@ -1156,7 +1220,7 @@ void FullscreenUI::DoChangeDisc()
           }
           else if (index > 0)
           {
-        System::SwitchMediaSubImage(static_cast<u32>(index - 1));
+            Host::RunOnCPUThread([index = static_cast<u32>(index - 1)]() { System::SwitchMediaSubImage(index); });
           }
 
           QueueResetFocus();
@@ -1166,6 +1230,7 @@ void FullscreenUI::DoChangeDisc()
 
         OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
                          std::move(callback));
+      });
 
       return;
     }
@@ -1189,6 +1254,7 @@ void FullscreenUI::DoChangeDisc()
           paths.push_back(glentry->path);
         }
 
+        GPUThread::RunOnThread([options = std::move(options), paths = std::move(paths)]() mutable {
           auto callback = [paths = std::move(paths)](s32 index, const std::string& title, bool checked) {
             if (index == 0)
             {
@@ -1198,7 +1264,7 @@ void FullscreenUI::DoChangeDisc()
             }
             else if (index > 0)
             {
-          System::InsertMedia(paths[index - 1].c_str());
+              Host::RunOnCPUThread([path = std::move(paths[index - 1])]() { System::InsertMedia(path.c_str()); });
             }
 
             QueueResetFocus();
@@ -1208,16 +1274,19 @@ void FullscreenUI::DoChangeDisc()
 
           OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
                            std::move(callback));
+        });
 
         return;
       }
     }
 
-  DoChangeDiscFromFile();
+    GPUThread::RunOnThread([]() { DoChangeDiscFromFile(); });
+  });
 }
 
 void FullscreenUI::DoCheatsMenu()
 {
+  Host::RunOnCPUThread([]() {
     CheatList* cl = System::GetCheatList();
     if (!cl)
     {
@@ -1238,6 +1307,7 @@ void FullscreenUI::DoCheatsMenu()
       options.emplace_back(cc.description.c_str(), cc.enabled);
     }
 
+    GPUThread::RunOnThread([options = std::move(options)]() mutable {
       auto callback = [](s32 index, const std::string& title, bool checked) {
         if (index < 0)
         {
@@ -1245,6 +1315,7 @@ void FullscreenUI::DoCheatsMenu()
           return;
         }
 
+        Host::RunOnCPUThread([index, checked]() {
           CheatList* cl = System::GetCheatList();
           if (!cl)
             return;
@@ -1254,13 +1325,17 @@ void FullscreenUI::DoCheatsMenu()
             cl->ApplyCode(static_cast<u32>(index));
           else
             System::SetCheatCodeState(static_cast<u32>(index), checked);
+        });
       };
       OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_FROWN, "Cheat List"), true, std::move(options), std::move(callback));
+    });
+  });
 }
 
 void FullscreenUI::DoToggleAnalogMode()
 {
   // hacky way to toggle analog mode
+  Host::RunOnCPUThread([]() {
     for (u32 i = 0; i < NUM_CONTROLLER_AND_CARD_PORTS; i++)
     {
       Controller* ctrl = System::GetController(i);
@@ -1281,6 +1356,7 @@ void FullscreenUI::DoToggleAnalogMode()
         }
       }
     }
+  });
 }
 
 void FullscreenUI::DoRequestExit()
@@ -3720,12 +3796,9 @@ void FullscreenUI::DrawControllerSettingsPage()
                   &Settings::GetMultitapModeName, &Settings::GetMultitapModeDisplayName, MultitapMode::Count);
 
   // load mtap settings
-  MultitapMode mtap_mode = g_settings.multitap_mode;
-  if (IsEditingGameSettings(bsi))
-  {
-    mtap_mode = Settings::ParseMultitapModeName(bsi->GetTinyStringValue("ControllerPorts", "MultitapMode", "").c_str())
-                  .value_or(g_settings.multitap_mode);
-  }
+  const MultitapMode mtap_mode =
+    Settings::ParseMultitapModeName(bsi->GetTinyStringValue("ControllerPorts", "MultitapMode", "").c_str())
+      .value_or(Settings::DEFAULT_MULTITAP_MODE);
   const std::array<bool, 2> mtap_enabled = {
     {(mtap_mode == MultitapMode::Port1Only || mtap_mode == MultitapMode::BothPorts),
      (mtap_mode == MultitapMode::Port2Only || mtap_mode == MultitapMode::BothPorts)}};
@@ -7076,22 +7149,25 @@ void FullscreenUI::DrawAboutWindow()
 
 void FullscreenUI::OpenAchievementsWindow()
 {
+  if (!System::IsValid())
+    return;
+
   if (!Achievements::IsActive())
   {
     Host::AddKeyedOSDMessage("achievements_disabled", FSUI_STR("Achievements are not enabled."),
                              Host::OSD_INFO_DURATION);
     return;
   }
-
-  if (!System::IsValid() || !Initialize())
-    return;
-
-  if (!Achievements::HasAchievements() || !Achievements::PrepareAchievementsWindow())
+  else if (!Achievements::HasAchievements())
   {
     ShowToast(std::string(), FSUI_STR("This game has no achievements."));
     return;
   }
 
+  GPUThread::RunOnThread([]() {
+    if (!Initialize() || !Achievements::PrepareAchievementsWindow())
+      return;
+
     if (s_current_main_window != MainWindowType::PauseMenu)
     {
       PauseForMenuOpen(false);
@@ -7101,6 +7177,7 @@ void FullscreenUI::OpenAchievementsWindow()
     s_current_main_window = MainWindowType::Achievements;
     QueueResetFocus();
     FixStateIfPaused();
+  });
 }
 
 bool FullscreenUI::IsAchievementsWindowOpen()
@@ -7110,22 +7187,25 @@ bool FullscreenUI::IsAchievementsWindowOpen()
 
 void FullscreenUI::OpenLeaderboardsWindow()
 {
+  if (!System::IsValid())
+    return;
+
   if (!Achievements::IsActive())
   {
     Host::AddKeyedOSDMessage("achievements_disabled", FSUI_STR("Leaderboards are not enabled."),
                              Host::OSD_INFO_DURATION);
     return;
   }
-
-  if (!System::IsValid() || !Initialize())
-    return;
-
-  if (!Achievements::HasLeaderboards() || !Achievements::PrepareLeaderboardsWindow())
+  else if (!Achievements::HasLeaderboards())
   {
     ShowToast(std::string(), FSUI_STR("This game has no leaderboards."));
     return;
   }
 
+  GPUThread::RunOnThread([]() {
+    if (!Initialize() || !Achievements::PrepareLeaderboardsWindow())
+      return;
+
     if (s_current_main_window != MainWindowType::PauseMenu)
     {
       PauseForMenuOpen(false);
@@ -7135,6 +7215,7 @@ void FullscreenUI::OpenLeaderboardsWindow()
     s_current_main_window = MainWindowType::Leaderboards;
     QueueResetFocus();
     FixStateIfPaused();
+  });
 }
 
 bool FullscreenUI::IsLeaderboardsWindowOpen()

src/core/gpu.h

@@ -28,14 +28,11 @@ class GPUDevice;
 class GPUTexture;
 class GPUPipeline;
 
+class GPUBackend;
 struct Settings;
 class TimingEvent;
 
-namespace Threading {
-class Thread;
-}
-
-class GPU
+class GPU final
 {
 public:
   enum class BlitterState : u8
@@ -60,7 +57,6 @@ public:
     DOT_TIMER_INDEX = 0,
     HBLANK_TIMER_INDEX = 1,
     MAX_RESOLUTION_SCALE = 32,
-    DEINTERLACE_BUFFER_COUNT = 4,
     DRAWING_AREA_COORD_MASK = 1023,
   };
 
@@ -86,25 +82,14 @@ public:
 
   // Base class constructor.
   GPU();
-  virtual ~GPU();
+  ~GPU();
 
-  virtual const Threading::Thread* GetSWThread() const = 0;
-  virtual bool IsHardwareRenderer() const = 0;
-
-  virtual bool Initialize();
-  virtual void Reset(bool clear_vram);
-  virtual bool DoState(StateWrapper& sw, GPUTexture** save_to_texture, bool update_display);
-
-  // Graphics API state reset/restore - call when drawing the UI etc.
-  // TODO: replace with "invalidate cached state"
-  virtual void RestoreDeviceContext();
+  void Initialize();
+  void Reset(bool clear_vram);
+  bool DoState(StateWrapper& sw, GPUTexture** save_to_texture, bool update_display);
 
   // Render statistics debug window.
   void DrawDebugStateWindow();
-  void GetStatsString(SmallStringBase& str);
-  void GetMemoryStatsString(SmallStringBase& str);
-  void ResetStatistics();
-  void UpdateStatistics(u32 frame_count);
 
   void CPUClockChanged();
 
@@ -160,24 +145,12 @@ public:
   void SynchronizeCRTC();
 
   /// Recompile shaders/recreate framebuffers when needed.
-  virtual void UpdateSettings(const Settings& old_settings);
-
-  /// Updates the resolution scale when it's set to automatic.
-  virtual void UpdateResolutionScale();
-
-  /// Returns the effective display resolution of the GPU.
-  virtual std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true);
-
-  /// Returns the full display resolution of the GPU, including padding.
-  virtual std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true);
+  void UpdateSettings(const Settings& old_settings);
 
   float ComputeHorizontalFrequency() const;
   float ComputeVerticalFrequency() const;
   float ComputeDisplayAspectRatio() const;
 
-  static std::unique_ptr<GPU> CreateHardwareRenderer();
-  static std::unique_ptr<GPU> CreateSoftwareRenderer();
-
   // Converts window coordinates into horizontal ticks and scanlines. Returns false if out of range. Used for lightguns.
   void ConvertScreenCoordinatesToDisplayCoordinates(float window_x, float window_y, float* display_x,
                                                     float* display_y) const;
@@ -203,30 +176,7 @@ public:
   // Dumps raw VRAM to a file.
   bool DumpVRAMToFile(const char* filename);
 
-  // Ensures all buffered vertices are drawn.
-  virtual void FlushRender() = 0;
-
-  /// Helper function for computing the draw rectangle in a larger window.
-  GSVector4i CalculateDrawRect(s32 window_width, s32 window_height, bool apply_aspect_ratio = true) const;
-
-  /// Helper function to save current display texture to PNG.
-  bool WriteDisplayTextureToFile(std::string filename, bool compress_on_thread = false);
-
-  /// Renders the display, optionally with postprocessing to the specified image.
-  bool RenderScreenshotToBuffer(u32 width, u32 height, const GSVector4i draw_rect, bool postfx,
-                                std::vector<u32>* out_pixels, u32* out_stride, GPUTexture::Format* out_format);
-
-  /// Helper function to save screenshot to PNG.
-  bool RenderScreenshotToFile(std::string filename, DisplayScreenshotMode mode, u8 quality, bool compress_on_thread,
-                              bool show_osd_message);
-
-  /// Draws the current display texture, with any post-processing.
-  bool PresentDisplay();
-
-  /// Reads the CLUT from the specified coordinates, accounting for wrap-around.
-  static void ReadCLUT(u16* dest, GPUTexturePaletteReg reg, bool clut_is_8bit);
-
-protected:
+private:
   TickCount CRTCTicksToSystemTicks(TickCount crtc_ticks, TickCount fractional_ticks) const;
   TickCount SystemTicksToCRTCTicks(TickCount sysclk_ticks, TickCount* fractional_ticks) const;
 
@@ -237,16 +187,6 @@ protected:
   }
   ALWAYS_INLINE static constexpr TickCount SystemTicksToGPUTicks(TickCount sysclk_ticks) { return sysclk_ticks << 1; }
 
-  static constexpr std::tuple<u8, u8> UnpackTexcoord(u16 texcoord)
-  {
-    return std::make_tuple(static_cast<u8>(texcoord), static_cast<u8>(texcoord >> 8));
-  }
-
-  static constexpr std::tuple<u8, u8, u8> UnpackColorRGB24(u32 rgb24)
-  {
-    return std::make_tuple(static_cast<u8>(rgb24), static_cast<u8>(rgb24 >> 8), static_cast<u8>(rgb24 >> 16));
-  }
-
   static bool DumpVRAMToFile(const char* filename, u32 width, u32 height, u32 stride, const void* buffer,
                              bool remove_alpha);
 
@@ -270,10 +210,10 @@ protected:
   void CommandTickEvent(TickCount ticks);
 
   /// Returns 0 if the currently-displayed field is on odd lines (1,3,5,...) or 1 if even (2,4,6,...).
-  ALWAYS_INLINE u32 GetInterlacedDisplayField() const { return ZeroExtend32(m_crtc_state.interlaced_field); }
+  ALWAYS_INLINE u8 GetInterlacedDisplayField() const { return m_crtc_state.interlaced_field; }
 
   /// Returns 0 if the currently-displayed field is on an even line in VRAM, otherwise 1.
-  ALWAYS_INLINE u32 GetActiveLineLSB() const { return ZeroExtend32(m_crtc_state.active_line_lsb); }
+  ALWAYS_INLINE u8 GetActiveLineLSB() const { return m_crtc_state.active_line_lsb; }
 
   /// Updates drawing area that's suitablef or clamping.
   void SetClampedDrawingArea();
@@ -308,16 +248,15 @@ protected:
   void InvalidateCLUT();
   bool IsCLUTValid() const;
 
-  // Rendering in the backend
-  virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height);
-  virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color);
-  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask);
-  virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height);
-  virtual void DispatchRenderCommand() = 0;
-  virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
-  virtual void UpdateDisplay() = 0;
-  virtual void DrawRendererStats();
-  virtual void OnBufferSwapped();
+  void ReadVRAM(u16 x, u16 y, u16 width, u16 height);
+  void UpdateVRAM(u16 x, u16 y, u16 width, u16 height, const void* data, bool set_mask, bool check_mask);
+
+  void UpdateDisplay(bool present_frame);
+
+  void PrepareForDraw();
+  void FinishPolyline();
+  void FillBackendCommandParameters(GPUBackendCommand* cmd) const;
+  void FillDrawCommand(GPUBackendDrawCommand* cmd, GPURenderCommand rc) const;
 
   ALWAYS_INLINE_RELEASE void AddDrawTriangleTicks(GSVector2i v1, GSVector2i v2, GSVector2i v3, bool shaded,
                                                   bool textured, bool semitransparent)
@@ -443,19 +382,10 @@ protected:
     u32 texture_window_value;
 
     // decoded values
+    // TODO: Make this a command
     GPUTextureWindow texture_window;
     bool texture_x_flip;
     bool texture_y_flip;
-    bool texture_page_changed;
-    bool texture_window_changed;
-
-    ALWAYS_INLINE bool IsTexturePageChanged() const { return texture_page_changed; }
-    ALWAYS_INLINE void SetTexturePageChanged() { texture_page_changed = true; }
-    ALWAYS_INLINE void ClearTexturePageChangedFlag() { texture_page_changed = false; }
-
-    ALWAYS_INLINE bool IsTextureWindowChanged() const { return texture_window_changed; }
-    ALWAYS_INLINE void SetTextureWindowChanged() { texture_window_changed = true; }
-    ALWAYS_INLINE void ClearTextureWindowChangedFlag() { texture_window_changed = false; }
   } m_draw_mode = {};
 
   GPUDrawingArea m_drawing_area = {};
@@ -587,64 +517,7 @@ protected:
   TickCount m_max_run_ahead = 128;
   u32 m_fifo_size = 128;
 
-  void ClearDisplayTexture();
-  void SetDisplayTexture(GPUTexture* texture, GPUTexture* depth_texture, s32 view_x, s32 view_y, s32 view_width,
-                         s32 view_height);
-
-  bool RenderDisplay(GPUTexture* target, const GSVector4i draw_rect, bool postfx);
-
-  bool Deinterlace(u32 field, u32 line_skip);
-  bool DeinterlaceExtractField(u32 dst_bufidx, GPUTexture* src, u32 x, u32 y, u32 width, u32 height, u32 line_skip);
-  bool DeinterlaceSetTargetSize(u32 width, u32 height, bool preserve);
-  void DestroyDeinterlaceTextures();
-  bool ApplyChromaSmoothing();
-
-  u32 m_current_deinterlace_buffer = 0;
-  std::unique_ptr<GPUPipeline> m_deinterlace_pipeline;
-  std::unique_ptr<GPUPipeline> m_deinterlace_extract_pipeline;
-  std::array<std::unique_ptr<GPUTexture>, DEINTERLACE_BUFFER_COUNT> m_deinterlace_buffers;
-  std::unique_ptr<GPUTexture> m_deinterlace_texture;
-
-  std::unique_ptr<GPUPipeline> m_chroma_smoothing_pipeline;
-  std::unique_ptr<GPUTexture> m_chroma_smoothing_texture;
-
-  std::unique_ptr<GPUPipeline> m_display_pipeline;
-  GPUTexture* m_display_texture = nullptr;
-  GPUTexture* m_display_depth_buffer = nullptr;
-  s32 m_display_texture_view_x = 0;
-  s32 m_display_texture_view_y = 0;
-  s32 m_display_texture_view_width = 0;
-  s32 m_display_texture_view_height = 0;
-
-  struct Counters
-  {
-    u32 num_reads;
-    u32 num_writes;
-    u32 num_copies;
-    u32 num_vertices;
-    u32 num_primitives;
-
-    // u32 num_read_texture_updates;
-    // u32 num_ubo_updates;
-  };
-
-  struct Stats : Counters
-  {
-    size_t host_buffer_streamed;
-    u32 host_num_draws;
-    u32 host_num_barriers;
-    u32 host_num_render_passes;
-    u32 host_num_copies;
-    u32 host_num_downloads;
-    u32 host_num_uploads;
-  };
-
-  Counters m_counters = {};
-  Stats m_stats = {};
-
 private:
-  bool CompileDisplayPipelines(bool display, bool deinterlace, bool chroma_smoothing);
-
   using GP0CommandHandler = bool (GPU::*)();
   using GP0CommandHandlerTable = std::array<GP0CommandHandler, 256>;
   static GP0CommandHandlerTable GenerateGP0CommandHandlerTable();

src/core/gpu_backend.h

@@ -2,91 +2,182 @@
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
-#include "common/heap_array.h"
-#include "common/threading.h"
-#include "gpu_types.h"
-#include <atomic>
-#include <condition_variable>
-#include <memory>
-#include <mutex>
-#include <thread>
 
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable : 4324) // warning C4324: 'GPUBackend': structure was padded due to alignment specifier
-#endif
+#include "gpu_types.h"
+
+#include "util/gpu_texture.h"
+
+#include <tuple>
+
+class Error;
+class SmallStringBase;
+
+class GPUFramebuffer;
+class GPUPipeline;
+
+struct Settings;
+class StateWrapper;
+
+// DESIGN NOTE: Only static methods should be called on the CPU thread.
+// You specifically don't have a global pointer available for this reason.
 
 class GPUBackend
 {
+public:
+  static GPUThreadCommand* NewClearVRAMCommand();
+  static GPUBackendDoStateCommand* NewDoStateCommand();
+  static GPUThreadCommand* NewClearDisplayCommand();
+  static GPUBackendUpdateDisplayCommand* NewUpdateDisplayCommand();
+  static GPUThreadCommand* NewClearCacheCommand();
+  static GPUThreadCommand* NewBufferSwappedCommand();
+  static GPUThreadCommand* NewFlushRenderCommand();
+  static GPUThreadCommand* NewUpdateResolutionScaleCommand();
+  static GPUBackendReadVRAMCommand* NewReadVRAMCommand();
+  static GPUBackendFillVRAMCommand* NewFillVRAMCommand();
+  static GPUBackendUpdateVRAMCommand* NewUpdateVRAMCommand(u32 num_words);
+  static GPUBackendCopyVRAMCommand* NewCopyVRAMCommand();
+  static GPUBackendSetDrawingAreaCommand* NewSetDrawingAreaCommand();
+  static GPUBackendUpdateCLUTCommand* NewUpdateCLUTCommand();
+  static GPUBackendDrawPolygonCommand* NewDrawPolygonCommand(u32 num_vertices);
+  static GPUBackendDrawPrecisePolygonCommand* NewDrawPrecisePolygonCommand(u32 num_vertices);
+  static GPUBackendDrawRectangleCommand* NewDrawRectangleCommand();
+  static GPUBackendDrawLineCommand* NewDrawLineCommand(u32 num_vertices);
+  static void PushCommand(GPUThreadCommand* cmd);
+  static void PushCommandAndWakeThread(GPUThreadCommand* cmd);
+  static void PushCommandAndSync(GPUThreadCommand* cmd, bool spin);
+
+  static bool IsUsingHardwareBackend();
+
+  static std::unique_ptr<GPUBackend> CreateHardwareBackend();
+  static std::unique_ptr<GPUBackend> CreateSoftwareBackend();
+
+  static bool BeginQueueFrame();
+  static void WaitForOneQueuedFrame();
+
+  static bool RenderScreenshotToBuffer(u32 width, u32 height, const GSVector4i draw_rect, bool postfx,
+                                       std::vector<u32>* out_pixels, u32* out_stride, GPUTexture::Format* out_format);
+
+  static std::tuple<u32, u32> GetLastDisplaySourceSize();
+
+  static void GetStatsString(SmallStringBase& str);
+  static void GetMemoryStatsString(SmallStringBase& str);
+  static void ResetStatistics();
+  static void UpdateStatistics(u32 frame_count);
+
 public:
   GPUBackend();
   virtual ~GPUBackend();
 
-  ALWAYS_INLINE const Threading::Thread* GetThread() const { return m_use_gpu_thread ? &m_gpu_thread : nullptr; }
+  ALWAYS_INLINE const void* GetDisplayTextureHandle() const { return m_display_texture; }
+  ALWAYS_INLINE s32 GetDisplayWidth() const { return m_display_width; }
+  ALWAYS_INLINE s32 GetDisplayHeight() const { return m_display_height; }
+  ALWAYS_INLINE s32 GetDisplayViewWidth() const { return m_display_texture_view_width; }
+  ALWAYS_INLINE s32 GetDisplayViewHeight() const { return m_display_texture_view_height; }
+  ALWAYS_INLINE float GetDisplayAspectRatio() const { return m_display_aspect_ratio; }
+  ALWAYS_INLINE bool HasDisplayTexture() const { return static_cast<bool>(m_display_texture); }
 
-  virtual bool Initialize(bool force_thread);
-  virtual void UpdateSettings();
-  virtual void Reset();
-  virtual void Shutdown();
+  virtual bool Initialize(bool clear_vram, Error* error);
 
-  GPUBackendFillVRAMCommand* NewFillVRAMCommand();
-  GPUBackendUpdateVRAMCommand* NewUpdateVRAMCommand(u32 num_words);
-  GPUBackendCopyVRAMCommand* NewCopyVRAMCommand();
-  GPUBackendSetDrawingAreaCommand* NewSetDrawingAreaCommand();
-  GPUBackendUpdateCLUTCommand* NewUpdateCLUTCommand();
-  GPUBackendDrawPolygonCommand* NewDrawPolygonCommand(u32 num_vertices);
-  GPUBackendDrawRectangleCommand* NewDrawRectangleCommand();
-  GPUBackendDrawLineCommand* NewDrawLineCommand(u32 num_vertices);
+  virtual void ClearVRAM() = 0;
+  virtual bool DoState(GPUTexture** host_texture, bool is_reading, bool update_display) = 0;
 
-  void PushCommand(GPUBackendCommand* cmd);
-  void Sync(bool allow_sleep);
-
-  /// Processes all pending GPU commands.
-  void RunGPULoop();
-
-protected:
-  void* AllocateCommand(GPUBackendCommandType command, u32 size);
-  u32 GetPendingCommandSize() const;
-  void WakeGPUThread();
-  void StartGPUThread();
-  void StopGPUThread();
-
-  virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params) = 0;
-  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data,
-                          GPUBackendCommandParameters params) = 0;
+  virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height) = 0;
+  virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params);
+  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, GPUBackendCommandParameters params);
   virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
-                        GPUBackendCommandParameters params) = 0;
+                        GPUBackendCommandParameters params);
+
   virtual void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) = 0;
-  virtual void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd) = 0;
+  virtual void DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd) = 0;
+  virtual void DrawSprite(const GPUBackendDrawRectangleCommand* cmd) = 0;
   virtual void DrawLine(const GPUBackendDrawLineCommand* cmd) = 0;
-  virtual void FlushRender() = 0;
+
   virtual void DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area) = 0;
   virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
+  virtual void ClearCache() = 0;
+  virtual void OnBufferSwapped() = 0;
 
-  void HandleCommand(const GPUBackendCommand* cmd);
+  virtual void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) = 0;
 
-  Threading::KernelSemaphore m_sync_semaphore;
-  std::atomic_bool m_gpu_thread_sleeping{false};
-  std::atomic_bool m_gpu_loop_done{false};
-  Threading::Thread m_gpu_thread;
-  bool m_use_gpu_thread = false;
+  virtual void UpdateSettings(const Settings& old_settings);
 
-  std::mutex m_sync_mutex;
-  std::condition_variable m_sync_cpu_thread_cv;
-  std::condition_variable m_wake_gpu_thread_cv;
-  bool m_sync_done = false;
+  /// Returns the effective display resolution of the GPU.
+  virtual std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) const = 0;
 
+  /// Returns the full display resolution of the GPU, including padding.
+  virtual std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) const = 0;
+
+  /// TODO: Updates the resolution scale when it's set to automatic.
+  virtual void UpdateResolutionScale() = 0;
+
+  /// Ensures all pending draws are flushed to the host GPU.
+  virtual void FlushRender() = 0;
+
+  // Graphics API state reset/restore - call when drawing the UI etc.
+  // TODO: replace with "invalidate cached state"
+  virtual void RestoreDeviceContext() = 0;
+
+  void HandleCommand(const GPUThreadCommand* cmd);
+
+  /// Draws the current display texture, with any post-processing.
+  bool PresentDisplay();
+
+protected:
   enum : u32
   {
-    COMMAND_QUEUE_SIZE = 4 * 1024 * 1024,
-    THRESHOLD_TO_WAKE_GPU = 256
+    DEINTERLACE_BUFFER_COUNT = 4,
   };
 
-  FixedHeapArray<u8, COMMAND_QUEUE_SIZE> m_command_fifo_data;
-  alignas(HOST_CACHE_LINE_SIZE) std::atomic<u32> m_command_fifo_read_ptr{0};
-  alignas(HOST_CACHE_LINE_SIZE) std::atomic<u32> m_command_fifo_write_ptr{0};
-};
+  /// Helper function for computing the draw rectangle in a larger window.
+  GSVector4i CalculateDrawRect(s32 window_width, s32 window_height, bool apply_aspect_ratio = true) const;
 
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif
+  /// Helper function to save current display texture to PNG.
+  bool WriteDisplayTextureToFile(std::string filename, bool compress_on_thread = false);
+
+  /// Renders the display, optionally with postprocessing to the specified image.
+  void HandleRenderScreenshotToBuffer(const GPUThreadRenderScreenshotToBufferCommand* cmd);
+
+  /// Helper function to save screenshot to PNG.
+  bool RenderScreenshotToFile(std::string filename, DisplayScreenshotMode mode, u8 quality, bool compress_on_thread,
+                              bool show_osd_message);
+
+  bool CompileDisplayPipelines(bool display, bool deinterlace, bool chroma_smoothing);
+
+  void ClearDisplay();
+  void ClearDisplayTexture();
+  void SetDisplayTexture(GPUTexture* texture, GPUTexture* depth_texture, s32 view_x, s32 view_y, s32 view_width,
+                         s32 view_height);
+
+  bool RenderDisplay(GPUTexture* target, const GSVector4i draw_rect, bool postfx);
+
+  bool Deinterlace(u32 field, u32 line_skip);
+  bool DeinterlaceExtractField(u32 dst_bufidx, GPUTexture* src, u32 x, u32 y, u32 width, u32 height, u32 line_skip);
+  bool DeinterlaceSetTargetSize(u32 width, u32 height, bool preserve);
+  void DestroyDeinterlaceTextures();
+  bool ApplyChromaSmoothing();
+
+  s32 m_display_width = 0;
+  s32 m_display_height = 0;
+  s32 m_display_origin_left = 0;
+  s32 m_display_origin_top = 0;
+  s32 m_display_vram_width = 0;
+  s32 m_display_vram_height = 0;
+  float m_display_aspect_ratio = 1.0f;
+
+  u32 m_current_deinterlace_buffer = 0;
+  std::unique_ptr<GPUPipeline> m_deinterlace_pipeline;
+  std::unique_ptr<GPUPipeline> m_deinterlace_extract_pipeline;
+  std::array<std::unique_ptr<GPUTexture>, DEINTERLACE_BUFFER_COUNT> m_deinterlace_buffers;
+  std::unique_ptr<GPUTexture> m_deinterlace_texture;
+
+  std::unique_ptr<GPUPipeline> m_chroma_smoothing_pipeline;
+  std::unique_ptr<GPUTexture> m_chroma_smoothing_texture;
+
+  std::unique_ptr<GPUPipeline> m_display_pipeline;
+  GPUTexture* m_display_texture = nullptr;
+  GPUTexture* m_display_depth_buffer = nullptr;
+  s32 m_display_texture_view_x = 0;
+  s32 m_display_texture_view_y = 0;
+  s32 m_display_texture_view_width = 0;
+  s32 m_display_texture_view_height = 0;
+};

src/core/gpu_commands.cpp

@@ -1,13 +1,18 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2023 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
-#include "common/assert.h"
-#include "common/log.h"
-#include "common/string_util.h"
+#include "cpu_pgxp.h"
 #include "gpu.h"
+#include "gpu_backend.h"
 #include "interrupt_controller.h"
 #include "system.h"
 #include "texture_replacements.h"
+
+#include "common/assert.h"
+#include "common/gsvector_formatter.h"
+#include "common/log.h"
+#include "common/string_util.h"
+
 Log_SetChannel(GPU);
 
 #define CHECK_COMMAND_SIZE(num_words)                                                                                  \
@@ -90,7 +95,7 @@ void GPU::TryExecuteCommands()
           // drop terminator
           m_fifo.RemoveOne();
           DEBUG_LOG("Drawing poly-line with {} vertices", GetPolyLineVertexCount());
-          DispatchRenderCommand();
+          FinishPolyline();
           m_blit_buffer.clear();
           EndCommand();
           continue;
@@ -197,8 +202,8 @@ bool GPU::HandleNOPCommand()
 bool GPU::HandleClearCacheCommand()
 {
   DEBUG_LOG("GP0 clear cache");
-  m_draw_mode.SetTexturePageChanged();
   InvalidateCLUT();
+  GPUBackend::PushCommand(GPUBackend::NewClearCacheCommand());
   m_fifo.RemoveOne();
   AddCommandTicks(1);
   EndCommand();
@@ -245,8 +250,6 @@ bool GPU::HandleSetDrawingAreaTopLeftCommand()
   DEBUG_LOG("Set drawing area top-left: ({}, {})", left, top);
   if (m_drawing_area.left != left || m_drawing_area.top != top)
   {
-    FlushRender();
-
     m_drawing_area.left = left;
     m_drawing_area.top = top;
     m_drawing_area_changed = true;
@@ -267,8 +270,6 @@ bool GPU::HandleSetDrawingAreaBottomRightCommand()
   DEBUG_LOG("Set drawing area bottom-right: ({}, {})", m_drawing_area.right, m_drawing_area.bottom);
   if (m_drawing_area.right != right || m_drawing_area.bottom != bottom)
   {
-    FlushRender();
-
     m_drawing_area.right = right;
     m_drawing_area.bottom = bottom;
     m_drawing_area_changed = true;
@@ -288,8 +289,6 @@ bool GPU::HandleSetDrawingOffsetCommand()
   DEBUG_LOG("Set drawing offset ({}, {})", m_drawing_offset.x, m_drawing_offset.y);
   if (m_drawing_offset.x != x || m_drawing_offset.y != y)
   {
-    FlushRender();
-
     m_drawing_offset.x = x;
     m_drawing_offset.y = y;
   }
@@ -305,11 +304,7 @@ bool GPU::HandleSetMaskBitCommand()
 
   constexpr u32 gpustat_mask = (1 << 11) | (1 << 12);
   const u32 gpustat_bits = (param & 0x03) << 11;
-  if ((m_GPUSTAT.bits & gpustat_mask) != gpustat_bits)
-  {
-    FlushRender();
   m_GPUSTAT.bits = (m_GPUSTAT.bits & ~gpustat_mask) | gpustat_bits;
-  }
   DEBUG_LOG("Set mask bit {} {}", BoolToUInt32(m_GPUSTAT.set_mask_while_drawing),
             BoolToUInt32(m_GPUSTAT.check_mask_before_draw));
 
@@ -318,6 +313,36 @@ bool GPU::HandleSetMaskBitCommand()
   return true;
 }
 
+void GPU::PrepareForDraw()
+{
+  if (m_drawing_area_changed)
+  {
+    m_drawing_area_changed = false;
+    GPUBackendSetDrawingAreaCommand* cmd = GPUBackend::NewSetDrawingAreaCommand();
+    cmd->new_area = m_drawing_area;
+    GSVector4i::store<false>(cmd->new_clamped_area, m_clamped_drawing_area);
+    GPUBackend::PushCommand(cmd);
+  }
+}
+
+void GPU::FillBackendCommandParameters(GPUBackendCommand* cmd) const
+{
+  cmd->params.bits = 0;
+  cmd->params.check_mask_before_draw = m_GPUSTAT.check_mask_before_draw;
+  cmd->params.set_mask_while_drawing = m_GPUSTAT.set_mask_while_drawing;
+  cmd->params.active_line_lsb = m_crtc_state.active_line_lsb;
+  cmd->params.interlaced_rendering = IsInterlacedRenderingEnabled();
+}
+
+void GPU::FillDrawCommand(GPUBackendDrawCommand* cmd, GPURenderCommand rc) const
+{
+  FillBackendCommandParameters(cmd);
+  cmd->rc.bits = rc.bits;
+  cmd->draw_mode.bits = m_draw_mode.mode_reg.bits;
+  cmd->palette.bits = m_draw_mode.palette_reg.bits;
+  cmd->window = m_draw_mode.texture_window;
+}
+
 bool GPU::HandleRenderPolygonCommand()
 {
   const GPURenderCommand rc{FifoPeek(0)};
@@ -343,6 +368,7 @@ bool GPU::HandleRenderPolygonCommand()
             words_per_vertex, setup_ticks);
 
   // set draw state up
+  // TODO: Get rid of SetTexturePalette() and just fill it as needed
   if (rc.texture_enable)
   {
     const u16 texpage_attribute = Truncate16((rc.shading_enable ? FifoPeek(5) : FifoPeek(4)) >> 16);
@@ -352,12 +378,219 @@ bool GPU::HandleRenderPolygonCommand()
     UpdateCLUTIfNeeded(m_draw_mode.mode_reg.texture_mode, m_draw_mode.palette_reg);
   }
 
-  m_counters.num_vertices += num_vertices;
-  m_counters.num_primitives++;
   m_render_command.bits = rc.bits;
   m_fifo.RemoveOne();
 
-  DispatchRenderCommand();
+  PrepareForDraw();
+
+  if (g_gpu_settings.gpu_pgxp_enable)
+  {
+    GPUBackendDrawPrecisePolygonCommand* cmd = GPUBackend::NewDrawPrecisePolygonCommand(num_vertices);
+    FillDrawCommand(cmd, rc);
+
+    const u32 first_color = rc.color_for_first_vertex;
+    const bool shaded = rc.shading_enable;
+    const bool textured = rc.texture_enable;
+    bool valid_w = g_gpu_settings.gpu_pgxp_texture_correction;
+    for (u32 i = 0; i < num_vertices; i++)
+    {
+      GPUBackendDrawPrecisePolygonCommand::Vertex* vert = &cmd->vertices[i];
+      vert->color = (shaded && i > 0) ? (FifoPop() & UINT32_C(0x00FFFFFF)) : first_color;
+      const u64 maddr_and_pos = m_fifo.Pop();
+      const GPUVertexPosition vp{Truncate32(maddr_and_pos)};
+      vert->native_x = m_drawing_offset.x + vp.x;
+      vert->native_y = m_drawing_offset.y + vp.y;
+      vert->texcoord = textured ? Truncate16(FifoPop()) : 0;
+
+      valid_w &= CPU::PGXP::GetPreciseVertex(Truncate32(maddr_and_pos >> 32), vp.bits, vert->native_x, vert->native_y,
+                                             m_drawing_offset.x, m_drawing_offset.y, &vert->x, &vert->y, &vert->w);
+    }
+
+    cmd->valid_w = valid_w;
+    if (!valid_w)
+    {
+      if (g_settings.gpu_pgxp_disable_2d)
+      {
+        // NOTE: This reads uninitialized data, but it's okay, it doesn't get used.
+        for (u32 i = 0; i < num_vertices; i++)
+        {
+          GPUBackendDrawPrecisePolygonCommand::Vertex& v = cmd->vertices[i];
+          GSVector2::store(&v.x, GSVector2(GSVector2i::load(&v.native_x)));
+          v.w = 1.0f;
+        }
+      }
+      else
+      {
+        for (u32 i = 0; i < num_vertices; i++)
+          cmd->vertices[i].w = 1.0f;
+      }
+    }
+
+    // Cull polygons which are too large.
+    const GSVector2 v0f = GSVector2::load(&cmd->vertices[0].x);
+    const GSVector2 v1f = GSVector2::load(&cmd->vertices[1].x);
+    const GSVector2 v2f = GSVector2::load(&cmd->vertices[2].x);
+    const GSVector2 min_pos_12 = v1f.min(v2f);
+    const GSVector2 max_pos_12 = v1f.max(v2f);
+    const GSVector4i draw_rect_012 = GSVector4i(GSVector4(min_pos_12.min(v0f)).upld(GSVector4(max_pos_12.max(v0f))))
+                                       .add32(GSVector4i::cxpr(0, 0, 1, 1));
+    const bool first_tri_culled =
+      (draw_rect_012.width() > MAX_PRIMITIVE_WIDTH || draw_rect_012.height() > MAX_PRIMITIVE_HEIGHT ||
+       !draw_rect_012.rintersects(m_clamped_drawing_area));
+    if (first_tri_culled)
+    {
+      // TODO: GPU events... somehow.
+      DEBUG_LOG("Culling off-screen/too-large polygon: {},{} {},{} {},{}", cmd->vertices[0].native_x,
+                cmd->vertices[0].native_y, cmd->vertices[1].native_x, cmd->vertices[1].native_y,
+                cmd->vertices[2].native_x, cmd->vertices[2].native_y);
+
+      if (!rc.quad_polygon)
+      {
+        EndCommand();
+        return true;
+      }
+    }
+    else
+    {
+      AddDrawTriangleTicks(GSVector2i::load(&cmd->vertices[0].native_x), GSVector2i::load(&cmd->vertices[1].native_x),
+                           GSVector2i::load(&cmd->vertices[2].native_x), rc.shading_enable, rc.texture_enable,
+                           rc.transparency_enable);
+    }
+
+    // quads
+    if (rc.quad_polygon)
+    {
+      const GSVector2 v3f = GSVector2::load(&cmd->vertices[3].x);
+      const GSVector4i draw_rect_123 = GSVector4i(GSVector4(min_pos_12.min(v3f)).upld(GSVector4(max_pos_12.max(v3f))))
+                                         .add32(GSVector4i::cxpr(0, 0, 1, 1));
+
+      // Cull polygons which are too large.
+      const bool second_tri_culled =
+        (draw_rect_123.width() > MAX_PRIMITIVE_WIDTH || draw_rect_123.height() > MAX_PRIMITIVE_HEIGHT ||
+         !draw_rect_123.rintersects(m_clamped_drawing_area));
+      if (second_tri_culled)
+      {
+        DEBUG_LOG("Culling off-screen/too-large polygon (quad second half): {},{} {},{} {},{}",
+                  cmd->vertices[2].native_x, cmd->vertices[2].native_y, cmd->vertices[1].native_x,
+                  cmd->vertices[1].native_y, cmd->vertices[0].native_x, cmd->vertices[0].native_y);
+
+        if (first_tri_culled)
+        {
+          EndCommand();
+          return true;
+        }
+
+        // Remove second part of quad.
+        cmd->num_vertices = 3;
+      }
+      else
+      {
+        AddDrawTriangleTicks(GSVector2i::load(&cmd->vertices[2].native_x), GSVector2i::load(&cmd->vertices[1].native_x),
+                             GSVector2i::load(&cmd->vertices[3].native_x), rc.shading_enable, rc.texture_enable,
+                             rc.transparency_enable);
+
+        // If first part was culled, move the second part to the first.
+        if (first_tri_culled)
+        {
+          std::memcpy(&cmd->vertices[0], &cmd->vertices[2], sizeof(GPUBackendDrawPrecisePolygonCommand::Vertex));
+          std::memcpy(&cmd->vertices[2], &cmd->vertices[3], sizeof(GPUBackendDrawPrecisePolygonCommand::Vertex));
+          cmd->num_vertices = 3;
+        }
+      }
+    }
+
+    GPUBackend::PushCommand(cmd);
+  }
+  else
+  {
+    GPUBackendDrawPolygonCommand* cmd = GPUBackend::NewDrawPolygonCommand(num_vertices);
+    FillDrawCommand(cmd, rc);
+
+    const u32 first_color = rc.color_for_first_vertex;
+    const bool shaded = rc.shading_enable;
+    const bool textured = rc.texture_enable;
+    for (u32 i = 0; i < num_vertices; i++)
+    {
+      GPUBackendDrawPolygonCommand::Vertex* vert = &cmd->vertices[i];
+      vert->color = (shaded && i > 0) ? (FifoPop() & UINT32_C(0x00FFFFFF)) : first_color;
+      const u64 maddr_and_pos = m_fifo.Pop();
+      const GPUVertexPosition vp{Truncate32(maddr_and_pos)};
+      vert->x = m_drawing_offset.x + vp.x;
+      vert->y = m_drawing_offset.y + vp.y;
+      vert->texcoord = textured ? Truncate16(FifoPop()) : 0;
+    }
+
+    // Cull polygons which are too large.
+    const GSVector2i v0 = GSVector2i::load(&cmd->vertices[0].x);
+    const GSVector2i v1 = GSVector2i::load(&cmd->vertices[1].x);
+    const GSVector2i v2 = GSVector2i::load(&cmd->vertices[2].x);
+    const GSVector2i min_pos_12 = v1.min_i32(v2);
+    const GSVector2i max_pos_12 = v1.max_i32(v2);
+    const GSVector4i draw_rect_012 =
+      GSVector4i(min_pos_12.min_i32(v0)).upl64(GSVector4i(max_pos_12.max_i32(v0))).add32(GSVector4i::cxpr(0, 0, 1, 1));
+    const bool first_tri_culled =
+      (draw_rect_012.width() > MAX_PRIMITIVE_WIDTH || draw_rect_012.height() > MAX_PRIMITIVE_HEIGHT ||
+       !draw_rect_012.rintersects(m_clamped_drawing_area));
+    if (first_tri_culled)
+    {
+      DEBUG_LOG("Culling off-screen/too-large polygon: {},{} {},{} {},{}", cmd->vertices[0].x, cmd->vertices[0].y,
+                cmd->vertices[1].x, cmd->vertices[1].y, cmd->vertices[2].x, cmd->vertices[2].y);
+
+      if (!rc.quad_polygon)
+      {
+        EndCommand();
+        return true;
+      }
+
+    }
+    else
+    {
+      AddDrawTriangleTicks(v0, v1, v2, rc.shading_enable, rc.texture_enable, rc.transparency_enable);
+    }
+
+    // quads
+    if (rc.quad_polygon)
+    {
+      const GSVector2i v3 = GSVector2i::load(&cmd->vertices[3].x);
+      const GSVector4i draw_rect_123 = GSVector4i(min_pos_12.min_i32(v3))
+                                         .upl64(GSVector4i(max_pos_12.max_i32(v3)))
+                                         .add32(GSVector4i::cxpr(0, 0, 1, 1));
+
+      // Cull polygons which are too large.
+      const bool second_tri_culled =
+        (draw_rect_123.width() > MAX_PRIMITIVE_WIDTH || draw_rect_123.height() > MAX_PRIMITIVE_HEIGHT ||
+         !draw_rect_123.rintersects(m_clamped_drawing_area));
+      if (second_tri_culled)
+      {
+        DEBUG_LOG("Culling too-large polygon (quad second half): {},{} {},{} {},{}", cmd->vertices[2].x,
+                  cmd->vertices[2].y, cmd->vertices[1].x, cmd->vertices[1].y, cmd->vertices[0].x, cmd->vertices[0].y);
+
+        if (first_tri_culled)
+        {
+          EndCommand();
+          return true;
+        }
+
+        // Remove second part of quad.
+        cmd->num_vertices = 3;
+      }
+      else
+      {
+        AddDrawTriangleTicks(v2, v1, v3, rc.shading_enable, rc.texture_enable, rc.transparency_enable);
+
+        // If first part was culled, move the second part to the first.
+        if (first_tri_culled)
+        {
+          std::memcpy(&cmd->vertices[0], &cmd->vertices[2], sizeof(GPUBackendDrawPolygonCommand::Vertex));
+          std::memcpy(&cmd->vertices[2], &cmd->vertices[3], sizeof(GPUBackendDrawPolygonCommand::Vertex));
+          cmd->num_vertices = 3;
+        }
+      }
+    }
+
+    GPUBackend::PushCommand(cmd);
+  }
+
   EndCommand();
   return true;
 }
@@ -386,12 +619,65 @@ bool GPU::HandleRenderRectangleCommand()
             rc.transparency_enable ? "semi-transparent" : "opaque", rc.texture_enable ? "textured" : "non-textured",
             rc.shading_enable ? "shaded" : "monochrome", total_words, setup_ticks);
 
-  m_counters.num_vertices++;
-  m_counters.num_primitives++;
   m_render_command.bits = rc.bits;
   m_fifo.RemoveOne();
 
-  DispatchRenderCommand();
+  PrepareForDraw();
+  GPUBackendDrawRectangleCommand* cmd = GPUBackend::NewDrawRectangleCommand();
+  FillDrawCommand(cmd, rc);
+  cmd->color = rc.color_for_first_vertex;
+
+  const GPUVertexPosition vp{FifoPop()};
+  cmd->x = TruncateGPUVertexPosition(m_drawing_offset.x + vp.x);
+  cmd->y = TruncateGPUVertexPosition(m_drawing_offset.y + vp.y);
+
+  if (rc.texture_enable)
+  {
+    const u32 texcoord_and_palette = FifoPop();
+    cmd->palette.bits = Truncate16(texcoord_and_palette >> 16);
+    cmd->texcoord = Truncate16(texcoord_and_palette);
+  }
+  else
+  {
+    cmd->palette.bits = 0;
+    cmd->texcoord = 0;
+  }
+
+  switch (rc.rectangle_size)
+  {
+    case GPUDrawRectangleSize::R1x1:
+      cmd->width = 1;
+      cmd->height = 1;
+      break;
+    case GPUDrawRectangleSize::R8x8:
+      cmd->width = 8;
+      cmd->height = 8;
+      break;
+    case GPUDrawRectangleSize::R16x16:
+      cmd->width = 16;
+      cmd->height = 16;
+      break;
+    default:
+    {
+      const u32 width_and_height = FifoPop();
+      cmd->width = static_cast<u16>(width_and_height & VRAM_WIDTH_MASK);
+      cmd->height = static_cast<u16>((width_and_height >> 16) & VRAM_HEIGHT_MASK);
+    }
+    break;
+  }
+
+  const GSVector4i rect = GSVector4i(cmd->x, cmd->y, cmd->x + cmd->width, cmd->y + cmd->height);
+  const GSVector4i clamped_rect = m_clamped_drawing_area.rintersect(rect);
+  if (clamped_rect.rempty()) [[unlikely]]
+  {
+    DEBUG_LOG("Culling off-screen rectangle {}", rect);
+    EndCommand();
+    return true;
+  }
+
+  AddDrawRectangleTicks(clamped_rect, rc.texture_enable, rc.transparency_enable);
+
+  GPUBackend::PushCommand(cmd);
   EndCommand();
   return true;
 }
@@ -408,12 +694,55 @@ bool GPU::HandleRenderLineCommand()
   TRACE_LOG("Render {} {} line ({} total words)", rc.transparency_enable ? "semi-transparent" : "opaque",
             rc.shading_enable ? "shaded" : "monochrome", total_words);
 
-  m_counters.num_vertices += 2;
-  m_counters.num_primitives++;
   m_render_command.bits = rc.bits;
   m_fifo.RemoveOne();
 
-  DispatchRenderCommand();
+  PrepareForDraw();
+  GPUBackendDrawLineCommand* cmd = GPUBackend::NewDrawLineCommand(2);
+  FillDrawCommand(cmd, rc);
+  cmd->palette.bits = 0;
+
+  if (rc.shading_enable)
+  {
+    cmd->vertices[0].color = rc.color_for_first_vertex;
+    const GPUVertexPosition start_pos{FifoPop()};
+    cmd->vertices[0].x = m_drawing_offset.x + start_pos.x;
+    cmd->vertices[0].y = m_drawing_offset.y + start_pos.y;
+
+    cmd->vertices[1].color = FifoPop() & UINT32_C(0x00FFFFFF);
+    const GPUVertexPosition end_pos{FifoPop()};
+    cmd->vertices[1].x = m_drawing_offset.x + end_pos.x;
+    cmd->vertices[1].y = m_drawing_offset.y + end_pos.y;
+  }
+  else
+  {
+    cmd->vertices[0].color = rc.color_for_first_vertex;
+    cmd->vertices[1].color = rc.color_for_first_vertex;
+
+    const GPUVertexPosition start_pos{FifoPop()};
+    cmd->vertices[0].x = m_drawing_offset.x + start_pos.x;
+    cmd->vertices[0].y = m_drawing_offset.y + start_pos.y;
+
+    const GPUVertexPosition end_pos{FifoPop()};
+    cmd->vertices[1].x = m_drawing_offset.x + end_pos.x;
+    cmd->vertices[1].y = m_drawing_offset.y + end_pos.y;
+  }
+
+  const GSVector4i v0 = GSVector4i::loadl(&cmd->vertices[0].x);
+  const GSVector4i v1 = GSVector4i::loadl(&cmd->vertices[1].x);
+  const GSVector4i rect = v0.min_i32(v1).xyxy(v0.max_i32(v1)).add32(GSVector4i::cxpr(0, 0, 1, 1));
+  const GSVector4i clamped_rect = rect.rintersect(m_clamped_drawing_area);
+
+  if (rect.width() > MAX_PRIMITIVE_WIDTH || rect.height() > MAX_PRIMITIVE_HEIGHT || clamped_rect.rempty())
+  {
+    DEBUG_LOG("Culling too-large/off-screen line: {},{} - {},{}", cmd->vertices[0].y, cmd->vertices[0].y,
+              cmd->vertices[1].x, cmd->vertices[1].y);
+    EndCommand();
+    return true;
+  }
+
+  AddDrawLineTicks(clamped_rect, rc.shading_enable);
+  GPUBackend::PushCommand(cmd);
   EndCommand();
   return true;
 }
@@ -450,6 +779,48 @@ bool GPU::HandleRenderPolyLineCommand()
   return true;
 }
 
+void GPU::FinishPolyline()
+{
+  PrepareForDraw();
+
+  const u32 num_vertices = GetPolyLineVertexCount();
+
+  GPUBackendDrawLineCommand* cmd = GPUBackend::NewDrawLineCommand(num_vertices);
+  FillDrawCommand(cmd, m_render_command);
+
+  u32 buffer_pos = 0;
+  const GPUVertexPosition start_vp{m_blit_buffer[buffer_pos++]};
+  cmd->vertices[0].x = start_vp.x + m_drawing_offset.x;
+  cmd->vertices[0].y = start_vp.y + m_drawing_offset.y;
+  cmd->vertices[0].color = m_render_command.color_for_first_vertex;
+
+  const bool shaded = m_render_command.shading_enable;
+  for (u32 i = 1; i < num_vertices; i++)
+  {
+    cmd->vertices[i].color =
+      shaded ? (m_blit_buffer[buffer_pos++] & UINT32_C(0x00FFFFFF)) : m_render_command.color_for_first_vertex;
+    const GPUVertexPosition vp{m_blit_buffer[buffer_pos++]};
+    cmd->vertices[i].x = m_drawing_offset.x + vp.x;
+    cmd->vertices[i].y = m_drawing_offset.y + vp.y;
+
+    const GSVector4i v0 = GSVector4i::loadl(&cmd->vertices[0].x);
+    const GSVector4i v1 = GSVector4i::loadl(&cmd->vertices[1].x);
+    const GSVector4i rect = v0.min_i32(v1).xyxy(v0.max_i32(v1)).add32(GSVector4i::cxpr(0, 0, 1, 1));
+    const GSVector4i clamped_rect = rect.rintersect(m_clamped_drawing_area);
+
+    if (rect.width() > MAX_PRIMITIVE_WIDTH || rect.height() > MAX_PRIMITIVE_HEIGHT || clamped_rect.rempty())
+    {
+      DEBUG_LOG("Culling too-large/off-screen line: {},{} - {},{}", cmd->vertices[i - 1].x, cmd->vertices[i - 1].y,
+                cmd->vertices[i].x, cmd->vertices[i].y);
+      return;
+    }
+    else
+    {
+      AddDrawLineTicks(clamped_rect, m_render_command.shading_enable);
+    }
+  }
+}
+
 bool GPU::HandleFillRectangleCommand()
 {
   CHECK_COMMAND_SIZE(3);
@@ -457,8 +828,6 @@ bool GPU::HandleFillRectangleCommand()
   if (IsInterlacedRenderingEnabled() && IsCRTCScanlinePending())
     SynchronizeCRTC();
 
-  FlushRender();
-
   const u32 color = FifoPop() & 0x00FFFFFF;
   const u32 dst_x = FifoPeek() & 0x3F0;
   const u32 dst_y = (FifoPop() >> 16) & VRAM_HEIGHT_MASK;
@@ -468,9 +837,17 @@ bool GPU::HandleFillRectangleCommand()
   DEBUG_LOG("Fill VRAM rectangle offset=({},{}), size=({},{})", dst_x, dst_y, width, height);
 
   if (width > 0 && height > 0)
-    FillVRAM(dst_x, dst_y, width, height, color);
+  {
+    GPUBackendFillVRAMCommand* cmd = GPUBackend::NewFillVRAMCommand();
+    FillBackendCommandParameters(cmd);
+    cmd->x = static_cast<u16>(dst_x);
+    cmd->y = static_cast<u16>(dst_y);
+    cmd->width = static_cast<u16>(width);
+    cmd->height = static_cast<u16>(height);
+    cmd->color = color;
+    GPUBackend::PushCommand(cmd);
+  }
 
-  m_counters.num_writes++;
   AddCommandTicks(46 + ((width / 8) + 9) * height);
   EndCommand();
   return true;
@@ -520,8 +897,6 @@ void GPU::FinishVRAMWrite()
   if (IsInterlacedRenderingEnabled() && IsCRTCScanlinePending())
     SynchronizeCRTC();
 
-  FlushRender();
-
   if (m_blit_remaining_words == 0)
   {
     if (g_settings.debugging.dump_cpu_to_vram_copies)
@@ -554,18 +929,18 @@ void GPU::FinishVRAMWrite()
     const u8* blit_ptr = reinterpret_cast<const u8*>(m_blit_buffer.data());
     if (transferred_full_rows > 0)
     {
-      UpdateVRAM(m_vram_transfer.x, m_vram_transfer.y, m_vram_transfer.width, transferred_full_rows, blit_ptr,
-                 m_GPUSTAT.set_mask_while_drawing, m_GPUSTAT.check_mask_before_draw);
+      UpdateVRAM(m_vram_transfer.x, m_vram_transfer.y, m_vram_transfer.width, static_cast<u16>(transferred_full_rows),
+                 blit_ptr, m_GPUSTAT.set_mask_while_drawing, m_GPUSTAT.check_mask_before_draw);
       blit_ptr += (ZeroExtend32(m_vram_transfer.width) * transferred_full_rows) * sizeof(u16);
     }
     if (transferred_width_last_row > 0)
     {
-      UpdateVRAM(m_vram_transfer.x, m_vram_transfer.y + transferred_full_rows, transferred_width_last_row, 1, blit_ptr,
-                 m_GPUSTAT.set_mask_while_drawing, m_GPUSTAT.check_mask_before_draw);
+      UpdateVRAM(m_vram_transfer.x, static_cast<u16>(m_vram_transfer.y + transferred_full_rows),
+                 static_cast<u16>(transferred_width_last_row), 1, blit_ptr, m_GPUSTAT.set_mask_while_drawing,
+                 m_GPUSTAT.check_mask_before_draw);
     }
   }
 
-  m_counters.num_writes++;
   m_blit_buffer.clear();
   m_vram_transfer = {};
   m_blitter_state = BlitterState::Idle;
@@ -585,9 +960,6 @@ bool GPU::HandleCopyRectangleVRAMToCPUCommand()
             m_vram_transfer.width, m_vram_transfer.height);
   DebugAssert(m_vram_transfer.col == 0 && m_vram_transfer.row == 0);
 
-  // all rendering should be done first...
-  FlushRender();
-
   // ensure VRAM shadow is up to date
   ReadVRAM(m_vram_transfer.x, m_vram_transfer.y, m_vram_transfer.width, m_vram_transfer.height);
 
@@ -599,7 +971,6 @@ bool GPU::HandleCopyRectangleVRAMToCPUCommand()
   }
 
   // switch to pixel-by-pixel read state
-  m_counters.num_reads++;
   m_blitter_state = BlitterState::ReadingVRAM;
   m_command_total_words = 0;
   return true;
@@ -625,10 +996,15 @@ bool GPU::HandleCopyRectangleVRAMToVRAMCommand()
     width == 0 || height == 0 || (src_x == dst_x && src_y == dst_y && !m_GPUSTAT.set_mask_while_drawing);
   if (!skip_copy)
   {
-    m_counters.num_copies++;
-
-    FlushRender();
-    CopyVRAM(src_x, src_y, dst_x, dst_y, width, height);
+    GPUBackendCopyVRAMCommand* cmd = GPUBackend::NewCopyVRAMCommand();
+    FillBackendCommandParameters(cmd);
+    cmd->src_x = static_cast<u16>(src_x);
+    cmd->src_y = static_cast<u16>(src_y);
+    cmd->dst_x = static_cast<u16>(dst_x);
+    cmd->dst_y = static_cast<u16>(dst_y);
+    cmd->width = static_cast<u16>(width);
+    cmd->height = static_cast<u16>(height);
+    GPUBackend::PushCommand(cmd);
   }
 
   AddCommandTicks(width * height * 2);

src/core/gpu_hw.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "gpu.h"
+#include "gpu_backend.h"
 #include "texture_replacements.h"
 
 #include "util/gpu_device.h"
@@ -21,7 +21,9 @@ class GPU_SW_Backend;
 struct GPUBackendCommand;
 struct GPUBackendDrawCommand;
 
-class GPU_HW final : public GPU
+// TODO: Move to cpp
+// TODO: Rename to GPUHWBackend, preserved to avoid conflicts.
+class GPU_HW final : public GPUBackend
 {
 public:
   enum class BatchRenderMode : u8
@@ -55,21 +57,37 @@ public:
   GPU_HW();
   ~GPU_HW() override;
 
-  const Threading::Thread* GetSWThread() const override;
-  bool IsHardwareRenderer() const override;
+  bool Initialize(bool clear_vram, Error* error) override;
 
-  bool Initialize() override;
-  void Reset(bool clear_vram) override;
-  bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display) override;
+  void ClearVRAM() override;
+  bool DoState(GPUTexture** host_texture, bool is_reading, bool update_display) override;
 
   void RestoreDeviceContext() override;
 
   void UpdateSettings(const Settings& old_settings) override;
-  void UpdateResolutionScale() override final;
-  std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) override;
-  std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) override;
 
-  void UpdateDisplay() override;
+  void UpdateResolutionScale() override;
+  std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) const override;
+  std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) const override;
+
+  void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params) override;
+  void ReadVRAM(u32 x, u32 y, u32 width, u32 height) override;
+  void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, GPUBackendCommandParameters params) override;
+  void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
+                GPUBackendCommandParameters params) override;
+  void ClearCache() override;
+  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
+  void OnBufferSwapped() override;
+
+  void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) override;
+  void DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd) override;
+  void DrawSprite(const GPUBackendDrawRectangleCommand* cmd) override;
+  void DrawLine(const GPUBackendDrawLineCommand* cmd) override;
+
+  void FlushRender() override;
+  void DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area) override;
+
+  void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) override;
 
 private:
   enum : u32
@@ -78,6 +96,7 @@ private:
     MAX_VERTICES_FOR_RECTANGLE = 6 * (((MAX_PRIMITIVE_WIDTH + (TEXTURE_PAGE_WIDTH - 1)) / TEXTURE_PAGE_WIDTH) + 1u) *
                                  (((MAX_PRIMITIVE_HEIGHT + (TEXTURE_PAGE_HEIGHT - 1)) / TEXTURE_PAGE_HEIGHT) + 1u),
     NUM_TEXTURE_MODES = static_cast<u32>(BatchTextureMode::MaxCount),
+    INVALID_DRAW_MODE_BITS = 0xFFFFFFFFu,
   };
   enum : u8
   {
@@ -152,8 +171,6 @@ private:
   bool CompilePipelines();
   void DestroyPipelines();
 
-  void LoadVertices();
-
   void PrintSettingsToLog();
   void CheckSettings();
 
@@ -171,7 +188,7 @@ private:
   GPUDownsampleMode GetDownsampleMode(u32 resolution_scale) const;
 
   bool IsUsingMultisampling() const;
-  bool IsUsingDownsampling() const;
+  bool IsUsingDownsampling(const GPUBackendUpdateDisplayCommand* cmd) const;
 
   void SetFullVRAMDirtyRectangle();
   void ClearVRAMDirtyRectangle();
@@ -181,11 +198,14 @@ private:
   void AddUnclampedDrawnRectangle(const GSVector4i rect);
   void SetTexPageChangedOnOverlap(const GSVector4i update_rect);
 
-  void CheckForTexPageOverlap(GSVector4i uv_rect);
+  void CheckForTexPageOverlap(const GPUBackendDrawCommand* cmd, GSVector4i uv_rect);
 
   bool IsFlushed() const;
   void EnsureVertexBufferSpace(u32 required_vertices, u32 required_indices);
-  void EnsureVertexBufferSpaceForCurrentCommand();
+  void EnsureVertexBufferSpaceForCommand(const GPUBackendDrawCommand* cmd);
+  void PrepareDraw(const GPUBackendDrawCommand* cmd);
+  void FinishPolygonDraw(const GPUBackendDrawCommand* cmd, std::array<BatchVertex, 4>& vertices, u32 num_vertices,
+                         bool is_3d);
   void ResetBatchVertexDepth();
 
   /// Returns the value to be written to the depth buffer for the current operation for mask bit emulation.
@@ -197,20 +217,8 @@ private:
   /// Returns true if the draw is going to use shader blending/framebuffer fetch.
   bool NeedsShaderBlending(GPUTransparencyMode transparency, bool check_mask) const;
 
-  void FillBackendCommandParameters(GPUBackendCommand* cmd) const;
-  void FillDrawCommand(GPUBackendDrawCommand* cmd, GPURenderCommand rc) const;
   void UpdateSoftwareRenderer(bool copy_vram_from_hw);
 
-  void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) override;
-  void ReadVRAM(u32 x, u32 y, u32 width, u32 height) override;
-  void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) override;
-  void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
-  void DispatchRenderCommand() override;
-  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
-  void FlushRender() override;
-  void DrawRendererStats() override;
-  void OnBufferSwapped() override;
-
   void UpdateVRAMOnGPU(u32 x, u32 y, u32 width, u32 height, const void* data, u32 data_pitch, bool set_mask,
                        bool check_mask, const GSVector4i bounds);
   bool BlitVRAMReplacementTexture(const TextureReplacements::ReplacementImage* tex, u32 dst_x, u32 dst_y, u32 width,
@@ -220,17 +228,17 @@ private:
   void DrawLine(const GSVector4 bounds, u32 col0, u32 col1, float depth);
 
   /// Handles quads with flipped texture coordinate directions.
-  void HandleFlippedQuadTextureCoordinates(BatchVertex* vertices);
+  void HandleFlippedQuadTextureCoordinates(const GPUBackendDrawCommand* cmd, BatchVertex* vertices);
   bool IsPossibleSpritePolygon(const BatchVertex* vertices) const;
   bool ExpandLineTriangles(BatchVertex* vertices);
 
   /// Computes polygon U/V boundaries, and for overlap with the current texture page.
-  void ComputePolygonUVLimits(BatchVertex* vertices, u32 num_vertices);
+  void ComputePolygonUVLimits(const GPUBackendDrawCommand* cmd, BatchVertex* vertices, u32 num_vertices);
 
   /// Sets the depth test flag for PGXP depth buffering.
-  void SetBatchDepthBuffer(bool enabled);
-  void CheckForDepthClear(const BatchVertex* vertices, u32 num_vertices);
-  void SetBatchSpriteMode(bool enabled);
+  void SetBatchDepthBuffer(const GPUBackendDrawCommand* cmd, bool enabled);
+  void CheckForDepthClear(const GPUBackendDrawCommand* cmd, const BatchVertex* vertices, u32 num_vertices);
+  void SetBatchSpriteMode(const GPUBackendDrawCommand* cmd, bool enabled);
 
   void UpdateDownsamplingLevels();
 
@@ -249,7 +257,7 @@ private:
   std::unique_ptr<GPUTextureBuffer> m_vram_upload_buffer;
   std::unique_ptr<GPUTexture> m_vram_write_texture;
 
-  std::unique_ptr<GPU_SW_Backend> m_sw_renderer;
+  // std::unique_ptr<GPU_SW_Backend> m_sw_renderer;
 
   BatchVertex* m_batch_vertex_ptr = nullptr;
   u16* m_batch_index_ptr = nullptr;
@@ -287,15 +295,29 @@ private:
   BatchConfig m_batch;
 
   // Changed state
+  bool m_drawing_area_changed = true;
   bool m_batch_ubo_dirty = true;
   BatchUBOData m_batch_ubo_data = {};
 
   // Bounding box of VRAM area that the GPU has drawn into.
+  GSVector4i m_clamped_drawing_area = {};
   GSVector4i m_vram_dirty_draw_rect = INVALID_RECT;
   GSVector4i m_vram_dirty_write_rect = INVALID_RECT;
   GSVector4i m_current_uv_rect = INVALID_RECT;
   s32 m_current_texture_page_offset[2] = {};
 
+  union
+  {
+    struct
+    {
+      GPUDrawModeReg mode_reg;
+      GPUTexturePaletteReg palette_reg;
+    };
+
+    u32 bits = INVALID_DRAW_MODE_BITS;
+  } m_draw_mode = {};
+  u32 m_last_texture_window_bits = 0;
+
   std::unique_ptr<GPUPipeline> m_wireframe_pipeline;
 
   // [wrapped][interlaced]

src/core/gpu_hw_shadergen.h

@@ -2,7 +2,9 @@
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
+
 #include "gpu_hw.h"
+
 #include "util/shadergen.h"
 
 class GPU_HW_ShaderGen : public ShaderGen

src/core/gpu_sw.cpp

@@ -2,14 +2,15 @@
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #include "gpu_sw.h"
+#include "gpu.h"
+#include "gpu_sw_rasterizer.h"
 #include "system.h"
 
 #include "util/gpu_device.h"
 
 #include "common/align.h"
 #include "common/assert.h"
-#include "common/gsvector.h"
-#include "common/gsvector_formatter.h"
+#include "common/intrin.h"
 #include "common/log.h"
 
 #include <algorithm>
@@ -18,27 +19,141 @@ Log_SetChannel(GPU_SW);
 
 GPU_SW::GPU_SW() = default;
 
-GPU_SW::~GPU_SW()
-{
-  g_gpu_device->RecycleTexture(std::move(m_upload_texture));
-  m_backend.Shutdown();
-}
+GPU_SW::~GPU_SW() = default;
 
-const Threading::Thread* GPU_SW::GetSWThread() const
+bool GPU_SW::Initialize(bool clear_vram, Error* error)
 {
-  return m_backend.GetThread();
-}
-
-bool GPU_SW::IsHardwareRenderer() const
-{
-  return false;
-}
-
-bool GPU_SW::Initialize()
-{
-  if (!GPU::Initialize() || !m_backend.Initialize(false))
+  if (!GPUBackend::Initialize(clear_vram, error))
     return false;
 
+  GPU_SW_Rasterizer::SelectImplementation();
+
+  // if we're using "new" vram, clear it out here
+  if (clear_vram)
+    std::memset(g_vram, 0, sizeof(g_vram));
+
+  SetDisplayTextureFormat();
+  return true;
+}
+
+bool GPU_SW::DoState(GPUTexture** host_texture, bool is_reading, bool update_display)
+{
+  // TODO: FIXME
+  // ignore the host texture for software mode, since we want to save vram here
+  return true;
+}
+
+void GPU_SW::ClearVRAM()
+{
+  std::memset(g_vram, 0, sizeof(g_vram));
+  std::memset(g_gpu_clut, 0, sizeof(g_gpu_clut));
+}
+
+std::tuple<u32, u32> GPU_SW::GetEffectiveDisplayResolution(bool scaled /* = true */) const
+{
+  return std::tie(m_display_vram_width, m_display_vram_height);
+}
+
+std::tuple<u32, u32> GPU_SW::GetFullDisplayResolution(bool scaled /* = true */) const
+{
+  return std::tie(m_display_width, m_display_height);
+}
+
+void GPU_SW::UpdateResolutionScale()
+{
+}
+
+void GPU_SW::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
+{
+}
+
+void GPU_SW::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)
+{
+  const GPURenderCommand rc{cmd->rc.bits};
+  const bool dithering_enable = rc.IsDitheringEnabled() && cmd->draw_mode.dither_enable;
+
+  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
+    rc.shading_enable, rc.texture_enable, rc.raw_texture_enable, rc.transparency_enable, dithering_enable);
+
+  DrawFunction(cmd, &cmd->vertices[0], &cmd->vertices[1], &cmd->vertices[2]);
+  if (rc.quad_polygon)
+    DrawFunction(cmd, &cmd->vertices[2], &cmd->vertices[1], &cmd->vertices[3]);
+}
+
+void GPU_SW::DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd)
+{
+  const GPURenderCommand rc{cmd->rc.bits};
+  const bool dithering_enable = rc.IsDitheringEnabled() && cmd->draw_mode.dither_enable;
+
+  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
+    rc.shading_enable, rc.texture_enable, rc.raw_texture_enable, rc.transparency_enable, dithering_enable);
+
+  // Need to cut out the irrelevant bits.
+  // TODO: In _theory_ we could use the fixed-point parts here.
+  GPUBackendDrawPolygonCommand::Vertex vertices[4];
+  for (u32 i = 0; i < cmd->num_vertices; i++)
+  {
+    const GPUBackendDrawPrecisePolygonCommand::Vertex& src = cmd->vertices[i];
+    GPUBackendDrawPolygonCommand::Vertex& dst = vertices[i];
+
+    dst.x = src.native_x;
+    dst.y = src.native_y;
+    dst.color = src.color;
+    dst.texcoord = src.texcoord;
+  }
+
+  DrawFunction(cmd, &vertices[0], &vertices[1], &vertices[2]);
+  if (rc.quad_polygon)
+    DrawFunction(cmd, &vertices[2], &vertices[1], &vertices[3]);
+}
+
+void GPU_SW::DrawSprite(const GPUBackendDrawRectangleCommand* cmd)
+{
+  const GPURenderCommand rc{cmd->rc.bits};
+
+  const GPU_SW_Rasterizer::DrawRectangleFunction DrawFunction =
+    GPU_SW_Rasterizer::GetDrawRectangleFunction(rc.texture_enable, rc.raw_texture_enable, rc.transparency_enable);
+
+  DrawFunction(cmd);
+}
+
+void GPU_SW::DrawLine(const GPUBackendDrawLineCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawLineFunction DrawFunction = GPU_SW_Rasterizer::GetDrawLineFunction(
+    cmd->rc.shading_enable, cmd->rc.transparency_enable, cmd->IsDitheringEnabled());
+
+  for (u16 i = 1; i < cmd->num_vertices; i++)
+    DrawFunction(cmd, &cmd->vertices[i - 1], &cmd->vertices[i]);
+}
+
+void GPU_SW::DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area)
+{
+  GPU_SW_Rasterizer::g_drawing_area = new_drawing_area;
+}
+
+void GPU_SW::ClearCache()
+{
+}
+
+void GPU_SW::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  GPU_SW_Rasterizer::UpdateCLUT(reg, clut_is_8bit);
+}
+
+void GPU_SW::OnBufferSwapped()
+{
+}
+
+void GPU_SW::FlushRender()
+{
+}
+
+void GPU_SW::RestoreDeviceContext()
+{
+}
+
+void GPU_SW::SetDisplayTextureFormat()
+{
   static constexpr const std::array formats_for_16bit = {GPUTexture::Format::RGB565, GPUTexture::Format::RGBA5551,
                                                          GPUTexture::Format::RGBA8, GPUTexture::Format::BGRA8};
   static constexpr const std::array formats_for_24bit = {GPUTexture::Format::RGBA8, GPUTexture::Format::BGRA8,
@@ -59,32 +174,7 @@ bool GPU_SW::Initialize()
       break;
     }
   }
-
-  return true;
 }
-
-bool GPU_SW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display)
-{
-  // need to ensure the worker thread is done
-  m_backend.Sync(true);
-
-  // ignore the host texture for software mode, since we want to save vram here
-  return GPU::DoState(sw, nullptr, update_display);
-}
-
-void GPU_SW::Reset(bool clear_vram)
-{
-  GPU::Reset(clear_vram);
-
-  m_backend.Reset();
-}
-
-void GPU_SW::UpdateSettings(const Settings& old_settings)
-{
-  GPU::UpdateSettings(old_settings);
-  m_backend.UpdateSettings();
-}
-
 GPUTexture* GPU_SW::GetDisplayTexture(u32 width, u32 height, GPUTexture::Format format)
 {
   if (!m_upload_texture || m_upload_texture->GetWidth() != width || m_upload_texture->GetHeight() != height ||
@@ -420,32 +510,28 @@ bool GPU_SW::CopyOut(u32 src_x, u32 src_y, u32 skip_x, u32 width, u32 height, u3
   }
 }
 
-void GPU_SW::UpdateDisplay()
+void GPU_SW::UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd)
 {
-  // fill display texture
-  m_backend.Sync(true);
-
   if (!g_settings.debugging.show_vram)
   {
-    if (IsDisplayDisabled())
+    if (cmd->display_disabled)
     {
       ClearDisplayTexture();
       return;
     }
 
-    const bool is_24bit = m_GPUSTAT.display_area_color_depth_24;
-    const bool interlaced = IsInterlacedDisplayEnabled();
-    const u32 field = GetInterlacedDisplayField();
-    const u32 vram_offset_x = is_24bit ? m_crtc_state.regs.X : m_crtc_state.display_vram_left;
-    const u32 vram_offset_y =
-      m_crtc_state.display_vram_top + ((interlaced && m_GPUSTAT.vertical_resolution) ? field : 0);
-    const u32 skip_x = is_24bit ? (m_crtc_state.display_vram_left - m_crtc_state.regs.X) : 0;
-    const u32 read_width = m_crtc_state.display_vram_width;
-    const u32 read_height = interlaced ? (m_crtc_state.display_vram_height / 2) : m_crtc_state.display_vram_height;
+    const bool is_24bit = cmd->display_24bit;
+    const bool interlaced = cmd->interlaced_display_enabled;
+    const u32 field = cmd->interlaced_display_field;
+    const u32 vram_offset_x = is_24bit ? cmd->X : cmd->display_vram_left;
+    const u32 vram_offset_y = cmd->display_vram_top + ((interlaced && cmd->interlaced_display_interleaved) ? field : 0);
+    const u32 skip_x = is_24bit ? (cmd->display_vram_left - cmd->X) : 0;
+    const u32 read_width = cmd->display_vram_width;
+    const u32 read_height = interlaced ? (cmd->display_vram_height / 2) : cmd->display_vram_height;
 
-    if (IsInterlacedDisplayEnabled())
+    if (cmd->interlaced_display_enabled)
     {
-      const u32 line_skip = m_GPUSTAT.vertical_resolution;
+      const u32 line_skip = cmd->interlaced_display_interleaved;
       if (CopyOut(vram_offset_x, vram_offset_y, skip_x, read_width, read_height, line_skip, is_24bit))
       {
         SetDisplayTexture(m_upload_texture.get(), nullptr, 0, 0, read_width, read_height);
@@ -477,336 +563,7 @@ void GPU_SW::UpdateDisplay()
   }
 }
 
-void GPU_SW::FillBackendCommandParameters(GPUBackendCommand* cmd) const
+std::unique_ptr<GPUBackend> GPUBackend::CreateSoftwareBackend()
 {
-  cmd->params.bits = 0;
-  cmd->params.check_mask_before_draw = m_GPUSTAT.check_mask_before_draw;
-  cmd->params.set_mask_while_drawing = m_GPUSTAT.set_mask_while_drawing;
-  cmd->params.active_line_lsb = m_crtc_state.active_line_lsb;
-  cmd->params.interlaced_rendering = IsInterlacedRenderingEnabled();
-}
-
-void GPU_SW::FillDrawCommand(GPUBackendDrawCommand* cmd, GPURenderCommand rc) const
-{
-  FillBackendCommandParameters(cmd);
-  cmd->rc.bits = rc.bits;
-  cmd->draw_mode.bits = m_draw_mode.mode_reg.bits;
-  cmd->palette.bits = m_draw_mode.palette_reg.bits;
-  cmd->window = m_draw_mode.texture_window;
-}
-
-void GPU_SW::DispatchRenderCommand()
-{
-  if (m_drawing_area_changed)
-  {
-    GPUBackendSetDrawingAreaCommand* cmd = m_backend.NewSetDrawingAreaCommand();
-    cmd->new_area = m_drawing_area;
-    GSVector4i::store<false>(cmd->new_clamped_area, m_clamped_drawing_area);
-    m_backend.PushCommand(cmd);
-    m_drawing_area_changed = false;
-  }
-
-  const GPURenderCommand rc{m_render_command.bits};
-
-  switch (rc.primitive)
-  {
-    case GPUPrimitive::Polygon:
-    {
-      const u32 num_vertices = rc.quad_polygon ? 4 : 3;
-      GPUBackendDrawPolygonCommand* cmd = m_backend.NewDrawPolygonCommand(num_vertices);
-      FillDrawCommand(cmd, rc);
-
-      std::array<GSVector2i, 4> positions;
-      const u32 first_color = rc.color_for_first_vertex;
-      const bool shaded = rc.shading_enable;
-      const bool textured = rc.texture_enable;
-      for (u32 i = 0; i < num_vertices; i++)
-      {
-        GPUBackendDrawPolygonCommand::Vertex* vert = &cmd->vertices[i];
-        vert->color = (shaded && i > 0) ? (FifoPop() & UINT32_C(0x00FFFFFF)) : first_color;
-        const u64 maddr_and_pos = m_fifo.Pop();
-        const GPUVertexPosition vp{Truncate32(maddr_and_pos)};
-        vert->x = m_drawing_offset.x + vp.x;
-        vert->y = m_drawing_offset.y + vp.y;
-        vert->texcoord = textured ? Truncate16(FifoPop()) : 0;
-        positions[i] = GSVector2i::load(&vert->x);
-      }
-
-      // Cull polygons which are too large.
-      const GSVector2i min_pos_12 = positions[1].min_i32(positions[2]);
-      const GSVector2i max_pos_12 = positions[1].max_i32(positions[2]);
-      const GSVector4i draw_rect_012 = GSVector4i(min_pos_12.min_i32(positions[0]))
-                                         .upl64(GSVector4i(max_pos_12.max_i32(positions[0])))
-                                         .add32(GSVector4i::cxpr(0, 0, 1, 1));
-      const bool first_tri_culled =
-        (draw_rect_012.width() > MAX_PRIMITIVE_WIDTH || draw_rect_012.height() > MAX_PRIMITIVE_HEIGHT ||
-         !m_clamped_drawing_area.rintersects(draw_rect_012));
-      if (first_tri_culled)
-      {
-        DEBUG_LOG("Culling off-screen/too-large polygon: {},{} {},{} {},{}", cmd->vertices[0].x, cmd->vertices[0].y,
-                  cmd->vertices[1].x, cmd->vertices[1].y, cmd->vertices[2].x, cmd->vertices[2].y);
-
-        if (!rc.quad_polygon)
-          return;
-      }
-      else
-      {
-        AddDrawTriangleTicks(positions[0], positions[1], positions[2], rc.shading_enable, rc.texture_enable,
-                             rc.transparency_enable);
-      }
-
-      // quads
-      if (rc.quad_polygon)
-      {
-        const GSVector4i draw_rect_123 = GSVector4i(min_pos_12.min_i32(positions[3]))
-                                           .upl64(GSVector4i(max_pos_12.max_i32(positions[3])))
-                                           .add32(GSVector4i::cxpr(0, 0, 1, 1));
-
-        // Cull polygons which are too large.
-        const bool second_tri_culled =
-          (draw_rect_123.width() > MAX_PRIMITIVE_WIDTH || draw_rect_123.height() > MAX_PRIMITIVE_HEIGHT ||
-           !m_clamped_drawing_area.rintersects(draw_rect_123));
-        if (second_tri_culled)
-        {
-          DEBUG_LOG("Culling too-large polygon (quad second half): {},{} {},{} {},{}", cmd->vertices[2].x,
-                    cmd->vertices[2].y, cmd->vertices[1].x, cmd->vertices[1].y, cmd->vertices[0].x, cmd->vertices[0].y);
-
-          if (first_tri_culled)
-            return;
-        }
-        else
-        {
-          AddDrawTriangleTicks(positions[2], positions[1], positions[3], rc.shading_enable, rc.texture_enable,
-                               rc.transparency_enable);
-        }
-      }
-
-      m_backend.PushCommand(cmd);
-    }
-    break;
-
-    case GPUPrimitive::Rectangle:
-    {
-      GPUBackendDrawRectangleCommand* cmd = m_backend.NewDrawRectangleCommand();
-      FillDrawCommand(cmd, rc);
-      cmd->color = rc.color_for_first_vertex;
-
-      const GPUVertexPosition vp{FifoPop()};
-      cmd->x = TruncateGPUVertexPosition(m_drawing_offset.x + vp.x);
-      cmd->y = TruncateGPUVertexPosition(m_drawing_offset.y + vp.y);
-
-      if (rc.texture_enable)
-      {
-        const u32 texcoord_and_palette = FifoPop();
-        cmd->palette.bits = Truncate16(texcoord_and_palette >> 16);
-        cmd->texcoord = Truncate16(texcoord_and_palette);
-      }
-      else
-      {
-        cmd->palette.bits = 0;
-        cmd->texcoord = 0;
-      }
-
-      switch (rc.rectangle_size)
-      {
-        case GPUDrawRectangleSize::R1x1:
-          cmd->width = 1;
-          cmd->height = 1;
-          break;
-        case GPUDrawRectangleSize::R8x8:
-          cmd->width = 8;
-          cmd->height = 8;
-          break;
-        case GPUDrawRectangleSize::R16x16:
-          cmd->width = 16;
-          cmd->height = 16;
-          break;
-        default:
-        {
-          const u32 width_and_height = FifoPop();
-          cmd->width = static_cast<u16>(width_and_height & VRAM_WIDTH_MASK);
-          cmd->height = static_cast<u16>((width_and_height >> 16) & VRAM_HEIGHT_MASK);
-        }
-        break;
-      }
-
-      const GSVector4i rect = GSVector4i(cmd->x, cmd->y, cmd->x + cmd->width, cmd->y + cmd->height);
-      const GSVector4i clamped_rect = m_clamped_drawing_area.rintersect(rect);
-      if (clamped_rect.rempty()) [[unlikely]]
-      {
-        DEBUG_LOG("Culling off-screen rectangle {}", rect);
-        return;
-      }
-
-      AddDrawRectangleTicks(clamped_rect, rc.texture_enable, rc.transparency_enable);
-
-      m_backend.PushCommand(cmd);
-    }
-    break;
-
-    case GPUPrimitive::Line:
-    {
-      if (!rc.polyline)
-      {
-        GPUBackendDrawLineCommand* cmd = m_backend.NewDrawLineCommand(2);
-        FillDrawCommand(cmd, rc);
-        cmd->palette.bits = 0;
-
-        if (rc.shading_enable)
-        {
-          cmd->vertices[0].color = rc.color_for_first_vertex;
-          const GPUVertexPosition start_pos{FifoPop()};
-          cmd->vertices[0].x = m_drawing_offset.x + start_pos.x;
-          cmd->vertices[0].y = m_drawing_offset.y + start_pos.y;
-
-          cmd->vertices[1].color = FifoPop() & UINT32_C(0x00FFFFFF);
-          const GPUVertexPosition end_pos{FifoPop()};
-          cmd->vertices[1].x = m_drawing_offset.x + end_pos.x;
-          cmd->vertices[1].y = m_drawing_offset.y + end_pos.y;
-        }
-        else
-        {
-          cmd->vertices[0].color = rc.color_for_first_vertex;
-          cmd->vertices[1].color = rc.color_for_first_vertex;
-
-          const GPUVertexPosition start_pos{FifoPop()};
-          cmd->vertices[0].x = m_drawing_offset.x + start_pos.x;
-          cmd->vertices[0].y = m_drawing_offset.y + start_pos.y;
-
-          const GPUVertexPosition end_pos{FifoPop()};
-          cmd->vertices[1].x = m_drawing_offset.x + end_pos.x;
-          cmd->vertices[1].y = m_drawing_offset.y + end_pos.y;
-        }
-
-        const GSVector4i v0 = GSVector4i::loadl(&cmd->vertices[0].x);
-        const GSVector4i v1 = GSVector4i::loadl(&cmd->vertices[1].x);
-        const GSVector4i rect = v0.min_i32(v1).xyxy(v0.max_i32(v1)).add32(GSVector4i::cxpr(0, 0, 1, 1));
-        const GSVector4i clamped_rect = rect.rintersect(m_clamped_drawing_area);
-
-        if (rect.width() > MAX_PRIMITIVE_WIDTH || rect.height() > MAX_PRIMITIVE_HEIGHT || clamped_rect.rempty())
-        {
-          DEBUG_LOG("Culling too-large/off-screen line: {},{} - {},{}", cmd->vertices[0].y, cmd->vertices[0].y,
-                    cmd->vertices[1].x, cmd->vertices[1].y);
-          return;
-        }
-
-        AddDrawLineTicks(clamped_rect, rc.shading_enable);
-
-        m_backend.PushCommand(cmd);
-      }
-      else
-      {
-        const u32 num_vertices = GetPolyLineVertexCount();
-
-        GPUBackendDrawLineCommand* cmd = m_backend.NewDrawLineCommand(num_vertices);
-        FillDrawCommand(cmd, m_render_command);
-
-        u32 buffer_pos = 0;
-        const GPUVertexPosition start_vp{m_blit_buffer[buffer_pos++]};
-        cmd->vertices[0].x = start_vp.x + m_drawing_offset.x;
-        cmd->vertices[0].y = start_vp.y + m_drawing_offset.y;
-        cmd->vertices[0].color = m_render_command.color_for_first_vertex;
-
-        const bool shaded = m_render_command.shading_enable;
-        for (u32 i = 1; i < num_vertices; i++)
-        {
-          cmd->vertices[i].color =
-            shaded ? (m_blit_buffer[buffer_pos++] & UINT32_C(0x00FFFFFF)) : m_render_command.color_for_first_vertex;
-          const GPUVertexPosition vp{m_blit_buffer[buffer_pos++]};
-          cmd->vertices[i].x = m_drawing_offset.x + vp.x;
-          cmd->vertices[i].y = m_drawing_offset.y + vp.y;
-
-          const GSVector4i v0 = GSVector4i::loadl(&cmd->vertices[0].x);
-          const GSVector4i v1 = GSVector4i::loadl(&cmd->vertices[1].x);
-          const GSVector4i rect = v0.min_i32(v1).xyxy(v0.max_i32(v1)).add32(GSVector4i::cxpr(0, 0, 1, 1));
-          const GSVector4i clamped_rect = rect.rintersect(m_clamped_drawing_area);
-
-          if (rect.width() > MAX_PRIMITIVE_WIDTH || rect.height() > MAX_PRIMITIVE_HEIGHT || clamped_rect.rempty())
-          {
-            DEBUG_LOG("Culling too-large/off-screen line: {},{} - {},{}", cmd->vertices[i - 1].x,
-                      cmd->vertices[i - 1].y, cmd->vertices[i].x, cmd->vertices[i].y);
-            return;
-          }
-          else
-          {
-            AddDrawLineTicks(clamped_rect, rc.shading_enable);
-          }
-        }
-
-        m_backend.PushCommand(cmd);
-      }
-    }
-    break;
-
-    default:
-      UnreachableCode();
-      break;
-  }
-}
-
-void GPU_SW::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
-{
-  m_backend.Sync(false);
-}
-
-void GPU_SW::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color)
-{
-  GPUBackendFillVRAMCommand* cmd = m_backend.NewFillVRAMCommand();
-  FillBackendCommandParameters(cmd);
-  cmd->x = static_cast<u16>(x);
-  cmd->y = static_cast<u16>(y);
-  cmd->width = static_cast<u16>(width);
-  cmd->height = static_cast<u16>(height);
-  cmd->color = color;
-  m_backend.PushCommand(cmd);
-}
-
-void GPU_SW::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask)
-{
-  const u32 num_words = width * height;
-  GPUBackendUpdateVRAMCommand* cmd = m_backend.NewUpdateVRAMCommand(num_words);
-  FillBackendCommandParameters(cmd);
-  cmd->params.set_mask_while_drawing = set_mask;
-  cmd->params.check_mask_before_draw = check_mask;
-  cmd->x = static_cast<u16>(x);
-  cmd->y = static_cast<u16>(y);
-  cmd->width = static_cast<u16>(width);
-  cmd->height = static_cast<u16>(height);
-  std::memcpy(cmd->data, data, sizeof(u16) * num_words);
-  m_backend.PushCommand(cmd);
-}
-
-void GPU_SW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height)
-{
-  GPUBackendCopyVRAMCommand* cmd = m_backend.NewCopyVRAMCommand();
-  FillBackendCommandParameters(cmd);
-  cmd->src_x = static_cast<u16>(src_x);
-  cmd->src_y = static_cast<u16>(src_y);
-  cmd->dst_x = static_cast<u16>(dst_x);
-  cmd->dst_y = static_cast<u16>(dst_y);
-  cmd->width = static_cast<u16>(width);
-  cmd->height = static_cast<u16>(height);
-  m_backend.PushCommand(cmd);
-}
-
-void GPU_SW::FlushRender()
-{
-}
-
-void GPU_SW::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
-{
-  GPUBackendUpdateCLUTCommand* cmd = m_backend.NewUpdateCLUTCommand();
-  FillBackendCommandParameters(cmd);
-  cmd->reg.bits = reg.bits;
-  cmd->clut_is_8bit = clut_is_8bit;
-  m_backend.PushCommand(cmd);
-}
-
-std::unique_ptr<GPU> GPU::CreateSoftwareRenderer()
-{
-  std::unique_ptr<GPU_SW> gpu(std::make_unique<GPU_SW>());
-  if (!gpu->Initialize())
-    return nullptr;
-
-  return gpu;
+  return std::make_unique<GPU_SW>();
 }

src/core/gpu_sw.h

@@ -1,9 +1,10 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
+
 #include "gpu.h"
-#include "gpu_sw_backend.h"
+#include "gpu_backend.h"
 
 #include "util/gpu_device.h"
 
@@ -13,36 +14,42 @@
 #include <memory>
 #include <vector>
 
-namespace Threading {
-class Thread;
-}
-
-class GPUTexture;
-
-class GPU_SW final : public GPU
+// TODO: Move to cpp
+// TODO: Rename to GPUSWBackend, preserved to avoid conflicts.
+class GPU_SW final : public GPUBackend
 {
 public:
   GPU_SW();
   ~GPU_SW() override;
 
-  ALWAYS_INLINE const GPU_SW_Backend& GetBackend() const { return m_backend; }
+  bool Initialize(bool clear_vram, Error* error) override;
 
-  const Threading::Thread* GetSWThread() const override;
-  bool IsHardwareRenderer() const override;
+  std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) const override;
+  std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) const override;
 
-  bool Initialize() override;
-  bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display) override;
-  void Reset(bool clear_vram) override;
-  void UpdateSettings(const Settings& old_settings) override;
+  void UpdateResolutionScale() override;
 
-protected:
-  void ReadVRAM(u32 x, u32 y, u32 width, u32 height) override;
-  void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) override;
-  void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) override;
-  void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
   void FlushRender() override;
-  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
 
+  void RestoreDeviceContext() override;
+
+  bool DoState(GPUTexture** host_texture, bool is_reading, bool update_display) override;
+  void ClearVRAM() override;
+
+  void ReadVRAM(u32 x, u32 y, u32 width, u32 height) override;
+
+  void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) override;
+  void DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd) override;
+  void DrawLine(const GPUBackendDrawLineCommand* cmd) override;
+  void DrawSprite(const GPUBackendDrawRectangleCommand * cmd) override;
+  void DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area) override;
+  void ClearCache() override;
+  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
+  void OnBufferSwapped() override;
+
+  void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) override;
+
+private:
   template<GPUTexture::Format display_format>
   bool CopyOut15Bit(u32 src_x, u32 src_y, u32 width, u32 height, u32 line_skip);
 
@@ -51,19 +58,11 @@ protected:
 
   bool CopyOut(u32 src_x, u32 src_y, u32 skip_x, u32 width, u32 height, u32 line_skip, bool is_24bit);
 
-  void UpdateDisplay() override;
-
-  void DispatchRenderCommand() override;
-
-  void FillBackendCommandParameters(GPUBackendCommand* cmd) const;
-  void FillDrawCommand(GPUBackendDrawCommand* cmd, GPURenderCommand rc) const;
-
+  void SetDisplayTextureFormat();
   GPUTexture* GetDisplayTexture(u32 width, u32 height, GPUTexture::Format format);
 
   FixedHeapArray<u8, GPU_MAX_DISPLAY_WIDTH * GPU_MAX_DISPLAY_HEIGHT * sizeof(u32)> m_upload_buffer;
   GPUTexture::Format m_16bit_display_format = GPUTexture::Format::RGB565;
   GPUTexture::Format m_24bit_display_format = GPUTexture::Format::RGBA8;
   std::unique_ptr<GPUTexture> m_upload_texture;
-
-  GPU_SW_Backend m_backend;
 };

src/core/gpu_sw_backend.cpp

@@ -1,227 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
-
-#include "gpu_sw_backend.h"
-#include "gpu.h"
-#include "gpu_sw_rasterizer.h"
-#include "system.h"
-
-#include "util/gpu_device.h"
-
-#include <algorithm>
-
-GPU_SW_Backend::GPU_SW_Backend() = default;
-
-GPU_SW_Backend::~GPU_SW_Backend() = default;
-
-bool GPU_SW_Backend::Initialize(bool force_thread)
-{
-  GPU_SW_Rasterizer::SelectImplementation();
-
-  return GPUBackend::Initialize(force_thread);
-}
-
-void GPU_SW_Backend::Reset()
-{
-  GPUBackend::Reset();
-}
-
-void GPU_SW_Backend::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)
-{
-  const GPURenderCommand rc{cmd->rc.bits};
-  const bool dithering_enable = rc.IsDitheringEnabled() && cmd->draw_mode.dither_enable;
-
-  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
-    rc.shading_enable, rc.texture_enable, rc.raw_texture_enable, rc.transparency_enable, dithering_enable);
-
-  DrawFunction(cmd, &cmd->vertices[0], &cmd->vertices[1], &cmd->vertices[2]);
-  if (rc.quad_polygon)
-    DrawFunction(cmd, &cmd->vertices[2], &cmd->vertices[1], &cmd->vertices[3]);
-}
-
-void GPU_SW_Backend::DrawRectangle(const GPUBackendDrawRectangleCommand* cmd)
-{
-  const GPURenderCommand rc{cmd->rc.bits};
-
-  const GPU_SW_Rasterizer::DrawRectangleFunction DrawFunction =
-    GPU_SW_Rasterizer::GetDrawRectangleFunction(rc.texture_enable, rc.raw_texture_enable, rc.transparency_enable);
-
-  DrawFunction(cmd);
-}
-
-void GPU_SW_Backend::DrawLine(const GPUBackendDrawLineCommand* cmd)
-{
-  const GPU_SW_Rasterizer::DrawLineFunction DrawFunction = GPU_SW_Rasterizer::GetDrawLineFunction(
-    cmd->rc.shading_enable, cmd->rc.transparency_enable, cmd->IsDitheringEnabled());
-
-  for (u16 i = 1; i < cmd->num_vertices; i++)
-    DrawFunction(cmd, &cmd->vertices[i - 1], &cmd->vertices[i]);
-}
-
-void GPU_SW_Backend::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params)
-{
-  const u16 color16 = VRAMRGBA8888ToRGBA5551(color);
-  if ((x + width) <= VRAM_WIDTH && !params.interlaced_rendering)
-  {
-    for (u32 yoffs = 0; yoffs < height; yoffs++)
-    {
-      const u32 row = (y + yoffs) % VRAM_HEIGHT;
-      std::fill_n(&g_vram[row * VRAM_WIDTH + x], width, color16);
-    }
-  }
-  else if (params.interlaced_rendering)
-  {
-    // Hardware tests show that fills seem to break on the first two lines when the offset matches the displayed field.
-    const u32 active_field = params.active_line_lsb;
-    for (u32 yoffs = 0; yoffs < height; yoffs++)
-    {
-      const u32 row = (y + yoffs) % VRAM_HEIGHT;
-      if ((row & u32(1)) == active_field)
-        continue;
-
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH];
-      for (u32 xoffs = 0; xoffs < width; xoffs++)
-      {
-        const u32 col = (x + xoffs) % VRAM_WIDTH;
-        row_ptr[col] = color16;
-      }
-    }
-  }
-  else
-  {
-    for (u32 yoffs = 0; yoffs < height; yoffs++)
-    {
-      const u32 row = (y + yoffs) % VRAM_HEIGHT;
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH];
-      for (u32 xoffs = 0; xoffs < width; xoffs++)
-      {
-        const u32 col = (x + xoffs) % VRAM_WIDTH;
-        row_ptr[col] = color16;
-      }
-    }
-  }
-}
-
-void GPU_SW_Backend::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data,
-                                GPUBackendCommandParameters params)
-{
-  // Fast path when the copy is not oversized.
-  if ((x + width) <= VRAM_WIDTH && (y + height) <= VRAM_HEIGHT && !params.IsMaskingEnabled())
-  {
-    const u16* src_ptr = static_cast<const u16*>(data);
-    u16* dst_ptr = &g_vram[y * VRAM_WIDTH + x];
-    for (u32 yoffs = 0; yoffs < height; yoffs++)
-    {
-      std::copy_n(src_ptr, width, dst_ptr);
-      src_ptr += width;
-      dst_ptr += VRAM_WIDTH;
-    }
-  }
-  else
-  {
-    // Slow path when we need to handle wrap-around.
-    const u16* src_ptr = static_cast<const u16*>(data);
-    const u16 mask_and = params.GetMaskAND();
-    const u16 mask_or = params.GetMaskOR();
-
-    for (u32 row = 0; row < height;)
-    {
-      u16* dst_row_ptr = &g_vram[((y + row++) % VRAM_HEIGHT) * VRAM_WIDTH];
-      for (u32 col = 0; col < width;)
-      {
-        // TODO: Handle unaligned reads...
-        u16* pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
-        if (((*pixel_ptr) & mask_and) == 0)
-          *pixel_ptr = *(src_ptr++) | mask_or;
-      }
-    }
-  }
-}
-
-void GPU_SW_Backend::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
-                              GPUBackendCommandParameters params)
-{
-  // Break up oversized copies. This behavior has not been verified on console.
-  if ((src_x + width) > VRAM_WIDTH || (dst_x + width) > VRAM_WIDTH)
-  {
-    u32 remaining_rows = height;
-    u32 current_src_y = src_y;
-    u32 current_dst_y = dst_y;
-    while (remaining_rows > 0)
-    {
-      const u32 rows_to_copy =
-        std::min<u32>(remaining_rows, std::min<u32>(VRAM_HEIGHT - current_src_y, VRAM_HEIGHT - current_dst_y));
-
-      u32 remaining_columns = width;
-      u32 current_src_x = src_x;
-      u32 current_dst_x = dst_x;
-      while (remaining_columns > 0)
-      {
-        const u32 columns_to_copy =
-          std::min<u32>(remaining_columns, std::min<u32>(VRAM_WIDTH - current_src_x, VRAM_WIDTH - current_dst_x));
-        CopyVRAM(current_src_x, current_src_y, current_dst_x, current_dst_y, columns_to_copy, rows_to_copy, params);
-        current_src_x = (current_src_x + columns_to_copy) % VRAM_WIDTH;
-        current_dst_x = (current_dst_x + columns_to_copy) % VRAM_WIDTH;
-        remaining_columns -= columns_to_copy;
-      }
-
-      current_src_y = (current_src_y + rows_to_copy) % VRAM_HEIGHT;
-      current_dst_y = (current_dst_y + rows_to_copy) % VRAM_HEIGHT;
-      remaining_rows -= rows_to_copy;
-    }
-
-    return;
-  }
-
-  // This doesn't have a fast path, but do we really need one? It's not common.
-  const u16 mask_and = params.GetMaskAND();
-  const u16 mask_or = params.GetMaskOR();
-
-  // Copy in reverse when src_x < dst_x, this is verified on console.
-  if (src_x < dst_x || ((src_x + width - 1) % VRAM_WIDTH) < ((dst_x + width - 1) % VRAM_WIDTH))
-  {
-    for (u32 row = 0; row < height; row++)
-    {
-      const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
-      u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
-
-      for (s32 col = static_cast<s32>(width - 1); col >= 0; col--)
-      {
-        const u16 src_pixel = src_row_ptr[(src_x + static_cast<u32>(col)) % VRAM_WIDTH];
-        u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + static_cast<u32>(col)) % VRAM_WIDTH];
-        if ((*dst_pixel_ptr & mask_and) == 0)
-          *dst_pixel_ptr = src_pixel | mask_or;
-      }
-    }
-  }
-  else
-  {
-    for (u32 row = 0; row < height; row++)
-    {
-      const u16* src_row_ptr = &g_vram[((src_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
-      u16* dst_row_ptr = &g_vram[((dst_y + row) % VRAM_HEIGHT) * VRAM_WIDTH];
-
-      for (u32 col = 0; col < width; col++)
-      {
-        const u16 src_pixel = src_row_ptr[(src_x + col) % VRAM_WIDTH];
-        u16* dst_pixel_ptr = &dst_row_ptr[(dst_x + col) % VRAM_WIDTH];
-        if ((*dst_pixel_ptr & mask_and) == 0)
-          *dst_pixel_ptr = src_pixel | mask_or;
-      }
-    }
-  }
-}
-
-void GPU_SW_Backend::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
-{
-  GPU::ReadCLUT(g_gpu_clut, reg, clut_is_8bit);
-}
-
-void GPU_SW_Backend::DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area)
-{
-  GPU_SW_Rasterizer::g_drawing_area = new_drawing_area;
-}
-
-void GPU_SW_Backend::FlushRender()
-{
-}

src/core/gpu_sw_backend.h

@@ -1,34 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
-
-#pragma once
-
-#include "gpu.h"
-#include "gpu_backend.h"
-
-#include <array>
-#include <memory>
-#include <vector>
-
-class GPU_SW_Backend final : public GPUBackend
-{
-public:
-  GPU_SW_Backend();
-  ~GPU_SW_Backend() override;
-
-  bool Initialize(bool force_thread) override;
-  void Reset() override;
-
-protected:
-  void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, GPUBackendCommandParameters params) override;
-  void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, GPUBackendCommandParameters params) override;
-  void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
-                GPUBackendCommandParameters params) override;
-
-  void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) override;
-  void DrawLine(const GPUBackendDrawLineCommand* cmd) override;
-  void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd) override;
-  void DrawingAreaChanged(const GPUDrawingArea& new_drawing_area, const GSVector4i clamped_drawing_area) override;
-  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
-  void FlushRender() override;
-};

src/core/gpu_sw_rasterizer.cpp

@@ -36,6 +36,32 @@ constinit const DitherLUT g_dither_lut = []() constexpr {
 GPUDrawingArea g_drawing_area = {};
 } // namespace GPU_SW_Rasterizer
 
+void GPU_SW_Rasterizer::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  u16* const dest = g_gpu_clut;
+  const u16* const src_row = &g_vram[reg.GetYBase() * VRAM_WIDTH];
+  const u32 start_x = reg.GetXBase();
+  if (!clut_is_8bit)
+  {
+    // Wraparound can't happen in 4-bit mode.
+    std::memcpy(g_gpu_clut, &src_row[start_x], sizeof(u16) * 16);
+  }
+  else
+  {
+    if ((start_x + 256) > VRAM_WIDTH) [[unlikely]]
+    {
+      const u32 end = VRAM_WIDTH - start_x;
+      const u32 start = 256 - end;
+      std::memcpy(dest, &src_row[start_x], sizeof(u16) * end);
+      std::memcpy(dest + end, src_row, sizeof(u16) * start);
+    }
+    else
+    {
+      std::memcpy(dest, &src_row[start_x], sizeof(u16) * 256);
+    }
+  }
+}
+
 // Default implementation definitions.
 namespace GPU_SW_Rasterizer {
 #include "gpu_sw_rasterizer.inl"

src/core/gpu_sw_rasterizer.h

@@ -20,12 +20,15 @@ static constexpr u32 DITHER_LUT_SIZE = 512;
 using DitherLUT = std::array<std::array<std::array<u8, DITHER_LUT_SIZE>, DITHER_MATRIX_SIZE>, DITHER_MATRIX_SIZE>;
 extern const DitherLUT g_dither_lut;
 
+// TODO: Pack in struct
 extern GPUDrawingArea g_drawing_area;
 
+extern void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit);
+
 using DrawRectangleFunction = void (*)(const GPUBackendDrawRectangleCommand* cmd);
 typedef const DrawRectangleFunction DrawRectangleFunctionTable[2][2][2];
 
-using DrawTriangleFunction = void (*)(const GPUBackendDrawPolygonCommand* cmd,
+using DrawTriangleFunction = void (*)(const GPUBackendDrawCommand* cmd,
                                       const GPUBackendDrawPolygonCommand::Vertex* v0,
                                       const GPUBackendDrawPolygonCommand::Vertex* v1,
                                       const GPUBackendDrawPolygonCommand::Vertex* v2);

src/core/gpu_sw_rasterizer.inl

@@ -809,7 +809,7 @@ ALWAYS_INLINE_RELEASE static void AddIDeltas_DY(i_group& ig, const i_deltas& idl
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable,
          bool dithering_enable>
-ALWAYS_INLINE_RELEASE static void DrawSpan(const GPUBackendDrawPolygonCommand* cmd, s32 y, s32 x_start, s32 x_bound,
+ALWAYS_INLINE_RELEASE static void DrawSpan(const GPUBackendDrawCommand* cmd, s32 y, s32 x_start, s32 x_bound,
                                            i_group ig, const i_deltas& idl)
 {
   if (cmd->params.interlaced_rendering && cmd->params.active_line_lsb == (Truncate8(static_cast<u32>(y)) & 1u))
@@ -988,7 +988,7 @@ ALWAYS_INLINE_RELEASE static void DrawSpan(const GPUBackendDrawCommand* cmd, s32
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable,
          bool dithering_enable>
-static void DrawTriangle(const GPUBackendDrawPolygonCommand* cmd, const GPUBackendDrawPolygonCommand::Vertex* v0,
+static void DrawTriangle(const GPUBackendDrawCommand* cmd, const GPUBackendDrawPolygonCommand::Vertex* v0,
                          const GPUBackendDrawPolygonCommand::Vertex* v1, const GPUBackendDrawPolygonCommand::Vertex* v2)
 {
 #if 0

src/core/gpu_thread.cpp

@@ -0,0 +1,862 @@
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
+
+#include "gpu_thread.h"
+#include "fullscreen_ui.h"
+#include "gpu_backend.h"
+#include "host.h"
+#include "imgui_overlays.h"
+#include "settings.h"
+#include "shader_cache_version.h"
+#include "system.h"
+
+#include "util/gpu_device.h"
+#include "util/imgui_manager.h"
+#include "util/postprocessing.h"
+#include "util/state_wrapper.h"
+
+#include "common/align.h"
+#include "common/error.h"
+#include "common/log.h"
+#include "common/threading.h"
+#include "common/timer.h"
+
+#include "IconsFontAwesome5.h"
+#include "imgui.h"
+
+#include <optional>
+
+Log_SetChannel(GPUThread);
+
+namespace GPUThread {
+enum : u32
+{
+  COMMAND_QUEUE_SIZE = 4 * 1024 * 1024,
+  THRESHOLD_TO_WAKE_GPU = 256
+};
+
+/// Starts the thread, if it hasn't already been started.
+/// TODO: Persist thread
+static bool Start(std::optional<GPURenderer> api, Error* error);
+
+static void RunGPULoop();
+static u32 GetPendingCommandSize();
+static void WakeGPUThread();
+
+static bool CreateDeviceOnThread(RenderAPI api, Error* error);
+static void DestroyDeviceOnThread();
+
+static void CreateGPUBackendOnThread(bool initialize_vram);
+static void ChangeGPUBackendOnThread();
+static void DestroyGPUBackendOnThread();
+
+static void UpdateSettingsOnThread(const Settings& old_settings);
+static void UpdateVSyncOnThread();
+static void UpdatePerformanceCountersOnThread();
+
+// TODO: Pack this crap together, don't trust LTO...
+
+static RenderAPI s_render_api = RenderAPI::None;
+static std::unique_ptr<GPUBackend> s_gpu_backend;
+static std::optional<GPURenderer> s_requested_renderer;
+static bool s_start_fullscreen_ui = false;
+static GPUVSyncMode s_requested_vsync = GPUVSyncMode::Disabled;
+static bool s_requested_allow_present_throttle = false;
+
+static bool s_last_frame_skipped = false;
+
+static Common::Timer::Value s_last_performance_counter_update_time = 0;
+static u32 s_presents_since_last_update = 0;
+static float s_accumulated_gpu_time = 0.0f;
+static float s_average_gpu_time = 0.0f;
+static float s_gpu_usage = 0.0f;
+
+static Threading::KernelSemaphore m_sync_semaphore;
+static Threading::Thread m_gpu_thread;
+static Error s_open_error;
+static std::atomic_bool s_open_flag{false};
+static std::atomic_bool s_shutdown_flag{false};
+static std::atomic_bool s_run_idle_flag{false};
+static std::atomic_flag s_performance_counters_updated = ATOMIC_FLAG_INIT;
+
+static Threading::WorkSema s_work_sema;
+
+static FixedHeapArray<u8, COMMAND_QUEUE_SIZE> m_command_fifo_data;
+alignas(HOST_CACHE_LINE_SIZE) static std::atomic<u32> m_command_fifo_read_ptr{0};
+alignas(HOST_CACHE_LINE_SIZE) static std::atomic<u32> m_command_fifo_write_ptr{0};
+} // namespace GPUThread
+
+const Threading::ThreadHandle& GPUThread::GetThreadHandle()
+{
+  return m_gpu_thread;
+}
+
+RenderAPI GPUThread::GetRenderAPI()
+{
+  std::atomic_thread_fence(std::memory_order_acquire);
+  return s_render_api;
+}
+
+bool GPUThread::IsStarted()
+{
+  return m_gpu_thread.Joinable();
+}
+
+bool GPUThread::WasFullscreenUIRequested()
+{
+  return s_start_fullscreen_ui;
+}
+
+bool GPUThread::Start(std::optional<GPURenderer> renderer, Error* error)
+{
+  Assert(!IsStarted());
+
+  INFO_LOG("Starting GPU thread...");
+
+  s_requested_renderer = renderer;
+  g_gpu_settings = g_settings;
+
+  s_last_performance_counter_update_time = Common::Timer::GetCurrentValue();
+  s_presents_since_last_update = 0;
+  s_average_gpu_time = 0.0f;
+  s_gpu_usage = 0.0f;
+  GPUBackend::ResetStatistics();
+
+  s_shutdown_flag.store(false, std::memory_order_release);
+  s_run_idle_flag.store(false, std::memory_order_release);
+  m_gpu_thread.Start(&GPUThread::RunGPULoop);
+  m_sync_semaphore.Wait();
+
+  if (!s_open_flag.load(std::memory_order_acquire))
+  {
+    ERROR_LOG("Failed to create GPU thread.");
+    if (error)
+      *error = s_open_error;
+
+    m_gpu_thread.Join();
+    return false;
+  }
+
+  VERBOSE_LOG("GPU thread started.");
+  return true;
+}
+
+bool GPUThread::StartFullscreenUI(Error* error)
+{
+  // NOTE: Racey read.
+  if (FullscreenUI::IsInitialized())
+    return true;
+
+  if (IsStarted())
+  {
+    RunOnThread([]() {
+      // TODO: Error handling.
+      if (!FullscreenUI::Initialize())
+        Panic("Failed");
+    });
+
+    return true;
+  }
+
+  s_start_fullscreen_ui = true;
+  if (!Start(std::nullopt, error))
+  {
+    s_start_fullscreen_ui = false;
+    return false;
+  }
+
+  return true;
+}
+
+std::optional<GPURenderer> GPUThread::GetRequestedRenderer()
+{
+  return s_requested_renderer;
+}
+
+bool GPUThread::CreateGPUBackend(GPURenderer renderer, Error* error)
+{
+  if (IsStarted())
+  {
+    s_requested_renderer = renderer;
+    std::atomic_thread_fence(std::memory_order_release);
+    PushCommandAndSync(AllocateCommand(GPUBackendCommandType::ChangeBackend, sizeof(GPUThreadCommand)), false);
+    return true;
+  }
+  else
+  {
+    return Start(renderer, error);
+  }
+}
+
+bool GPUThread::SwitchGPUBackend(GPURenderer renderer, bool force_recreate_device, Error* error)
+{
+  if (!force_recreate_device)
+  {
+    s_requested_renderer = renderer;
+    std::atomic_thread_fence(std::memory_order_release);
+    PushCommandAndSync(AllocateCommand(GPUBackendCommandType::ChangeBackend, sizeof(GPUThreadCommand)), false);
+    return true;
+  }
+
+  const bool was_running_fsui = s_start_fullscreen_ui;
+  Shutdown();
+  s_requested_renderer = renderer;
+  s_start_fullscreen_ui = was_running_fsui;
+  if (!Start(renderer, error))
+  {
+    s_requested_renderer.reset();
+    s_start_fullscreen_ui = false;
+    return false;
+  }
+
+  return true;
+}
+
+void GPUThread::DestroyGPUBackend()
+{
+  if (!IsStarted())
+    return;
+
+  if (s_start_fullscreen_ui)
+  {
+    VERBOSE_LOG("Keeping GPU thread open for fullscreen UI");
+    s_requested_renderer.reset();
+    std::atomic_thread_fence(std::memory_order_release);
+    PushCommandAndSync(AllocateCommand(GPUBackendCommandType::ChangeBackend, sizeof(GPUThreadCommand)), false);
+    return;
+  }
+
+  Shutdown();
+}
+
+void GPUThread::Shutdown()
+{
+  if (!IsStarted())
+    return;
+
+  s_shutdown_flag.store(true, std::memory_order_release);
+  s_start_fullscreen_ui = false;
+  s_requested_renderer.reset();
+
+  WakeGPUThread();
+  m_gpu_thread.Join();
+  INFO_LOG("GPU thread stopped.");
+}
+
+GPUThreadCommand* GPUThread::AllocateCommand(GPUBackendCommandType command, u32 size)
+{
+  // Ensure size is a multiple of 4 so we don't end up with an unaligned command.
+  size = Common::AlignUpPow2(size, 4);
+
+  for (;;)
+  {
+    u32 read_ptr = m_command_fifo_read_ptr.load(std::memory_order_acquire);
+    u32 write_ptr = m_command_fifo_write_ptr.load(std::memory_order_relaxed);
+    if (read_ptr > write_ptr)
+    {
+      u32 available_size = read_ptr - write_ptr;
+      while (available_size < (size + sizeof(GPUBackendCommandType)))
+      {
+        WakeGPUThread();
+        read_ptr = m_command_fifo_read_ptr.load(std::memory_order_acquire);
+        available_size = (read_ptr > write_ptr) ? (read_ptr - write_ptr) : (COMMAND_QUEUE_SIZE - write_ptr);
+      }
+    }
+    else
+    {
+      const u32 available_size = COMMAND_QUEUE_SIZE - write_ptr;
+      if ((size + sizeof(GPUBackendCommand)) > available_size)
+      {
+        // allocate a dummy command to wrap the buffer around
+        GPUBackendCommand* dummy_cmd = reinterpret_cast<GPUBackendCommand*>(&m_command_fifo_data[write_ptr]);
+        dummy_cmd->type = GPUBackendCommandType::Wraparound;
+        dummy_cmd->size = available_size;
+        dummy_cmd->params.bits = 0;
+        m_command_fifo_write_ptr.store(0, std::memory_order_release);
+        continue;
+      }
+    }
+
+    GPUThreadCommand* cmd = reinterpret_cast<GPUThreadCommand*>(&m_command_fifo_data[write_ptr]);
+    cmd->type = command;
+    cmd->size = size;
+    return cmd;
+  }
+}
+
+u32 GPUThread::GetPendingCommandSize()
+{
+  const u32 read_ptr = m_command_fifo_read_ptr.load();
+  const u32 write_ptr = m_command_fifo_write_ptr.load();
+  return (write_ptr >= read_ptr) ? (write_ptr - read_ptr) : (COMMAND_QUEUE_SIZE - read_ptr + write_ptr);
+}
+
+void GPUThread::PushCommand(GPUThreadCommand* cmd)
+{
+  const u32 new_write_ptr = m_command_fifo_write_ptr.fetch_add(cmd->size, std::memory_order_release) + cmd->size;
+  DebugAssert(new_write_ptr <= COMMAND_QUEUE_SIZE);
+  UNREFERENCED_VARIABLE(new_write_ptr);
+  if (GetPendingCommandSize() >= THRESHOLD_TO_WAKE_GPU)
+    WakeGPUThread();
+}
+
+void GPUThread::PushCommandAndWakeThread(GPUThreadCommand* cmd)
+{
+  const u32 new_write_ptr = m_command_fifo_write_ptr.fetch_add(cmd->size, std::memory_order_release) + cmd->size;
+  DebugAssert(new_write_ptr <= COMMAND_QUEUE_SIZE);
+  UNREFERENCED_VARIABLE(new_write_ptr);
+  WakeGPUThread();
+}
+
+void GPUThread::PushCommandAndSync(GPUThreadCommand* cmd, bool spin)
+{
+  const u32 new_write_ptr = m_command_fifo_write_ptr.fetch_add(cmd->size, std::memory_order_release) + cmd->size;
+  DebugAssert(new_write_ptr <= COMMAND_QUEUE_SIZE);
+  UNREFERENCED_VARIABLE(new_write_ptr);
+  WakeGPUThread();
+
+  if (spin)
+    s_work_sema.WaitForEmptyWithSpin();
+  else
+    s_work_sema.WaitForEmpty();
+}
+
+void GPUThread::WakeGPUThread()
+{
+  s_work_sema.NotifyOfWork();
+}
+
+void GPUThread::RunGPULoop()
+{
+  Threading::SetNameOfCurrentThread("GPUThread");
+
+  if (!CreateDeviceOnThread(
+        Settings::GetRenderAPIForRenderer(s_requested_renderer.value_or(g_gpu_settings.gpu_renderer)), &s_open_error))
+  {
+    Host::ReleaseRenderWindow();
+    s_open_flag.store(false, std::memory_order_release);
+    m_sync_semaphore.Post();
+    return;
+  }
+
+  CreateGPUBackendOnThread(true);
+
+  s_open_flag.store(true, std::memory_order_release);
+  m_sync_semaphore.Post();
+
+  for (;;)
+  {
+    u32 write_ptr = m_command_fifo_write_ptr.load(std::memory_order_acquire);
+    u32 read_ptr = m_command_fifo_read_ptr.load(std::memory_order_relaxed);
+    if (read_ptr == write_ptr)
+    {
+      if (s_shutdown_flag.load(std::memory_order_relaxed))
+      {
+        break;
+      }
+      else if (s_run_idle_flag.load(std::memory_order_relaxed))
+      {
+        if (!s_work_sema.CheckForWork())
+        {
+          Internal::PresentFrame(false, 0);
+          if (!g_gpu_device->IsVSyncModeBlocking())
+            g_gpu_device->ThrottlePresentation();
+
+          continue;
+        }
+
+        // we should have something to do, since we got woken...
+      }
+      else
+      {
+        s_work_sema.WaitForWork();
+        continue;
+      }
+    }
+
+    write_ptr = (write_ptr < read_ptr) ? COMMAND_QUEUE_SIZE : write_ptr;
+    while (read_ptr < write_ptr)
+    {
+      GPUThreadCommand* cmd = reinterpret_cast<GPUThreadCommand*>(&m_command_fifo_data[read_ptr]);
+      DebugAssert((read_ptr + cmd->size) <= COMMAND_QUEUE_SIZE);
+      read_ptr += cmd->size;
+
+      switch (cmd->type)
+      {
+        case GPUBackendCommandType::Wraparound:
+        {
+          DebugAssert(read_ptr == COMMAND_QUEUE_SIZE);
+          write_ptr = m_command_fifo_write_ptr.load(std::memory_order_acquire);
+          read_ptr = 0;
+
+          // let the CPU thread know as early as possible that we're here
+          m_command_fifo_read_ptr.store(read_ptr, std::memory_order_release);
+        }
+        break;
+
+        case GPUBackendCommandType::AsyncCall:
+        {
+          GPUThreadAsyncCallCommand* acmd = static_cast<GPUThreadAsyncCallCommand*>(cmd);
+          acmd->func();
+          acmd->~GPUThreadAsyncCallCommand();
+        }
+        break;
+
+        case GPUBackendCommandType::ChangeBackend:
+        {
+          ChangeGPUBackendOnThread();
+        }
+        break;
+
+        case GPUBackendCommandType::UpdateVSync:
+        {
+          UpdateVSyncOnThread();
+        }
+        break;
+
+        default:
+        {
+          DebugAssert(s_gpu_backend);
+          s_gpu_backend->HandleCommand(cmd);
+        }
+        break;
+      }
+    }
+
+    m_command_fifo_read_ptr.store(read_ptr, std::memory_order_release);
+  }
+
+  DestroyGPUBackendOnThread();
+  DestroyDeviceOnThread();
+  Host::ReleaseRenderWindow();
+}
+
+bool GPUThread::CreateDeviceOnThread(RenderAPI api, Error* error)
+{
+  DebugAssert(!g_gpu_device);
+
+  INFO_LOG("Trying to create a {} GPU device...", GPUDevice::RenderAPIToString(api));
+  g_gpu_device = GPUDevice::CreateDeviceForAPI(api);
+
+  std::optional<bool> exclusive_fullscreen_control;
+  if (g_settings.display_exclusive_fullscreen_control != DisplayExclusiveFullscreenControl::Automatic)
+  {
+    exclusive_fullscreen_control =
+      (g_settings.display_exclusive_fullscreen_control == DisplayExclusiveFullscreenControl::Allowed);
+  }
+
+  u32 disabled_features = 0;
+  if (g_settings.gpu_disable_dual_source_blend)
+    disabled_features |= GPUDevice::FEATURE_MASK_DUAL_SOURCE_BLEND;
+  if (g_settings.gpu_disable_framebuffer_fetch)
+    disabled_features |= GPUDevice::FEATURE_MASK_FRAMEBUFFER_FETCH;
+  if (g_settings.gpu_disable_texture_buffers)
+    disabled_features |= GPUDevice::FEATURE_MASK_TEXTURE_BUFFERS;
+  if (g_settings.gpu_disable_texture_copy_to_self)
+    disabled_features |= GPUDevice::FEATURE_MASK_TEXTURE_COPY_TO_SELF;
+
+  Error create_error;
+  if (!g_gpu_device ||
+      !g_gpu_device->Create(
+        g_gpu_settings.gpu_adapter,
+        g_gpu_settings.gpu_disable_shader_cache ? std::string_view() : std::string_view(EmuFolders::Cache),
+        SHADER_CACHE_VERSION, g_gpu_settings.gpu_use_debug_device, s_requested_vsync,
+        s_requested_allow_present_throttle, g_gpu_settings.gpu_threaded_presentation, exclusive_fullscreen_control,
+        static_cast<GPUDevice::FeatureMask>(disabled_features), &create_error))
+  {
+    ERROR_LOG("Failed to create GPU device: {}", create_error.GetDescription());
+    if (g_gpu_device)
+      g_gpu_device->Destroy();
+    g_gpu_device.reset();
+
+    Error::SetStringFmt(
+      error,
+      TRANSLATE_FS("System", "Failed to create render device:\n\n{0}\n\nThis may be due to your GPU not supporting the "
+                             "chosen renderer ({1}), or because your graphics drivers need to be updated."),
+      create_error.GetDescription(), GPUDevice::RenderAPIToString(api));
+
+    s_render_api = RenderAPI::None;
+    std::atomic_thread_fence(std::memory_order_release);
+    return false;
+  }
+
+  if (!ImGuiManager::Initialize(g_settings.display_osd_scale / 100.0f, g_settings.display_show_osd_messages,
+                                &create_error) ||
+      (s_start_fullscreen_ui && !FullscreenUI::Initialize()))
+  {
+    ERROR_LOG("Failed to initialize ImGuiManager: {}", create_error.GetDescription());
+    Error::SetStringFmt(error, "Failed to initialize ImGuiManager: {}", create_error.GetDescription());
+    FullscreenUI::Shutdown();
+    ImGuiManager::Shutdown();
+    g_gpu_device->Destroy();
+    g_gpu_device.reset();
+    s_render_api = RenderAPI::None;
+    std::atomic_thread_fence(std::memory_order_release);
+    return false;
+  }
+
+  s_accumulated_gpu_time = 0.0f;
+  s_presents_since_last_update = 0;
+  s_render_api = g_gpu_device->GetRenderAPI();
+  std::atomic_thread_fence(std::memory_order_release);
+
+  return true;
+}
+
+void GPUThread::DestroyDeviceOnThread()
+{
+  if (!g_gpu_device)
+    return;
+
+  ImGuiManager::DestroyOverlayTextures();
+  FullscreenUI::Shutdown();
+  ImGuiManager::Shutdown();
+
+  INFO_LOG("Destroying {} GPU device...", GPUDevice::RenderAPIToString(g_gpu_device->GetRenderAPI()));
+  g_gpu_device->Destroy();
+  g_gpu_device.reset();
+}
+
+void GPUThread::CreateGPUBackendOnThread(bool clear_vram)
+{
+  Assert(!s_gpu_backend);
+  if (!s_requested_renderer.has_value())
+    return;
+
+  const bool is_hardware = (s_requested_renderer.value() != GPURenderer::Software);
+
+  if (is_hardware)
+    s_gpu_backend = GPUBackend::CreateHardwareBackend();
+  else
+    s_gpu_backend = GPUBackend::CreateSoftwareBackend();
+
+  Error error;
+  DebugAssert(s_gpu_backend);
+  if (!s_gpu_backend->Initialize(clear_vram, &error))
+  {
+    ERROR_LOG("Failed to create {} renderer: {}", Settings::GetRendererName(s_requested_renderer.value()),
+              error.GetDescription());
+
+    if (is_hardware)
+    {
+      Host::AddIconOSDMessage(
+        "GPUBackendCreationFailed", ICON_FA_PAINT_ROLLER,
+        fmt::format(TRANSLATE_FS("OSDMessage", "Failed to initialize {} renderer, falling back to software renderer."),
+                    Settings::GetRendererName(s_requested_renderer.value())),
+        Host::OSD_CRITICAL_ERROR_DURATION);
+
+      s_requested_renderer = GPURenderer::Software;
+      s_gpu_backend = GPUBackend::CreateSoftwareBackend();
+      if (!s_gpu_backend)
+        Panic("Failed to initialize software backend.");
+    }
+  }
+}
+
+ALWAYS_INLINE_RELEASE void GPUThread::ChangeGPUBackendOnThread()
+{
+  std::atomic_thread_fence(std::memory_order_acquire);
+  if (!s_requested_renderer.has_value())
+  {
+    if (s_gpu_backend)
+      DestroyGPUBackendOnThread();
+
+    return;
+  }
+
+  // Readback old VRAM for hardware renderers.
+  s_gpu_backend->ReadVRAM(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+
+  if (s_requested_renderer.value() == GPURenderer::Software)
+  {
+    // Just recreate the backend, software works with everything.
+    DestroyGPUBackendOnThread();
+    CreateGPUBackendOnThread(false);
+    return;
+  }
+
+  DestroyGPUBackendOnThread();
+
+  DebugAssert(g_gpu_device);
+  const RenderAPI current_api = s_render_api;
+  const RenderAPI expected_api = Settings::GetRenderAPIForRenderer(s_requested_renderer.value());
+  if (!GPUDevice::IsSameRenderAPI(current_api, expected_api))
+  {
+    WARNING_LOG("Recreating GPU device, expecting {} got {}", GPUDevice::RenderAPIToString(expected_api),
+                GPUDevice::RenderAPIToString(current_api));
+
+    DestroyDeviceOnThread();
+
+    // Things tend to break when you don't recreate the window, after switching APIs.
+    Host::ReleaseRenderWindow();
+
+    Error error;
+    if (!CreateDeviceOnThread(expected_api, &error))
+    {
+      Host::AddIconOSDMessage(
+        "DeviceSwitchFailed", ICON_FA_PAINT_ROLLER,
+        fmt::format(TRANSLATE_FS("OSDMessage", "Failed to create {} GPU device, reverting to {}.\n{}"),
+                    GPUDevice::RenderAPIToString(expected_api), GPUDevice::RenderAPIToString(current_api),
+                    error.GetDescription()),
+        Host::OSD_CRITICAL_ERROR_DURATION);
+
+      Host::ReleaseRenderWindow();
+      if (!CreateDeviceOnThread(current_api, &error))
+        Panic("Failed to switch back to old API after creation failure");
+    }
+  }
+
+  CreateGPUBackendOnThread(false);
+}
+
+void GPUThread::DestroyGPUBackendOnThread()
+{
+  if (!s_gpu_backend)
+    return;
+
+  VERBOSE_LOG("Shutting down GPU backend...");
+
+  PostProcessing::Shutdown();
+  s_gpu_backend.reset();
+}
+
+void GPUThread::UpdateSettingsOnThread(const Settings& old_settings)
+{
+  if (g_gpu_settings.display_show_gpu_usage != old_settings.display_show_gpu_usage ||
+      g_gpu_settings.display_show_gpu_stats != old_settings.display_show_gpu_stats)
+  {
+    s_performance_counters_updated.clear(std::memory_order_relaxed);
+    s_last_performance_counter_update_time = Common::Timer::GetCurrentValue();
+    s_presents_since_last_update = 0;
+  }
+
+  if (g_gpu_settings.display_show_gpu_usage != old_settings.display_show_gpu_usage)
+  {
+    s_accumulated_gpu_time = 0.0f;
+    s_average_gpu_time = 0.0f;
+    s_gpu_usage = 0.0f;
+  }
+
+  if (s_gpu_backend)
+    s_gpu_backend->UpdateSettings(old_settings);
+}
+
+void GPUThread::UpdateVSyncOnThread()
+{
+  std::atomic_thread_fence(std::memory_order_acquire);
+
+  g_gpu_device->SetVSyncMode(s_requested_vsync, s_requested_allow_present_throttle);
+}
+
+void GPUThread::RunOnThread(AsyncCallType func)
+{
+  GPUThreadAsyncCallCommand* cmd = static_cast<GPUThreadAsyncCallCommand*>(
+    AllocateCommand(GPUBackendCommandType::AsyncCall, sizeof(GPUThreadAsyncCallCommand)));
+  new (cmd) GPUThreadAsyncCallCommand;
+  cmd->func = std::move(func);
+  PushCommandAndWakeThread(cmd);
+}
+
+void GPUThread::UpdateSettings()
+{
+  AssertMsg(IsStarted(), "GPU Thread is running");
+
+  RunOnThread([settings = g_settings]() {
+    VERBOSE_LOG("Updating GPU settings on thread...");
+
+    Settings old_settings = std::move(g_gpu_settings);
+    g_gpu_settings = std::move(settings);
+
+    UpdateSettingsOnThread(old_settings);
+  });
+}
+
+void GPUThread::ResizeDisplayWindow(s32 width, s32 height, float scale)
+{
+  AssertMsg(IsStarted(), "GPU Thread is running");
+  RunOnThread([width, height, scale]() {
+    if (!g_gpu_device)
+      return;
+
+    DEV_LOG("Display window resized to {}x{}", width, height);
+
+    g_gpu_device->ResizeWindow(width, height, scale);
+    ImGuiManager::WindowResized();
+
+    // If we're paused, re-present the current frame at the new window size.
+    if (System::IsValid())
+    {
+      if (System::IsPaused())
+      {
+        // Hackity hack, on some systems, presenting a single frame isn't enough to actually get it
+        // displayed. Two seems to be good enough. Maybe something to do with direct scanout.
+        PresentCurrentFrame();
+        PresentCurrentFrame();
+      }
+    }
+  });
+
+  // TODO: The window size for GTE and stuff isn't going to be correct here.
+  System::HostDisplayResized();
+}
+
+void GPUThread::UpdateDisplayWindow()
+{
+  AssertMsg(IsStarted(), "MTGS is running");
+  RunOnThread([]() {
+    if (!g_gpu_device)
+      return;
+
+    if (!g_gpu_device->UpdateWindow())
+    {
+      Host::ReportErrorAsync("Error", "Failed to change window after update. The log may contain more information.");
+      return;
+    }
+
+    ImGuiManager::WindowResized();
+
+    if (System::IsValid())
+    {
+      // Fix up vsync etc.
+      System::UpdateSpeedLimiterState();
+
+      // If we're paused, re-present the current frame at the new window size.
+      if (System::IsPaused())
+        PresentCurrentFrame();
+    }
+  });
+}
+
+void GPUThread::SetVSync(GPUVSyncMode mode, bool allow_present_throttle)
+{
+  Assert(IsStarted());
+
+  if (s_requested_vsync == mode && s_requested_allow_present_throttle == allow_present_throttle)
+    return;
+
+  s_requested_vsync = mode;
+  s_requested_allow_present_throttle = allow_present_throttle;
+  std::atomic_thread_fence(std::memory_order_release);
+  PushCommandAndWakeThread(AllocateCommand(GPUBackendCommandType::UpdateVSync, sizeof(GPUThreadCommand)));
+}
+
+void GPUThread::PresentCurrentFrame()
+{
+  if (s_run_idle_flag.load(std::memory_order_relaxed))
+  {
+    // If we're running idle, we're going to re-present anyway.
+    return;
+  }
+
+  RunOnThread([]() { Internal::PresentFrame(false, 0); });
+}
+
+void GPUThread::Internal::PresentFrame(bool allow_skip_present, Common::Timer::Value present_time)
+{
+  // Make sure the GPU is flushed, otherwise the VB might still be mapped.
+  // TODO: Make this suck less...
+  if (s_gpu_backend)
+    s_gpu_backend->FlushRender();
+
+  s_presents_since_last_update++;
+  if (!s_performance_counters_updated.test_and_set(std::memory_order_acq_rel))
+    UpdatePerformanceCountersOnThread();
+
+  const bool skip_present = (allow_skip_present && g_gpu_device->ShouldSkipPresentingFrame());
+  const bool explicit_present = (present_time != 0 && g_gpu_device->GetFeatures().explicit_present);
+
+  // TODO: Maybe?
+  //(present_time != 0 && Common::Timer::GetCurrentValue() > present_time && !s_last_frame_skipped)));
+
+  // acquire for IO.MousePos.
+  std::atomic_thread_fence(std::memory_order_acquire);
+
+  if (!skip_present)
+  {
+    FullscreenUI::Render();
+    ImGuiManager::RenderTextOverlays();
+    ImGuiManager::RenderOSDMessages();
+
+    if (System::GetState() == System::State::Running)
+      ImGuiManager::RenderSoftwareCursors();
+  }
+
+  // Debug windows are always rendered, otherwise mouse input breaks on skip.
+  ImGuiManager::RenderOverlayWindows();
+  ImGuiManager::RenderDebugWindows();
+
+  if (s_gpu_backend && !skip_present)
+    s_last_frame_skipped = !s_gpu_backend->PresentDisplay();
+  else
+    s_last_frame_skipped = !g_gpu_device->BeginPresent(skip_present);
+
+  if (!s_last_frame_skipped)
+  {
+    g_gpu_device->RenderImGui();
+    g_gpu_device->EndPresent(explicit_present);
+
+    if (g_gpu_device->IsGPUTimingEnabled())
+      s_accumulated_gpu_time += g_gpu_device->GetAndResetAccumulatedGPUTime();
+
+    if (explicit_present)
+    {
+      // See note in System::Throttle().
+#if !defined(__linux__) && !defined(__ANDROID__)
+      Common::Timer::SleepUntil(present_time, false);
+#else
+      Common::Timer::SleepUntil(present_time, true);
+#endif
+
+      g_gpu_device->SubmitPresent();
+    }
+  }
+  else
+  {
+    // Still need to kick ImGui or it gets cranky.
+    ImGui::Render();
+  }
+
+  ImGuiManager::NewFrame();
+
+  if (s_gpu_backend)
+    s_gpu_backend->RestoreDeviceContext();
+}
+
+void GPUThread::SetRunIdle(bool enabled)
+{
+  s_run_idle_flag.store(enabled, std::memory_order_release);
+  DEV_LOG("GPU thread now {} idle", enabled ? "running" : "NOT running");
+}
+
+float GPUThread::GetGPUUsage()
+{
+  return s_gpu_usage;
+}
+
+float GPUThread::GetGPUAverageTime()
+{
+  return s_average_gpu_time;
+}
+
+void GPUThread::SetPerformanceCounterUpdatePending()
+{
+  s_performance_counters_updated.clear(std::memory_order_release);
+}
+
+void GPUThread::UpdatePerformanceCountersOnThread()
+{
+  const Common::Timer::Value current_time = Common::Timer::GetCurrentValue();
+  const u32 frames = std::exchange(s_presents_since_last_update, 0);
+  const float time = static_cast<float>(Common::Timer::ConvertValueToSeconds(
+    current_time - std::exchange(s_last_performance_counter_update_time, current_time)));
+
+  if (g_gpu_device->IsGPUTimingEnabled())
+  {
+    s_average_gpu_time = s_accumulated_gpu_time / static_cast<float>(std::max(frames, 1u));
+    s_gpu_usage = static_cast<float>(s_accumulated_gpu_time / (time * 10.0));
+    s_accumulated_gpu_time = 0.0f;
+  }
+
+  if (g_settings.display_show_gpu_stats)
+    GPUBackend::UpdateStatistics(frames);
+}

src/core/gpu_thread.h

@@ -0,0 +1,66 @@
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
+
+#pragma once
+
+#include "gpu_types.h"
+
+#include "common/threading.h"
+
+#include <optional>
+
+class Error;
+
+enum class RenderAPI : u32;
+enum class GPUVSyncMode : u8;
+
+namespace GPUThread {
+using AsyncCallType = std::function<void()>;
+
+const Threading::ThreadHandle& GetThreadHandle();
+RenderAPI GetRenderAPI();
+bool IsStarted();
+bool WasFullscreenUIRequested();
+
+/// Starts Big Picture UI.
+bool StartFullscreenUI(Error* error);
+
+/// Backend control.
+std::optional<GPURenderer> GetRequestedRenderer();
+bool CreateGPUBackend(GPURenderer renderer, Error* error);
+bool SwitchGPUBackend(GPURenderer renderer, bool force_recreate_device, Error* error);
+void DestroyGPUBackend();
+
+/// Fully stops the thread, closing in the process if needed.
+void Shutdown();
+
+/// Re-presents the current frame. Call when things like window resizes happen to re-display
+/// the current frame with the correct proportions. Should only be called from the CPU thread.
+void PresentCurrentFrame();
+
+/// Handles fullscreen transitions and such.
+void UpdateDisplayWindow();
+
+/// Called when the window is resized.
+void ResizeDisplayWindow(s32 width, s32 height, float scale);
+
+void UpdateSettings();
+
+void RunOnThread(AsyncCallType func);
+void SetVSync(GPUVSyncMode mode, bool allow_present_throttle);
+void SetRunIdle(bool enabled);
+
+float GetGPUUsage();
+float GetGPUAverageTime();
+void SetPerformanceCounterUpdatePending();
+
+GPUThreadCommand* AllocateCommand(GPUBackendCommandType command, u32 size);
+void PushCommand(GPUThreadCommand* cmd);
+void PushCommandAndWakeThread(GPUThreadCommand* cmd);
+void PushCommandAndSync(GPUThreadCommand* cmd, bool spin);
+
+// NOTE: Only called by GPUBackend
+namespace Internal {
+void PresentFrame(bool allow_skip_present, Common::Timer::Value present_time);
+}
+} // namespace GPUThread

src/core/gpu_types.h

@@ -2,11 +2,21 @@
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
+
+#include "types.h"
+
+#include "util/gpu_texture.h"
+
 #include "common/bitfield.h"
 #include "common/bitutils.h"
 #include "common/gsvector.h"
-#include "types.h"
+#include "common/timer.h"
+
 #include <array>
+#include <functional>
+#include <vector>
+
+class StateWrapper;
 
 enum : u32
 {
@@ -166,7 +176,7 @@ static constexpr s32 TruncateGPUVertexPosition(s32 x)
 union GPUDrawModeReg
 {
   static constexpr u16 MASK = 0b1111111111111;
-  static constexpr u16 TEXTURE_PAGE_MASK = UINT16_C(0b0000000000011111);
+  static constexpr u16 TEXTURE_PAGE_AND_MODE_MASK = UINT16_C(0b0000000110011111);
 
   // Polygon texpage commands only affect bits 0-8, 11
   static constexpr u16 POLYGON_TEXPAGE_MASK = 0b0000100111111111;
@@ -227,7 +237,9 @@ union GPUTexturePaletteReg
   }
 };
 
-struct GPUTextureWindow
+union GPUTextureWindow
+{
+  struct
   {
     u8 and_x;
     u8 and_y;
@@ -235,6 +247,9 @@ struct GPUTextureWindow
     u8 or_y;
   };
 
+  u32 bits;
+};
+
 // 4x4 dither matrix.
 static constexpr s32 DITHER_MATRIX[DITHER_MATRIX_SIZE][DITHER_MATRIX_SIZE] = {{-4, +0, -3, +1},  // row 0
                                                                               {+2, -2, +3, -1},  // row 1
@@ -249,17 +264,100 @@ static constexpr s32 DITHER_MATRIX[DITHER_MATRIX_SIZE][DITHER_MATRIX_SIZE] = {{-
 enum class GPUBackendCommandType : u8
 {
   Wraparound,
-  Sync,
+  AsyncCall,
+  ChangeBackend,
+  UpdateVSync,
+  ClearVRAM,
+  DoState,
+  ClearDisplay,
+  UpdateDisplay,
+  ClearCache,
+  BufferSwapped,
+  FlushRender,
+  UpdateResolutionScale,
+  RenderScreenshotToBuffer,
+  ReadVRAM,
   FillVRAM,
   UpdateVRAM,
   CopyVRAM,
   SetDrawingArea,
   UpdateCLUT,
   DrawPolygon,
+  DrawPrecisePolygon,
   DrawRectangle,
   DrawLine,
 };
 
+struct GPUThreadCommand
+{
+  u32 size;
+  GPUBackendCommandType type;
+};
+
+struct GPUThreadAsyncCallCommand : public GPUThreadCommand
+{
+  std::function<void()> func;
+};
+
+struct GPUThreadRenderScreenshotToBufferCommand : public GPUThreadCommand
+{
+  u32 width;
+  u32 height;
+  s32 draw_rect[4];
+  std::vector<u32>* out_pixels;
+  u32* out_stride;
+  GPUTexture::Format* out_format;
+  bool* out_result;
+  bool postfx;
+};
+
+struct GPUBackendDoStateCommand : public GPUThreadCommand
+{
+  GPUTexture** host_texture;
+  bool is_reading;
+  bool update_display;
+};
+
+struct GPUBackendUpdateDisplayCommand : public GPUThreadCommand
+{
+  u16 display_width;
+  u16 display_height;
+  u16 display_origin_left;
+  u16 display_origin_top;
+  u16 display_vram_left;
+  u16 display_vram_top;
+  u16 display_vram_width;
+  u16 display_vram_height;
+
+  u16 X; // TODO: Can we get rid of this?
+
+  union
+  {
+    u8 bits;
+
+    BitField<u8, bool, 0, 1> interlaced_display_enabled;
+    BitField<u8, u8, 1, 1> interlaced_display_field;
+    BitField<u8, bool, 2, 1> interlaced_display_interleaved;
+    BitField<u8, bool, 3, 1> display_24bit;
+    BitField<u8, bool, 4, 1> display_disabled;
+
+    BitField<u8, bool, 6, 1> allow_present_skip;
+    BitField<u8, bool, 7, 1> present_frame;
+  };
+
+  float display_aspect_ratio;
+
+  Common::Timer::Value present_time;
+};
+
+struct GPUBackendReadVRAMCommand : public GPUThreadCommand
+{
+  u16 x;
+  u16 y;
+  u16 width;
+  u16 height;
+};
+
 union GPUBackendCommandParameters
 {
   u8 bits;
@@ -287,18 +385,12 @@ union GPUBackendCommandParameters
   }
 };
 
-struct GPUBackendCommand
+// TODO: Merge this into the other structs, saves padding bytes
+struct GPUBackendCommand : public GPUThreadCommand
 {
-  u32 size;
-  GPUBackendCommandType type;
   GPUBackendCommandParameters params;
 };
 
-struct GPUBackendSyncCommand : public GPUBackendCommand
-{
-  bool allow_sleep;
-};
-
 struct GPUBackendFillVRAMCommand : public GPUBackendCommand
 {
   u16 x;
@@ -339,8 +431,10 @@ struct GPUBackendUpdateCLUTCommand : public GPUBackendCommand
   bool clut_is_8bit;
 };
 
+// TODO: Pack texpage
 struct GPUBackendDrawCommand : public GPUBackendCommand
 {
+  // TODO: Cut this down
   GPUDrawModeReg draw_mode;
   GPURenderCommand rc;
   GPUTexturePaletteReg palette;
@@ -351,7 +445,7 @@ struct GPUBackendDrawCommand : public GPUBackendCommand
 
 struct GPUBackendDrawPolygonCommand : public GPUBackendDrawCommand
 {
-  u16 num_vertices;
+  u8 num_vertices;
 
   struct Vertex
   {
@@ -372,14 +466,22 @@ struct GPUBackendDrawPolygonCommand : public GPUBackendDrawCommand
       };
       u16 texcoord;
     };
+  };
 
-    ALWAYS_INLINE void Set(s32 x_, s32 y_, u32 color_, u16 texcoord_)
+  Vertex vertices[0];
+};
+
+struct GPUBackendDrawPrecisePolygonCommand : public GPUBackendDrawCommand
 {
-      x = x_;
-      y = y_;
-      color = color_;
-      texcoord = texcoord_;
-    }
+  u8 num_vertices;
+  bool valid_w;
+
+  struct Vertex
+  {
+    float x, y, w;
+    s32 native_x, native_y;
+    u32 color;
+    u16 texcoord;
   };
 
   Vertex vertices[0];
@@ -387,9 +489,9 @@ struct GPUBackendDrawPolygonCommand : public GPUBackendDrawCommand
 
 struct GPUBackendDrawRectangleCommand : public GPUBackendDrawCommand
 {
-  s32 x, y;
   u16 width, height;
   u16 texcoord;
+  s32 x, y;
   u32 color;
 };
 

src/core/host.cpp

@@ -2,17 +2,11 @@
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #include "host.h"
-#include "fullscreen_ui.h"
 #include "gpu.h"
-#include "imgui_overlays.h"
-#include "shader_cache_version.h"
 #include "system.h"
 
 #include "scmversion/scmversion.h"
 
-#include "util/gpu_device.h"
-#include "util/imgui_manager.h"
-
 #include "common/assert.h"
 #include "common/error.h"
 #include "common/layered_settings_interface.h"
@@ -259,125 +253,3 @@ void Host::ReportFormattedDebuggerMessage(const char* format, ...)
 
   ReportDebuggerMessage(message);
 }
-
-bool Host::CreateGPUDevice(RenderAPI api, Error* error)
-{
-  DebugAssert(!g_gpu_device);
-
-  INFO_LOG("Trying to create a {} GPU device...", GPUDevice::RenderAPIToString(api));
-  g_gpu_device = GPUDevice::CreateDeviceForAPI(api);
-
-  std::optional<bool> exclusive_fullscreen_control;
-  if (g_settings.display_exclusive_fullscreen_control != DisplayExclusiveFullscreenControl::Automatic)
-  {
-    exclusive_fullscreen_control =
-      (g_settings.display_exclusive_fullscreen_control == DisplayExclusiveFullscreenControl::Allowed);
-  }
-
-  u32 disabled_features = 0;
-  if (g_settings.gpu_disable_dual_source_blend)
-    disabled_features |= GPUDevice::FEATURE_MASK_DUAL_SOURCE_BLEND;
-  if (g_settings.gpu_disable_framebuffer_fetch)
-    disabled_features |= GPUDevice::FEATURE_MASK_FRAMEBUFFER_FETCH;
-  if (g_settings.gpu_disable_texture_buffers)
-    disabled_features |= GPUDevice::FEATURE_MASK_TEXTURE_BUFFERS;
-  if (g_settings.gpu_disable_texture_copy_to_self)
-    disabled_features |= GPUDevice::FEATURE_MASK_TEXTURE_COPY_TO_SELF;
-
-  Error create_error;
-  if (!g_gpu_device || !g_gpu_device->Create(g_settings.gpu_adapter,
-                                             g_settings.gpu_disable_shader_cache ? std::string_view() :
-                                                                                   std::string_view(EmuFolders::Cache),
-                                             SHADER_CACHE_VERSION, g_settings.gpu_use_debug_device,
-                                             System::GetEffectiveVSyncMode(), System::ShouldAllowPresentThrottle(),
-                                             g_settings.gpu_threaded_presentation, exclusive_fullscreen_control,
-                                             static_cast<GPUDevice::FeatureMask>(disabled_features), &create_error))
-  {
-    ERROR_LOG("Failed to create GPU device: {}", create_error.GetDescription());
-    if (g_gpu_device)
-      g_gpu_device->Destroy();
-    g_gpu_device.reset();
-
-    Error::SetStringFmt(
-      error,
-      TRANSLATE_FS("System", "Failed to create render device:\n\n{0}\n\nThis may be due to your GPU not supporting the "
-                             "chosen renderer ({1}), or because your graphics drivers need to be updated."),
-      create_error.GetDescription(), GPUDevice::RenderAPIToString(api));
-    return false;
-  }
-
-  if (!ImGuiManager::Initialize(g_settings.display_osd_scale / 100.0f, g_settings.display_show_osd_messages,
-                                &create_error))
-  {
-    ERROR_LOG("Failed to initialize ImGuiManager: {}", create_error.GetDescription());
-    Error::SetStringFmt(error, "Failed to initialize ImGuiManager: {}", create_error.GetDescription());
-    g_gpu_device->Destroy();
-    g_gpu_device.reset();
-    return false;
-  }
-
-  return true;
-}
-
-void Host::UpdateDisplayWindow()
-{
-  if (!g_gpu_device)
-    return;
-
-  if (!g_gpu_device->UpdateWindow())
-  {
-    Host::ReportErrorAsync("Error", "Failed to change window after update. The log may contain more information.");
-    return;
-  }
-
-  ImGuiManager::WindowResized();
-
-  if (System::IsValid())
-  {
-    // Fix up vsync etc.
-    System::UpdateSpeedLimiterState();
-
-    // If we're paused, re-present the current frame at the new window size.
-    if (System::IsPaused())
-      System::InvalidateDisplay();
-  }
-}
-
-void Host::ResizeDisplayWindow(s32 width, s32 height, float scale)
-{
-  if (!g_gpu_device)
-    return;
-
-  DEV_LOG("Display window resized to {}x{}", width, height);
-
-  g_gpu_device->ResizeWindow(width, height, scale);
-  ImGuiManager::WindowResized();
-
-  // If we're paused, re-present the current frame at the new window size.
-  if (System::IsValid())
-  {
-    if (System::IsPaused())
-    {
-      // Hackity hack, on some systems, presenting a single frame isn't enough to actually get it
-      // displayed. Two seems to be good enough. Maybe something to do with direct scanout.
-      System::InvalidateDisplay();
-      System::InvalidateDisplay();
-    }
-
-    System::HostDisplayResized();
-  }
-}
-
-void Host::ReleaseGPUDevice()
-{
-  if (!g_gpu_device)
-    return;
-
-  ImGuiManager::DestroyOverlayTextures();
-  FullscreenUI::Shutdown();
-  ImGuiManager::Shutdown();
-
-  INFO_LOG("Destroying {} GPU device...", GPUDevice::RenderAPIToString(g_gpu_device->GetRenderAPI()));
-  g_gpu_device->Destroy();
-  g_gpu_device.reset();
-}

src/core/host.h

@@ -87,18 +87,6 @@ void DisplayLoadingScreen(const char* message, int progress_min = -1, int progre
 /// Safely executes a function on the VM thread.
 void RunOnCPUThread(std::function<void()> function, bool block = false);
 
-/// Attempts to create the rendering device backend.
-bool CreateGPUDevice(RenderAPI api, Error* error);
-
-/// Handles fullscreen transitions and such.
-void UpdateDisplayWindow();
-
-/// Called when the window is resized.
-void ResizeDisplayWindow(s32 width, s32 height, float scale);
-
-/// Destroys any active rendering device.
-void ReleaseGPUDevice();
-
 /// Called at the end of the frame, before presentation.
 void FrameDone();
 

src/core/hotkeys.cpp

@@ -7,6 +7,7 @@
 #include "cpu_pgxp.h"
 #include "fullscreen_ui.h"
 #include "gpu.h"
+#include "gpu_thread.h"
 #include "host.h"
 #include "imgui_overlays.h"
 #include "settings.h"
@@ -55,8 +56,7 @@ static void HotkeyModifyResolutionScale(s32 increment)
 
   if (System::IsValid())
   {
-    g_gpu->RestoreDeviceContext();
-    g_gpu->UpdateSettings(old_settings);
+    GPUThread::UpdateSettings();
     System::ClearMemorySaveStates();
   }
 }
@@ -367,11 +367,10 @@ DEFINE_HOTKEY("TogglePGXP", TRANSLATE_NOOP("Hotkeys", "Graphics"), TRANSLATE_NOO
               [](s32 pressed) {
                 if (!pressed && System::IsValid())
                 {
-                  Settings old_settings = g_settings;
                   g_settings.gpu_pgxp_enable = !g_settings.gpu_pgxp_enable;
-                  g_gpu->RestoreDeviceContext();
-                  g_gpu->UpdateSettings(old_settings);
                   System::ClearMemorySaveStates();
+                  GPUThread::UpdateSettings();
+
                   Host::AddKeyedOSDMessage("TogglePGXP",
                                            g_settings.gpu_pgxp_enable ?
                                              TRANSLATE_STR("OSDMessage", "PGXP is now enabled.") :
@@ -444,12 +443,11 @@ DEFINE_HOTKEY("TogglePGXPDepth", TRANSLATE_NOOP("Hotkeys", "Graphics"),
                   if (!g_settings.gpu_pgxp_enable)
                     return;
 
-                  const Settings old_settings = g_settings;
                   g_settings.gpu_pgxp_depth_buffer = !g_settings.gpu_pgxp_depth_buffer;
 
-                  g_gpu->RestoreDeviceContext();
-                  g_gpu->UpdateSettings(old_settings);
                   System::ClearMemorySaveStates();
+                  GPUThread::UpdateSettings();
+
                   Host::AddKeyedOSDMessage("TogglePGXPDepth",
                                            g_settings.gpu_pgxp_depth_buffer ?
                                              TRANSLATE_STR("OSDMessage", "PGXP Depth Buffer is now enabled.") :
@@ -465,12 +463,11 @@ DEFINE_HOTKEY("TogglePGXPCPU", TRANSLATE_NOOP("Hotkeys", "Graphics"), TRANSLATE_
                   if (!g_settings.gpu_pgxp_enable)
                     return;
 
-                  const Settings old_settings = g_settings;
                   g_settings.gpu_pgxp_cpu = !g_settings.gpu_pgxp_cpu;
 
-                  g_gpu->RestoreDeviceContext();
-                  g_gpu->UpdateSettings(old_settings);
+                  // GPU thread is unchanged
                   System::ClearMemorySaveStates();
+
                   Host::AddKeyedOSDMessage("TogglePGXPCPU",
                                            g_settings.gpu_pgxp_cpu ?
                                              TRANSLATE_STR("OSDMessage", "PGXP CPU mode is now enabled.") :
@@ -558,29 +555,29 @@ DEFINE_HOTKEY("AudioVolumeDown", TRANSLATE_NOOP("Hotkeys", "Audio"), TRANSLATE_N
 DEFINE_HOTKEY("LoadSelectedSaveState", TRANSLATE_NOOP("Hotkeys", "Save States"),
               TRANSLATE_NOOP("Hotkeys", "Load From Selected Slot"), [](s32 pressed) {
                 if (!pressed)
-                  Host::RunOnCPUThread(SaveStateSelectorUI::LoadCurrentSlot);
+                  GPUThread::RunOnThread(SaveStateSelectorUI::LoadCurrentSlot);
               })
 DEFINE_HOTKEY("SaveSelectedSaveState", TRANSLATE_NOOP("Hotkeys", "Save States"),
               TRANSLATE_NOOP("Hotkeys", "Save To Selected Slot"), [](s32 pressed) {
                 if (!pressed)
-                  Host::RunOnCPUThread(SaveStateSelectorUI::SaveCurrentSlot);
+                  GPUThread::RunOnThread(SaveStateSelectorUI::SaveCurrentSlot);
               })
 DEFINE_HOTKEY("SelectPreviousSaveStateSlot", TRANSLATE_NOOP("Hotkeys", "Save States"),
               TRANSLATE_NOOP("Hotkeys", "Select Previous Save Slot"), [](s32 pressed) {
                 if (!pressed)
-                  Host::RunOnCPUThread([]() { SaveStateSelectorUI::SelectPreviousSlot(true); });
+                  GPUThread::RunOnThread([]() { SaveStateSelectorUI::SelectPreviousSlot(true); });
               })
 DEFINE_HOTKEY("SelectNextSaveStateSlot", TRANSLATE_NOOP("Hotkeys", "Save States"),
               TRANSLATE_NOOP("Hotkeys", "Select Next Save Slot"), [](s32 pressed) {
                 if (!pressed)
-                  Host::RunOnCPUThread([]() { SaveStateSelectorUI::SelectNextSlot(true); });
+                  GPUThread::RunOnThread([]() { SaveStateSelectorUI::SelectNextSlot(true); });
               })
 DEFINE_HOTKEY("SaveStateAndSelectNextSlot", TRANSLATE_NOOP("Hotkeys", "Save States"),
               TRANSLATE_NOOP("Hotkeys", "Save State and Select Next Slot"), [](s32 pressed) {
                 if (!pressed && System::IsValid())
                 {
                   SaveStateSelectorUI::SaveCurrentSlot();
-                  SaveStateSelectorUI::SelectNextSlot(false);
+                  GPUThread::RunOnThread([]() { SaveStateSelectorUI::SelectNextSlot(false); });
                 }
               })
 

src/core/imgui_overlays.cpp

@@ -9,6 +9,8 @@
 #include "dma.h"
 #include "fullscreen_ui.h"
 #include "gpu.h"
+#include "gpu_backend.h"
+#include "gpu_thread.h"
 #include "host.h"
 #include "mdec.h"
 #include "resources.h"
@@ -176,17 +178,17 @@ void ImGuiManager::RenderDebugWindows()
 {
   if (System::IsValid())
   {
-    if (g_settings.debugging.show_gpu_state)
+    if (g_gpu_settings.debugging.show_gpu_state)
       g_gpu->DrawDebugStateWindow();
-    if (g_settings.debugging.show_cdrom_state)
+    if (g_gpu_settings.debugging.show_cdrom_state)
       CDROM::DrawDebugWindow();
-    if (g_settings.debugging.show_timers_state)
+    if (g_gpu_settings.debugging.show_timers_state)
       Timers::DrawDebugStateWindow();
-    if (g_settings.debugging.show_spu_state)
+    if (g_gpu_settings.debugging.show_spu_state)
       SPU::DrawDebugStateWindow();
-    if (g_settings.debugging.show_mdec_state)
+    if (g_gpu_settings.debugging.show_mdec_state)
       MDEC::DrawDebugStateWindow();
-    if (g_settings.debugging.show_dma_state)
+    if (g_gpu_settings.debugging.show_dma_state)
       DMA::DrawDebugStateWindow();
   }
 }
@@ -194,14 +196,14 @@ void ImGuiManager::RenderDebugWindows()
 void ImGuiManager::RenderTextOverlays()
 {
   const System::State state = System::GetState();
-  if (state != System::State::Shutdown)
+  if (state == System::State::Paused || state == System::State::Running)
   {
     DrawPerformanceOverlay();
 
-    if (g_settings.display_show_enhancements && state != System::State::Paused)
+    if (g_gpu_settings.display_show_enhancements && state != System::State::Paused)
       DrawEnhancementsOverlay();
 
-    if (g_settings.display_show_inputs && state != System::State::Paused)
+    if (g_gpu_settings.display_show_inputs && state != System::State::Paused)
       DrawInputsOverlay();
   }
 }
@@ -219,9 +221,9 @@ void ImGuiManager::FormatProcessorStat(SmallStringBase& text, double usage, doub
 
 void ImGuiManager::DrawPerformanceOverlay()
 {
-  if (!(g_settings.display_show_fps || g_settings.display_show_speed || g_settings.display_show_gpu_stats ||
-        g_settings.display_show_resolution || g_settings.display_show_cpu_usage ||
-        (g_settings.display_show_status_indicators &&
+  if (!(g_gpu_settings.display_show_fps || g_gpu_settings.display_show_speed || g_gpu_settings.display_show_gpu_stats ||
+        g_gpu_settings.display_show_resolution || g_gpu_settings.display_show_cpu_usage ||
+        (g_gpu_settings.display_show_status_indicators &&
          (System::IsPaused() || System::IsFastForwardEnabled() || System::IsTurboEnabled()))))
   {
     return;
@@ -258,12 +260,12 @@ void ImGuiManager::DrawPerformanceOverlay()
   if (state == System::State::Running)
   {
     const float speed = System::GetEmulationSpeed();
-    if (g_settings.display_show_fps)
+    if (g_gpu_settings.display_show_fps)
     {
       text.append_format("G: {:.2f} | V: {:.2f}", System::GetFPS(), System::GetVPS());
       first = false;
     }
-    if (g_settings.display_show_speed)
+    if (g_gpu_settings.display_show_speed)
     {
       text.append_format("{}{}%", first ? "" : " | ", static_cast<u32>(std::round(speed)));
 
@@ -288,33 +290,33 @@ void ImGuiManager::DrawPerformanceOverlay()
       DRAW_LINE(fixed_font, text, color);
     }
 
-    if (g_settings.display_show_gpu_stats)
+    if (g_gpu_settings.display_show_gpu_stats)
     {
-      g_gpu->GetStatsString(text);
+      GPUBackend::GetStatsString(text);
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
 
-      g_gpu->GetMemoryStatsString(text);
+      GPUBackend::GetMemoryStatsString(text);
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
     }
 
-    if (g_settings.display_show_resolution)
+    if (g_gpu_settings.display_show_resolution)
     {
       // TODO: this seems wrong?
-      const auto [effective_width, effective_height] = g_gpu->GetEffectiveDisplayResolution();
+      const auto [src_width, src_height] = GPUBackend::GetLastDisplaySourceSize();
       const bool interlaced = g_gpu->IsInterlacedDisplayEnabled();
       const bool pal = g_gpu->IsInPALMode();
-      text.format("{}x{} {} {}", effective_width, effective_height, pal ? "PAL" : "NTSC",
+      text.format("{}x{} {} {}", src_width, src_height, pal ? "PAL" : "NTSC",
                   interlaced ? "Interlaced" : "Progressive");
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
     }
 
-    if (g_settings.display_show_latency_stats)
+    if (g_gpu_settings.display_show_latency_stats)
     {
       System::FormatLatencyStats(text);
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
     }
 
-    if (g_settings.display_show_cpu_usage)
+    if (g_gpu_settings.display_show_cpu_usage)
     {
       text.format("{:.2f}ms | {:.2f}ms | {:.2f}ms", System::GetMinimumFrameTime(), System::GetAverageFrameTime(),
         System::GetMaximumFrameTime());
@@ -369,22 +371,19 @@ void ImGuiManager::DrawPerformanceOverlay()
       FormatProcessorStat(text, System::GetCPUThreadUsage(), System::GetCPUThreadAverageTime());
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
 
-      if (g_gpu->GetSWThread())
-      {
-        text.assign("SW: ");
-        FormatProcessorStat(text, System::GetSWThreadUsage(), System::GetSWThreadAverageTime());
+      text.assign("RNDR: ");
+      FormatProcessorStat(text, System::GetGPUThreadUsage(), System::GetGPUThreadAverageTime());
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
     }
-    }
 
-    if (g_settings.display_show_gpu_usage && g_gpu_device->IsGPUTimingEnabled())
+    if (g_gpu_settings.display_show_gpu_usage)
     {
       text.assign("GPU: ");
-      FormatProcessorStat(text, System::GetGPUUsage(), System::GetGPUAverageTime());
+      FormatProcessorStat(text, GPUThread::GetGPUUsage(), GPUThread::GetGPUAverageTime());
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
     }
 
-    if (g_settings.display_show_status_indicators)
+    if (g_gpu_settings.display_show_status_indicators)
     {
       const bool rewinding = System::IsRewinding();
       if (rewinding || System::IsFastForwardEnabled() || System::IsTurboEnabled())
@@ -394,7 +393,7 @@ void ImGuiManager::DrawPerformanceOverlay()
       }
     }
 
-    if (g_settings.display_show_frame_times)
+    if (g_gpu_settings.display_show_frame_times)
     {
       const ImVec2 history_size(200.0f * scale, 50.0f * scale);
       ImGui::SetNextWindowSize(ImVec2(history_size.x, history_size.y));
@@ -455,7 +454,7 @@ void ImGuiManager::DrawPerformanceOverlay()
       ImGui::PopStyleColor(3);
     }
   }
-  else if (g_settings.display_show_status_indicators && state == System::State::Paused &&
+  else if (g_gpu_settings.display_show_status_indicators && state == System::State::Paused &&
            !FullscreenUI::HasActiveWindow())
   {
     text.assign(ICON_FA_PAUSE);
@@ -470,7 +469,7 @@ void ImGuiManager::DrawEnhancementsOverlay()
   LargeString text;
   text.append_format("{} {}-{}", Settings::GetConsoleRegionName(System::GetRegion()),
                      GPUDevice::RenderAPIToString(g_gpu_device->GetRenderAPI()),
-                     g_gpu->IsHardwareRenderer() ? "HW" : "SW");
+                     GPUBackend::IsUsingHardwareBackend() ? "HW" : "SW");
 
   if (g_settings.rewind_enable)
     text.append_format(" RW={}/{}", g_settings.rewind_save_frequency, g_settings.rewind_save_slots);
@@ -784,7 +783,11 @@ void SaveStateSelectorUI::ClearList()
   for (ListEntry& li : s_slots)
   {
     if (li.preview_texture)
-      g_gpu_device->RecycleTexture(std::move(li.preview_texture));
+    {
+      GPUThread::RunOnThread([tex = li.preview_texture.release()]() {
+        g_gpu_device->RecycleTexture(std::unique_ptr<GPUTexture>(tex));
+      });
+    }
   }
   s_slots.clear();
 }
@@ -1102,7 +1105,7 @@ void SaveStateSelectorUI::LoadCurrentSlot()
     }
   }
 
-  Close();
+  GPUThread::RunOnThread(&Close);
 }
 
 void SaveStateSelectorUI::SaveCurrentSlot()
@@ -1119,7 +1122,7 @@ void SaveStateSelectorUI::SaveCurrentSlot()
     }
   }
 
-  Close();
+  GPUThread::RunOnThread(&Close);
 }
 
 void SaveStateSelectorUI::ShowSlotOSDMessage()
@@ -1143,7 +1146,7 @@ void SaveStateSelectorUI::ShowSlotOSDMessage()
 void ImGuiManager::RenderOverlayWindows()
 {
   const System::State state = System::GetState();
-  if (state != System::State::Shutdown)
+  if (state == System::State::Paused || state == System::State::Running)
   {
     if (SaveStateSelectorUI::s_open)
       SaveStateSelectorUI::Draw();

src/core/settings.cpp

@@ -25,6 +25,7 @@
 Log_SetChannel(Settings);
 
 Settings g_settings;
+Settings g_gpu_settings;
 
 const char* SettingInfo::StringDefaultValue() const
 {
@@ -191,7 +192,7 @@ void Settings::Load(SettingsInterface& si)
   gpu_disable_texture_buffers = si.GetBoolValue("GPU", "DisableTextureBuffers", false);
   gpu_disable_texture_copy_to_self = si.GetBoolValue("GPU", "DisableTextureCopyToSelf", false);
   gpu_per_sample_shading = si.GetBoolValue("GPU", "PerSampleShading", false);
-  gpu_use_thread = si.GetBoolValue("GPU", "UseThread", true);
+  gpu_max_queued_frames = static_cast<u8>(si.GetUIntValue("GPU", "MaxQueuedFrames", DEFAULT_GPU_MAX_QUEUED_FRAMES));
   gpu_use_software_renderer_for_readbacks = si.GetBoolValue("GPU", "UseSoftwareRendererForReadbacks", false);
   gpu_threaded_presentation = si.GetBoolValue("GPU", "ThreadedPresentation", DEFAULT_THREADED_PRESENTATION);
   gpu_true_color = si.GetBoolValue("GPU", "TrueColor", true);
@@ -495,7 +496,7 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
   }
 
   si.SetBoolValue("GPU", "PerSampleShading", gpu_per_sample_shading);
-  si.SetBoolValue("GPU", "UseThread", gpu_use_thread);
+  si.SetUIntValue("GPU", "MaxQueuedFrames", gpu_max_queued_frames);
   si.SetBoolValue("GPU", "ThreadedPresentation", gpu_threaded_presentation);
   si.SetBoolValue("GPU", "UseSoftwareRendererForReadbacks", gpu_use_software_renderer_for_readbacks);
   si.SetBoolValue("GPU", "TrueColor", gpu_true_color);

src/core/settings.h

@@ -104,7 +104,7 @@ struct Settings
   std::string gpu_adapter;
   u8 gpu_resolution_scale = 1;
   u8 gpu_multisamples = 1;
-  bool gpu_use_thread : 1 = true;
+  u8 gpu_max_queued_frames = 2;
   bool gpu_use_software_renderer_for_readbacks : 1 = false;
   bool gpu_threaded_presentation : 1 = DEFAULT_THREADED_PRESENTATION;
   bool gpu_use_debug_device : 1 = false;
@@ -457,6 +457,8 @@ struct Settings
   static constexpr ConsoleRegion DEFAULT_CONSOLE_REGION = ConsoleRegion::Auto;
   static constexpr float DEFAULT_GPU_PGXP_DEPTH_THRESHOLD = 300.0f;
   static constexpr float GPU_PGXP_DEPTH_THRESHOLD_SCALE = 4096.0f;
+  static constexpr u8 DEFAULT_GPU_MAX_QUEUED_FRAMES = 2; // TODO: Maybe lower? But that means fast CPU threads would
+                                                         // always stall, could be a problem for power management.
 
   // Prefer oldrec over newrec for now. Except on RISC-V, where there is no oldrec.
 #if defined(CPU_ARCH_RISCV64)
@@ -526,7 +528,9 @@ struct Settings
   static constexpr u16 DEFAULT_PINE_SLOT = 28011;
 };
 
-extern Settings g_settings;
+// TODO: Use smaller copy for GPU thread copy.
+extern Settings g_settings;     // CPU thread copy.
+extern Settings g_gpu_settings; // GPU thread copy.
 
 namespace EmuFolders {
 extern std::string AppRoot;

src/core/system.cpp

@@ -17,6 +17,8 @@
 #include "game_database.h"
 #include "game_list.h"
 #include "gpu.h"
+#include "gpu_backend.h"
+#include "gpu_thread.h"
 #include "gte.h"
 #include "host.h"
 #include "host_interface_progress_callback.h"
@@ -122,13 +124,12 @@ static void ClearRunningGame();
 static void DestroySystem();
 static std::string GetMediaPathFromSaveState(const char* path);
 static bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display, bool is_memory_state);
-static bool CreateGPU(GPURenderer renderer, bool is_switching, Error* error);
 static bool SaveUndoLoadState();
 static void WarnAboutUnsafeSettings();
 static void LogUnsafeSettingsToConsole(const SmallStringBase& messages);
 
 /// Throttles the system, i.e. sleeps until it's time to execute the next frame.
-static void Throttle(Common::Timer::Value current_time);
+static void Throttle(Common::Timer::Value current_time, Common::Timer::Value sleep_until);
 static void UpdatePerformanceCounters();
 static void AccumulatePreFrameSleepTime();
 static void UpdatePreFrameSleepTime();
@@ -170,7 +171,6 @@ static std::string s_input_profile_name;
 
 static System::State s_state = System::State::Shutdown;
 static std::atomic_bool s_startup_cancelled{false};
-static bool s_keep_gpu_device_on_shutdown = false;
 
 static ConsoleRegion s_region = ConsoleRegion::NTSC_U;
 TickCount System::g_ticks_per_second = System::MASTER_CLOCK;
@@ -226,11 +226,8 @@ static float s_maximum_frame_time = 0.0f;
 static float s_average_frame_time = 0.0f;
 static float s_cpu_thread_usage = 0.0f;
 static float s_cpu_thread_time = 0.0f;
-static float s_sw_thread_usage = 0.0f;
-static float s_sw_thread_time = 0.0f;
-static float s_average_gpu_time = 0.0f;
-static float s_accumulated_gpu_time = 0.0f;
-static float s_gpu_usage = 0.0f;
+static float s_gpu_thread_usage = 0.0f;
+static float s_gpu_thread_time = 0.0f;
 static System::FrameTimeHistory s_frame_time_history;
 static u32 s_frame_time_history_pos = 0;
 static u32 s_last_frame_number = 0;
@@ -238,7 +235,6 @@ static u32 s_last_internal_frame_number = 0;
 static u32 s_last_global_tick_counter = 0;
 static u64 s_last_cpu_time = 0;
 static u64 s_last_sw_time = 0;
-static u32 s_presents_since_last_update = 0;
 static Common::Timer s_fps_timer;
 static Common::Timer s_frame_timer;
 static Threading::ThreadHandle s_cpu_thread_handle;
@@ -648,21 +644,13 @@ float System::GetCPUThreadAverageTime()
 {
   return s_cpu_thread_time;
 }
-float System::GetSWThreadUsage()
+float System::GetGPUThreadUsage()
 {
-  return s_sw_thread_usage;
+  return s_gpu_thread_usage;
 }
-float System::GetSWThreadAverageTime()
+float System::GetGPUThreadAverageTime()
 {
-  return s_sw_thread_time;
-}
-float System::GetGPUUsage()
-{
-  return s_gpu_usage;
-}
-float System::GetGPUAverageTime()
-{
-  return s_average_gpu_time;
+  return s_gpu_thread_time;
 }
 const System::FrameTimeHistory& System::GetFrameTimeHistory()
 {
@@ -1059,7 +1047,6 @@ std::string System::GetInputProfilePath(std::string_view name)
 bool System::RecreateGPU(GPURenderer renderer, bool force_recreate_device, bool update_display /* = true*/)
 {
   ClearMemorySaveStates();
-  g_gpu->RestoreDeviceContext();
 
   // save current state
   std::unique_ptr<ByteStream> state_stream = ByteStream::CreateGrowableMemoryStream();
@@ -1068,16 +1055,8 @@ bool System::RecreateGPU(GPURenderer renderer, bool force_recreate_device, bool
   if (!state_valid)
     ERROR_LOG("Failed to save old GPU state when switching renderers");
 
-  // create new renderer
-  g_gpu.reset();
-  if (force_recreate_device)
-  {
-    PostProcessing::Shutdown();
-    Host::ReleaseGPUDevice();
-  }
-
   Error error;
-  if (!CreateGPU(renderer, true, &error))
+  if (!GPUThread::SwitchGPUBackend(renderer, force_recreate_device, &error))
   {
     if (!IsStartupCancelled())
       Host::ReportErrorAsync("Error", error.GetDescription());
@@ -1090,7 +1069,6 @@ bool System::RecreateGPU(GPURenderer renderer, bool force_recreate_device, bool
   {
     state_stream->SeekAbsolute(0);
     sw.SetMode(StateWrapper::Mode::Read);
-    g_gpu->RestoreDeviceContext();
     g_gpu->DoState(sw, nullptr, update_display);
     TimingEvents::DoState(sw);
   }
@@ -1326,9 +1304,6 @@ void System::PauseSystem(bool paused)
 
   if (paused)
   {
-    // Make sure the GPU is flushed, otherwise the VB might still be mapped.
-    g_gpu->FlushRender();
-
     FullscreenUI::OnSystemPaused();
 
     InputManager::PauseVibration();
@@ -1345,7 +1320,7 @@ void System::PauseSystem(bool paused)
     Host::OnSystemPaused();
     Host::OnIdleStateChanged();
     UpdateDisplayVSync();
-    InvalidateDisplay();
+    GPUThread::PresentCurrentFrame();
   }
   else
   {
@@ -1420,7 +1395,7 @@ bool System::LoadState(const char* filename, Error* error)
   ResetThrottler();
 
   if (IsPaused())
-    InvalidateDisplay();
+    GPUThread::PresentCurrentFrame();
 
   VERBOSE_LOG("Loading state took {:.2f} msec", load_timer.GetTimeMilliseconds());
   return true;
@@ -1511,7 +1486,6 @@ bool System::BootSystem(SystemBootParameters parameters, Error* error)
   Assert(s_state == State::Shutdown);
   s_state = State::Starting;
   s_startup_cancelled.store(false);
-  s_keep_gpu_device_on_shutdown = static_cast<bool>(g_gpu_device);
   s_region = g_settings.region;
   Host::OnSystemStarting();
 
@@ -1801,15 +1775,11 @@ bool System::Initialize(bool force_software_renderer, Error* error)
   s_average_frame_time = 0.0f;
   s_cpu_thread_usage = 0.0f;
   s_cpu_thread_time = 0.0f;
-  s_sw_thread_usage = 0.0f;
-  s_sw_thread_time = 0.0f;
-  s_average_gpu_time = 0.0f;
-  s_accumulated_gpu_time = 0.0f;
-  s_gpu_usage = 0.0f;
+  s_gpu_thread_usage = 0.0f;
+  s_gpu_thread_time = 0.0f;
   s_last_frame_number = 0;
   s_last_internal_frame_number = 0;
   s_last_global_tick_counter = 0;
-  s_presents_since_last_update = 0;
   s_last_cpu_time = 0;
   s_fps_timer.Reset();
   s_frame_timer.Reset();
@@ -1828,34 +1798,33 @@ bool System::Initialize(bool force_software_renderer, Error* error)
 
   CPU::CodeCache::Initialize();
 
-  if (!CreateGPU(force_software_renderer ? GPURenderer::Software : g_settings.gpu_renderer, false, error))
+  // TODO: Drop pointer
+  g_gpu = std::make_unique<GPU>();
+  g_gpu->Initialize();
+
+  if (!GPUThread::CreateGPUBackend(force_software_renderer ? GPURenderer::Software : g_settings.gpu_renderer, error))
   {
+    g_gpu.reset();
     Bus::Shutdown();
     CPU::Shutdown();
     return false;
   }
 
+  // Was startup cancelled? (e.g. shading compilers took too long and the user closed the application)
+  if (IsStartupCancelled())
+  {
+    GPUThread::DestroyGPUBackend();
+    g_gpu.reset();
+    CPU::Shutdown();
+    Bus::Shutdown();
+    return false;
+  }
+
   GTE::UpdateAspectRatio();
 
   if (g_settings.gpu_pgxp_enable)
     CPU::PGXP::Initialize();
 
-  // Was startup cancelled? (e.g. shading compilers took too long and the user closed the application)
-  if (IsStartupCancelled())
-  {
-    g_gpu.reset();
-    if (!s_keep_gpu_device_on_shutdown)
-    {
-      Host::ReleaseGPUDevice();
-      Host::ReleaseRenderWindow();
-    }
-    if (g_settings.gpu_pgxp_enable)
-      CPU::PGXP::Shutdown();
-    CPU::Shutdown();
-    Bus::Shutdown();
-    return false;
-  }
-
   DMA::Initialize();
   CDROM::Initialize();
   Pad::Initialize();
@@ -1919,16 +1888,9 @@ void System::DestroySystem()
   TimingEvents::Shutdown();
   ClearRunningGame();
 
-  // Restore present-all-frames behavior.
-  if (s_keep_gpu_device_on_shutdown && g_gpu_device)
-  {
+  if (GPUThread::WasFullscreenUIRequested())
     UpdateDisplayVSync();
-  }
-  else
-  {
-    Host::ReleaseGPUDevice();
-    Host::ReleaseRenderWindow();
-  }
+  GPUThread::DestroyGPUBackend();
 
   s_bios_hash = {};
   s_bios_image_info = nullptr;
@@ -1967,9 +1929,6 @@ void System::Execute()
       case State::Running:
       {
         s_system_executing = true;
-
-        // TODO: Purge reset/restore
-        g_gpu->RestoreDeviceContext();
         TimingEvents::UpdateCPUDowncount();
 
         if (s_rewind_load_counter >= 0)
@@ -1998,9 +1957,6 @@ void System::FrameDone()
 {
   s_frame_number++;
 
-  // Vertex buffer is shared, need to flush what we have.
-  g_gpu->FlushRender();
-
   // Generate any pending samples from the SPU before sleeping, this way we reduce the chances of underruns.
   // TODO: when running ahead, we can skip this (and the flush above)
   SPU::GeneratePendingSamples();
@@ -2051,7 +2007,6 @@ void System::FrameDone()
       // counter-acts that.
       Host::PumpMessagesOnCPUThread();
       InputManager::PollSources();
-      g_gpu->RestoreDeviceContext();
 
       if (IsExecutionInterrupted())
       {
@@ -2078,48 +2033,6 @@ void System::FrameDone()
   if (s_pre_frame_sleep)
     AccumulatePreFrameSleepTime();
 
-  // explicit present (frame pacing)
-  const bool is_unique_frame = (s_last_presented_internal_frame_number != s_internal_frame_number);
-  s_last_presented_internal_frame_number = s_internal_frame_number;
-
-  const bool skip_this_frame = (((s_skip_presenting_duplicate_frames && !is_unique_frame &&
-                                  s_skipped_frame_count < MAX_SKIPPED_DUPLICATE_FRAME_COUNT) ||
-                                 (!s_optimal_frame_pacing && current_time > s_next_frame_time &&
-                                  s_skipped_frame_count < MAX_SKIPPED_TIMEOUT_FRAME_COUNT) ||
-                                 g_gpu_device->ShouldSkipPresentingFrame()) &&
-                                !s_syncing_to_host_with_vsync && !IsExecutionInterrupted());
-  if (!skip_this_frame)
-  {
-    s_skipped_frame_count = 0;
-
-    const bool throttle_before_present = (s_optimal_frame_pacing && s_throttler_enabled && !IsExecutionInterrupted());
-    const bool explicit_present = (throttle_before_present && g_gpu_device->GetFeatures().explicit_present);
-    if (explicit_present)
-    {
-      const bool do_present = PresentDisplay(false, true);
-      Throttle(current_time);
-      if (do_present)
-        g_gpu_device->SubmitPresent();
-    }
-    else
-    {
-      if (throttle_before_present)
-        Throttle(current_time);
-
-      PresentDisplay(false, false);
-
-      if (!throttle_before_present && s_throttler_enabled && !IsExecutionInterrupted())
-        Throttle(current_time);
-    }
-  }
-  else
-  {
-    DEBUG_LOG("Skipping displaying frame");
-    s_skipped_frame_count++;
-    if (s_throttler_enabled)
-      Throttle(current_time);
-  }
-
   // pre-frame sleep (input lag reduction)
   current_time = Common::Timer::GetCurrentValue();
   if (s_pre_frame_sleep)
@@ -2128,10 +2041,15 @@ void System::FrameDone()
     if (pre_frame_sleep_until > current_time &&
         Common::Timer::ConvertValueToMilliseconds(pre_frame_sleep_until - current_time) >= 1)
     {
-      Common::Timer::SleepUntil(pre_frame_sleep_until, true);
+      Throttle(current_time, pre_frame_sleep_until);
       current_time = Common::Timer::GetCurrentValue();
     }
   }
+  else
+  {
+    if (s_throttler_enabled)
+      Throttle(current_time, s_next_frame_time);
+  }
 
   s_frame_start_time = current_time;
 
@@ -2149,14 +2067,42 @@ void System::FrameDone()
     }
   }
 
-  g_gpu->RestoreDeviceContext();
-
   // Update perf counters *after* throttling, we want to measure from start-of-frame
   // to start-of-frame, not end-of-frame to end-of-frame (will be noisy due to different
   // amounts of computation happening in each frame).
   System::UpdatePerformanceCounters();
 }
 
+void System::GetFramePresentationDetails(bool* present_frame, bool* allow_present_skip,
+                                         Common::Timer::Value* present_time)
+{
+  const Common::Timer::Value current_time = Common::Timer::GetCurrentValue();
+
+  // explicit present (frame pacing)
+  const bool is_unique_frame = (s_last_presented_internal_frame_number != s_internal_frame_number);
+  s_last_presented_internal_frame_number = s_internal_frame_number;
+
+  const bool skip_this_frame = (((s_skip_presenting_duplicate_frames && !is_unique_frame &&
+                                  s_skipped_frame_count < MAX_SKIPPED_DUPLICATE_FRAME_COUNT) ||
+                                 (!s_optimal_frame_pacing && current_time > s_next_frame_time &&
+                                  s_skipped_frame_count < MAX_SKIPPED_TIMEOUT_FRAME_COUNT)) &&
+                                !s_syncing_to_host_with_vsync && !IsExecutionInterrupted());
+  const bool should_allow_present_skip = !s_syncing_to_host_with_vsync && !s_optimal_frame_pacing;
+  *present_frame = !skip_this_frame;
+  *allow_present_skip = should_allow_present_skip;
+  *present_time = (should_allow_present_skip && !IsExecutionInterrupted()) ? s_next_frame_time : 0;
+
+  if (!skip_this_frame)
+  {
+    s_skipped_frame_count = 0;
+  }
+  else
+  {
+    DEBUG_LOG("Skipping displaying frame");
+    s_skipped_frame_count++;
+  }
+}
+
 void System::SetThrottleFrequency(float frequency)
 {
   if (s_throttle_frequency == frequency)
@@ -2188,12 +2134,12 @@ void System::ResetThrottler()
   s_pre_frame_sleep_time = 0;
 }
 
-void System::Throttle(Common::Timer::Value current_time)
+void System::Throttle(Common::Timer::Value current_time, Common::Timer::Value sleep_until)
 {
   // If we're running too slow, advance the next frame time based on the time we lost. Effectively skips
   // running those frames at the intended time, because otherwise if we pause in the debugger, we'll run
   // hundreds of frames when we resume.
-  if (current_time > s_next_frame_time)
+  if (current_time > sleep_until)
   {
     const Common::Timer::Value diff = static_cast<s64>(current_time) - static_cast<s64>(s_next_frame_time);
     s_next_frame_time += (diff / s_frame_period) * s_frame_period + s_frame_period;
@@ -2208,11 +2154,10 @@ void System::Throttle(Common::Timer::Value current_time)
     Common::Timer::Value poll_start_time = current_time;
     for (;;)
     {
-      const u32 sleep_ms =
-        static_cast<u32>(Common::Timer::ConvertValueToMilliseconds(s_next_frame_time - poll_start_time));
+      const u32 sleep_ms = static_cast<u32>(Common::Timer::ConvertValueToMilliseconds(sleep_until - poll_start_time));
       s_socket_multiplexer->PollEventsWithTimeout(sleep_ms);
       poll_start_time = Common::Timer::GetCurrentValue();
-      if (poll_start_time >= s_next_frame_time || (!g_settings.display_optimal_frame_pacing && sleep_ms == 0))
+      if (poll_start_time >= sleep_until || (!g_settings.display_optimal_frame_pacing && sleep_ms == 0))
         break;
     }
   }
@@ -2221,21 +2166,21 @@ void System::Throttle(Common::Timer::Value current_time)
     // Use a spinwait if we undersleep for all platforms except android.. don't want to burn battery.
     // Linux also seems to do a much better job of waking up at the requested time.
 #if !defined(__linux__)
-    Common::Timer::SleepUntil(s_next_frame_time, g_settings.display_optimal_frame_pacing);
+    Common::Timer::SleepUntil(sleep_until, g_settings.display_optimal_frame_pacing);
 #else
-    Common::Timer::SleepUntil(s_next_frame_time, false);
+    Common::Timer::SleepUntil(sleep_until, false);
 #endif
   }
 #else
   // No spinwait on Android, see above.
-  Common::Timer::SleepUntil(s_next_frame_time, false);
+  Common::Timer::SleepUntil(sleep_until, false);
 #endif
 
 #if 0
   DEV_LOG("Asked for {:.2f} ms, slept for {:.2f} ms, {:.2f} ms late",
-          Common::Timer::ConvertValueToMilliseconds(s_next_frame_time - current_time),
+          Common::Timer::ConvertValueToMilliseconds(sleep_until - current_time),
           Common::Timer::ConvertValueToMilliseconds(Common::Timer::GetCurrentValue() - current_time),
-          Common::Timer::ConvertValueToMilliseconds(Common::Timer::GetCurrentValue() - s_next_frame_time));
+          Common::Timer::ConvertValueToMilliseconds(Common::Timer::GetCurrentValue() - sleep_until));
 #endif
 
   s_next_frame_time += s_frame_period;
@@ -2293,62 +2238,6 @@ void System::RecreateSystem()
     PauseSystem(true);
 }
 
-bool System::CreateGPU(GPURenderer renderer, bool is_switching, Error* error)
-{
-  const RenderAPI api = Settings::GetRenderAPIForRenderer(renderer);
-
-  if (!g_gpu_device ||
-      (renderer != GPURenderer::Software && !GPUDevice::IsSameRenderAPI(g_gpu_device->GetRenderAPI(), api)))
-  {
-    if (g_gpu_device)
-    {
-      WARNING_LOG("Recreating GPU device, expecting {} got {}", GPUDevice::RenderAPIToString(api),
-                  GPUDevice::RenderAPIToString(g_gpu_device->GetRenderAPI()));
-      PostProcessing::Shutdown();
-    }
-
-    Host::ReleaseGPUDevice();
-    if (!Host::CreateGPUDevice(api, error))
-    {
-      Host::ReleaseRenderWindow();
-      return false;
-    }
-
-    if (is_switching)
-      PostProcessing::Initialize();
-  }
-
-  if (renderer == GPURenderer::Software)
-    g_gpu = GPU::CreateSoftwareRenderer();
-  else
-    g_gpu = GPU::CreateHardwareRenderer();
-
-  if (!g_gpu)
-  {
-    ERROR_LOG("Failed to initialize {} renderer, falling back to software renderer",
-              Settings::GetRendererName(renderer));
-    Host::AddOSDMessage(
-      fmt::format(TRANSLATE_FS("System", "Failed to initialize {} renderer, falling back to software renderer."),
-                  Settings::GetRendererName(renderer)),
-      Host::OSD_CRITICAL_ERROR_DURATION);
-    g_gpu.reset();
-    g_gpu = GPU::CreateSoftwareRenderer();
-    if (!g_gpu)
-    {
-      ERROR_LOG("Failed to create fallback software renderer.");
-      if (!s_keep_gpu_device_on_shutdown)
-      {
-        PostProcessing::Shutdown();
-        Host::ReleaseGPUDevice();
-        Host::ReleaseRenderWindow();
-      }
-      return false;
-    }
-  }
-
-  return true;
-}
-
 bool System::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display, bool is_memory_state)
 {
   if (!sw.DoMarker("System"))
@@ -2398,7 +2287,6 @@ bool System::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_di
   if (!sw.DoMarker("InterruptController") || !InterruptController::DoState(sw))
     return false;
 
-  g_gpu->RestoreDeviceContext();
   if (!sw.DoMarker("GPU") || !g_gpu->DoState(sw, host_texture, update_display))
     return false;
 
@@ -2745,7 +2633,7 @@ bool System::SaveStateToStream(ByteStream* state, Error* error, u32 screenshot_s
     std::vector<u32> screenshot_buffer;
     u32 screenshot_stride;
     GPUTexture::Format screenshot_format;
-    if (g_gpu->RenderScreenshotToBuffer(screenshot_width, screenshot_height,
+    if (GPUBackend::RenderScreenshotToBuffer(screenshot_width, screenshot_height,
                                              GSVector4i(0, 0, screenshot_width, screenshot_height), false,
                                              &screenshot_buffer, &screenshot_stride, &screenshot_format) &&
         GPUTexture::ConvertTextureDataToRGBA8(screenshot_width, screenshot_height, screenshot_buffer, screenshot_stride,
@@ -2782,9 +2670,6 @@ bool System::SaveStateToStream(ByteStream* state, Error* error, u32 screenshot_s
   // write data
   {
     header.offset_to_data = static_cast<u32>(state->GetPosition());
-
-    g_gpu->RestoreDeviceContext();
-
     header.data_compression_type = compression_method;
 
     bool result = false;
@@ -2869,9 +2754,9 @@ void System::UpdatePerformanceCounters()
             100.0f;
   s_last_global_tick_counter = global_tick_counter;
 
-  const Threading::Thread* sw_thread = g_gpu->GetSWThread();
+  const Threading::ThreadHandle& gpu_thread = GPUThread::GetThreadHandle();
   const u64 cpu_time = s_cpu_thread_handle ? s_cpu_thread_handle.GetCPUTime() : 0;
-  const u64 sw_time = sw_thread ? sw_thread->GetCPUTime() : 0;
+  const u64 sw_time = gpu_thread ? gpu_thread.GetCPUTime() : 0;
   const u64 cpu_delta = cpu_time - s_last_cpu_time;
   const u64 sw_delta = sw_time - s_last_sw_time;
   s_last_cpu_time = cpu_time;
@@ -2879,27 +2764,19 @@ void System::UpdatePerformanceCounters()
 
   s_cpu_thread_usage = static_cast<float>(static_cast<double>(cpu_delta) * pct_divider);
   s_cpu_thread_time = static_cast<float>(static_cast<double>(cpu_delta) * time_divider);
-  s_sw_thread_usage = static_cast<float>(static_cast<double>(sw_delta) * pct_divider);
-  s_sw_thread_time = static_cast<float>(static_cast<double>(sw_delta) * time_divider);
+  s_gpu_thread_usage = static_cast<float>(static_cast<double>(sw_delta) * pct_divider);
+  s_gpu_thread_time = static_cast<float>(static_cast<double>(sw_delta) * time_divider);
 
   s_fps_timer.ResetTo(now_ticks);
 
-  if (g_gpu_device->IsGPUTimingEnabled())
-  {
-    s_average_gpu_time = s_accumulated_gpu_time / static_cast<float>(std::max(s_presents_since_last_update, 1u));
-    s_gpu_usage = s_accumulated_gpu_time / (time * 10.0f);
-  }
-  s_accumulated_gpu_time = 0.0f;
-  s_presents_since_last_update = 0;
-
-  if (g_settings.display_show_gpu_stats)
-    g_gpu->UpdateStatistics(frames_run);
+  if (g_settings.display_show_gpu_stats || g_settings.display_show_gpu_stats)
+    GPUThread::SetPerformanceCounterUpdatePending();
 
   if (s_pre_frame_sleep)
     UpdatePreFrameSleepTime();
 
-  VERBOSE_LOG("FPS: {:.2f} VPS: {:.2f} CPU: {:.2f} GPU: {:.2f} Average: {:.2f}ms Min: {:.2f}ms Max: {:.2f}ms", s_fps,
-              s_vps, s_cpu_thread_usage, s_gpu_usage, s_average_frame_time, s_minimum_frame_time, s_maximum_frame_time);
+  VERBOSE_LOG("FPS: {:.2f} VPS: {:.2f} CPU: {:.2f} Average: {:.2f}ms Min: {:.2f}ms Max: {:.2f}ms", s_fps, s_vps,
+              s_cpu_thread_usage, s_average_frame_time, s_minimum_frame_time, s_maximum_frame_time);
 
   Host::OnPerformanceCountersUpdated();
 }
@@ -2910,8 +2787,8 @@ void System::ResetPerformanceCounters()
   s_last_internal_frame_number = s_internal_frame_number;
   s_last_global_tick_counter = GetGlobalTickCounter();
   s_last_cpu_time = s_cpu_thread_handle ? s_cpu_thread_handle.GetCPUTime() : 0;
-  if (const Threading::Thread* sw_thread = g_gpu->GetSWThread(); sw_thread)
-    s_last_sw_time = sw_thread->GetCPUTime();
+  if (const Threading::ThreadHandle& sw_thread = GPUThread::GetThreadHandle(); sw_thread)
+    s_last_sw_time = sw_thread.GetCPUTime();
   else
     s_last_sw_time = 0;
 
@@ -3045,7 +2922,7 @@ void System::UpdateDisplayVSync()
               s_syncing_to_host_with_vsync ? " (for throttling)" : "",
               allow_present_throttle ? " (present throttle allowed)" : "");
 
-  g_gpu_device->SetVSyncMode(vsync_mode, allow_present_throttle);
+  GPUThread::SetVSync(vsync_mode, allow_present_throttle);
 }
 
 GPUVSyncMode System::GetEffectiveVSyncMode()
@@ -3653,7 +3530,6 @@ bool System::DumpVRAM(const char* filename)
   if (!IsValid())
     return false;
 
-  g_gpu->RestoreDeviceContext();
   return g_gpu->DumpVRAMToFile(filename);
 }
 
@@ -3805,10 +3681,11 @@ void System::UpdateRunningGame(const char* path, CDImage* image, bool booting)
   if (s_running_game_serial != prev_serial)
     UpdateSessionTime(prev_serial);
 
+  // TODO GPU-THREAD: Racey...
   if (SaveStateSelectorUI::IsOpen())
-    SaveStateSelectorUI::RefreshList(s_running_game_serial);
-  else
-    SaveStateSelectorUI::ClearList();
+  {
+    GPUThread::RunOnThread([serial = s_running_game_serial]() { SaveStateSelectorUI::RefreshList(serial); });
+  }
 
   UpdateRichPresence(booting);
 
@@ -4082,7 +3959,7 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
     if (g_settings.gpu_resolution_scale != old_settings.gpu_resolution_scale ||
         g_settings.gpu_multisamples != old_settings.gpu_multisamples ||
         g_settings.gpu_per_sample_shading != old_settings.gpu_per_sample_shading ||
-        g_settings.gpu_use_thread != old_settings.gpu_use_thread ||
+        g_settings.gpu_max_queued_frames != old_settings.gpu_max_queued_frames ||
         g_settings.gpu_use_software_renderer_for_readbacks != old_settings.gpu_use_software_renderer_for_readbacks ||
         g_settings.gpu_fifo_size != old_settings.gpu_fifo_size ||
         g_settings.gpu_max_run_ahead != old_settings.gpu_max_run_ahead ||
@@ -4116,9 +3993,9 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.rewind_enable != old_settings.rewind_enable ||
         g_settings.runahead_frames != old_settings.runahead_frames)
     {
-      g_gpu->UpdateSettings(old_settings);
+      GPUThread::UpdateSettings();
       if (IsPaused())
-        InvalidateDisplay();
+        GPUThread::PresentCurrentFrame();
     }
 
     if (g_settings.gpu_widescreen_hack != old_settings.gpu_widescreen_hack ||
@@ -4144,9 +4021,6 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
       CPU::CodeCache::Reset();
     }
 
-    if (g_settings.display_show_gpu_stats != old_settings.display_show_gpu_stats)
-      g_gpu->ResetStatistics();
-
     if (g_settings.cdrom_readahead_sectors != old_settings.cdrom_readahead_sectors)
       CDROM::SetReadaheadSectors(g_settings.cdrom_readahead_sectors);
 
@@ -4392,6 +4266,9 @@ void System::CalculateRewindMemoryUsage(u32 num_saves, u32 resolution_scale, u64
 
 void System::ClearMemorySaveStates()
 {
+  if (!s_rewind_states.empty() || !s_runahead_states.empty())
+    Panic("FIXME TEXTURE CLEAR");
+
   s_rewind_states.clear();
   s_runahead_states.clear();
 }
@@ -4563,11 +4440,12 @@ void System::DoRewind()
     s_rewind_load_counter--;
   }
 
-  InvalidateDisplay();
+  // TODO FIXME InvalidateDisplay();
+
   Host::PumpMessagesOnCPUThread();
   Internal::IdlePollUpdate();
 
-  Throttle(Common::Timer::GetCurrentValue());
+  Throttle(Common::Timer::GetCurrentValue(), s_next_frame_time);
 }
 
 void System::SaveRunaheadState()
@@ -4858,7 +4736,9 @@ bool System::SaveScreenshot(const char* filename, DisplayScreenshotMode mode, Di
     filename = auto_filename.c_str();
   }
 
-  return g_gpu->RenderScreenshotToFile(filename, mode, quality, compress_on_thread, true);
+  Panic("TODO FIXME");
+  // return g_gpu->RenderScreenshotToFile(filename, mode, quality, compress_on_thread, true);
+  return false;
 }
 
 std::string System::GetGameSaveStateFileName(std::string_view serial, s32 slot)
@@ -5322,7 +5202,8 @@ void System::ToggleSoftwareRendering()
   if (IsShutdown() || g_settings.gpu_renderer == GPURenderer::Software)
     return;
 
-  const GPURenderer new_renderer = g_gpu->IsHardwareRenderer() ? GPURenderer::Software : g_settings.gpu_renderer;
+  const GPURenderer new_renderer =
+    GPUBackend::IsUsingHardwareBackend() ? GPURenderer::Software : g_settings.gpu_renderer;
 
   Host::AddIconOSDMessage("SoftwareRendering", ICON_FA_PAINT_ROLLER,
                           fmt::format(TRANSLATE_FS("OSDMessage", "Switching to {} renderer..."),
@@ -5338,7 +5219,7 @@ void System::RequestDisplaySize(float scale /*= 0.0f*/)
     return;
 
   if (scale == 0.0f)
-    scale = g_gpu->IsHardwareRenderer() ? static_cast<float>(g_settings.gpu_resolution_scale) : 1.0f;
+    scale = GPUBackend::IsUsingHardwareBackend() ? static_cast<float>(g_settings.gpu_resolution_scale) : 1.0f;
 
   const float y_scale =
     (static_cast<float>(g_gpu->GetCRTCDisplayWidth()) / static_cast<float>(g_gpu->GetCRTCDisplayHeight())) /
@@ -5360,62 +5241,7 @@ void System::HostDisplayResized()
   if (g_settings.gpu_widescreen_hack && g_settings.display_aspect_ratio == DisplayAspectRatio::MatchWindow)
     GTE::UpdateAspectRatio();
 
-  g_gpu->UpdateResolutionScale();
-}
-
-bool System::PresentDisplay(bool skip_present, bool explicit_present)
-{
-  // acquire for IO.MousePos.
-  std::atomic_thread_fence(std::memory_order_acquire);
-
-  if (!skip_present)
-  {
-    FullscreenUI::Render();
-    ImGuiManager::RenderTextOverlays();
-    ImGuiManager::RenderOSDMessages();
-
-    if (s_state == State::Running)
-      ImGuiManager::RenderSoftwareCursors();
-  }
-
-  // Debug windows are always rendered, otherwise mouse input breaks on skip.
-  ImGuiManager::RenderOverlayWindows();
-  ImGuiManager::RenderDebugWindows();
-
-  bool do_present;
-  if (g_gpu && !skip_present)
-    do_present = g_gpu->PresentDisplay();
-  else
-    do_present = g_gpu_device->BeginPresent(skip_present);
-
-  if (do_present)
-  {
-    g_gpu_device->RenderImGui();
-    g_gpu_device->EndPresent(explicit_present);
-
-    if (g_gpu_device->IsGPUTimingEnabled())
-    {
-      s_accumulated_gpu_time += g_gpu_device->GetAndResetAccumulatedGPUTime();
-      s_presents_since_last_update++;
-    }
-  }
-  else
-  {
-    // Still need to kick ImGui or it gets cranky.
-    ImGui::Render();
-  }
-
-  ImGuiManager::NewFrame();
-
-  return do_present;
-}
-
-void System::InvalidateDisplay()
-{
-  PresentDisplay(false, false);
-
-  if (g_gpu)
-    g_gpu->RestoreDeviceContext();
+  GPUBackend::PushCommand(GPUBackend::NewUpdateResolutionScaleCommand());
 }
 
 void System::SetTimerResolutionIncreased(bool enabled)

src/core/system.h

@@ -232,10 +232,8 @@ float GetMaximumFrameTime();
 float GetThrottleFrequency();
 float GetCPUThreadUsage();
 float GetCPUThreadAverageTime();
-float GetSWThreadUsage();
-float GetSWThreadAverageTime();
-float GetGPUUsage();
-float GetGPUAverageTime();
+float GetGPUThreadUsage();
+float GetGPUThreadAverageTime();
 const FrameTimeHistory& GetFrameTimeHistory();
 u32 GetFrameTimeHistoryPos();
 void FormatLatencyStats(SmallStringBase& str);
@@ -299,6 +297,7 @@ void SetThrottleFrequency(float frequency);
 void UpdateThrottlePeriod();
 void ResetThrottler();
 void ResetPerformanceCounters();
+void GetFramePresentationDetails(bool* present_frame, bool* allow_present_skip, Common::Timer::Value* present_time);
 
 // Access controllers for simulating input.
 Controller* GetController(u32 slot);
@@ -479,10 +478,6 @@ void RequestDisplaySize(float scale = 0.0f);
 /// Call when host display size changes, use with "match display" aspect ratio setting.
 void HostDisplayResized();
 
-/// Renders the display.
-bool PresentDisplay(bool skip_present, bool explicit_present);
-void InvalidateDisplay();
-
 //////////////////////////////////////////////////////////////////////////
 // Memory Save States (Rewind and Runahead)
 //////////////////////////////////////////////////////////////////////////

src/duckstation-qt/qthost.cpp

@@ -18,6 +18,8 @@
 #include "core/game_list.h"
 #include "core/gdb_server.h"
 #include "core/gpu.h"
+#include "core/gpu_backend.h"
+#include "core/gpu_thread.h"
 #include "core/host.h"
 #include "core/imgui_overlays.h"
 #include "core/memory_card.h"
@@ -762,12 +764,9 @@ void EmuThread::startFullscreenUI()
   m_run_fullscreen_ui = true;
 
   Error error;
-  if (!Host::CreateGPUDevice(Settings::GetRenderAPIForRenderer(g_settings.gpu_renderer), &error) ||
-      !FullscreenUI::Initialize())
+  if (!GPUThread::StartFullscreenUI(&error))
   {
     Host::ReportErrorAsync("Error", error.GetDescription());
-    Host::ReleaseGPUDevice();
-    Host::ReleaseRenderWindow();
     m_run_fullscreen_ui = false;
     return;
   }
@@ -803,8 +802,7 @@ void EmuThread::stopFullscreenUI()
   if (!g_gpu_device)
     return;
 
-  Host::ReleaseGPUDevice();
-  Host::ReleaseRenderWindow();
+  GPUThread::Shutdown();
 }
 
 void EmuThread::bootSystem(std::shared_ptr<SystemBootParameters> params)
@@ -911,7 +909,10 @@ void EmuThread::onDisplayWindowMouseWheelEvent(const QPoint& delta_angle)
 
 void EmuThread::onDisplayWindowResized(int width, int height, float scale)
 {
-  Host::ResizeDisplayWindow(width, height, scale);
+  if (!GPUThread::IsStarted())
+    return;
+
+  GPUThread::ResizeDisplayWindow(width, height, scale);
 }
 
 void EmuThread::redrawDisplayWindow()
@@ -925,7 +926,7 @@ void EmuThread::redrawDisplayWindow()
   if (!g_gpu_device || System::IsShutdown())
     return;
 
-  System::InvalidateDisplay();
+  GPUThread::PresentCurrentFrame();
 }
 
 void EmuThread::toggleFullscreen()
@@ -953,7 +954,7 @@ void EmuThread::setFullscreen(bool fullscreen, bool allow_render_to_main)
 
   m_is_fullscreen = fullscreen;
   m_is_rendering_to_main = allow_render_to_main && shouldRenderToMain();
-  Host::UpdateDisplayWindow();
+  GPUThread::UpdateDisplayWindow();
 }
 
 bool Host::IsFullscreen()
@@ -978,7 +979,7 @@ void EmuThread::setSurfaceless(bool surfaceless)
     return;
 
   m_is_surfaceless = surfaceless;
-  Host::UpdateDisplayWindow();
+  GPUThread::UpdateDisplayWindow();
 }
 
 void EmuThread::requestDisplaySize(float scale)
@@ -1757,33 +1758,16 @@ void EmuThread::run()
   // main loop
   while (!m_shutdown_flag)
   {
+    // TODO: Maybe make this better?
     if (System::IsRunning())
-    {
       System::Execute();
-    }
     else
-    {
-      // we want to keep rendering the UI when paused and fullscreen UI is enabled
-      if (!FullscreenUI::HasActiveWindow() && !System::IsRunning())
-      {
-        // wait until we have a system before running
       m_event_loop->exec();
-        continue;
-      }
-
-      m_event_loop->processEvents(QEventLoop::AllEvents);
-      System::Internal::IdlePollUpdate();
-      if (g_gpu_device)
-      {
-        System::PresentDisplay(false, false);
-        if (!g_gpu_device->IsVSyncModeBlocking())
-          g_gpu_device->ThrottlePresentation();
-      }
-    }
   }
 
   if (System::IsValid())
     System::ShutdownSystem(false);
+  GPUThread::Shutdown();
 
   destroyBackgroundControllerPollTimer();
   System::Internal::CPUThreadShutdown();
@@ -1979,13 +1963,13 @@ void Host::ReleaseRenderWindow()
 
 void EmuThread::updatePerformanceCounters()
 {
-  const RenderAPI render_api = g_gpu_device ? g_gpu_device->GetRenderAPI() : RenderAPI::None;
-  const bool hardware_renderer = g_gpu && g_gpu->IsHardwareRenderer();
+  const RenderAPI render_api = GPUThread::GetRenderAPI();
+  const bool hardware_renderer = GPUBackend::IsUsingHardwareBackend();
   u32 render_width = 0;
   u32 render_height = 0;
 
-  if (g_gpu)
-    std::tie(render_width, render_height) = g_gpu->GetEffectiveDisplayResolution();
+  if (System::IsValid())
+    std::tie(render_width, render_height) = GPUBackend::GetLastDisplaySourceSize();
 
   if (render_api != m_last_render_api || hardware_renderer != m_last_hardware_renderer)
   {

src/duckstation-qt/qthost.h

@@ -96,6 +96,8 @@ public:
   ALWAYS_INLINE bool isFullscreen() const { return m_is_fullscreen; }
   ALWAYS_INLINE bool isRenderingToMain() const { return m_is_rendering_to_main; }
   ALWAYS_INLINE bool isSurfaceless() const { return m_is_surfaceless; }
+
+  // TODO: Maybe remove this...
   ALWAYS_INLINE bool isRunningFullscreenUI() const { return m_run_fullscreen_ui; }
 
   std::optional<WindowInfo> acquireRenderWindow(bool recreate_window);