GPU: CLUT should not be saved on CPU thread memory state

Improves sprite jitter.

data/resources/gamedb.yaml

@@ -21454,6 +21454,7 @@ SCES-02105:
     - AnalogController
     - DigitalController
     - NeGcon
+    - JogCon
   traits:
     - DisablePGXPPreserveProjFP
   libcrypt: true
@@ -21485,6 +21486,7 @@ SCUS-94426:
     - AnalogController
     - DigitalController
     - NeGcon
+    - JogCon
   traits:
     - DisablePGXPPreserveProjFP
   metadata:
@@ -106685,6 +106687,8 @@ SCES-02403:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -106704,6 +106708,8 @@ SCES-03090:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -106723,6 +106729,8 @@ SCES-03091:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -106742,6 +106750,8 @@ SCES-03092:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -106761,6 +106771,8 @@ SCES-03093:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -106780,6 +106792,8 @@ SLUS-01238:
   controllers:
     - AnalogController
     - DigitalController
+  traits:
+    - ForcePGXPCPUMode # Fixes gaps in geometry.
   metadata:
     publisher: "Midway"
     developer: "Magenta Software Ltd. / Traveller's Tales"
@@ -186469,6 +186483,8 @@ SCES-00010:
   controllers:
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"
@@ -186491,6 +186507,8 @@ SIPS-60003:
   controllers:
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"
@@ -186513,6 +186531,8 @@ SCUS-94301:
   controllers:
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"
@@ -186534,6 +186554,8 @@ SLES-00327:
   controllers:
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"
@@ -186550,6 +186572,8 @@ SLES-00327:
     linkCable: true
 SLED-00498:
   name: "WipEout 2097 (Europe) (Demo)"
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     genre: "** DEMO **"
 SCPS-10098:
@@ -186562,6 +186586,8 @@ SCPS-10098:
     - AnalogController
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Sony"
     developer: "Psygnosis Leeds Studio / The Designers Republic"
@@ -186586,6 +186612,8 @@ SLUS-00865:
     - AnalogController
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Sony"
     developer: "Psygnosis Leeds Studio / The Designers Republic"
@@ -186602,6 +186630,8 @@ SLUS-00865:
     linkCable: false
 SLUS-90063:
   name: "WipEout 3 (USA) (Demo)"
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     genre: "** DEMO **"
 SCES-02845:
@@ -186613,6 +186643,8 @@ SCES-02845:
     - AnalogController
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Sony"
     developer: "Psygnosis Leeds Studio / The Designers Republic"
@@ -186642,6 +186674,8 @@ SCPS-45078:
   codes:
     - SCPS-45078
     - SIPS-60010
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"
@@ -186664,6 +186698,8 @@ SCUS-94351:
   controllers:
     - DigitalController
     - NeGcon
+  traits:
+    - ForcePGXPCPUMode # Improves sprite jitter.
   metadata:
     publisher: "Psygnosis"
     developer: "Psygnosis"

scripts/packaging/appimage/inject-libc-apprun.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+
+REQUIRED_GLIBC_VERSION="__REQ_GLIBC_VERSION__"
+
+this_dir="$(readlink -f "$(dirname "$0")")"
+
+APPBIN="${this_dir}/usr/bin/__APPNAME__"
+RUNTIME_DIR="${this_dir}/libc-runtime"
+LOADER_BIN="${this_dir}/usr/bin/ld-linux"
+
+GLIBC_VERSION=$(ldd --version | head -1 | sed -e 's/.* \([0-9.]\)/\1/')
+
+echo "Detected glibc version ${GLIBC_VERSION}."
+
+if [[ -z "${GLIBC_VERSION}" || ! "${GLIBC_VERSION}" < "${REQUIRED_GLIBC_VERSION}" ]]; then
+	echo "Using system libc/libstdc++."
+	exec "${APPBIN}" "$@"
+fi
+
+
+echo "Using bundled libc/libstdc++ from ${RUNTIME_DIR}."
+if [ -z "$LD_LIBRARY_PATH" ]; then
+	export LD_LIBRARY_PATH="${RUNTIME_DIR}"
+else
+	export LD_LIBRARY_PATH="${RUNTIME_DIR}:${LD_LIBRARY_PATH}"
+fi
+
+exec "${LOADER_BIN}" "${APPBIN}" "$@"
+

scripts/packaging/appimage/inject-libc.sh

@@ -0,0 +1,81 @@
+#!/usr/bin/env bash
+
+set -e
+
+function retry_command {
+  # Package servers tend to be unreliable at times..
+  # Retry a bunch of times.
+  local RETRIES=10
+
+  for i in $(seq 1 "$RETRIES"); do
+    "$@" && break
+    if [ "$i" == "$RETRIES" ]; then
+      echo "Command \"$@\" failed after ${RETRIES} retries."
+      exit 1
+    fi
+  done
+}
+
+this_dir="$(readlink -f "$(dirname "$0")")"
+
+if [ "$#" -ne 5 ]; then
+	echo "Syntax: $0 <path to AppDir> <.deb arch> <triple> <ubuntu mirror> <binary to run>"
+	echo "e.g. $0 DuckStation.AppDir amd64 x86_64-linux-gnu https://archive.ubuntu.com/ubuntu/ duckstation-qt"
+	exit 1
+fi
+
+
+APPDIR=$1
+DEBARCH=$2
+TRIPLE=$3
+MIRROR=$4
+APPNAME=$5
+
+LIBC_PACKAGE_URL="${MIRROR}/pool/main/g/glibc/libc6_2.35-0ubuntu3.8_${DEBARCH}.deb"
+LIBGCCS_PACKAGE_URL="${MIRROR}/pool/main/g/gcc-12/libgcc-s1_12.3.0-1ubuntu1~22.04_${DEBARCH}.deb"
+LIBSTDCXX_PACKAGE_URL="${MIRROR}/pool/main/g/gcc-12/libstdc++6_12.3.0-1ubuntu1~22.04_${DEBARCH}.deb"
+GLIBC_VERSION=2.35
+
+mkdir "temp"
+cd "temp"
+retry_command wget -O "libc.deb" "${LIBC_PACKAGE_URL}"
+retry_command wget -O "libgccs.deb" "${LIBGCCS_PACKAGE_URL}"
+retry_command wget -O "libstdc++.deb" "${LIBSTDCXX_PACKAGE_URL}"
+dpkg -x "libc.deb" .
+dpkg -x "libgccs.deb" .
+dpkg -x "libstdc++.deb" .
+
+# Copy everything into AppDir
+RUNTIME="${APPDIR}/libc-runtime"
+mkdir -p "${RUNTIME}"
+
+# libc.so.6 and friends
+cd "lib/${TRIPLE}"
+cp -v * "${RUNTIME}"
+cd ../../
+
+# libstdc++
+cd "usr/lib/${TRIPLE}"
+cp -v * "${RUNTIME}" || true
+cd ../../..
+
+# done with temps now
+cd ..
+rm -fr temp
+
+# Not risking mixing resolvers...
+cd "${RUNTIME}"
+rm -vf libnss_*
+
+# Move ld-linux.so.2 into the binary directory so we can preserve arg0's directory
+mv -v "ld-linux-"*.so.2 "${APPDIR}/usr/bin/ld-linux"
+
+# Set up the replacement apprun script
+cd "${APPDIR}"
+rm -f AppRun.wrapped
+cp "${this_dir}/inject-libc-apprun.sh" AppRun.wrapped
+sed -i -e "s/__APPNAME__/${APPNAME}/" AppRun.wrapped
+sed -i -e "s/__REQ_GLIBC_VERSION__/${GLIBC_VERSION}/" AppRun.wrapped
+
+echo Done.
+

src/common/gsvector_neon.h

@@ -3131,6 +3131,11 @@ public:
 
   ALWAYS_INLINE GSVector2 zw() const { return GSVector2(vget_high_s32(v4s)); }
 
+  ALWAYS_INLINE static GSVector4 xyxy(const GSVector2& l, const GSVector2& h)
+  {
+    return GSVector4(vcombine_f32(l.v2s, h.v2s));
+  }
+
 #define VECTOR4_SHUFFLE_4(xs, xn, ys, yn, zs, zn, ws, wn)                                                              \
   ALWAYS_INLINE GSVector4 xs##ys##zs##ws() const                                                                       \
   {                                                                                                                    \

src/common/gsvector_nosimd.h

@@ -2316,6 +2316,11 @@ public:
 
   ALWAYS_INLINE GSVector2 zw() const { return GSVector2(z, w); }
 
+  ALWAYS_INLINE static GSVector4 xyxy(const GSVector2& l, const GSVector2& h)
+  {
+    return GSVector4(l.x, l.y, h.x, h.y);
+  }
+
 #define VECTOR4_SHUFFLE_4(xs, xn, ys, yn, zs, zn, ws, wn)                                                              \
   ALWAYS_INLINE GSVector4 xs##ys##zs##ws() const { return GSVector4(F32[xn], F32[yn], F32[zn], F32[wn]); }
 

src/common/intrin.h

@@ -14,9 +14,9 @@
 #define CPU_ARCH_SIMD 1
 #define CPU_ARCH_SSE 1
 #include <emmintrin.h>
-#include <tmmintrin.h>
-#include <smmintrin.h>
 #include <immintrin.h>
+#include <smmintrin.h>
+#include <tmmintrin.h>
 
 #if defined(__AVX2__)
 #define CPU_ARCH_AVX 1
@@ -96,3 +96,40 @@ ALWAYS_INLINE_RELEASE static void MemsetPtrs(T* ptr, T value, u32 count)
   for (u32 i = 0; i < remaining_count; i++)
     *(dest++) = value;
 }
+
+ALWAYS_INLINE 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
+}

src/common/log_channels.h

@@ -27,6 +27,7 @@
   X(GPU)                                                                                                               \
   X(GPUDevice)                                                                                                         \
   X(GPUDump)                                                                                                           \
+  X(GPUThread)                                                                                                         \
   X(GPU_SW)                                                                                                            \
   X(GPU_HW)                                                                                                            \
   X(GameDatabase)                                                                                                      \

src/common/types.h

@@ -93,6 +93,13 @@ char (&__countof_ArraySizeHelper(T (&array)[N]))[N];
   } while (0)
 #endif
 
+// __restrict, potentially enables optimization by hinting the compiler that the object is unique.
+#ifdef _MSC_VER
+#define RESTRICT __restrict
+#else
+#define RESTRICT __restrict__
+#endif
+
 // disable warnings that show up at warning level 4
 // TODO: Move to build system instead
 #ifdef _MSC_VER

src/core/CMakeLists.txt

@@ -61,10 +61,11 @@ 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_thread_commands.h
   gpu_types.h
   guncon.cpp
   guncon.h
@@ -73,8 +74,6 @@ add_library(core
   gte_types.h
   host.cpp
   host.h
-  host_interface_progress_callback.cpp
-  host_interface_progress_callback.h
   hotkeys.cpp
   input_types.h
   imgui_overlays.cpp

src/core/achievements.cpp

@@ -9,7 +9,9 @@
 #include "bus.h"
 #include "cpu_core.h"
 #include "fullscreen_ui.h"
+#include "gpu_thread.h"
 #include "host.h"
+#include "imgui_overlays.h"
 #include "system.h"
 
 #include "scmversion/scmversion.h"
@@ -480,7 +482,9 @@ void Achievements::UpdateGlyphRanges()
   std::sort(sorted_codepoints.begin(), sorted_codepoints.end());
 
   // Compact codepoints to ranges.
-  ImGuiManager::SetEmojiFontRange(ImGuiManager::CompactFontRange(sorted_codepoints));
+  GPUThread::RunOnThread([sorted_codepoints = std::move(sorted_codepoints)]() {
+    ImGuiManager::SetEmojiFontRange(ImGuiManager::CompactFontRange(sorted_codepoints));
+  });
 }
 
 bool Achievements::IsActive()
@@ -828,15 +832,6 @@ void Achievements::IdleUpdate()
   rc_client_idle(s_state.client);
 }
 
-bool Achievements::NeedsIdleUpdate()
-{
-  if (!IsActive())
-    return false;
-
-  const auto lock = GetLock();
-  return (s_state.http_downloader && s_state.http_downloader->HasAnyRequests());
-}
-
 void Achievements::FrameUpdate()
 {
   if (!IsActive())
@@ -1182,7 +1177,7 @@ 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);
 
   char url_buf[URL_BUFFER_SIZE];
   if (int err = rc_client_game_get_image_url(info, url_buf, std::size(url_buf)); err == RC_OK)
@@ -1238,7 +1233,7 @@ void Achievements::ClearGameHash()
 
 void Achievements::DisplayAchievementSummary()
 {
-  if (g_settings.achievements_notifications && FullscreenUI::Initialize())
+  if (g_settings.achievements_notifications)
   {
     std::string title;
     if (IsHardcoreModeActive())
@@ -1263,8 +1258,14 @@ void Achievements::DisplayAchievementSummary()
       summary = TRANSLATE_STR("Achievements", "This game has no achievements.");
     }
 
-    ImGuiFullscreen::AddNotification("achievement_summary", ACHIEVEMENT_SUMMARY_NOTIFICATION_TIME, std::move(title),
+    GPUThread::RunOnThread(
-                                     std::move(summary), s_state.game_icon);
+      [title = std::move(title), summary = std::move(summary), icon = s_state.game_icon]() mutable {
+        if (!FullscreenUI::Initialize())
+          return;
+
+        ImGuiFullscreen::AddNotification("achievement_summary", ACHIEVEMENT_SUMMARY_NOTIFICATION_TIME, std::move(title),
+                                         std::move(summary), std::move(icon));
+      });
   }
 
   // Technically not going through the resource API, but since we're passing this to something else, we can't.
@@ -1274,12 +1275,16 @@ void Achievements::DisplayAchievementSummary()
 
 void Achievements::DisplayHardcoreDeferredMessage()
 {
-  if (g_settings.achievements_hardcore_mode && !s_state.hardcore_mode && System::IsValid() &&
+  if (g_settings.achievements_hardcore_mode && !s_state.hardcore_mode && System::IsValid())
-      FullscreenUI::Initialize())
   {
-    ImGuiFullscreen::ShowToast(std::string(),
+    GPUThread::RunOnThread([]() {
-                               TRANSLATE_STR("Achievements", "Hardcore mode will be enabled on system reset."),
+      if (!FullscreenUI::Initialize())
-                               Host::OSD_WARNING_DURATION);
+        return;
+
+      ImGuiFullscreen::ShowToast(std::string(),
+                                 TRANSLATE_STR("Achievements", "Hardcore mode will be enabled on system reset."),
+                                 Host::OSD_WARNING_DURATION);
+    });
   }
 }
 
@@ -1301,7 +1306,7 @@ void Achievements::HandleUnlockEvent(const rc_client_event_t* event)
   INFO_LOG("Achievement {} ({}) for game {} unlocked", cheevo->title, cheevo->id, s_state.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)
@@ -1311,9 +1316,15 @@ 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),
+    GPUThread::RunOnThread([id = cheevo->id, duration = g_settings.achievements_notification_duration,
-                                     static_cast<float>(g_settings.achievements_notification_duration),
+                            title = std::move(title), description = std::string(cheevo->description),
-                                     std::move(title), cheevo->description, std::move(badge_path));
+                            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)
@@ -1325,7 +1336,7 @@ void Achievements::HandleGameCompleteEvent(const rc_client_event_t* event)
   INFO_LOG("Game {} complete", s_state.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_state.game_title);
     std::string message = fmt::format(
@@ -1334,8 +1345,14 @@ void Achievements::HandleGameCompleteEvent(const rc_client_event_t* event)
                            s_state.game_summary.num_unlocked_achievements),
       TRANSLATE_PLURAL_STR("Achievements", "%n points", "Achievement points", s_state.game_summary.points_unlocked));
 
-    ImGuiFullscreen::AddNotification("achievement_mastery", GAME_COMPLETE_NOTIFICATION_TIME, std::move(title),
+    GPUThread::RunOnThread(
-                                     std::move(message), s_state.game_icon);
+      [title = std::move(title), message = std::move(message), icon = s_state.game_icon]() mutable {
+        if (!FullscreenUI::Initialize())
+          return;
+
+        ImGuiFullscreen::AddNotification("achievement_mastery", GAME_COMPLETE_NOTIFICATION_TIME, std::move(title),
+                                         std::move(message), std::move(icon));
+      });
   }
 }
 
@@ -1343,14 +1360,19 @@ 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),
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
-                                     LEADERBOARD_STARTED_NOTIFICATION_TIME, std::move(title), std::move(message),
+                            icon = s_state.game_icon]() mutable {
-                                     s_state.game_icon);
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id), LEADERBOARD_STARTED_NOTIFICATION_TIME,
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1358,14 +1380,19 @@ 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),
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
-                                     LEADERBOARD_FAILED_NOTIFICATION_TIME, std::move(title), std::move(message),
+                            icon = s_state.game_icon]() mutable {
-                                     s_state.game_icon);
+      if (!FullscreenUI::Initialize())
+        return;
+
+      ImGuiFullscreen::AddNotification(fmt::format("leaderboard_{}", id), LEADERBOARD_FAILED_NOTIFICATION_TIME,
+                                       std::move(title), std::move(message), std::move(icon));
+    });
   }
 }
 
@@ -1373,7 +1400,7 @@ 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: {}{}"),
@@ -1389,9 +1416,14 @@ 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),
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
-                                     static_cast<float>(g_settings.achievements_leaderboard_duration), std::move(title),
+                            icon = s_state.game_icon]() mutable {
-                                     std::move(message), s_state.game_icon);
+      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)
@@ -1403,7 +1435,7 @@ 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: {})"),
@@ -1420,9 +1452,15 @@ 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),
+    GPUThread::RunOnThread([id = event->leaderboard->id, title = std::move(title), message = std::move(message),
-                                     static_cast<float>(g_settings.achievements_leaderboard_duration), std::move(title),
+                            icon = s_state.game_icon]() mutable {
-                                     std::move(message), s_state.game_icon);
+      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));
+    });
   }
 }
 
@@ -1552,26 +1590,30 @@ 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()
@@ -1649,12 +1691,17 @@ void Achievements::SetHardcoreMode(bool enabled, bool force_display_message)
   // new mode
   s_state.hardcore_mode = enabled;
 
-  if (System::IsValid() && (HasActiveGame() || force_display_message) && FullscreenUI::Initialize())
+  if (System::IsValid() && (HasActiveGame() || force_display_message))
   {
-    ImGuiFullscreen::ShowToast(std::string(),
+    GPUThread::RunOnThread([enabled]() {
-                               enabled ? TRANSLATE_STR("Achievements", "Hardcore mode is now enabled.") :
+      if (!FullscreenUI::Initialize())
-                                         TRANSLATE_STR("Achievements", "Hardcore mode is now disabled."),
+        return;
-                               Host::OSD_INFO_DURATION);
+
+      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_state.client, enabled);
@@ -1696,8 +1743,15 @@ bool Achievements::DoState(StateWrapper& sw)
     // before deserializing, otherwise that state's going to get lost.
     if (!IsUsingRAIntegration() && s_state.load_game_request)
     {
-      Host::DisplayLoadingScreen("Downloading achievements data...");
+      // Messy because GPU-thread, but at least it looks pretty.
+      GPUThread::RunOnThread([]() {
+        FullscreenUI::OpenLoadingScreen(ImGuiManager::LOGO_IMAGE_NAME,
+                                        TRANSLATE_SV("Achievements", "Downloading achievements data..."));
+      });
+
       s_state.http_downloader->WaitForAllRequests();
+
+      GPUThread::RunOnThread([]() { FullscreenUI::CloseLoadingScreen(); });
     }
 
     u32 data_size = 0;
@@ -1966,7 +2020,7 @@ 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;
@@ -1975,8 +2029,14 @@ void Achievements::ShowLoginNotification()
     std::string summary = fmt::format(TRANSLATE_FS("Achievements", "Score: {} ({} softcore)\nUnread messages: {}"),
                                       user->score, user->score_softcore, user->num_unread_messages);
 
-    ImGuiFullscreen::AddNotification("achievements_login", LOGIN_NOTIFICATION_TIME, std::move(title),
+    GPUThread::RunOnThread(
-                                     std::move(summary), std::move(badge_path));
+      [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));
+      });
   }
 }
 
@@ -2075,6 +2135,15 @@ bool Achievements::ConfirmHardcoreModeDisable(const char* trigger)
 
 void Achievements::ConfirmHardcoreModeDisableAsync(const char* trigger, std::function<void(bool)> callback)
 {
+  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]() {
+      if (res)
+        DisableHardcoreMode();
+      callback(res);
+    });
+  };
+
 #ifndef __ANDROID__
 #ifdef ENABLE_RAINTEGRATION
   if (IsUsingRAIntegration())
@@ -2085,45 +2154,38 @@ void Achievements::ConfirmHardcoreModeDisableAsync(const char* trigger, std::fun
   }
 #endif
 
-  if (!FullscreenUI::Initialize())
+  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);
+      Host::AddOSDMessage(
-    callback(false);
+        fmt::format(TRANSLATE_FS("Achievements", "Cannot {} while hardcode mode is active."), trigger),
-    return;
+        Host::OSD_WARNING_DURATION);
-  }
+      real_callback(false);
+      return;
+    }
 
-  auto real_callback = [callback = std::move(callback)](bool res) mutable {
+    ImGuiFullscreen::OpenConfirmMessageDialog(
-    // don't run the callback in the middle of rendering the UI
+      TRANSLATE_STR("Achievements", "Confirm Hardcore Mode"),
-    Host::RunOnCPUThread([callback = std::move(callback), res]() {
+      fmt::format(TRANSLATE_FS("Achievements",
-      if (res)
+                               "{0} cannot be performed while hardcore mode is active. Do you "
-        DisableHardcoreMode();
+                               "want to disable hardcore mode? {0} will be cancelled if you select No."),
-      callback(res);
+                  trigger),
-    });
+      std::move(real_callback), fmt::format(ICON_FA_CHECK " {}", TRANSLATE_SV("Achievements", "Yes")),
-  };
+      fmt::format(ICON_FA_TIMES " {}", TRANSLATE_SV("Achievements", "No")));
-
+  });
-  ImGuiFullscreen::OpenConfirmMessageDialog(
-    TRANSLATE_STR("Achievements", "Confirm Hardcore Mode"),
-    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::ConfirmMessageAsync(
-                      Host::OSD_WARNING_DURATION);
+    TRANSLATE_STR("Achievements", "Confirm Hardcore Mode"),
-  callback(false);
+    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));
 #endif
 }
 
 void Achievements::ClearUIState()
 {
-#ifndef __ANDROID__
-  if (FullscreenUI::IsAchievementsWindowOpen() || FullscreenUI::IsLeaderboardsWindowOpen())
-    FullscreenUI::ReturnToPreviousWindow();
-
   CloseLeaderboard();
-#endif
 
   s_state.achievement_badge_paths = {};
 
@@ -2403,11 +2465,15 @@ void Achievements::DrawAchievementsWindow()
   using ImGuiFullscreen::LayoutScale;
   using ImGuiFullscreen::UIStyle;
 
-  if (!s_state.achievement_list)
-    return;
-
   auto lock = Achievements::GetLock();
 
+  // achievements can get turned off via the main UI
+  if (!s_state.achievement_list)
+  {
+    FullscreenUI::ReturnToPreviousWindow();
+    return;
+  }
+
   static constexpr float alpha = 0.8f;
   static constexpr float heading_alpha = 0.95f;
   static constexpr float heading_height_unscaled = 110.0f;
@@ -2779,7 +2845,12 @@ void Achievements::DrawLeaderboardsWindow()
   static constexpr float heading_height_unscaled = 110.0f;
   static constexpr float tab_height_unscaled = 50.0f;
 
-  auto lock = Achievements::GetLock();
+  const auto lock = Achievements::GetLock();
+  if (!s_state.leaderboard_list)
+  {
+    FullscreenUI::ReturnToPreviousWindow();
+    return;
+  }
 
   const bool is_leaderboard_open = (s_state.open_leaderboard != nullptr);
   bool close_leaderboard_on_exit = false;
@@ -3370,7 +3441,6 @@ void Achievements::CloseLeaderboard()
   }
 
   s_state.open_leaderboard = nullptr;
-  ImGuiFullscreen::QueueResetFocus(ImGuiFullscreen::FocusResetType::Other);
 }
 
 #if defined(_WIN32)

src/core/achievements.h

@@ -58,9 +58,6 @@ void FrameUpdate();
 /// Called when the system is paused, because FrameUpdate() won't be getting called.
 void IdleUpdate();
 
-/// Returns true if idle updates are necessary (e.g. outstanding requests).
-bool NeedsIdleUpdate();
-
 /// Saves/loads state.
 bool DoState(StateWrapper& sw);
 

src/core/analog_joystick.cpp

@@ -10,6 +10,7 @@
 
 #include "common/bitutils.h"
 #include "common/log.h"
+#include "common/settings_interface.h"
 #include "common/string_util.h"
 
 #include "IconsFontAwesome5.h"
@@ -340,14 +341,14 @@ std::unique_ptr<AnalogJoystick> AnalogJoystick::Create(u32 index)
 
 static const Controller::ControllerBindingInfo s_binding_info[] = {
 #define BUTTON(name, display_name, icon_name, button, genb)                                                            \
-  {                                                                                                                    \
+  {name, display_name, icon_name, static_cast<u32>(button), InputBindingInfo::Type::Button, genb}
-    name, display_name, icon_name, static_cast<u32>(button), InputBindingInfo::Type::Button, genb                      \
-  }
 #define AXIS(name, display_name, icon_name, halfaxis, genb)                                                            \
-  {                                                                                                                    \
+  {name,                                                                                                               \
-    name, display_name, icon_name, static_cast<u32>(AnalogJoystick::Button::Count) + static_cast<u32>(halfaxis),       \
+   display_name,                                                                                                       \
-      InputBindingInfo::Type::HalfAxis, genb                                                                           \
+   icon_name,                                                                                                          \
-  }
+   static_cast<u32>(AnalogJoystick::Button::Count) + static_cast<u32>(halfaxis),                                       \
+   InputBindingInfo::Type::HalfAxis,                                                                                   \
+   genb}
 
   // clang-format off
   BUTTON("Up", TRANSLATE_NOOP("AnalogJoystick", "D-Pad Up"), ICON_PF_DPAD_UP, AnalogJoystick::Button::Up, GenericInputBinding::DPadUp),

src/core/bus.cpp

@@ -1875,7 +1875,7 @@ template<MemoryAccessSize size>
 u32 Bus::HWHandlers::GPURead(PhysicalMemoryAddress address)
 {
   const u32 offset = address & GPU_MASK;
-  u32 value = g_gpu->ReadRegister(FIXUP_WORD_OFFSET(size, offset));
+  u32 value = g_gpu.ReadRegister(FIXUP_WORD_OFFSET(size, offset));
   value = FIXUP_WORD_READ_VALUE(size, offset, value);
   BUS_CYCLES(2);
   return value;
@@ -1885,7 +1885,7 @@ template<MemoryAccessSize size>
 void Bus::HWHandlers::GPUWrite(PhysicalMemoryAddress address, u32 value)
 {
   const u32 offset = address & GPU_MASK;
-  g_gpu->WriteRegister(FIXUP_WORD_OFFSET(size, offset), FIXUP_WORD_WRITE_VALUE(size, offset, value));
+  g_gpu.WriteRegister(FIXUP_WORD_OFFSET(size, offset), FIXUP_WORD_WRITE_VALUE(size, offset, value));
 }
 
 template<MemoryAccessSize size>

src/core/cdrom.cpp

@@ -5,8 +5,8 @@
 #include "cdrom_async_reader.h"
 #include "cdrom_subq_replacement.h"
 #include "dma.h"
+#include "fullscreen_ui.h"
 #include "host.h"
-#include "host_interface_progress_callback.h"
 #include "interrupt_controller.h"
 #include "settings.h"
 #include "spu.h"
@@ -999,7 +999,7 @@ bool CDROM::PrecacheMedia()
     return false;
   }
 
-  HostInterfaceProgressCallback callback;
+  LoadingScreenProgressCallback callback;
   if (!s_reader.Precache(&callback))
   {
     Host::AddOSDMessage(TRANSLATE_STR("OSDMessage", "Precaching CD image failed, it may be unreliable."),

src/core/cheats.cpp

@@ -17,6 +17,7 @@
 #include "common/file_system.h"
 #include "common/log.h"
 #include "common/path.h"
+#include "common/settings_interface.h"
 #include "common/small_string.h"
 #include "common/string_util.h"
 #include "common/zip_helpers.h"

src/core/core.vcxproj

@@ -40,14 +40,13 @@
     <ClCompile Include="gpu_hw_texture_cache.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_thread.cpp" />
     <ClCompile Include="gte.cpp" />
     <ClCompile Include="dma.cpp" />
     <ClCompile Include="gpu.cpp" />
     <ClCompile Include="gpu_hw.cpp" />
     <ClCompile Include="host.cpp" />
-    <ClCompile Include="host_interface_progress_callback.cpp" />
     <ClCompile Include="hotkeys.cpp" />
     <ClCompile Include="imgui_overlays.cpp" />
     <ClCompile Include="interrupt_controller.cpp" />
@@ -119,8 +118,9 @@
     <ClInclude Include="gpu_hw_texture_cache.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_thread_commands.h" />
     <ClInclude Include="gpu_types.h" />
     <ClInclude Include="gte.h" />
     <ClInclude Include="cpu_types.h" />
@@ -129,7 +129,6 @@
     <ClInclude Include="gpu_hw.h" />
     <ClInclude Include="gte_types.h" />
     <ClInclude Include="host.h" />
-    <ClInclude Include="host_interface_progress_callback.h" />
     <ClInclude Include="imgui_overlays.h" />
     <ClInclude Include="input_types.h" />
     <ClInclude Include="interrupt_controller.h" />

src/core/core.vcxproj.filters

@@ -33,13 +33,11 @@
     <ClCompile Include="guncon.cpp" />
     <ClCompile Include="playstation_mouse.cpp" />
     <ClCompile Include="negcon.cpp" />
-    <ClCompile Include="host_interface_progress_callback.cpp" />
     <ClCompile Include="cpu_pgxp.cpp" />
     <ClCompile Include="cheats.cpp" />
     <ClCompile Include="memory_card_image.cpp" />
     <ClCompile Include="analog_joystick.cpp" />
     <ClCompile Include="gpu_backend.cpp" />
-    <ClCompile Include="gpu_sw_backend.cpp" />
     <ClCompile Include="multitap.cpp" />
     <ClCompile Include="host.cpp" />
     <ClCompile Include="game_database.cpp" />
@@ -67,6 +65,7 @@
     <ClCompile Include="performance_counters.cpp" />
     <ClCompile Include="jogcon.cpp" />
     <ClCompile Include="pio.cpp" />
+    <ClCompile Include="gpu_thread.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="types.h" />
@@ -102,7 +101,6 @@
     <ClInclude Include="guncon.h" />
     <ClInclude Include="playstation_mouse.h" />
     <ClInclude Include="negcon.h" />
-    <ClInclude Include="host_interface_progress_callback.h" />
     <ClInclude Include="gte_types.h" />
     <ClInclude Include="cpu_pgxp.h" />
     <ClInclude Include="cpu_core_private.h" />
@@ -111,7 +109,6 @@
     <ClInclude Include="analog_joystick.h" />
     <ClInclude Include="gpu_types.h" />
     <ClInclude Include="gpu_backend.h" />
-    <ClInclude Include="gpu_sw_backend.h" />
     <ClInclude Include="multitap.h" />
     <ClInclude Include="host.h" />
     <ClInclude Include="achievements.h" />
@@ -143,6 +140,8 @@
     <ClInclude Include="system_private.h" />
     <ClInclude Include="jogcon.h" />
     <ClInclude Include="pio.h" />
+    <ClInclude Include="gpu_thread.h" />
+    <ClInclude Include="gpu_thread_commands.h" />
   </ItemGroup>
   <ItemGroup>
     <None Include="gpu_sw_rasterizer.inl" />

src/core/cpu_code_cache.cpp

@@ -1692,9 +1692,10 @@ PageFaultHandler::HandlerResult CPU::CodeCache::HandleFastmemException(void* exc
 
     // if we're writing to ram, let it go through a few times, and use manual block protection to sort it out
     // TODO: path for manual protection to return back to read-only pages
-    if (is_write && !g_state.cop0_regs.sr.Isc && GetSegmentForAddress(guest_address) != CPU::Segment::KSEG2 &&
+    if (!g_state.cop0_regs.sr.Isc && GetSegmentForAddress(guest_address) != CPU::Segment::KSEG2 &&
         AddressInRAM(guest_address))
     {
+      DebugAssert(is_write);
       DEV_LOG("Ignoring fault due to RAM write @ 0x{:08X}", guest_address);
       InvalidateBlocksWithPageIndex(Bus::GetRAMCodePageIndex(guest_address));
       return PageFaultHandler::HandlerResult::ContinueExecution;
@@ -1707,15 +1708,12 @@ PageFaultHandler::HandlerResult CPU::CodeCache::HandleFastmemException(void* exc
     guest_address = std::numeric_limits<PhysicalMemoryAddress>::max();
   }
 
-  DEV_LOG("Page fault handler invoked at PC={} Address={} {}, fastmem offset {:08X}", exception_pc, fault_address,
-          is_write ? "(write)" : "(read)", guest_address);
-
   auto iter = s_fastmem_backpatch_info.find(exception_pc);
   if (iter == s_fastmem_backpatch_info.end())
-  {
-    ERROR_LOG("No backpatch info found for {}", exception_pc);
     return PageFaultHandler::HandlerResult::ExecuteNextHandler;
-  }
+
+  DEV_LOG("Page fault handler invoked at PC={} Address={} {}, fastmem offset {:08X}", exception_pc, fault_address,
+          is_write ? "(write)" : "(read)", guest_address);
 
   LoadstoreBackpatchInfo& info = iter->second;
   DEV_LOG("Backpatching {} at {}[{}] (pc {:08X} addr {:08X}): Bitmask {:08X} Addr {} Data {} Size {} Signed {:02X}",

src/core/cpu_recompiler_arm32.cpp

@@ -31,10 +31,6 @@ LOG_CHANNEL(Recompiler);
 #define PTR(x) vixl::aarch32::MemOperand(RSTATE, (((u8*)(x)) - ((u8*)&g_state)))
 #define RMEMBASE vixl::aarch32::r3
 
-static constexpr u32 FUNCTION_CALLEE_SAVED_SPACE_RESERVE = 80;  // 8 registers
-static constexpr u32 FUNCTION_CALLER_SAVED_SPACE_RESERVE = 144; // 18 registers -> 224 bytes
-static constexpr u32 FUNCTION_STACK_SIZE = FUNCTION_CALLEE_SAVED_SPACE_RESERVE + FUNCTION_CALLER_SAVED_SPACE_RESERVE;
-
 #define RRET vixl::aarch32::r0
 #define RRETHI vixl::aarch32::r1
 #define RARG1 vixl::aarch32::r0
@@ -266,9 +262,6 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 
   g_enter_recompiler = armAsm->GetCursorAddress<decltype(g_enter_recompiler)>();
   {
-    // reserve some space for saving caller-saved registers
-    armAsm->sub(sp, sp, FUNCTION_STACK_SIZE);
-
     // Need the CPU state for basically everything :-)
     armMoveAddressToReg(armAsm, RSTATE, &g_state);
   }
@@ -2272,16 +2265,19 @@ void CPU::ARM32Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
-    SwitchToFarCodeIfBitSet(changed_bits, 16);
+    Label caches_unchanged;
-    armAsm->push(RegisterList(RARG1));
+    armAsm->tst(changed_bits, 1u << 16);
+    armAsm->b(eq, &caches_unchanged);
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
-    armAsm->pop(RegisterList(RARG1));
+    armAsm->ldr(RARG1, PTR(ptr)); // reload value for interrupt test below
+    armAsm->bind(&caches_unchanged);
+
+    // might need to reload fastmem base too
     if (CodeCache::IsUsingFastmem() && m_block->HasFlag(CodeCache::BlockFlags::ContainsLoadStoreInstructions) &&
         IsHostRegAllocated(RMEMBASE.GetCode()))
     {
       FreeHostReg(RMEMBASE.GetCode());
     }
-    SwitchToNearCode(true);
 
     TestInterrupts(RARG1);
   }

src/core/cpu_recompiler_arm64.cpp

@@ -28,10 +28,6 @@ LOG_CHANNEL(Recompiler);
 
 #define PTR(x) vixl::aarch64::MemOperand(RSTATE, (((u8*)(x)) - ((u8*)&g_state)))
 
-static constexpr u64 FUNCTION_CALLEE_SAVED_SPACE_RESERVE = 80;  // 8 registers
-static constexpr u64 FUNCTION_CALLER_SAVED_SPACE_RESERVE = 144; // 18 registers -> 224 bytes
-static constexpr u64 FUNCTION_STACK_SIZE = FUNCTION_CALLEE_SAVED_SPACE_RESERVE + FUNCTION_CALLER_SAVED_SPACE_RESERVE;
-
 #define RWRET vixl::aarch64::w0
 #define RXRET vixl::aarch64::x0
 #define RWARG1 vixl::aarch64::w0
@@ -448,9 +444,6 @@ u32 CPU::CodeCache::EmitASMFunctions(void* code, u32 code_size)
 
   g_enter_recompiler = armAsm->GetCursorAddress<decltype(g_enter_recompiler)>();
   {
-    // reserve some space for saving caller-saved registers
-    armAsm->sub(sp, sp, FUNCTION_STACK_SIZE);
-
     // Need the CPU state for basically everything :-)
     armMoveAddressToReg(armAsm, RSTATE, &g_state);
 
@@ -2435,14 +2428,13 @@ void CPU::ARM64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
-    SwitchToFarCodeIfBitSet(changed_bits, 16);
+    Label caches_unchanged;
-    armAsm->sub(sp, sp, 16);
+    armAsm->tbz(changed_bits, 16, &caches_unchanged);
-    armAsm->str(RWARG1, MemOperand(sp));
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
-    armAsm->ldr(RWARG1, MemOperand(sp));
+    armAsm->ldr(RWARG1, PTR(ptr)); // reload value for interrupt test below
-    armAsm->add(sp, sp, 16);
+    if (CodeCache::IsUsingFastmem())
-    armAsm->ldr(RMEMBASE, PTR(&g_state.fastmem_base));
+      armAsm->ldr(RMEMBASE, PTR(&g_state.fastmem_base));
-    SwitchToNearCode(true);
+    armAsm->bind(&caches_unchanged);
 
     TestInterrupts(RWARG1);
   }

src/core/cpu_recompiler_riscv64.cpp

@@ -2283,16 +2283,15 @@ void CPU::RISCV64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
+    Label caches_unchanged;
     rvAsm->SRLIW(RSCRATCH, changed_bits, 16);
     rvAsm->ANDI(RSCRATCH, RSCRATCH, 1);
-    SwitchToFarCode(true, &Assembler::BEQ, RSCRATCH, zero);
+    rvAsm->BEQ(RSCRATCH, zero, &caches_unchanged);
-    rvAsm->ADDI(sp, sp, -16);
-    rvAsm->SW(RARG1, 0, sp);
     EmitCall(reinterpret_cast<const void*>(&CPU::UpdateMemoryPointers));
-    rvAsm->LW(RARG1, 0, sp);
+    rvAsm->LW(new_value, PTR(ptr));
-    rvAsm->ADDI(sp, sp, 16);
+    if (CodeCache::IsUsingFastmem())
-    rvAsm->LD(RMEMBASE, PTR(&g_state.fastmem_base));
+      rvAsm->LD(RMEMBASE, PTR(&g_state.fastmem_base));
-    SwitchToNearCode(true);
+    rvAsm->Bind(&caches_unchanged);
 
     TestInterrupts(RARG1);
   }

src/core/cpu_recompiler_x64.cpp

@@ -12,6 +12,7 @@
 #include "common/align.h"
 #include "common/assert.h"
 #include "common/log.h"
+#include "common/small_string.h"
 #include "common/string_util.h"
 
 #include <limits>
@@ -2241,15 +2242,15 @@ void CPU::X64Recompiler::Compile_mtc0(CompileFlags cf)
     // We could just inline the whole thing..
     Flush(FLUSH_FOR_C_CALL);
 
+    Label caches_unchanged;
     cg->test(changed_bits, 1u << 16);
-    SwitchToFarCode(true, &CodeGenerator::jnz);
+    cg->jz(caches_unchanged);
-    cg->mov(cg->dword[cg->rsp], RWARG2);
-    cg->sub(cg->rsp, STACK_SHADOW_SIZE + 8);
     cg->call(&CPU::UpdateMemoryPointers);
-    cg->add(cg->rsp, STACK_SHADOW_SIZE + 8);
+    cg->mov(RWARG2, cg->dword[PTR(ptr)]); // reload value for interrupt test below
-    cg->mov(RWARG2, cg->dword[cg->rsp]);
+    if (CodeCache::IsUsingFastmem())
-    cg->mov(RMEMBASE, cg->qword[PTR(&g_state.fastmem_base)]);
+      cg->mov(RMEMBASE, cg->qword[PTR(&g_state.fastmem_base)]);
-    SwitchToNearCode(true);
+
+    cg->L(caches_unchanged);
 
     TestInterrupts(RWARG2);
   }
@@ -2475,7 +2476,7 @@ u32 CPU::Recompiler::CompileLoadStoreThunk(void* thunk_code, u32 thunk_space, vo
       num_gprs++;
   }
 
-  const u32 stack_size = (((num_gprs + 1) & ~1u) * GPR_SIZE) + STACK_SHADOW_SIZE;
+  const u32 stack_size = (((num_gprs + 1) & ~1u) * GPR_SIZE);
 
   if (stack_size > 0)
   {

src/core/dma.cpp

@@ -802,9 +802,9 @@ TickCount DMA::TransferMemoryToDevice(u32 address, u32 increment, u32 word_count
   {
     case Channel::GPU:
     {
-      if (g_gpu->BeginDMAWrite()) [[likely]]
+      if (g_gpu.BeginDMAWrite()) [[likely]]
       {
-        if (GPUDump::Recorder* dump = g_gpu->GetGPUDump()) [[unlikely]]
+        if (GPUDump::Recorder* dump = g_gpu.GetGPUDump()) [[unlikely]]
         {
           // No wraparound?
           dump->BeginGP0Packet(word_count);
@@ -831,10 +831,10 @@ TickCount DMA::TransferMemoryToDevice(u32 address, u32 increment, u32 word_count
         {
           u32 value;
           std::memcpy(&value, &ram_pointer[address], sizeof(u32));
-          g_gpu->DMAWrite(address, value);
+          g_gpu.DMAWrite(address, value);
           address = (address + increment) & mask;
         }
-        g_gpu->EndDMAWrite();
+        g_gpu.EndDMAWrite();
       }
     }
     break;
@@ -900,7 +900,7 @@ TickCount DMA::TransferDeviceToMemory(u32 address, u32 increment, u32 word_count
   switch (channel)
   {
     case Channel::GPU:
-      g_gpu->DMARead(dest_pointer, word_count);
+      g_gpu.DMARead(dest_pointer, word_count);
       break;
 
     case Channel::CDROM:

src/core/fullscreen_ui.cpp

@@ -8,7 +8,9 @@
 #include "controller.h"
 #include "game_list.h"
 #include "gpu.h"
+#include "gpu_thread.h"
 #include "host.h"
+#include "imgui_overlays.h"
 #include "settings.h"
 #include "system.h"
 #include "system_private.h"
@@ -203,6 +205,7 @@ struct PostProcessingStageInfo
 //////////////////////////////////////////////////////////////////////////
 // Main
 //////////////////////////////////////////////////////////////////////////
+static void UpdateRunIdleState();
 static void PauseForMenuOpen(bool set_pause_menu_open);
 static bool AreAnyDialogsOpen();
 static void ClosePauseMenu();
@@ -602,12 +605,12 @@ bool FullscreenUI::Initialize()
   s_state.about_window_open = false;
   s_state.hotkey_list_cache = InputManager::GetHotkeyList();
 
-  if (!System::IsValid())
+  Host::RunOnCPUThread([]() { Host::OnFullscreenUIStartedOrStopped(true); });
+
+  if (!GPUThread::HasGPUBackend() && !GPUThread::IsGPUBackendRequested())
     SwitchToLanding();
 
-  if (!System::IsRunning())
+  UpdateRunIdleState();
-    Host::OnIdleStateChanged();
-
   ForceKeyNavEnabled();
   return true;
 }
@@ -629,64 +632,90 @@ bool FullscreenUI::AreAnyDialogsOpen()
           ImGuiFullscreen::IsFileSelectorOpen());
 }
 
-void FullscreenUI::CheckForConfigChanges(const Settings& old_settings)
+void FullscreenUI::CheckForConfigChanges(const GPUSettings& old_settings)
 {
-  if (!IsInitialized())
+  // NOTE: Called on GPU thread.
-    return;
+}
 
-  // If achievements got disabled, we might have the menu open...
+void FullscreenUI::UpdateRunIdleState()
-  // That means we're going to be reading achievement state.
+{
-  if (old_settings.achievements_enabled && !g_settings.achievements_enabled)
+  const bool new_run_idle = HasActiveWindow();
-  {
+  GPUThread::SetRunIdleReason(GPUThread::RunIdleReason::FullscreenUIActive, new_run_idle);
-    if (s_state.current_main_window == MainWindowType::Achievements ||
-        s_state.current_main_window == MainWindowType::Leaderboards)
-      ReturnToPreviousWindow();
-  }
 }
 
 void FullscreenUI::OnSystemStarted()
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
-  s_state.current_main_window = MainWindowType::None;
+  GPUThread::RunOnThread([]() {
-  QueueResetFocus(FocusResetType::ViewChanged);
+    if (!IsInitialized())
-}
+      return;
 
-void FullscreenUI::OnSystemPaused()
+    s_state.current_main_window = MainWindowType::None;
-{
+    QueueResetFocus(FocusResetType::ViewChanged);
-  // noop
+    UpdateRunIdleState();
+  });
 }
 
 void FullscreenUI::OnSystemResumed()
 {
-  // get rid of pause menu if we unpaused another way
+  // NOTE: Called on CPU thread.
-  if (s_state.current_main_window == MainWindowType::PauseMenu)
+  if (!IsInitialized())
-    ClosePauseMenu();
+    return;
+
+  GPUThread::RunOnThread([]() {
+    if (!IsInitialized())
+      return;
+
+    // get rid of pause menu if we unpaused another way
+    if (s_state.current_main_window == MainWindowType::PauseMenu)
+      ClosePauseMenu();
+
+    UpdateRunIdleState();
+  });
 }
 
 void FullscreenUI::OnSystemDestroyed()
 {
+  // NOTE: Called on CPU thread.
   if (!IsInitialized())
     return;
 
-  s_state.pause_menu_was_open = false;
+  GPUThread::RunOnThread([]() {
-  s_state.was_paused_on_quick_menu_open = false;
+    if (!IsInitialized())
-  s_state.current_pause_submenu = PauseSubMenu::None;
+      return;
-  SwitchToLanding();
+
+    s_state.pause_menu_was_open = false;
+    s_state.was_paused_on_quick_menu_open = false;
+    s_state.current_pause_submenu = PauseSubMenu::None;
+    SwitchToLanding();
+    UpdateRunIdleState();
+  });
 }
 
 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_state.current_game_subtitle = fmt::format("{0} - {1}", serial, Path::GetFileName(path));
+    subtitle = fmt::format("{0} - {1}", serial, Path::GetFileName(path));
   else
-    s_state.current_game_subtitle = {};
+    subtitle = {};
+
+  GPUThread::RunOnThread([subtitle = std::move(subtitle)]() mutable {
+    if (!IsInitialized())
+      return;
+
+    s_state.current_game_subtitle = std::move(subtitle);
+  });
 }
 
 void FullscreenUI::PauseForMenuOpen(bool set_pause_menu_open)
@@ -703,15 +732,18 @@ void FullscreenUI::OpenPauseMenu()
   if (!System::IsValid())
     return;
 
-  if (!Initialize() || s_state.current_main_window != MainWindowType::None)
+  GPUThread::RunOnThread([]() {
-    return;
+    if (!Initialize() || s_state.current_main_window != MainWindowType::None)
+      return;
 
-  PauseForMenuOpen(true);
+    PauseForMenuOpen(true);
-  s_state.current_main_window = MainWindowType::PauseMenu;
+    s_state.current_main_window = MainWindowType::PauseMenu;
-  s_state.current_pause_submenu = PauseSubMenu::None;
+    s_state.current_pause_submenu = PauseSubMenu::None;
-  QueueResetFocus(FocusResetType::ViewChanged);
+    QueueResetFocus(FocusResetType::ViewChanged);
-  ForceKeyNavEnabled();
+    ForceKeyNavEnabled();
-  FixStateIfPaused();
+    UpdateRunIdleState();
+    FixStateIfPaused();
+  });
 }
 
 void FullscreenUI::OpenCheatsMenu()
@@ -725,41 +757,39 @@ void FullscreenUI::OpenCheatsMenu()
   s_state.settings_page = SettingsPage::Cheats;
   PauseForMenuOpen(true);
   ForceKeyNavEnabled();
+  UpdateRunIdleState();
   FixStateIfPaused();
 }
 
 void FullscreenUI::FixStateIfPaused()
 {
-  if (!System::IsValid() || System::IsRunning())
+  if (!GPUThread::HasGPUBackend() || System::IsRunning())
     return;
 
   // 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;
 
-  if (System::GetState() == System::State::Paused && !s_state.was_paused_on_quick_menu_open)
+  const bool paused = System::IsPaused();
-    Host::RunOnCPUThread([]() { System::PauseSystem(false); });
+  GPUThread::RunOnThread([paused]() {
+    if (!IsInitialized())
+      return;
 
-  s_state.current_main_window = MainWindowType::None;
+    if (paused && !s_state.was_paused_on_quick_menu_open)
-  s_state.current_pause_submenu = PauseSubMenu::None;
+      Host::RunOnCPUThread([]() { System::PauseSystem(false); });
-  s_state.pause_menu_was_open = false;
+
-  QueueResetFocus(FocusResetType::ViewChanged);
+    s_state.current_main_window = MainWindowType::None;
-  FixStateIfPaused();
+    s_state.current_pause_submenu = PauseSubMenu::None;
+    s_state.pause_menu_was_open = false;
+    QueueResetFocus(FocusResetType::ViewChanged);
+    UpdateRunIdleState();
+    FixStateIfPaused();
+  });
 }
 
 void FullscreenUI::OpenPauseSubMenu(PauseSubMenu submenu)
@@ -790,14 +820,21 @@ void FullscreenUI::Shutdown()
   s_state.current_game_subtitle = {};
   DestroyResources();
   ImGuiFullscreen::Shutdown();
+  if (s_state.initialized)
+    Host::RunOnCPUThread([]() { Host::OnFullscreenUIStartedOrStopped(false); });
+
   s_state.initialized = false;
   s_state.tried_to_initialize = false;
+  UpdateRunIdleState();
 }
 
 void FullscreenUI::Render()
 {
   if (!s_state.initialized)
+  {
+    ImGuiFullscreen::RenderLoadingScreen();
     return;
+  }
 
   ImGuiFullscreen::UploadAsyncTextures();
 
@@ -856,6 +893,8 @@ void FullscreenUI::Render()
 
   ImGuiFullscreen::EndLayout();
 
+  ImGuiFullscreen::RenderLoadingScreen();
+
   if (s_state.settings_changed.load(std::memory_order_relaxed))
   {
     Host::CommitBaseSettingChanges();
@@ -889,7 +928,7 @@ void FullscreenUI::Render()
         }
       }
 
-      if (System::IsValid())
+      if (GPUThread::HasGPUBackend())
         Host::RunOnCPUThread([]() { System::ReloadGameSettings(false); });
     }
     s_state.game_settings_changed.store(false, std::memory_order_release);
@@ -908,7 +947,7 @@ void FullscreenUI::InvalidateCoverCache()
 
 void FullscreenUI::ReturnToPreviousWindow()
 {
-  if (System::IsValid() && s_state.pause_menu_was_open)
+  if (GPUThread::HasGPUBackend() && s_state.pause_menu_was_open)
   {
     s_state.current_main_window = MainWindowType::PauseMenu;
     QueueResetFocus(FocusResetType::ViewChanged);
@@ -922,7 +961,8 @@ void FullscreenUI::ReturnToPreviousWindow()
 void FullscreenUI::ReturnToMainWindow()
 {
   ClosePauseMenu();
-  s_state.current_main_window = System::IsValid() ? MainWindowType::None : MainWindowType::Landing;
+  s_state.current_main_window = GPUThread::HasGPUBackend() ? MainWindowType::None : MainWindowType::Landing;
+  UpdateRunIdleState();
   FixStateIfPaused();
 }
 
@@ -958,9 +998,14 @@ ImGuiFullscreen::FileSelectorFilters FullscreenUI::GetDiscImageFilters()
 
 void FullscreenUI::DoStartPath(std::string path, std::string state, std::optional<bool> fast_boot)
 {
-  if (System::IsValid())
+  if (GPUThread::HasGPUBackend())
     return;
 
+  // Switch to nothing, we'll get called back via OnSystemDestroyed() if startup fails.
+  s_state.current_main_window = MainWindowType::None;
+  QueueResetFocus(FocusResetType::ViewChanged);
+  UpdateRunIdleState();
+
   SystemBootParameters params;
   params.filename = std::move(path);
   params.save_state = std::move(state);
@@ -1155,108 +1200,116 @@ void FullscreenUI::DoChangeDiscFromFile()
 
 void FullscreenUI::DoChangeDisc()
 {
-  ImGuiFullscreen::ChoiceDialogOptions options;
+  Host::RunOnCPUThread([]() {
+    ImGuiFullscreen::ChoiceDialogOptions options;
 
-  if (System::HasMediaSubImages())
+    if (System::HasMediaSubImages())
-  {
-    const u32 current_index = System::GetMediaSubImageIndex();
-    const u32 count = System::GetMediaSubImageCount();
-    options.reserve(count + 1);
-    options.emplace_back(FSUI_STR("From File..."), false);
-
-    for (u32 i = 0; i < count; i++)
-      options.emplace_back(System::GetMediaSubImageTitle(i), i == current_index);
-
-    auto callback = [](s32 index, const std::string& title, bool checked) {
-      if (index == 0)
-      {
-        CloseChoiceDialog();
-        DoChangeDiscFromFile();
-        return;
-      }
-      else if (index > 0)
-      {
-        System::SwitchMediaSubImage(static_cast<u32>(index - 1));
-      }
-
-      CloseChoiceDialog();
-      ReturnToPreviousWindow();
-    };
-
-    OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
-                     std::move(callback));
-
-    return;
-  }
-
-  if (const GameDatabase::Entry* entry = System::GetGameDatabaseEntry(); entry && !entry->disc_set_serials.empty())
-  {
-    const auto lock = GameList::GetLock();
-    auto matches = GameList::GetMatchingEntriesForSerial(entry->disc_set_serials);
-    if (matches.size() > 1)
     {
-      options.reserve(matches.size() + 1);
+      const u32 current_index = System::GetMediaSubImageIndex();
+      const u32 count = System::GetMediaSubImageCount();
+      options.reserve(count + 1);
       options.emplace_back(FSUI_STR("From File..."), false);
 
-      std::vector<std::string> paths;
+      for (u32 i = 0; i < count; i++)
-      paths.reserve(matches.size());
+        options.emplace_back(System::GetMediaSubImageTitle(i), i == current_index);
 
-      const std::string& current_path = System::GetDiscPath();
+      GPUThread::RunOnThread([options = std::move(options)]() mutable {
-      for (auto& [title, glentry] : matches)
+        auto callback = [](s32 index, const std::string& title, bool checked) {
-      {
+          if (index == 0)
-        options.emplace_back(std::move(title), current_path == glentry->path);
+          {
-        paths.push_back(glentry->path);
+            CloseChoiceDialog();
-      }
+            DoChangeDiscFromFile();
+            return;
+          }
+          else if (index > 0)
+          {
+            System::SwitchMediaSubImage(static_cast<u32>(index - 1));
+          }
 
-      auto callback = [paths = std::move(paths)](s32 index, const std::string& title, bool checked) {
-        if (index == 0)
-        {
           CloseChoiceDialog();
-          DoChangeDiscFromFile();
+          ReturnToPreviousWindow();
-          return;
+        };
-        }
-        else if (index > 0)
-        {
-          System::InsertMedia(paths[index - 1].c_str());
-        }
 
-        CloseChoiceDialog();
+        OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
-        ReturnToMainWindow();
+                         std::move(callback));
-      };
+      });
-
-      OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
-                       std::move(callback));
 
       return;
     }
-  }
 
-  DoChangeDiscFromFile();
+    if (const GameDatabase::Entry* entry = System::GetGameDatabaseEntry(); entry && !entry->disc_set_serials.empty())
+    {
+      const auto lock = GameList::GetLock();
+      auto matches = GameList::GetMatchingEntriesForSerial(entry->disc_set_serials);
+      if (matches.size() > 1)
+      {
+        options.reserve(matches.size() + 1);
+        options.emplace_back(FSUI_STR("From File..."), false);
+
+        std::vector<std::string> paths;
+        paths.reserve(matches.size());
+
+        const std::string& current_path = System::GetDiscPath();
+        for (auto& [title, glentry] : matches)
+        {
+          options.emplace_back(std::move(title), current_path == glentry->path);
+          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)
+            {
+              CloseChoiceDialog();
+              DoChangeDiscFromFile();
+              return;
+            }
+            else if (index > 0)
+            {
+              System::InsertMedia(paths[index - 1].c_str());
+            }
+
+            CloseChoiceDialog();
+            ReturnToMainWindow();
+          };
+
+          OpenChoiceDialog(FSUI_ICONSTR(ICON_FA_COMPACT_DISC, "Select Disc Image"), true, std::move(options),
+                           std::move(callback));
+        });
+
+        return;
+      }
+    }
+
+    GPUThread::RunOnThread([]() { DoChangeDiscFromFile(); });
+  });
 }
 
 void FullscreenUI::DoToggleAnalogMode()
 {
   // hacky way to toggle analog mode
-  for (u32 i = 0; i < NUM_CONTROLLER_AND_CARD_PORTS; i++)
+  Host::RunOnCPUThread([]() {
-  {
+    for (u32 i = 0; i < NUM_CONTROLLER_AND_CARD_PORTS; i++)
-    Controller* ctrl = System::GetController(i);
-    if (!ctrl)
-      continue;
-
-    const Controller::ControllerInfo* cinfo = Controller::GetControllerInfo(ctrl->GetType());
-    if (!cinfo)
-      continue;
-
-    for (const Controller::ControllerBindingInfo& bi : cinfo->bindings)
     {
-      if (std::strcmp(bi.name, "Analog") == 0)
+      Controller* ctrl = System::GetController(i);
+      if (!ctrl)
+        continue;
+
+      const Controller::ControllerInfo* cinfo = Controller::GetControllerInfo(ctrl->GetType());
+      if (!cinfo)
+        continue;
+
+      for (const Controller::ControllerBindingInfo& bi : cinfo->bindings)
       {
-        ctrl->SetBindState(bi.bind_index, 1.0f);
+        if (std::strcmp(bi.name, "Analog") == 0)
-        ctrl->SetBindState(bi.bind_index, 0.0f);
+        {
-        break;
+          ctrl->SetBindState(bi.bind_index, 1.0f);
+          ctrl->SetBindState(bi.bind_index, 0.0f);
+          break;
+        }
       }
     }
-  }
+  });
 }
 
 void FullscreenUI::DoRequestExit()
@@ -2857,7 +2910,7 @@ void FullscreenUI::DrawSettingsWindow()
                                (LayoutScale(LAYOUT_MENU_BUTTON_Y_PADDING) * 2.0f) + LayoutScale(2.0f));
 
   const float bg_alpha =
-    System::IsValid() ? (s_state.settings_page == SettingsPage::PostProcessing ? 0.50f : 0.90f) : 1.0f;
+    GPUThread::HasGPUBackend() ? (s_state.settings_page == SettingsPage::PostProcessing ? 0.50f : 0.90f) : 1.0f;
 
   if (BeginFullscreenWindow(ImVec2(0.0f, 0.0f), heading_size, "settings_category",
                             ImVec4(UIStyle.PrimaryColor.x, UIStyle.PrimaryColor.y, UIStyle.PrimaryColor.z, bg_alpha)))
@@ -3802,12 +3855,9 @@ void FullscreenUI::DrawControllerSettingsPage()
                   &Settings::GetMultitapModeName, &Settings::GetMultitapModeDisplayName, MultitapMode::Count);
 
   // load mtap settings
-  MultitapMode mtap_mode = g_settings.multitap_mode;
+  const MultitapMode mtap_mode =
-  if (IsEditingGameSettings(bsi))
+    Settings::ParseMultitapModeName(bsi->GetTinyStringValue("ControllerPorts", "MultitapMode", "").c_str())
-  {
+      .value_or(Settings::DEFAULT_MULTITAP_MODE);
-    mtap_mode = Settings::ParseMultitapModeName(bsi->GetTinyStringValue("ControllerPorts", "MultitapMode", "").c_str())
-                  .value_or(g_settings.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)}};
@@ -4674,7 +4724,7 @@ void FullscreenUI::DrawPostProcessingSettingsPage()
                  FSUI_CSTR("Reloads the shaders from disk, applying any changes."),
                  bsi->GetBoolValue("PostProcessing", "Enabled", false)))
   {
-    if (System::IsValid() && PostProcessing::ReloadShaders())
+    if (GPUThread::HasGPUBackend() && PostProcessing::ReloadShaders())
       ShowToast(std::string(), FSUI_STR("Post-processing shaders reloaded."));
   }
 
@@ -5110,7 +5160,7 @@ void FullscreenUI::DrawAchievementsSettingsPage()
                   "cheats, and slowdown functions."),
         "Cheevos", "ChallengeMode", false, enabled))
   {
-    if (System::IsValid() && bsi->GetBoolValue("Cheevos", "ChallengeMode", false))
+    if (GPUThread::HasGPUBackend() && bsi->GetBoolValue("Cheevos", "ChallengeMode", false))
       ShowToast(std::string(), FSUI_STR("Hardcore mode will be enabled on next game restart."));
   }
   DrawToggleSetting(
@@ -5272,28 +5322,30 @@ void FullscreenUI::DrawAchievementsLoginWindow()
                                                     0, 0, 0);
 
       Host::RunOnCPUThread([username = std::string(username), password = std::string(password)]() {
-        ImGuiFullscreen::CloseBackgroundProgressDialog(LOGIN_PROGRESS_NAME);
-
         Error error;
-        if (Achievements::Login(username.c_str(), password.c_str(), &error))
+        const bool result = Achievements::Login(username.c_str(), password.c_str(), &error);
-        {
+        GPUThread::RunOnThread([result, error = std::move(error)]() {
-          // TODO-GPU-THREAD: Synchronize access to s_achievements_login_window_open.
+          ImGuiFullscreen::CloseBackgroundProgressDialog(LOGIN_PROGRESS_NAME);
-          actually_close_popup();
-          return;
-        }
 
-        // keep popup open on failure
+          if (result)
-        // because of the whole popup stack thing, we need to hide the dialog while this popup is visible
+          {
-        s_state.achievements_login_window_open = false;
+            actually_close_popup();
-        ImGuiFullscreen::OpenInfoMessageDialog(
+            return;
-          FSUI_STR("Login Error"),
+          }
-          fmt::format(FSUI_FSTR("Login Failed.\nError: {}\nPlease check your username and password, and try again."),
+
-                      error.GetDescription()),
+          // keep popup open on failure
-          []() {
+          // because of the whole popup stack thing, we need to hide the dialog while this popup is visible
-            s_state.achievements_login_window_open = true;
+          s_state.achievements_login_window_open = false;
-            QueueResetFocus(FocusResetType::PopupOpened);
+          ImGuiFullscreen::OpenInfoMessageDialog(
-          },
+            FSUI_STR("Login Error"),
-          FSUI_ICONSTR(ICON_FA_TIMES, "Close"));
+            fmt::format(FSUI_FSTR("Login Failed.\nError: {}\nPlease check your username and password, and try again."),
+                        error.GetDescription()),
+            []() {
+              s_state.achievements_login_window_open = true;
+              QueueResetFocus(FocusResetType::PopupOpened);
+            },
+            FSUI_ICONSTR(ICON_FA_TIMES, "Close"));
+        });
       });
     }
 
@@ -5804,7 +5856,7 @@ void FullscreenUI::DrawPauseMenu()
       case PauseSubMenu::None:
       {
         // NOTE: Menu close must come first, because otherwise VM destruction options will race.
-        const bool has_game = System::IsValid() && !System::GetGameSerial().empty();
+        const bool has_game = GPUThread::HasGPUBackend() && !System::GetGameSerial().empty();
 
         if (DefaultActiveButton(FSUI_ICONSTR(ICON_FA_PLAY, "Resume Game"), false) || WantsToCloseMenu())
           ClosePauseMenu();
@@ -6667,7 +6719,7 @@ void FullscreenUI::DrawGameListWindow()
     ImVec2(io.DisplaySize.x, LayoutScale(LAYOUT_MENU_BUTTON_HEIGHT_NO_SUMMARY) +
                                (LayoutScale(LAYOUT_MENU_BUTTON_Y_PADDING) * 2.0f) + LayoutScale(2.0f));
 
-  const float bg_alpha = System::IsValid() ? 0.90f : 1.0f;
+  const float bg_alpha = GPUThread::HasGPUBackend() ? 0.90f : 1.0f;
 
   if (BeginFullscreenWindow(ImVec2(0.0f, 0.0f), heading_size, "gamelist_view",
                             MulAlpha(UIStyle.PrimaryColor, bg_alpha)))
@@ -7217,7 +7269,7 @@ void FullscreenUI::DrawGameListSettingsWindow()
     ImVec2(io.DisplaySize.x, LayoutScale(LAYOUT_MENU_BUTTON_HEIGHT_NO_SUMMARY) +
                                (LayoutScale(LAYOUT_MENU_BUTTON_Y_PADDING) * 2.0f) + LayoutScale(2.0f));
 
-  const float bg_alpha = System::IsValid() ? 0.90f : 1.0f;
+  const float bg_alpha = GPUThread::HasGPUBackend() ? 0.90f : 1.0f;
 
   if (BeginFullscreenWindow(ImVec2(0.0f, 0.0f), heading_size, "gamelist_view",
                             MulAlpha(UIStyle.PrimaryColor, bg_alpha)))
@@ -7533,74 +7585,223 @@ 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;
   }
-
+  else if (!Achievements::HasAchievements())
-  if (!System::IsValid() || !Initialize())
-    return;
-
-  if (!Achievements::HasAchievements() || !Achievements::PrepareAchievementsWindow())
   {
     ShowToast(std::string(), FSUI_STR("This game has no achievements."));
     return;
   }
 
-  if (s_state.current_main_window != MainWindowType::PauseMenu)
+  GPUThread::RunOnThread([]() {
-  {
+    if (!Initialize() || !Achievements::PrepareAchievementsWindow())
-    PauseForMenuOpen(false);
+      return;
-    ForceKeyNavEnabled();
-  }
 
-  s_state.current_main_window = MainWindowType::Achievements;
+    if (s_state.current_main_window != MainWindowType::PauseMenu)
-  QueueResetFocus(FocusResetType::ViewChanged);
+    {
-  FixStateIfPaused();
+      PauseForMenuOpen(false);
-}
+      ForceKeyNavEnabled();
+    }
 
-bool FullscreenUI::IsAchievementsWindowOpen()
+    s_state.current_main_window = MainWindowType::Achievements;
-{
+    QueueResetFocus(FocusResetType::ViewChanged);
-  return (s_state.current_main_window == MainWindowType::Achievements);
+    UpdateRunIdleState();
+    FixStateIfPaused();
+  });
 }
 
 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;
   }
-
+  else if (!Achievements::HasLeaderboards())
-  if (!System::IsValid() || !Initialize())
-    return;
-
-  if (!Achievements::HasLeaderboards() || !Achievements::PrepareLeaderboardsWindow())
   {
     ShowToast(std::string(), FSUI_STR("This game has no leaderboards."));
     return;
   }
 
-  if (s_state.current_main_window != MainWindowType::PauseMenu)
+  GPUThread::RunOnThread([]() {
-  {
+    if (!Initialize() || !Achievements::PrepareLeaderboardsWindow())
-    PauseForMenuOpen(false);
+      return;
-    ForceKeyNavEnabled();
-  }
 
-  s_state.current_main_window = MainWindowType::Leaderboards;
+    if (s_state.current_main_window != MainWindowType::PauseMenu)
-  QueueResetFocus(FocusResetType::ViewChanged);
+    {
-  FixStateIfPaused();
+      PauseForMenuOpen(false);
-}
+      ForceKeyNavEnabled();
+    }
 
-bool FullscreenUI::IsLeaderboardsWindowOpen()
+    s_state.current_main_window = MainWindowType::Leaderboards;
-{
+    QueueResetFocus(FocusResetType::ViewChanged);
-  return (s_state.current_main_window == MainWindowType::Leaderboards);
+    UpdateRunIdleState();
+    FixStateIfPaused();
+  });
 }
 
 #endif // __ANDROID__
 
+LoadingScreenProgressCallback::LoadingScreenProgressCallback()
+  : ProgressCallback(), m_open_time(Timer::GetCurrentValue()), m_on_gpu_thread(GPUThread::IsOnThread())
+{
+}
+
+LoadingScreenProgressCallback::~LoadingScreenProgressCallback()
+{
+  // Did we activate?
+  if (m_last_progress_percent < 0)
+    return;
+
+  if (!m_on_gpu_thread)
+  {
+    GPUThread::RunOnThread([]() {
+      ImGuiFullscreen::CloseLoadingScreen();
+      Assert(GPUThread::GetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive));
+      GPUThread::SetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive, false);
+    });
+  }
+  else
+  {
+    // since this was pushing frames, we need to restore the context
+    GPUThread::Internal::RestoreContextAfterPresent();
+  }
+}
+
+void LoadingScreenProgressCallback::PushState()
+{
+  ProgressCallback::PushState();
+}
+
+void LoadingScreenProgressCallback::PopState()
+{
+  ProgressCallback::PopState();
+  Redraw(true);
+}
+
+void LoadingScreenProgressCallback::SetCancellable(bool cancellable)
+{
+  ProgressCallback::SetCancellable(cancellable);
+  Redraw(true);
+}
+
+void LoadingScreenProgressCallback::SetTitle(const std::string_view title)
+{
+  // todo?
+}
+
+void LoadingScreenProgressCallback::SetStatusText(const std::string_view text)
+{
+  ProgressCallback::SetStatusText(text);
+  Redraw(true);
+}
+
+void LoadingScreenProgressCallback::SetProgressRange(u32 range)
+{
+  u32 last_range = m_progress_range;
+
+  ProgressCallback::SetProgressRange(range);
+
+  if (m_progress_range != last_range)
+    Redraw(false);
+}
+
+void LoadingScreenProgressCallback::SetProgressValue(u32 value)
+{
+  u32 lastValue = m_progress_value;
+
+  ProgressCallback::SetProgressValue(value);
+
+  if (m_progress_value != lastValue)
+    Redraw(false);
+}
+
+void LoadingScreenProgressCallback::Redraw(bool force)
+{
+  if (m_last_progress_percent < 0 &&
+      Timer::ConvertValueToSeconds(Timer::GetCurrentValue() - m_open_time) < m_open_delay)
+  {
+    return;
+  }
+
+  const int percent =
+    static_cast<int>((static_cast<float>(m_progress_value) / static_cast<float>(m_progress_range)) * 100.0f);
+  DebugAssert(percent >= 0);
+  if (percent == m_last_progress_percent && !force)
+    return;
+
+  // activation?
+  if (m_last_progress_percent < 0 && !m_on_gpu_thread)
+  {
+    GPUThread::RunOnThread([]() {
+      Assert(!GPUThread::GetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive));
+      GPUThread::SetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive, true);
+    });
+  }
+
+  m_last_progress_percent = percent;
+  if (m_on_gpu_thread)
+  {
+    ImGuiFullscreen::RenderLoadingScreen(ImGuiManager::LOGO_IMAGE_NAME, m_status_text, 0,
+                                         static_cast<s32>(m_progress_range), static_cast<s32>(m_progress_value));
+  }
+  else
+  {
+    GPUThread::RunOnThread([status_text = SmallString(std::string_view(m_status_text)),
+                            range = static_cast<s32>(m_progress_range), value = static_cast<s32>(m_progress_value)]() {
+      ImGuiFullscreen::OpenOrUpdateLoadingScreen(ImGuiManager::LOGO_IMAGE_NAME, status_text, 0, range, value);
+    });
+  }
+}
+
+void LoadingScreenProgressCallback::ModalError(const std::string_view message)
+{
+  ERROR_LOG(message);
+  Host::ReportErrorAsync("Error", message);
+}
+
+bool LoadingScreenProgressCallback::ModalConfirmation(const std::string_view message)
+{
+  INFO_LOG(message);
+  return Host::ConfirmMessage("Confirm", message);
+}
+
+void FullscreenUI::OpenLoadingScreen(std::string_view image, std::string_view message, s32 progress_min /*= -1*/,
+                                     s32 progress_max /*= -1*/, s32 progress_value /*= -1*/)
+{
+  Assert(GPUThread::IsOnThread());
+  Assert(!GPUThread::GetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive));
+  GPUThread::SetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive, true);
+  ImGuiFullscreen::OpenOrUpdateLoadingScreen(image, message, progress_min, progress_max, progress_value);
+}
+
+void FullscreenUI::UpdateLoadingScreen(std::string_view image, std::string_view message, s32 progress_min /*= -1*/,
+                                       s32 progress_max /*= -1*/, s32 progress_value /*= -1*/)
+{
+  Assert(GPUThread::IsOnThread());
+  Assert(GPUThread::GetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive));
+  ImGuiFullscreen::OpenOrUpdateLoadingScreen(image, message, progress_min, progress_max, progress_value);
+}
+
+void FullscreenUI::CloseLoadingScreen()
+{
+  Assert(GPUThread::IsOnThread());
+  Assert(GPUThread::GetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive));
+  ImGuiFullscreen::CloseLoadingScreen();
+  GPUThread::SetRunIdleReason(GPUThread::RunIdleReason::LoadingScreenActive, false);
+}
+
 /////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Translation String Area
 // To avoid having to type T_RANSLATE("FullscreenUI", ...) everywhere, we use the shorter macros at the top

src/core/fullscreen_ui.h

@@ -13,15 +13,14 @@
 
 class SmallStringBase;
 
-struct Settings;
+struct GPUSettings;
 
 namespace FullscreenUI {
 bool Initialize();
 bool IsInitialized();
 bool HasActiveWindow();
-void CheckForConfigChanges(const Settings& old_settings);
+void CheckForConfigChanges(const GPUSettings& old_settings);
 void OnSystemStarted();
-void OnSystemPaused();
 void OnSystemResumed();
 void OnSystemDestroyed();
 void OnRunningGameChanged();
@@ -30,9 +29,7 @@ void OnRunningGameChanged();
 void OpenPauseMenu();
 void OpenCheatsMenu();
 void OpenAchievementsWindow();
-bool IsAchievementsWindowOpen();
 void OpenLeaderboardsWindow();
-bool IsLeaderboardsWindowOpen();
 void ReturnToMainWindow();
 void ReturnToPreviousWindow();
 void SetStandardSelectionFooterText(bool back_instead_of_cancel);
@@ -43,13 +40,51 @@ void Render();
 void InvalidateCoverCache();
 void TimeToPrintableString(SmallStringBase* str, time_t t);
 
+void OpenLoadingScreen(std::string_view image, std::string_view message, s32 progress_min = -1, s32 progress_max = -1,
+                       s32 progress_value = -1);
+void UpdateLoadingScreen(std::string_view image, std::string_view message, s32 progress_min = -1, s32 progress_max = -1,
+                         s32 progress_value = -1);
+void CloseLoadingScreen();
+
 } // namespace FullscreenUI
 
+class LoadingScreenProgressCallback final : public ProgressCallback
+{
+public:
+  LoadingScreenProgressCallback();
+  ~LoadingScreenProgressCallback() override;
+
+  ALWAYS_INLINE void SetOpenDelay(float delay) { m_open_delay = delay; }
+
+  void PushState() override;
+  void PopState() override;
+
+  void SetCancellable(bool cancellable) override;
+  void SetTitle(const std::string_view title) override;
+  void SetStatusText(const std::string_view text) override;
+  void SetProgressRange(u32 range) override;
+  void SetProgressValue(u32 value) override;
+
+  void ModalError(const std::string_view message) override;
+  bool ModalConfirmation(const std::string_view message) override;
+
+private:
+  void Redraw(bool force);
+
+  u64 m_open_time = 0;
+  float m_open_delay = 1.0f;
+  s32 m_last_progress_percent = -1;
+  bool m_on_gpu_thread = false;
+};
+
 // Host UI triggers from Big Picture mode.
 namespace Host {
 
 #ifndef __ANDROID__
 
+/// Called whenever fullscreen UI starts/stops.
+void OnFullscreenUIStartedOrStopped(bool started);
+
 /// Requests shut down and exit of the hosting application. This may not actually exit,
 /// if the user cancels the shutdown confirmation.
 void RequestExitApplication(bool allow_confirm);

src/core/gpu.h

@@ -39,13 +39,18 @@ enum class PacketType : u8;
 class Recorder;
 class Player;
 } // namespace GPUDump
+
+class GPUBackend;
 struct Settings;
 
-namespace Threading {
+namespace System {
-class Thread;
+struct MemorySaveState;
 }
 
-class GPU
+struct GPUBackendCommand;
+struct GPUBackendDrawCommand;
+
+class GPU final
 {
 public:
   enum class BlitterState : u8
@@ -62,7 +67,6 @@ public:
     DOT_TIMER_INDEX = 0,
     HBLANK_TIMER_INDEX = 1,
     MAX_RESOLUTION_SCALE = 32,
-    DEINTERLACE_BUFFER_COUNT = 4,
     DRAWING_AREA_COORD_MASK = 1023,
   };
 
@@ -80,33 +84,24 @@ public:
     NTSC_HORIZONTAL_ACTIVE_END = 3288,
     NTSC_VERTICAL_ACTIVE_START = 16,
     NTSC_VERTICAL_ACTIVE_END = 256,
-    PAL_HORIZONTAL_ACTIVE_START = 487,
+    PAL_HORIZONTAL_ACTIVE_START = 488,
-    PAL_HORIZONTAL_ACTIVE_END = 3282,
+    PAL_HORIZONTAL_ACTIVE_END = 3300,
     PAL_VERTICAL_ACTIVE_START = 20,
     PAL_VERTICAL_ACTIVE_END = 308,
   };
 
   // Base class constructor.
   GPU();
-  virtual ~GPU();
+  ~GPU();
 
-  virtual const Threading::Thread* GetSWThread() const = 0;
+  void Initialize();
-  virtual bool IsHardwareRenderer() const = 0;
+  void Shutdown();
-
+  void Reset(bool clear_vram);
-  virtual bool Initialize(Error* error);
+  bool DoState(StateWrapper& sw, bool update_display);
-  virtual void Reset(bool clear_vram);
+  void DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display);
-  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();
 
   // Render statistics debug window.
   void DrawDebugStateWindow(float scale);
-  void GetStatsString(SmallStringBase& str);
-  void GetMemoryStatsString(SmallStringBase& str);
-  void ResetStatistics();
-  void UpdateStatistics(u32 frame_count);
 
   void CPUClockChanged();
 
@@ -170,31 +165,26 @@ public:
   void SynchronizeCRTC();
 
   /// Recompile shaders/recreate framebuffers when needed.
-  virtual void UpdateSettings(const Settings& old_settings);
+  void UpdateSettings(const Settings& old_settings);
-
-  /// Returns the current resolution scale.
-  virtual u32 GetResolutionScale() const;
-
-  /// Updates the resolution scale when it's set to automatic.
-  virtual void UpdateResolutionScale();
 
   /// Returns the full display resolution of the GPU, including padding.
   std::tuple<u32, u32> GetFullDisplayResolution() const;
 
+  /// Computes clamped drawing area.
+  static GSVector4i GetClampedDrawingArea(const GPUDrawingArea& drawing_area);
+
   float ComputeHorizontalFrequency() const;
   float ComputeVerticalFrequency() const;
   float ComputeDisplayAspectRatio() const;
   float ComputeSourceAspectRatio() const;
+  float ComputePixelAspectRatio() const;
 
   /// Computes aspect ratio correction, i.e. the scale to apply to the source aspect ratio to preserve
   /// the original pixel aspect ratio regardless of how much cropping has been applied.
   float ComputeAspectRatioCorrection() const;
 
   /// Applies the pixel aspect ratio to a given size, preserving the larger dimension.
-  void ApplyPixelAspectRatioToSize(float* width, float* height) const;
+  static void ApplyPixelAspectRatioToSize(float par, float* width, float* height);
-
-  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,
@@ -226,38 +216,17 @@ public:
   // Dumps raw VRAM to a file.
   bool DumpVRAMToFile(const char* filename);
 
-  // Ensures all buffered vertices are drawn.
+  // Queues the current frame for presentation. Should only be used with runahead.
-  virtual void FlushRender() = 0;
+  void QueuePresentCurrentFrame();
 
   /// Helper function for computing the draw rectangle in a larger window.
-  void CalculateDrawRect(s32 window_width, s32 window_height, bool apply_rotation, bool apply_aspect_ratio,
+  static void CalculateDrawRect(u32 window_width, u32 window_height, u32 crtc_display_width, u32 crtc_display_height,
-                         GSVector4i* display_rect, GSVector4i* draw_rect) const;
+                                s32 display_origin_left, s32 display_origin_top, u32 display_vram_width,
+                                u32 display_vram_height, DisplayRotation rotation, DisplayAlignment alignment,
+                                float pixel_aspect_ratio, bool stretch_vertically, bool integer_scale,
+                                GSVector4i* display_rect, GSVector4i* draw_rect);
 
-  /// Helper function for computing screenshot bounds.
+private:
-  void CalculateScreenshotSize(DisplayScreenshotMode mode, u32* width, u32* height, GSVector4i* display_rect,
-                               GSVector4i* draw_rect) const;
-
-  /// Helper function to save current display texture to PNG.
-  bool WriteDisplayTextureToFile(std::string path);
-
-  /// Renders the display, optionally with postprocessing to the specified image.
-  bool RenderScreenshotToBuffer(u32 width, u32 height, const GSVector4i display_rect, const GSVector4i draw_rect,
-                                bool postfx, Image* out_image);
-
-  /// Helper function to save screenshot to PNG.
-  bool RenderScreenshotToFile(std::string path, DisplayScreenshotMode mode, u8 quality, bool compress_on_thread,
-                              bool show_osd_message);
-
-  /// Draws the current display texture, with any post-processing.
-  GPUDevice::PresentResult PresentDisplay();
-
-  /// Sends the current frame to media capture.
-  bool SendDisplayToMediaCapture(MediaCapture* cap);
-
-  /// Reads the CLUT from the specified coordinates, accounting for wrap-around.
-  static void ReadCLUT(u16* dest, GPUTexturePaletteReg reg, bool clut_is_8bit);
-
-protected:
   TickCount CRTCTicksToSystemTicks(TickCount crtc_ticks, TickCount fractional_ticks) const;
   TickCount SystemTicksToCRTCTicks(TickCount sysclk_ticks, TickCount* fractional_ticks) const;
 
@@ -268,16 +237,6 @@ 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);
 
@@ -297,10 +256,10 @@ protected:
   void UpdateGPUIdle();
 
   /// 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();
@@ -320,7 +279,7 @@ protected:
   /// Returns the number of vertices in the buffered poly-line.
   ALWAYS_INLINE u32 GetPolyLineVertexCount() const
   {
-    return (static_cast<u32>(m_blit_buffer.size()) + BoolToUInt32(m_render_command.shading_enable)) >>
+    return (static_cast<u32>(m_polyline_buffer.size()) + BoolToUInt32(m_render_command.shading_enable)) >>
            BoolToUInt8(m_render_command.shading_enable);
   }
 
@@ -335,16 +294,13 @@ protected:
   void InvalidateCLUT();
   bool IsCLUTValid() const;
 
-  // Rendering in the backend
+  void ReadVRAM(u16 x, u16 y, u16 width, u16 height);
-  virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height) = 0;
+  void UpdateVRAM(u16 x, u16 y, u16 width, u16 height, const void* data, bool set_mask, bool check_mask);
-  virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) = 0;
+  void UpdateDisplay(bool submit_frame);
-  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) = 0;
+
-  virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) = 0;
+  void PrepareForDraw();
-  virtual void DispatchRenderCommand() = 0;
+  void FinishPolyline();
-  virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
+  void FillDrawCommand(GPUBackendDrawCommand* RESTRICT cmd, GPURenderCommand rc) const;
-  virtual void UpdateDisplay() = 0;
-  virtual void DrawRendererStats();
-  virtual void OnBufferSwapped();
 
   ALWAYS_INLINE_RELEASE void AddDrawTriangleTicks(GSVector2i v1, GSVector2i v2, GSVector2i v3, bool shaded,
                                                   bool textured, bool semitransparent)
@@ -441,14 +397,10 @@ 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;
-
-    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; }
   } m_draw_mode = {};
 
   GPUDrawingArea m_drawing_area = {};
@@ -568,79 +520,21 @@ protected:
     u16 row;
   } m_vram_transfer = {};
 
+  std::unique_ptr<GPUDump::Recorder> m_gpu_dump;
+
   HeapFIFOQueue<u64, MAX_FIFO_SIZE> m_fifo;
-  std::vector<u32> m_blit_buffer;
+  TickCount m_max_run_ahead = 128;
+  u32 m_fifo_size = 128;
   u32 m_blit_remaining_words;
   GPURenderCommand m_render_command{};
-
+  std::vector<u32> m_blit_buffer;
-  std::unique_ptr<GPUDump::Recorder> m_gpu_dump;
+  std::vector<u64> m_polyline_buffer;
 
   ALWAYS_INLINE u32 FifoPop() { return Truncate32(m_fifo.Pop()); }
   ALWAYS_INLINE u32 FifoPeek() { return Truncate32(m_fifo.Peek()); }
   ALWAYS_INLINE u32 FifoPeek(u32 i) { return Truncate32(m_fifo.Peek(i)); }
 
-  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);
-
-  GPUDevice::PresentResult RenderDisplay(GPUTexture* target, const GSVector4i display_rect, 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, Error* error);
-
   using GP0CommandHandler = bool (GPU::*)();
   using GP0CommandHandlerTable = std::array<GP0CommandHandler, 256>;
   static GP0CommandHandlerTable GenerateGP0CommandHandlerTable();
@@ -668,6 +562,6 @@ private:
   static const GP0CommandHandlerTable s_GP0_command_handler_table;
 };
 
-extern std::unique_ptr<GPU> g_gpu;
+extern GPU g_gpu;
 extern u16 g_vram[VRAM_SIZE / sizeof(u16)];
 extern u16 g_gpu_clut[GPU_CLUT_SIZE];

src/core/gpu_backend.h

@@ -3,7 +3,9 @@
 
 #pragma once
 
-#include "gpu_types.h"
+#include "gpu_thread_commands.h"
+
+#include "util/gpu_device.h"
 
 #include "common/heap_array.h"
 #include "common/threading.h"
@@ -12,85 +14,195 @@
 #include <condition_variable>
 #include <memory>
 #include <mutex>
+#include <tuple>
 
-#ifdef _MSC_VER
+class Error;
-#pragma warning(push)
+class SmallStringBase;
-#pragma warning(disable : 4324) // warning C4324: 'GPUBackend': structure was padded due to alignment specifier
+
-#endif
+class GPUFramebuffer;
+class GPUPipeline;
+
+struct GPUSettings;
+class StateWrapper;
+
+namespace System {
+struct MemorySaveState;
+}
+
+// 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 GPUThreadCommand* NewClearDisplayCommand();
+  static GPUBackendUpdateDisplayCommand* NewUpdateDisplayCommand();
+  static GPUBackendSubmitFrameCommand* NewSubmitFrameCommand();
+  static GPUThreadCommand* NewClearCacheCommand();
+  static GPUThreadCommand* NewBufferSwappedCommand();
+  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 GPUBackendDrawPreciseLineCommand* NewDrawPreciseLineCommand(u32 num_vertices);
+  static void PushCommand(GPUThreadCommand* cmd);
+  static void PushCommandAndWakeThread(GPUThreadCommand* cmd);
+  static void PushCommandAndSync(GPUThreadCommand* cmd, bool spin);
+  static void SyncGPUThread(bool spin);
+
+  static bool IsUsingHardwareBackend();
+
+  static std::unique_ptr<GPUBackend> CreateHardwareBackend();
+  static std::unique_ptr<GPUBackend> CreateSoftwareBackend();
+
+  static bool RenderScreenshotToBuffer(u32 width, u32 height, bool postfx, Image* out_image);
+  static void RenderScreenshotToFile(const std::string_view path, DisplayScreenshotMode mode, u8 quality,
+                                     bool compress_on_thread, bool show_osd_message);
+
+  static bool BeginQueueFrame();
+  static void WaitForOneQueuedFrame();
+  static u32 GetQueuedFrameCount();
+
+  static bool AllocateMemorySaveStates(std::span<System::MemorySaveState> states, Error* error);
+
 public:
   GPUBackend();
   virtual ~GPUBackend();
 
-  ALWAYS_INLINE const Threading::Thread* GetThread() const { return m_use_gpu_thread ? &m_gpu_thread : nullptr; }
+  virtual bool Initialize(bool upload_vram, Error* error);
-  ALWAYS_INLINE bool IsUsingThread() const { return m_use_gpu_thread; }
 
-  virtual bool Initialize(bool use_thread);
+  virtual void UpdateSettings(const GPUSettings& old_settings);
-  virtual void Reset();
-  virtual void Shutdown();
 
-  void SetThreadEnabled(bool use_thread);
+  /// Returns the current resolution scale.
+  virtual u32 GetResolutionScale() const = 0;
 
-  GPUBackendFillVRAMCommand* NewFillVRAMCommand();
+  /// Updates the resolution scale when it's set to automatic.
-  GPUBackendUpdateVRAMCommand* NewUpdateVRAMCommand(u32 num_words);
+  virtual void UpdateResolutionScale() = 0;
-  GPUBackendCopyVRAMCommand* NewCopyVRAMCommand();
-  GPUBackendSetDrawingAreaCommand* NewSetDrawingAreaCommand();
-  GPUBackendUpdateCLUTCommand* NewUpdateCLUTCommand();
-  GPUBackendDrawPolygonCommand* NewDrawPolygonCommand(u32 num_vertices);
-  GPUBackendDrawRectangleCommand* NewDrawRectangleCommand();
-  GPUBackendDrawLineCommand* NewDrawLineCommand(u32 num_vertices);
 
-  void PushCommand(GPUBackendCommand* cmd);
+  // Graphics API state reset/restore - call when drawing the UI etc.
-  void Sync(bool allow_sleep);
+  // TODO: replace with "invalidate cached state"
+  virtual void RestoreDeviceContext() = 0;
 
-  /// Processes all pending GPU commands.
+  /// Main command handler for GPU thread.
-  void RunGPULoop();
+  void HandleCommand(const GPUThreadCommand* cmd);
+
+  /// Draws the current display texture, with any post-processing.
+  GPUDevice::PresentResult PresentDisplay();
+
+  /// Helper function to save current display texture to PNG. Used for regtest.
+  bool WriteDisplayTextureToFile(std::string filename);
+
+  /// Helper function for computing screenshot bounds.
+  void CalculateScreenshotSize(DisplayScreenshotMode mode, u32* width, u32* height, GSVector4i* display_rect,
+                               GSVector4i* draw_rect) const;
+
+  void GetStatsString(SmallStringBase& str) const;
+  void GetMemoryStatsString(SmallStringBase& str) const;
+
+  void ResetStatistics();
+  void UpdateStatistics(u32 frame_count);
 
 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 CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
-                        GPUBackendCommandParameters params) = 0;
-  virtual void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) = 0;
-  virtual void DrawRectangle(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;
-
-  void HandleCommand(const GPUBackendCommand* cmd);
-
-  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;
-
-  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;
-
   enum : u32
   {
-    COMMAND_QUEUE_SIZE = 4 * 1024 * 1024,
+    DEINTERLACE_BUFFER_COUNT = 4,
-    THRESHOLD_TO_WAKE_GPU = 256
   };
 
-  FixedHeapArray<u8, COMMAND_QUEUE_SIZE> m_command_fifo_data;
+  virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height) = 0;
-  alignas(HOST_CACHE_LINE_SIZE) std::atomic<u32> m_command_fifo_read_ptr{0};
+  virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced_rendering,
-  alignas(HOST_CACHE_LINE_SIZE) std::atomic<u32> m_command_fifo_write_ptr{0};
+                        u8 interlaced_display_field) = 0;
+  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) = 0;
+  virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask,
+                        bool check_mask) = 0;
+
+  virtual void DrawPolygon(const GPUBackendDrawPolygonCommand* 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 DrawPreciseLine(const GPUBackendDrawPreciseLineCommand* cmd) = 0;
+
+  virtual void DrawingAreaChanged() = 0;
+  virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
+  virtual void ClearCache() = 0;
+  virtual void OnBufferSwapped() = 0;
+  virtual void ClearVRAM() = 0;
+
+  virtual void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) = 0;
+
+  virtual void LoadState(const GPUBackendLoadStateCommand* cmd) = 0;
+
+  virtual bool AllocateMemorySaveState(System::MemorySaveState& mss, Error* error) = 0;
+  virtual void DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss) = 0;
+
+  /// Ensures all pending draws are flushed to the host GPU.
+  virtual void FlushRender() = 0;
+
+  /// Helper function for computing the draw rectangle in a larger window.
+  void CalculateDrawRect(s32 window_width, s32 window_height, bool apply_rotation, bool apply_aspect_ratio,
+                         GSVector4i* display_rect, GSVector4i* draw_rect) const;
+
+  /// Renders the display, optionally with postprocessing to the specified image.
+  bool RenderScreenshotToBuffer(u32 width, u32 height, const GSVector4i display_rect, const GSVector4i draw_rect,
+                                bool postfx, Image* out_image);
+
+  bool CompileDisplayPipelines(bool display, bool deinterlace, bool chroma_smoothing, Error* error);
+
+  void HandleUpdateDisplayCommand(const GPUBackendUpdateDisplayCommand* cmd);
+  void HandleSubmitFrameCommand(const GPUBackendFramePresentationParameters* cmd);
+
+  void ClearDisplay();
+  void ClearDisplayTexture();
+  void SetDisplayTexture(GPUTexture* texture, GPUTexture* depth_buffer, s32 view_x, s32 view_y, s32 view_width,
+                         s32 view_height);
+
+  GPUDevice::PresentResult RenderDisplay(GPUTexture* target, const GSVector4i display_rect, const GSVector4i draw_rect,
+                                         bool postfx);
+
+  /// Sends the current frame to media capture.
+  void SendDisplayToMediaCapture(MediaCapture* cap);
+
+  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_pixel_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;
 };
 
-#ifdef _MSC_VER
+namespace Host {
-#pragma warning(pop)
+
-#endif
+/// Called at the end of the frame, before presentation.
+void FrameDoneOnGPUThread(GPUBackend* gpu_backend, u32 frame_number);
+
+} // namespace Host

src/core/gpu_commands.cpp

@@ -1,13 +1,17 @@
 // SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: CC-BY-NC-ND-4.0
 
+#include "cpu_pgxp.h"
 #include "gpu.h"
+#include "gpu_backend.h"
 #include "gpu_dump.h"
 #include "gpu_hw_texture_cache.h"
+#include "gpu_thread_commands.h"
 #include "interrupt_controller.h"
 #include "system.h"
 
 #include "common/assert.h"
+#include "common/gsvector_formatter.h"
 #include "common/log.h"
 #include "common/string_util.h"
 
@@ -69,7 +73,7 @@ void GPU::TryExecuteCommands()
       {
         const u32 words_per_vertex = m_render_command.shading_enable ? 2 : 1;
         u32 terminator_index =
-          m_render_command.shading_enable ? ((static_cast<u32>(m_blit_buffer.size()) & 1u) ^ 1u) : 0u;
+          m_render_command.shading_enable ? ((static_cast<u32>(m_polyline_buffer.size()) & 1u) ^ 1u) : 0u;
         for (; terminator_index < m_fifo.GetSize(); terminator_index += words_per_vertex)
         {
           // polyline must have at least two vertices, and the terminator is (word & 0xf000f000) == 0x50005000.
@@ -82,9 +86,9 @@ void GPU::TryExecuteCommands()
         const u32 words_to_copy = std::min(terminator_index, m_fifo.GetSize());
         if (words_to_copy > 0)
         {
-          m_blit_buffer.reserve(m_blit_buffer.size() + words_to_copy);
+          m_polyline_buffer.reserve(m_polyline_buffer.size() + words_to_copy);
           for (u32 i = 0; i < words_to_copy; i++)
-            m_blit_buffer.push_back(FifoPop());
+            m_polyline_buffer.push_back(m_fifo.Pop());
         }
 
         DEBUG_LOG("Added {} words to polyline", words_to_copy);
@@ -93,8 +97,8 @@ void GPU::TryExecuteCommands()
           // drop terminator
           m_fifo.RemoveOne();
           DEBUG_LOG("Drawing poly-line with {} vertices", GetPolyLineVertexCount());
-          DispatchRenderCommand();
+          FinishPolyline();
-          m_blit_buffer.clear();
+          m_polyline_buffer.clear();
           EndCommand();
           continue;
         }
@@ -200,8 +204,8 @@ 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();
@@ -248,8 +252,6 @@ 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;
@@ -270,8 +272,6 @@ bool GPU::HandleSetDrawingAreaBottomRightCommand()
   DEBUG_LOG("Set drawing area bottom-right: ({}, {})", right, 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;
@@ -291,8 +291,6 @@ bool GPU::HandleSetDrawingOffsetCommand()
   DEBUG_LOG("Set drawing offset ({}, {})", x, y);
   if (m_drawing_offset.x != x || m_drawing_offset.y != y)
   {
-    FlushRender();
-
     m_drawing_offset.x = x;
     m_drawing_offset.y = y;
   }
@@ -308,11 +306,7 @@ 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)
+  m_GPUSTAT.bits = (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));
 
@@ -321,6 +315,35 @@ 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;
+    GPUBackend::PushCommand(cmd);
+  }
+}
+
+void GPU::FillDrawCommand(GPUBackendDrawCommand* RESTRICT cmd, GPURenderCommand rc) const
+{
+  cmd->interlaced_rendering = IsInterlacedRenderingEnabled();
+  cmd->active_line_lsb = ConvertToBoolUnchecked(m_crtc_state.active_line_lsb);
+  cmd->check_mask_before_draw = m_GPUSTAT.check_mask_before_draw;
+  cmd->set_mask_while_drawing = m_GPUSTAT.set_mask_while_drawing;
+  cmd->texture_enable = rc.IsTexturingEnabled();
+  cmd->raw_texture_enable = rc.raw_texture_enable;
+  cmd->transparency_enable = rc.transparency_enable;
+  cmd->shading_enable = rc.shading_enable;
+  cmd->quad_polygon = rc.quad_polygon;
+  cmd->dither_enable = rc.IsDitheringEnabled() && m_draw_mode.mode_reg.dither_enable;
+
+  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)};
@@ -346,6 +369,7 @@ 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);
@@ -355,12 +379,233 @@ 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_settings.gpu_pgxp_enable)
+  {
+    GPUBackendDrawPrecisePolygonCommand* RESTRICT cmd = GPUBackend::NewDrawPrecisePolygonCommand(num_vertices);
+    FillDrawCommand(cmd, rc);
+    cmd->num_vertices = Truncate16(num_vertices);
+
+    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_settings.gpu_pgxp_texture_correction;
+    for (u32 i = 0; i < num_vertices; i++)
+    {
+      GPUBackendDrawPrecisePolygonCommand::Vertex* RESTRICT 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<false>(&v.x, GSVector2(GSVector2i::load<false>(&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<false>(&cmd->vertices[0].x);
+    const GSVector2 v1f = GSVector2::load<false>(&cmd->vertices[1].x);
+    const GSVector2 v2f = GSVector2::load<false>(&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<false>(&cmd->vertices[0].native_x),
+                           GSVector2i::load<false>(&cmd->vertices[1].native_x),
+                           GSVector2i::load<false>(&cmd->vertices[2].native_x), rc.shading_enable, rc.texture_enable,
+                           rc.transparency_enable);
+    }
+
+    // quads
+    if (rc.quad_polygon)
+    {
+      const GSVector2 v3f = GSVector2::load<false>(&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.
+        // NOTE: Culling this way results in subtle differences with UV clamping, since the fourth vertex is no
+        // longer considered in the range. This is mainly apparent when the UV gradient is zero. Seems like it
+        // generally looks better this way, so I'm keeping it.
+        cmd->size = GPUThreadCommand::AlignCommandSize(sizeof(GPUBackendDrawPrecisePolygonCommand) +
+                                                       3 * sizeof(GPUBackendDrawPrecisePolygonCommand::Vertex));
+        cmd->num_vertices = 3;
+      }
+      else
+      {
+        AddDrawTriangleTicks(GSVector2i::load<false>(&cmd->vertices[2].native_x),
+                             GSVector2i::load<false>(&cmd->vertices[1].native_x),
+                             GSVector2i::load<false>(&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->size = GPUThreadCommand::AlignCommandSize(sizeof(GPUBackendDrawPrecisePolygonCommand) +
+                                                         3 * sizeof(GPUBackendDrawPrecisePolygonCommand::Vertex));
+          cmd->num_vertices = 3;
+        }
+      }
+    }
+
+    GPUBackend::PushCommand(cmd);
+  }
+  else
+  {
+    GPUBackendDrawPolygonCommand* RESTRICT cmd = GPUBackend::NewDrawPolygonCommand(num_vertices);
+    FillDrawCommand(cmd, rc);
+    cmd->num_vertices = Truncate16(num_vertices);
+
+    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* RESTRICT 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<false>(&cmd->vertices[0].x);
+    const GSVector2i v1 = GSVector2i::load<false>(&cmd->vertices[1].x);
+    const GSVector2i v2 = GSVector2i::load<false>(&cmd->vertices[2].x);
+    const GSVector2i min_pos_12 = v1.min_s32(v2);
+    const GSVector2i max_pos_12 = v1.max_s32(v2);
+    const GSVector4i draw_rect_012 =
+      GSVector4i::xyxy(min_pos_12.min_s32(v0), max_pos_12.max_s32(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<false>(&cmd->vertices[3].x);
+      const GSVector4i draw_rect_123 = GSVector4i(min_pos_12.min_s32(v3))
+                                         .upl64(GSVector4i(max_pos_12.max_s32(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->size = GPUThreadCommand::AlignCommandSize(sizeof(GPUBackendDrawPolygonCommand) +
+                                                       3 * sizeof(GPUBackendDrawPolygonCommand::Vertex));
+        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->size = GPUThreadCommand::AlignCommandSize(sizeof(GPUBackendDrawPolygonCommand) +
+                                                         3 * sizeof(GPUBackendDrawPolygonCommand::Vertex));
+          cmd->num_vertices = 3;
+        }
+      }
+    }
+
+    GPUBackend::PushCommand(cmd);
+  }
+
   EndCommand();
   return true;
 }
@@ -389,12 +634,65 @@ 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;
 }
@@ -411,12 +709,103 @@ 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();
+
+  if (g_settings.gpu_pgxp_enable)
+  {
+    GPUBackendDrawPreciseLineCommand* RESTRICT cmd = GPUBackend::NewDrawPreciseLineCommand(2);
+    FillDrawCommand(cmd, rc);
+    cmd->palette.bits = 0;
+
+    bool valid_w = g_settings.gpu_pgxp_texture_correction;
+    for (u32 i = 0; i < 2; i++)
+    {
+      const u32 color = ((i != 0 && rc.shading_enable) ? FifoPop() : rc.bits) & UINT32_C(0x00FFFFFF);
+      const u64 maddr_and_pos = m_fifo.Pop();
+      const GPUVertexPosition vp{Truncate32(maddr_and_pos)};
+      GPUBackendDrawPreciseLineCommand::Vertex* RESTRICT vert = &cmd->vertices[i];
+      vert->native_x = m_drawing_offset.x + vp.x;
+      vert->native_y = m_drawing_offset.y + vp.y;
+      vert->color = color;
+
+      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);
+    }
+    if (!(cmd->valid_w = valid_w))
+    {
+      for (u32 i = 0; i < 2; i++)
+        cmd->vertices[i].w = 1.0f;
+    }
+
+    const GSVector2 v0f = GSVector2::load<false>(&cmd->vertices[0].x);
+    const GSVector2 v1f = GSVector2::load<false>(&cmd->vertices[1].x);
+    const GSVector4i rect =
+      GSVector4i(GSVector4(v0f.min(v1f)).upld(GSVector4(v0f.max(v1f)))).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);
+  }
+  else
+  {
+    GPUBackendDrawLineCommand* RESTRICT 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 GSVector2i v0 = GSVector2i::load<false>(&cmd->vertices[0].x);
+    const GSVector2i v1 = GSVector2i::load<false>(&cmd->vertices[1].x);
+    const GSVector4i rect = GSVector4i::xyxy(v0.min_s32(v1), v0.max_s32(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;
 }
@@ -443,9 +832,9 @@ bool GPU::HandleRenderPolyLineCommand()
   const u32 words_to_pop = min_words - 1;
   // m_blit_buffer.resize(words_to_pop);
   // FifoPopRange(m_blit_buffer.data(), words_to_pop);
-  m_blit_buffer.reserve(words_to_pop);
+  m_polyline_buffer.reserve(words_to_pop);
   for (u32 i = 0; i < words_to_pop; i++)
-    m_blit_buffer.push_back(Truncate32(FifoPop()));
+    m_polyline_buffer.push_back(m_fifo.Pop());
 
   // polyline goes via a different path through the blit buffer
   m_blitter_state = BlitterState::DrawingPolyLine;
@@ -453,6 +842,127 @@ bool GPU::HandleRenderPolyLineCommand()
   return true;
 }
 
+void GPU::FinishPolyline()
+{
+  PrepareForDraw();
+
+  const u32 num_vertices = GetPolyLineVertexCount();
+  DebugAssert(num_vertices >= 2);
+
+  if (g_settings.gpu_pgxp_enable)
+  {
+    GPUBackendDrawPreciseLineCommand* RESTRICT cmd = GPUBackend::NewDrawPreciseLineCommand((num_vertices - 1) * 2);
+    FillDrawCommand(cmd, m_render_command);
+    cmd->palette.bits = 0;
+
+    u32 buffer_pos = 0;
+    u32 out_vertex_count = 0;
+    const bool shaded = m_render_command.shading_enable;
+    bool valid_w = g_settings.gpu_pgxp_texture_correction;
+    GPUBackendDrawPreciseLineCommand::Vertex start, end;
+
+    const auto read_vertex = [this, &buffer_pos, &valid_w](GPUBackendDrawPreciseLineCommand::Vertex& RESTRICT dest,
+                                                           u32 color) {
+      const u64 maddr_and_pos = m_polyline_buffer[buffer_pos++];
+      const GPUVertexPosition vp{Truncate32(maddr_and_pos)};
+      dest.native_x = m_drawing_offset.x + vp.x;
+      dest.native_y = m_drawing_offset.y + vp.y;
+      dest.color = color;
+      valid_w &= CPU::PGXP::GetPreciseVertex(Truncate32(maddr_and_pos >> 32), vp.bits, dest.native_x, dest.native_y,
+                                             m_drawing_offset.x, m_drawing_offset.y, &dest.x, &dest.y, &dest.w);
+    };
+
+    read_vertex(start, m_render_command.color_for_first_vertex);
+
+    for (u32 i = 1; i < num_vertices; i++)
+    {
+      const u32 color =
+        (shaded ? Truncate32(m_polyline_buffer[buffer_pos++]) : m_render_command.bits) & UINT32_C(0x00FFFFFF);
+      read_vertex(end, color);
+
+      const GSVector2 start_pos = GSVector2::load<false>(&start.x);
+      const GSVector2 end_pos = GSVector2::load<false>(&end.x);
+      const GSVector4i rect =
+        GSVector4i(GSVector4::xyxy(start_pos.min(end_pos), start_pos.max(end_pos))).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: {},{} - {},{}", start_pos.x, start_pos.y, end_pos.x, end_pos.y);
+      }
+      else
+      {
+        AddDrawLineTicks(clamped_rect, m_render_command.shading_enable);
+
+        cmd->vertices[out_vertex_count++] = start;
+        cmd->vertices[out_vertex_count++] = end;
+      }
+
+      start = end;
+    }
+
+    if (out_vertex_count > 0)
+    {
+      DebugAssert(out_vertex_count <= cmd->num_vertices);
+      cmd->num_vertices = Truncate16(out_vertex_count);
+      GPUBackend::PushCommand(cmd);
+    }
+  }
+  else
+  {
+    GPUBackendDrawLineCommand* RESTRICT cmd = GPUBackend::NewDrawLineCommand((num_vertices - 1) * 2);
+    FillDrawCommand(cmd, m_render_command);
+    cmd->palette.bits = 0;
+
+    u32 buffer_pos = 0;
+    const GPUVertexPosition start_vp{Truncate32(m_polyline_buffer[buffer_pos++])};
+    const GSVector2i draw_offset = GSVector2i::load<false>(&m_drawing_offset.x);
+    GSVector2i start_pos = GSVector2i(start_vp.x, start_vp.y).add32(draw_offset);
+    u32 start_color = m_render_command.color_for_first_vertex;
+
+    const bool shaded = m_render_command.shading_enable;
+    u32 out_vertex_count = 0;
+    for (u32 i = 1; i < num_vertices; i++)
+    {
+      const u32 end_color = shaded ? (Truncate32(m_polyline_buffer[buffer_pos++] & UINT32_C(0x00FFFFFF))) :
+                                     m_render_command.color_for_first_vertex;
+      const GPUVertexPosition vp{Truncate32(m_polyline_buffer[buffer_pos++])};
+      const GSVector2i end_pos = GSVector2i(vp.x, vp.y).add32(draw_offset);
+
+      const GSVector4i rect =
+        GSVector4i::xyxy(start_pos.min_s32(end_pos), start_pos.max_s32(end_pos)).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: {},{} - {},{}", start_pos.x, start_pos.y, end_pos.x, end_pos.y);
+      }
+      else
+      {
+        AddDrawLineTicks(clamped_rect, m_render_command.shading_enable);
+
+        GPUBackendDrawLineCommand::Vertex* out_vertex = &cmd->vertices[out_vertex_count];
+        out_vertex_count += 2;
+
+        GSVector2i::store<false>(&out_vertex[0].x, start_pos);
+        out_vertex[0].color = start_color;
+        GSVector2i::store<false>(&out_vertex[1].x, end_pos);
+        out_vertex[1].color = end_color;
+      }
+
+      start_pos = end_pos;
+      start_color = end_color;
+    }
+
+    if (out_vertex_count > 0)
+    {
+      DebugAssert(out_vertex_count <= cmd->num_vertices);
+      cmd->num_vertices = Truncate16(out_vertex_count);
+      GPUBackend::PushCommand(cmd);
+    }
+  }
+}
+
 bool GPU::HandleFillRectangleCommand()
 {
   CHECK_COMMAND_SIZE(3);
@@ -460,8 +970,6 @@ 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;
@@ -471,9 +979,18 @@ 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();
+    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;
+    cmd->interlaced_rendering = IsInterlacedRenderingEnabled();
+    cmd->active_line_lsb = m_crtc_state.active_line_lsb;
+    GPUBackend::PushCommand(cmd);
+  }
 
-  m_counters.num_writes++;
   AddCommandTicks(46 + ((width / 8) + 9) * height);
   EndCommand();
   return true;
@@ -523,11 +1040,9 @@ void GPU::FinishVRAMWrite()
   if (IsInterlacedRenderingEnabled() && IsCRTCScanlinePending())
     SynchronizeCRTC();
 
-  FlushRender();
-
   if (m_blit_remaining_words == 0)
   {
-    if (g_settings.debugging.dump_cpu_to_vram_copies)
+    if (g_settings.gpu_dump_cpu_to_vram_copies)
     {
       DumpVRAMToFile(TinyString::from_format("cpu_to_vram_copy_{}.png", s_cpu_to_vram_dump_id++), m_vram_transfer.width,
                      m_vram_transfer.height, sizeof(u16) * m_vram_transfer.width, m_blit_buffer.data(), true);
@@ -557,18 +1072,18 @@ 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,
+      UpdateVRAM(m_vram_transfer.x, m_vram_transfer.y, m_vram_transfer.width, static_cast<u16>(transferred_full_rows),
-                 m_GPUSTAT.set_mask_while_drawing, m_GPUSTAT.check_mask_before_draw);
+                 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,
+      UpdateVRAM(m_vram_transfer.x, static_cast<u16>(m_vram_transfer.y + transferred_full_rows),
-                 m_GPUSTAT.set_mask_while_drawing, m_GPUSTAT.check_mask_before_draw);
+                 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;
@@ -588,13 +1103,10 @@ 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);
 
-  if (g_settings.debugging.dump_vram_to_cpu_copies)
+  if (g_settings.gpu_dump_vram_to_cpu_copies)
   {
     DumpVRAMToFile(TinyString::from_format("vram_to_cpu_copy_{}.png", s_vram_to_cpu_dump_id++), m_vram_transfer.width,
                    m_vram_transfer.height, sizeof(u16) * VRAM_WIDTH,
@@ -602,7 +1114,6 @@ bool GPU::HandleCopyRectangleVRAMToCPUCommand()
   }
 
   // switch to pixel-by-pixel read state
-  m_counters.num_reads++;
   m_blitter_state = BlitterState::ReadingVRAM;
   m_command_total_words = 0;
 
@@ -633,10 +1144,16 @@ 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++;
+    GPUBackendCopyVRAMCommand* cmd = GPUBackend::NewCopyVRAMCommand();
-
+    cmd->src_x = static_cast<u16>(src_x);
-    FlushRender();
+    cmd->src_y = static_cast<u16>(src_y);
-    CopyVRAM(src_x, src_y, dst_x, dst_y, width, height);
+    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);
+    cmd->check_mask_before_draw = m_GPUSTAT.check_mask_before_draw;
+    cmd->set_mask_while_drawing = m_GPUSTAT.set_mask_while_drawing;
+    GPUBackend::PushCommand(cmd);
   }
 
   AddCommandTicks(width * height * 2);

src/core/gpu_dump.cpp

@@ -76,7 +76,7 @@ std::unique_ptr<GPUDump::Recorder> GPUDump::Recorder::Create(std::string path, s
 
   ret = std::unique_ptr<Recorder>(new Recorder(std::move(fp), num_frames, std::move(path)));
   ret->WriteHeaders(serial);
-  g_gpu->WriteCurrentVideoModeToDump(ret.get());
+  g_gpu.WriteCurrentVideoModeToDump(ret.get());
   ret->WriteCurrentVRAM();
 
   // Write start of stream.
@@ -285,7 +285,7 @@ void GPUDump::Recorder::WriteHeaders(std::string_view serial)
 
   // Write textual video mode.
   BeginPacket(PacketType::TextualVideoFormat);
-  WriteString(g_gpu->IsInPALMode() ? "PAL" : "NTSC");
+  WriteString(g_gpu.IsInPALMode() ? "PAL" : "NTSC");
   EndPacket();
 
   // Write DuckStation version.
@@ -520,7 +520,7 @@ void GPUDump::Player::ProcessPacket(const PacketRef& pkt)
   if (pkt.type <= PacketType::VSyncEvent)
   {
     // gp0/gp1/vsync => direct to gpu
-    g_gpu->ProcessGPUDumpPacket(pkt.type, pkt.data);
+    g_gpu.ProcessGPUDumpPacket(pkt.type, pkt.data);
     return;
   }
 }

src/core/gpu_hw.h

@@ -3,7 +3,7 @@
 
 #pragma once
 
-#include "gpu.h"
+#include "gpu_backend.h"
 #include "gpu_hw_texture_cache.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
@@ -63,21 +65,42 @@ public:
   GPU_HW();
   ~GPU_HW() override;
 
-  const Threading::Thread* GetSWThread() const override;
+  bool Initialize(bool upload_vram, Error* error) override;
-  bool IsHardwareRenderer() const override;
 
-  bool Initialize(Error* error) override;
+  u32 GetResolutionScale() const override;
-  void Reset(bool clear_vram) override;
-  bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display) override;
 
   void RestoreDeviceContext() override;
 
-  void UpdateSettings(const Settings& old_settings) override;
+protected:
+  void UpdateSettings(const GPUSettings& old_settings) override;
 
-  u32 GetResolutionScale() const override;
   void UpdateResolutionScale() override;
 
-  void UpdateDisplay() override;
+  void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced_rendering, u8 active_line_lsb) 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, bool set_mask,
+                bool check_mask) 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 DrawPreciseLine(const GPUBackendDrawPreciseLineCommand* cmd) override;
+
+  void FlushRender() override;
+  void DrawingAreaChanged() override;
+  void ClearVRAM() override;
+
+  void LoadState(const GPUBackendLoadStateCommand* cmd) override;
+
+  bool AllocateMemorySaveState(System::MemorySaveState& mss, Error* error) override;
+  void DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss) override;
+
+  void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) override;
 
 private:
   enum : u32
@@ -86,6 +109,7 @@ 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
   {
@@ -164,8 +188,6 @@ private:
   bool CompileResolutionDependentPipelines(Error* error);
   bool CompileDownsamplePipelines(Error* error);
 
-  void LoadVertices();
-
   void PrintSettingsToLog();
   void CheckSettings();
 
@@ -184,8 +206,10 @@ private:
   u32 CalculateResolutionScale() const;
   GPUDownsampleMode GetDownsampleMode(u32 resolution_scale) const;
 
+  bool ShouldDrawWithSoftwareRenderer() const;
+
   bool IsUsingMultisampling() const;
-  bool IsUsingDownsampling() const;
+  bool IsUsingDownsampling(const GPUBackendUpdateDisplayCommand* cmd) const;
 
   void SetFullVRAMDirtyRectangle();
   void ClearVRAMDirtyRectangle();
@@ -195,12 +219,15 @@ 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 ShouldCheckForTexPageOverlap() const;
 
   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_precise, bool is_3d);
   void ResetBatchVertexDepth();
 
   /// Returns the value to be written to the depth buffer for the current operation for mask bit emulation.
@@ -212,20 +239,6 @@ private:
   /// Returns true if the draw is going to use shader blending/framebuffer fetch.
   bool NeedsShaderBlending(GPUTransparencyMode transparency, BatchTextureMode texture, 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(GPUTexture* tex, u32 dst_x, u32 dst_y, u32 width, u32 height);
@@ -234,17 +247,17 @@ 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 SetBatchDepthBuffer(const GPUBackendDrawCommand* cmd, bool enabled);
-  void CheckForDepthClear(const BatchVertex* vertices, u32 num_vertices);
+  void CheckForDepthClear(const GPUBackendDrawCommand* cmd, const BatchVertex* vertices, u32 num_vertices);
-  void SetBatchSpriteMode(bool enabled);
+  void SetBatchSpriteMode(const GPUBackendDrawCommand* cmd, bool enabled);
 
   void UpdateDownsamplingLevels();
 
@@ -262,8 +275,6 @@ 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;
-
   BatchVertex* m_batch_vertex_ptr = nullptr;
   u16* m_batch_index_ptr = nullptr;
   u32 m_batch_base_vertex = 0;
@@ -305,18 +316,32 @@ private:
   u8 m_texpage_dirty = 0;
 
   bool m_batch_ubo_dirty = true;
+  bool m_drawing_area_changed = true;
   BatchConfig m_batch;
 
   // Changed state
   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; // TODO: Don't use in TC mode, should be kept at zero.
   GSVector4i m_current_uv_rect = INVALID_RECT;
   GSVector4i m_current_draw_rect = INVALID_RECT;
   alignas(8) s32 m_current_texture_page_offset[2] = {};
 
+  union
+  {
+    struct
+    {
+      // NOTE: Only the texture-related bits should be used here, the others are not validated.
+      GPUDrawModeReg mode_reg;
+      GPUTexturePaletteReg palette_reg;
+    };
+
+    u32 bits = INVALID_DRAW_MODE_BITS;
+  } m_draw_mode = {};
+
   std::unique_ptr<GPUPipeline> m_wireframe_pipeline;
 
   // [wrapped][interlaced]

src/core/gpu_hw_texture_cache.cpp

@@ -5,11 +5,14 @@
 #include "gpu_hw.h"
 #include "gpu_hw_shadergen.h"
 #include "gpu_sw_rasterizer.h"
+#include "gpu_thread.h"
 #include "host.h"
+#include "imgui_overlays.h"
 #include "settings.h"
 #include "system.h"
 
 #include "util/gpu_device.h"
+#include "util/imgui_fullscreen.h"
 #include "util/imgui_manager.h"
 #include "util/state_wrapper.h"
 
@@ -50,6 +53,9 @@ static constexpr const GSVector4i& INVALID_RECT = GPU_HW::INVALID_RECT;
 static constexpr const GPUTexture::Format REPLACEMENT_TEXTURE_FORMAT = GPUTexture::Format::RGBA8;
 static constexpr const char LOCAL_CONFIG_FILENAME[] = "config.yaml";
 
+static constexpr u32 STATE_PALETTE_RECORD_SIZE =
+  sizeof(GSVector4i) + sizeof(SourceKey) + sizeof(PaletteRecordFlags) + sizeof(HashType) + sizeof(u16) * MAX_CLUT_SIZE;
+
 // Has to be public because it's referenced in Source.
 struct HashCacheEntry
 {
@@ -518,6 +524,7 @@ struct GPUTextureCacheState
 
   GPUTexture::Format hash_cache_texture_format = GPUTexture::Format::Unknown;
   HashCache hash_cache;
+  GPU_HW* hw_backend = nullptr; // TODO:FIXME: remove me
 
   /// List of candidates for purging when the hash cache gets too large.
   std::vector<std::pair<HashCache::iterator, s32>> hash_cache_purge_list;
@@ -529,7 +536,6 @@ struct GPUTextureCacheState
   std::unique_ptr<GPUPipeline> replacement_draw_pipeline;                 // copies alpha as-is
   std::unique_ptr<GPUPipeline> replacement_semitransparent_draw_pipeline; // inverts alpha (i.e. semitransparent)
 
-  std::string game_id;
   VRAMReplacementMap vram_replacements;
 
   // TODO: Combine these into one map?
@@ -555,26 +561,28 @@ ALIGN_TO_CACHE_LINE GPUTextureCacheState s_state;
 
 bool GPUTextureCache::ShouldTrackVRAMWrites()
 {
-  if (!g_settings.gpu_texture_cache)
+  if (!g_gpu_settings.gpu_texture_cache)
     return false;
 
 #ifdef ALWAYS_TRACK_VRAM_WRITES
   return true;
 #else
   return (IsDumpingVRAMWriteTextures() ||
-          (g_settings.texture_replacements.enable_texture_replacements && HasVRAMWriteTextureReplacements()));
+          (g_gpu_settings.texture_replacements.enable_texture_replacements && HasVRAMWriteTextureReplacements()));
 #endif
 }
 
 bool GPUTextureCache::IsDumpingVRAMWriteTextures()
 {
-  return (g_settings.texture_replacements.dump_textures && !s_state.config.dump_texture_pages);
+  return (g_gpu_settings.texture_replacements.dump_textures && !s_state.config.dump_texture_pages);
 }
 
-bool GPUTextureCache::Initialize()
+bool GPUTextureCache::Initialize(GPU_HW* backend)
 {
+  s_state.hw_backend = backend;
+
   SetHashCacheTextureFormat();
-  LoadLocalConfiguration(false, false);
+  ReloadTextureReplacements(false);
   UpdateVRAMTrackingState();
   if (!CompilePipelines())
     return false;
@@ -582,13 +590,13 @@ bool GPUTextureCache::Initialize()
   return true;
 }
 
-void GPUTextureCache::UpdateSettings(bool use_texture_cache, const Settings& old_settings)
+void GPUTextureCache::UpdateSettings(bool use_texture_cache, const GPUSettings& old_settings)
 {
   if (use_texture_cache)
   {
     UpdateVRAMTrackingState();
 
-    if (g_settings.texture_replacements.enable_texture_replacements !=
+    if (g_gpu_settings.texture_replacements.enable_texture_replacements !=
         old_settings.texture_replacements.enable_texture_replacements)
     {
       Invalidate();
@@ -602,9 +610,9 @@ void GPUTextureCache::UpdateSettings(bool use_texture_cache, const Settings& old
   // Reload textures if configuration changes.
   const bool old_replacement_scale_linear_filter = s_state.config.replacement_scale_linear_filter;
   if (LoadLocalConfiguration(false, false) ||
-      g_settings.texture_replacements.enable_texture_replacements !=
+      g_gpu_settings.texture_replacements.enable_texture_replacements !=
         old_settings.texture_replacements.enable_texture_replacements ||
-      g_settings.texture_replacements.enable_vram_write_replacements !=
+      g_gpu_settings.texture_replacements.enable_vram_write_replacements !=
         old_settings.texture_replacements.enable_vram_write_replacements)
   {
     if (use_texture_cache)
@@ -620,6 +628,37 @@ void GPUTextureCache::UpdateSettings(bool use_texture_cache, const Settings& old
   }
 }
 
+bool GPUTextureCache::GetStateSize(StateWrapper& sw, u32* size)
+{
+  if (sw.GetVersion() < 73)
+  {
+    *size = 0;
+    return true;
+  }
+
+  const size_t start = sw.GetPosition();
+  if (!sw.DoMarker("GPUTextureCache")) [[unlikely]]
+    return false;
+
+  u32 num_vram_writes = 0;
+  sw.Do(&num_vram_writes);
+
+  for (u32 i = 0; i < num_vram_writes; i++)
+  {
+    sw.SkipBytes(sizeof(GSVector4i) * 2 + sizeof(HashType));
+
+    u32 num_palette_records = 0;
+    sw.Do(&num_palette_records);
+    sw.SkipBytes(num_palette_records * STATE_PALETTE_RECORD_SIZE);
+  }
+
+  if (sw.HasError()) [[unlikely]]
+    return false;
+
+  *size = static_cast<u32>(sw.GetPosition() - start);
+  return true;
+}
+
 bool GPUTextureCache::DoState(StateWrapper& sw, bool skip)
 {
   if (sw.GetVersion() < 73)
@@ -668,7 +707,7 @@ bool GPUTextureCache::DoState(StateWrapper& sw, bool skip)
         sw.Do(&num_palette_records);
 
         // Skip palette records if we're not dumping now.
-        if (g_settings.texture_replacements.dump_textures)
+        if (g_gpu_settings.texture_replacements.dump_textures)
         {
           vrw->palette_records.reserve(num_palette_records);
           for (u32 j = 0; j < num_palette_records; j++)
@@ -760,6 +799,7 @@ void GPUTextureCache::Shutdown()
   s_state.hash_cache_purge_list = {};
   s_state.temp_vram_write_list = {};
   s_state.track_vram_writes = false;
+  s_state.hw_backend = nullptr;
 
   for (auto it = s_state.gpu_replacement_image_cache.begin(); it != s_state.gpu_replacement_image_cache.end();)
   {
@@ -773,7 +813,6 @@ void GPUTextureCache::Shutdown()
   s_state.vram_write_texture_replacements.clear();
   s_state.texture_page_texture_replacements.clear();
   s_state.dumped_textures.clear();
-  s_state.game_id = {};
 }
 
 void GPUTextureCache::SetHashCacheTextureFormat()
@@ -791,7 +830,7 @@ void GPUTextureCache::SetHashCacheTextureFormat()
 
 bool GPUTextureCache::CompilePipelines()
 {
-  if (!g_settings.texture_replacements.enable_texture_replacements)
+  if (!g_gpu_settings.texture_replacements.enable_texture_replacements)
     return true;
 
   GPUPipeline::GraphicsConfig plconfig = {};
@@ -1390,7 +1429,7 @@ const GPUTextureCache::Source* GPUTextureCache::ReturnSource(Source* source, con
   source->from_hash_cache->last_used_frame = System::GetFrameNumber();
 
   // TODO: Cache var.
-  if (g_settings.texture_replacements.dump_textures)
+  if (g_gpu_settings.texture_replacements.dump_textures)
   {
     source->active_uv_rect = source->active_uv_rect.runion(uv_rect);
     source->palette_record_flags |= flags;
@@ -1548,7 +1587,7 @@ void GPUTextureCache::DestroySource(Source* src, bool remove_from_hash_cache)
 {
   GL_INS_FMT("Invalidate source {}", SourceToString(src));
 
-  if (g_settings.texture_replacements.dump_textures && !src->active_uv_rect.eq(INVALID_RECT))
+  if (g_gpu_settings.texture_replacements.dump_textures && !src->active_uv_rect.eq(INVALID_RECT))
   {
     if (!s_state.config.dump_texture_pages)
     {
@@ -1950,7 +1989,7 @@ void GPUTextureCache::RemoveVRAMWrite(VRAMWrite* entry)
 
 void GPUTextureCache::DumpTexturesFromVRAMWrite(VRAMWrite* entry)
 {
-  if (g_settings.texture_replacements.dump_textures && !s_state.config.dump_texture_pages)
+  if (g_gpu_settings.texture_replacements.dump_textures && !s_state.config.dump_texture_pages)
   {
     for (const VRAMWrite::PaletteRecord& prec : entry->palette_records)
     {
@@ -2200,7 +2239,7 @@ GPUTextureCache::HashCacheEntry* GPUTextureCache::LookupHashCache(SourceKey key,
 
   DecodeTexture(key.page, key.palette, key.mode, entry.texture.get());
 
-  if (g_settings.texture_replacements.enable_texture_replacements)
+  if (g_gpu_settings.texture_replacements.enable_texture_replacements)
     ApplyTextureReplacements(key, tex_hash, pal_hash, &entry);
 
   s_state.hash_cache_memory_usage += entry.texture->GetVRAMUsage();
@@ -2603,12 +2642,8 @@ size_t GPUTextureCache::DumpedTextureKeyHash::operator()(const DumpedTextureKey&
   return hash;
 }
 
-void GPUTextureCache::SetGameID(std::string game_id)
+void GPUTextureCache::GameSerialChanged()
 {
-  if (s_state.game_id == game_id)
-    return;
-
-  s_state.game_id = game_id;
   ReloadTextureReplacements(false);
 }
 
@@ -2625,7 +2660,8 @@ GPUTexture* GPUTextureCache::GetVRAMReplacement(u32 width, u32 height, const voi
 
 bool GPUTextureCache::ShouldDumpVRAMWrite(u32 width, u32 height)
 {
-  return (g_settings.texture_replacements.dump_vram_writes && width >= s_state.config.vram_write_dump_width_threshold &&
+  return (g_gpu_settings.texture_replacements.dump_vram_writes &&
+          width >= s_state.config.vram_write_dump_width_threshold &&
           height >= s_state.config.vram_write_dump_height_threshold);
 }
 
@@ -2716,7 +2752,7 @@ void GPUTextureCache::DumpTexture(TextureReplacementType type, u32 offset_x, u32
   };
 
   // skip if dumped already
-  if (!g_settings.texture_replacements.dump_replaced_textures)
+  if (!g_gpu_settings.texture_replacements.dump_replaced_textures)
   {
     const TextureReplacementMap& map = (type == TextureReplacementType::TextureFromPage) ?
                                          s_state.texture_page_texture_replacements :
@@ -2942,7 +2978,7 @@ bool GPUTextureCache::HasValidReplacementExtension(const std::string_view path)
 
 void GPUTextureCache::FindTextureReplacements(bool load_vram_write_replacements, bool load_texture_replacements)
 {
-  if (s_state.game_id.empty())
+  if (GPUThread::GetGameSerial().empty())
     return;
 
   FileSystem::FindResultsArray files;
@@ -3015,23 +3051,23 @@ void GPUTextureCache::FindTextureReplacements(bool load_vram_write_replacements,
     }
   }
 
-  if (g_settings.texture_replacements.enable_texture_replacements)
+  if (g_gpu_settings.texture_replacements.enable_texture_replacements)
   {
     INFO_LOG("Found {} replacement upload textures for '{}'", s_state.vram_write_texture_replacements.size(),
-             s_state.game_id);
+             GPUThread::GetGameSerial());
     INFO_LOG("Found {} replacement page textures for '{}'", s_state.texture_page_texture_replacements.size(),
-             s_state.game_id);
+             GPUThread::GetGameSerial());
   }
 
-  if (g_settings.texture_replacements.enable_vram_write_replacements)
+  if (g_gpu_settings.texture_replacements.enable_vram_write_replacements)
-    INFO_LOG("Found {} replacement VRAM for '{}'", s_state.vram_replacements.size(), s_state.game_id);
+    INFO_LOG("Found {} replacement VRAM for '{}'", s_state.vram_replacements.size(), GPUThread::GetGameSerial());
 }
 
 void GPUTextureCache::LoadTextureReplacementAliases(const ryml::ConstNodeRef& root,
                                                     bool load_vram_write_replacement_aliases,
                                                     bool load_texture_replacement_aliases)
 {
-  if (s_state.game_id.empty())
+  if (GPUThread::GetGameSerial().empty())
     return;
 
   const std::string source_dir = GetTextureReplacementDirectory();
@@ -3107,17 +3143,19 @@ void GPUTextureCache::LoadTextureReplacementAliases(const ryml::ConstNodeRef& ro
     }
   }
 
-  if (g_settings.texture_replacements.enable_texture_replacements)
+  if (g_gpu_settings.texture_replacements.enable_texture_replacements)
   {
     INFO_LOG("Found {} replacement upload textures after applying aliases for '{}'",
-             s_state.vram_write_texture_replacements.size(), s_state.game_id);
+             s_state.vram_write_texture_replacements.size(), GPUThread::GetGameSerial());
     INFO_LOG("Found {} replacement page textures after applying aliases for '{}'",
-             s_state.texture_page_texture_replacements.size(), s_state.game_id);
+             s_state.texture_page_texture_replacements.size(), GPUThread::GetGameSerial());
   }
 
-  if (g_settings.texture_replacements.enable_vram_write_replacements)
+  if (g_gpu_settings.texture_replacements.enable_vram_write_replacements)
+  {
     INFO_LOG("Found {} replacement VRAM after applying aliases for '{}'", s_state.vram_replacements.size(),
-             s_state.game_id);
+             GPUThread::GetGameSerial());
+  }
 }
 
 const GPUTextureCache::TextureReplacementImage* GPUTextureCache::GetTextureReplacementImage(const std::string& path)
@@ -3241,8 +3279,8 @@ void GPUTextureCache::PreloadReplacementTextures()
 #define UPDATE_PROGRESS()                                                                                              \
   if (last_update_time.GetTimeSeconds() >= UPDATE_INTERVAL)                                                            \
   {                                                                                                                    \
-    Host::DisplayLoadingScreen("Preloading replacement textures...", 0, static_cast<int>(total_textures),              \
+    ImGuiFullscreen::RenderLoadingScreen(ImGuiManager::LOGO_IMAGE_NAME, "Preloading replacement textures...", 0,       \
-                               static_cast<int>(num_textures_loaded));                                                 \
+                                         static_cast<int>(total_textures), static_cast<int>(num_textures_loaded));     \
     last_update_time.Reset();                                                                                          \
   }
 
@@ -3269,10 +3307,10 @@ void GPUTextureCache::PreloadReplacementTextures()
 
 bool GPUTextureCache::EnsureGameDirectoryExists()
 {
-  if (s_state.game_id.empty())
+  if (GPUThread::GetGameSerial().empty())
     return false;
 
-  const std::string game_directory = Path::Combine(EmuFolders::Textures, s_state.game_id);
+  const std::string game_directory = Path::Combine(EmuFolders::Textures, GPUThread::GetGameSerial());
   if (FileSystem::DirectoryExists(game_directory.c_str()))
     return true;
 
@@ -3309,12 +3347,13 @@ bool GPUTextureCache::EnsureGameDirectoryExists()
 
 std::string GPUTextureCache::GetTextureReplacementDirectory()
 {
-  std::string dir = Path::Combine(
+  std::string dir =
-    EmuFolders::Textures, SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "replacements", s_state.game_id));
+    Path::Combine(EmuFolders::Textures,
+                  SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "replacements", GPUThread::GetGameSerial()));
   if (!FileSystem::DirectoryExists(dir.c_str()))
   {
     // Check for the old directory structure without a replacements subdirectory.
-    std::string altdir = Path::Combine(EmuFolders::Textures, s_state.game_id);
+    std::string altdir = Path::Combine(EmuFolders::Textures, GPUThread::GetGameSerial());
     if (FileSystem::DirectoryExists(altdir.c_str()))
       WARNING_LOG("Using deprecated texture replacement directory {}", altdir);
     dir = std::move(altdir);
@@ -3326,7 +3365,7 @@ std::string GPUTextureCache::GetTextureReplacementDirectory()
 std::string GPUTextureCache::GetTextureDumpDirectory()
 {
   return Path::Combine(EmuFolders::Textures,
-                       SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "dumps", s_state.game_id));
+                       SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "dumps", GPUThread::GetGameSerial()));
 }
 
 GPUTextureCache::VRAMReplacementName GPUTextureCache::GetVRAMWriteHash(u32 width, u32 height, const void* pixels)
@@ -3354,14 +3393,15 @@ bool GPUTextureCache::LoadLocalConfiguration(bool load_vram_write_replacement_al
   const Settings::TextureReplacementSettings::Configuration old_config = s_state.config;
 
   // load settings from ini
-  s_state.config = g_settings.texture_replacements.config;
+  s_state.config = g_gpu_settings.texture_replacements.config;
 
-  if (s_state.game_id.empty())
+  const std::string& game_serial = GPUThread::GetGameSerial();
+  if (game_serial.empty())
     return (s_state.config != old_config);
 
   const std::optional<std::string> ini_data = FileSystem::ReadFileToString(
     Path::Combine(EmuFolders::Textures,
-                  SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "{}", s_state.game_id, LOCAL_CONFIG_FILENAME))
+                  SmallString::from_format("{}" FS_OSPATH_SEPARATOR_STR "{}", game_serial, LOCAL_CONFIG_FILENAME))
       .c_str());
   if (!ini_data.has_value() || ini_data->empty())
     return (s_state.config != old_config);
@@ -3430,20 +3470,19 @@ void GPUTextureCache::ReloadTextureReplacements(bool show_info)
   s_state.vram_write_texture_replacements.clear();
   s_state.texture_page_texture_replacements.clear();
 
-  const bool load_vram_write_replacements = (g_settings.texture_replacements.enable_vram_write_replacements);
+  const bool load_vram_write_replacements = (g_gpu_settings.texture_replacements.enable_vram_write_replacements);
   const bool load_texture_replacements =
-    (g_settings.gpu_texture_cache && g_settings.texture_replacements.enable_texture_replacements);
+    (g_gpu_settings.gpu_texture_cache && g_gpu_settings.texture_replacements.enable_texture_replacements);
   if (load_vram_write_replacements || load_texture_replacements)
     FindTextureReplacements(load_vram_write_replacements, load_texture_replacements);
 
   LoadLocalConfiguration(load_vram_write_replacements, load_texture_replacements);
 
-  if (g_settings.texture_replacements.preload_textures)
+  if (g_gpu_settings.texture_replacements.preload_textures)
     PreloadReplacementTextures();
 
   PurgeUnreferencedTexturesFromCache();
 
-  DebugAssert(g_gpu);
   UpdateVRAMTrackingState();
   InvalidateSources();
 
@@ -3596,5 +3635,5 @@ void GPUTextureCache::ApplyTextureReplacements(SourceKey key, HashType tex_hash,
   g_gpu_device->RecycleTexture(std::move(entry->texture));
   entry->texture = std::move(replacement_tex);
 
-  g_gpu->RestoreDeviceContext();
+  s_state.hw_backend->RestoreDeviceContext();
 }

src/core/gpu_hw_texture_cache.h

@@ -9,7 +9,8 @@ class Image;
 class GPUTexture;
 class StateWrapper;
 
-struct Settings;
+struct GPUSettings;
+class GPU_HW;
 
 //////////////////////////////////////////////////////////////////////////
 // Texture Cache
@@ -102,9 +103,12 @@ struct Source
   TListNode<Source> hash_cache_ref;
 };
 
-bool Initialize();
+bool Initialize(GPU_HW* backend);
-void UpdateSettings(bool use_texture_cache, const Settings& old_settings);
+void UpdateSettings(bool use_texture_cache, const GPUSettings& old_settings);
+
+bool GetStateSize(StateWrapper& sw, u32* size);
 bool DoState(StateWrapper& sw, bool skip);
+
 void Shutdown();
 
 void Invalidate();
@@ -124,7 +128,7 @@ bool AreSourcePagesDrawn(SourceKey key, const GSVector4i rect);
 
 void Compact();
 
-void SetGameID(std::string game_id);
+void GameSerialChanged();
 void ReloadTextureReplacements(bool show_info);
 
 // VRAM Write Replacements

src/core/gpu_sw.cpp

@@ -2,16 +2,17 @@
 // SPDX-License-Identifier: CC-BY-NC-ND-4.0
 
 #include "gpu_sw.h"
-#include "gpu_hw_texture_cache.h"
+#include "gpu.h"
+#include "gpu_sw_rasterizer.h"
 #include "settings.h"
-#include "system.h"
+#include "system_private.h"
 
 #include "util/gpu_device.h"
+#include "util/state_wrapper.h"
 
 #include "common/align.h"
 #include "common/assert.h"
-#include "common/gsvector.h"
+#include "common/intrin.h"
-#include "common/gsvector_formatter.h"
 #include "common/log.h"
 
 #include <algorithm>
@@ -20,25 +21,16 @@ LOG_CHANNEL(GPU);
 
 GPU_SW::GPU_SW() = default;
 
-GPU_SW::~GPU_SW()
+GPU_SW::~GPU_SW() = default;
+
+u32 GPU_SW::GetResolutionScale() const
 {
-  g_gpu_device->RecycleTexture(std::move(m_upload_texture));
+  return 1u;
-  m_backend.Shutdown();
 }
 
-const Threading::Thread* GPU_SW::GetSWThread() const
+bool GPU_SW::Initialize(bool upload_vram, Error* error)
 {
-  return m_backend.GetThread();
+  if (!GPUBackend::Initialize(upload_vram, error))
-}
-
-bool GPU_SW::IsHardwareRenderer() const
-{
-  return false;
-}
-
-bool GPU_SW::Initialize(Error* error)
-{
-  if (!GPU::Initialize(error) || !m_backend.Initialize(g_settings.gpu_use_thread))
     return false;
 
   static constexpr const std::array formats_for_16bit = {GPUTexture::Format::RGB5A1, GPUTexture::Format::A1BGR5,
@@ -55,34 +47,153 @@ bool GPU_SW::Initialize(Error* error)
   // RGBA8 will always be supported, hence we'll find one.
   INFO_LOG("Using {} format for 16-bit display", GPUTexture::GetFormatName(m_16bit_display_format));
   Assert(m_16bit_display_format != GPUTexture::Format::Unknown);
+
+  // if we're using "new" vram, clear it out here
+  if (!upload_vram)
+    std::memset(g_vram, 0, sizeof(g_vram));
+
   return true;
 }
 
-bool GPU_SW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display)
+void GPU_SW::ClearVRAM()
 {
-  // need to ensure the worker thread is done
+  std::memset(g_vram, 0, sizeof(g_vram));
-  m_backend.Sync(true);
+  std::memset(g_gpu_clut, 0, sizeof(g_gpu_clut));
-
-  // ignore the host texture for software mode, since we want to save vram here
-  if (!GPU::DoState(sw, nullptr, update_display))
-    return false;
-
-  // need to still call the TC, to toss any data in the state
-  return GPUTextureCache::DoState(sw, true);
 }
 
-void GPU_SW::Reset(bool clear_vram)
+void GPU_SW::UpdateResolutionScale()
 {
-  GPU::Reset(clear_vram);
-
-  m_backend.Reset();
 }
 
-void GPU_SW::UpdateSettings(const Settings& old_settings)
+void GPU_SW::LoadState(const GPUBackendLoadStateCommand* cmd)
+{
+  std::memcpy(g_vram, cmd->vram_data, sizeof(g_vram));
+  std::memcpy(g_gpu_clut, cmd->clut_data, sizeof(g_gpu_clut));
+}
+
+bool GPU_SW::AllocateMemorySaveState(System::MemorySaveState& mss, Error* error)
+{
+  mss.gpu_state_data.resize(sizeof(g_vram) + sizeof(g_gpu_clut));
+  return true;
+}
+
+void GPU_SW::DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss)
+{
+  sw.DoBytes(g_vram, sizeof(g_vram));
+  sw.DoBytes(g_gpu_clut, sizeof(g_gpu_clut));
+  DebugAssert(!sw.HasError());
+}
+
+void GPU_SW::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
+{
+}
+
+void GPU_SW::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced_rendering, u8 active_line_lsb)
+{
+  GPU_SW_Rasterizer::FillVRAM(x, y, width, height, color, interlaced_rendering, active_line_lsb);
+}
+
+void GPU_SW::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask)
+{
+  GPU_SW_Rasterizer::WriteVRAM(x, y, width, height, data, set_mask, check_mask);
+}
+
+void GPU_SW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask, bool check_mask)
+{
+  GPU_SW_Rasterizer::CopyVRAM(src_x, src_y, dst_x, dst_y, width, height, set_mask, check_mask);
+}
+
+void GPU_SW::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
+    cmd->shading_enable, cmd->texture_enable, cmd->raw_texture_enable, cmd->transparency_enable);
+
+  DrawFunction(cmd, &cmd->vertices[0], &cmd->vertices[1], &cmd->vertices[2]);
+  if (cmd->num_vertices > 3)
+    DrawFunction(cmd, &cmd->vertices[2], &cmd->vertices[1], &cmd->vertices[3]);
+}
+
+void GPU_SW::DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
+    cmd->shading_enable, cmd->texture_enable, cmd->raw_texture_enable, cmd->transparency_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];
+    vertices[i] = GPUBackendDrawPolygonCommand::Vertex{
+      .x = src.native_x, .y = src.native_y, .color = src.color, .texcoord = src.texcoord};
+  }
+
+  DrawFunction(cmd, &vertices[0], &vertices[1], &vertices[2]);
+  if (cmd->num_vertices > 3)
+    DrawFunction(cmd, &vertices[2], &vertices[1], &vertices[3]);
+}
+
+void GPU_SW::DrawSprite(const GPUBackendDrawRectangleCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawRectangleFunction DrawFunction =
+    GPU_SW_Rasterizer::GetDrawRectangleFunction(cmd->texture_enable, cmd->raw_texture_enable, cmd->transparency_enable);
+
+  DrawFunction(cmd);
+}
+
+void GPU_SW::DrawLine(const GPUBackendDrawLineCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawLineFunction DrawFunction =
+    GPU_SW_Rasterizer::GetDrawLineFunction(cmd->shading_enable, cmd->transparency_enable);
+
+  for (u16 i = 0; i < cmd->num_vertices; i += 2)
+    DrawFunction(cmd, &cmd->vertices[i], &cmd->vertices[i + 1]);
+}
+
+void GPU_SW::DrawPreciseLine(const GPUBackendDrawPreciseLineCommand* cmd)
+{
+  const GPU_SW_Rasterizer::DrawLineFunction DrawFunction =
+    GPU_SW_Rasterizer::GetDrawLineFunction(cmd->shading_enable, cmd->transparency_enable);
+
+  // Need to cut out the irrelevant bits.
+  // TODO: In _theory_ we could use the fixed-point parts here.
+  for (u32 i = 0; i < cmd->num_vertices; i += 2)
+  {
+    const GPUBackendDrawPreciseLineCommand::Vertex& RESTRICT start = cmd->vertices[i];
+    const GPUBackendDrawPreciseLineCommand::Vertex& RESTRICT end = cmd->vertices[i + 1];
+    const GPUBackendDrawLineCommand::Vertex vertices[2] = {
+      {.x = start.native_x, .y = start.native_y, .color = start.color},
+      {.x = end.native_x, .y = end.native_y, .color = end.color},
+    };
+
+    DrawFunction(cmd, &vertices[0], &vertices[1]);
+  }
+}
+
+void GPU_SW::DrawingAreaChanged()
+{
+  // GPU_SW_Rasterizer::g_drawing_area set by base class.
+}
+
+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()
 {
-  GPU::UpdateSettings(old_settings);
-  if (g_settings.gpu_use_thread != old_settings.gpu_use_thread)
-    m_backend.SetThreadEnabled(g_settings.gpu_use_thread);
 }
 
 GPUTexture* GPU_SW::GetDisplayTexture(u32 width, u32 height, GPUTexture::Format format)
@@ -263,32 +374,28 @@ 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
+  if (!g_settings.gpu_show_vram)
-  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 is_24bit = cmd->display_24bit;
-    const bool interlaced = IsInterlacedDisplayEnabled();
+    const bool interlaced = cmd->interlaced_display_enabled;
-    const u32 field = GetInterlacedDisplayField();
+    const u32 field = BoolToUInt32(cmd->interlaced_display_field);
-    const u32 vram_offset_x = is_24bit ? m_crtc_state.regs.X : m_crtc_state.display_vram_left;
+    const u32 vram_offset_x = is_24bit ? cmd->X : cmd->display_vram_left;
-    const u32 vram_offset_y =
+    const u32 vram_offset_y = cmd->display_vram_top + ((interlaced && cmd->interlaced_display_interleaved) ? field : 0);
-      m_crtc_state.display_vram_top + ((interlaced && m_GPUSTAT.vertical_resolution) ? field : 0);
+    const u32 skip_x = is_24bit ? (cmd->display_vram_left - cmd->X) : 0;
-    const u32 skip_x = is_24bit ? (m_crtc_state.display_vram_left - m_crtc_state.regs.X) : 0;
+    const u32 read_width = cmd->display_vram_width;
-    const u32 read_width = m_crtc_state.display_vram_width;
+    const u32 read_height = interlaced ? (cmd->display_vram_height / 2) : cmd->display_vram_height;
-    const u32 read_height = interlaced ? (m_crtc_state.display_vram_height / 2) : m_crtc_state.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);
@@ -320,347 +427,7 @@ 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->draw_mode.dither_enable = rc.IsDitheringEnabled() && cmd->draw_mode.dither_enable;
-  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<false>(&vert->x);
-      }
-
-      // Cull polygons which are too large.
-      const GSVector2i min_pos_12 = positions[1].min_s32(positions[2]);
-      const GSVector2i max_pos_12 = positions[1].max_s32(positions[2]);
-      const GSVector4i draw_rect_012 = GSVector4i(min_pos_12.min_s32(positions[0]))
-                                         .upl64(GSVector4i(max_pos_12.max_s32(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_s32(positions[3]))
-                                           .upl64(GSVector4i(max_pos_12.max_s32(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<false>(&cmd->vertices[0].x);
-        const GSVector4i v1 = GSVector4i::loadl<false>(&cmd->vertices[1].x);
-        const GSVector4i rect = v0.min_s32(v1).xyxy(v0.max_s32(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 - 1) * 2);
-        FillDrawCommand(cmd, m_render_command);
-
-        u32 buffer_pos = 0;
-        const GPUVertexPosition start_vp{m_blit_buffer[buffer_pos++]};
-        const GSVector2i draw_offset = GSVector2i::load<false>(&m_drawing_offset.x);
-        GSVector2i start_pos = GSVector2i(start_vp.x, start_vp.y).add32(draw_offset);
-        u32 start_color = m_render_command.color_for_first_vertex;
-
-        const bool shaded = m_render_command.shading_enable;
-        u32 out_vertex_count = 0;
-        for (u32 i = 1; i < num_vertices; i++)
-        {
-          const u32 end_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++]};
-          const GSVector2i end_pos = GSVector2i(vp.x, vp.y).add32(draw_offset);
-
-          const GSVector4i rect = GSVector4i::xyxy(start_pos.min_s32(end_pos), start_pos.max_s32(end_pos))
-                                    .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);
-          }
-          else
-          {
-            AddDrawLineTicks(clamped_rect, rc.shading_enable);
-
-            GPUBackendDrawLineCommand::Vertex* out_vertex = &cmd->vertices[out_vertex_count];
-            out_vertex_count += 2;
-
-            GSVector2i::store<false>(&out_vertex[0].x, start_pos);
-            out_vertex[0].color = start_color;
-            GSVector2i::store<false>(&out_vertex[1].x, end_pos);
-            out_vertex[1].color = end_color;
-          }
-
-          start_pos = end_pos;
-          start_color = end_color;
-        }
-
-        if (out_vertex_count > 0)
-        {
-          DebugAssert(out_vertex_count <= cmd->num_vertices);
-          cmd->num_vertices = Truncate16(out_vertex_count);
-          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()
 {
   return std::make_unique<GPU_SW>();
 }

src/core/gpu_sw.h

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "gpu.h"
-#include "gpu_sw_backend.h"
+#include "gpu_backend.h"
 
 #include "util/gpu_device.h"
 
@@ -12,35 +12,52 @@
 
 #include <memory>
 
-namespace Threading {
+// TODO: Move to cpp
-class Thread;
+// TODO: Rename to GPUSWBackend, preserved to avoid conflicts.
-}
+class GPU_SW final : public GPUBackend
-
-class GPUTexture;
-
-class GPU_SW final : public GPU
 {
 public:
   GPU_SW();
   ~GPU_SW() override;
 
-  ALWAYS_INLINE const GPU_SW_Backend& GetBackend() const { return m_backend; }
+  bool Initialize(bool upload_vram, Error* error) override;
 
-  const Threading::Thread* GetSWThread() const override;
+  void RestoreDeviceContext() override;
-  bool IsHardwareRenderer() const override;
 
-  bool Initialize(Error* error) override;
+  u32 GetResolutionScale() const 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;
 
 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 FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color, bool interlaced_rendering, u8 active_line_lsb) 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 CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask,
-  void FlushRender() override;
+                bool check_mask) override;
+
+  void DrawPolygon(const GPUBackendDrawPolygonCommand* cmd) override;
+  void DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd) override;
+  void DrawLine(const GPUBackendDrawLineCommand* cmd) override;
+  void DrawPreciseLine(const GPUBackendDrawPreciseLineCommand* cmd) override;
+  void DrawSprite(const GPUBackendDrawRectangleCommand* cmd) override;
+  void DrawingAreaChanged() override;
+  void ClearCache() override;
   void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
+  void OnBufferSwapped() override;
+
+  void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) override;
+
+  void ClearVRAM() override;
+
+  void FlushRender() override;
+
+  void UpdateResolutionScale() override;
+
+  void LoadState(const GPUBackendLoadStateCommand* cmd) override;
+
+  bool AllocateMemorySaveState(System::MemorySaveState& mss, Error* error) override;
+  void DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss) override;
+
+private:
+  static constexpr GPUTexture::Format FORMAT_FOR_24BIT = GPUTexture::Format::RGBA8; // RGBA8 always supported.
 
   template<GPUTexture::Format display_format>
   bool CopyOut15Bit(u32 src_x, u32 src_y, u32 width, u32 height, u32 line_skip);
@@ -49,21 +66,9 @@ 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;
-
-private:
-  static constexpr GPUTexture::Format FORMAT_FOR_24BIT = GPUTexture::Format::RGBA8; // RGBA8 always supported.
-
   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::Unknown;
   std::unique_ptr<GPUTexture> m_upload_texture;
-
-  GPU_SW_Backend m_backend;
 };

src/core/gpu_sw_backend.cpp

@@ -1,88 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: 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 use_thread)
-{
-  return GPUBackend::Initialize(use_thread);
-}
-
-void GPU_SW_Backend::Reset()
-{
-  GPUBackend::Reset();
-}
-
-void GPU_SW_Backend::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)
-{
-  const GPURenderCommand rc{cmd->rc.bits};
-
-  const GPU_SW_Rasterizer::DrawTriangleFunction DrawFunction = GPU_SW_Rasterizer::GetDrawTriangleFunction(
-    rc.shading_enable, rc.texture_enable, rc.raw_texture_enable, rc.transparency_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);
-
-  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)
-{
-  GPU_SW_Rasterizer::FillVRAM(x, y, width, height, color, params.interlaced_rendering, params.active_line_lsb);
-}
-
-void GPU_SW_Backend::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data,
-                                GPUBackendCommandParameters params)
-{
-  GPU_SW_Rasterizer::WriteVRAM(x, y, width, height, data, params.set_mask_while_drawing, params.check_mask_before_draw);
-}
-
-void GPU_SW_Backend::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height,
-                              GPUBackendCommandParameters params)
-{
-  GPU_SW_Rasterizer::CopyVRAM(src_x, src_y, dst_x, dst_y, width, height, params.set_mask_while_drawing,
-                              params.check_mask_before_draw);
-}
-
-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

@@ -1,32 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: CC-BY-NC-ND-4.0
-
-#pragma once
-
-#include "gpu.h"
-#include "gpu_backend.h"
-
-#include <array>
-
-class GPU_SW_Backend final : public GPUBackend
-{
-public:
-  GPU_SW_Backend();
-  ~GPU_SW_Backend() override;
-
-  bool Initialize(bool use_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

@@ -38,6 +38,31 @@ CopyVRAMFunction CopyVRAM = nullptr;
 GPUDrawingArea g_drawing_area = {};
 } // namespace GPU_SW_Rasterizer
 
+void GPU_SW_Rasterizer::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  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(g_gpu_clut, &src_row[start_x], sizeof(u16) * end);
+      std::memcpy(g_gpu_clut + end, src_row, sizeof(u16) * start);
+    }
+    else
+    {
+      std::memcpy(g_gpu_clut, &src_row[start_x], sizeof(u16) * 256);
+    }
+  }
+}
+
 // Default scalar implementation definitions.
 namespace GPU_SW_Rasterizer::Scalar {
 namespace {

src/core/gpu_sw_rasterizer.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "gpu.h"
+#include "gpu_thread_commands.h"
 #include "gpu_types.h"
 
 #include "common/intrin.h"
@@ -18,18 +19,20 @@ 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* v0,
                                       const GPUBackendDrawPolygonCommand::Vertex* v1,
                                       const GPUBackendDrawPolygonCommand::Vertex* v2);
 typedef const DrawTriangleFunction DrawTriangleFunctionTable[2][2][2][2];
 
-using DrawLineFunction = void (*)(const GPUBackendDrawLineCommand* cmd, const GPUBackendDrawLineCommand::Vertex* p0,
+using DrawLineFunction = void (*)(const GPUBackendDrawCommand* cmd, const GPUBackendDrawLineCommand::Vertex* p0,
                                   const GPUBackendDrawLineCommand::Vertex* p1);
 typedef const DrawLineFunction DrawLineFunctionTable[2][2];
 

src/core/gpu_sw_rasterizer.inl

@@ -129,7 +129,7 @@ template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
     }
     else
     {
-      const bool dithering_enable = cmd->draw_mode.dither_enable;
+      const bool dithering_enable = cmd->dither_enable;
       const u32 dither_y = (dithering_enable) ? (y & 3u) : 2u;
       const u32 dither_x = (dithering_enable) ? (x & 3u) : 3u;
 
@@ -143,7 +143,7 @@ template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
   }
   else
   {
-    const bool dithering_enable = cmd->draw_mode.dither_enable;
+    const bool dithering_enable = cmd->dither_enable;
     const u32 dither_y = (dithering_enable) ? (y & 3u) : 2u;
     const u32 dither_x = (dithering_enable) ? (x & 3u) : 3u;
 
@@ -215,12 +215,12 @@ template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
     }
   }
 
-  const u16 mask_and = cmd->params.GetMaskAND();
+  const u16 mask_and = cmd->GetMaskAND();
   if ((bg_color & mask_and) != 0)
     return;
 
   DebugAssert(static_cast<u32>(x) < VRAM_WIDTH && static_cast<u32>(y) < VRAM_HEIGHT);
-  SetPixel(static_cast<u32>(x), static_cast<u32>(y), color | cmd->params.GetMaskOR());
+  SetPixel(static_cast<u32>(x), static_cast<u32>(y), color | cmd->GetMaskOR());
 }
 
 #ifndef USE_VECTOR
@@ -237,7 +237,8 @@ static void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd)
   {
     const s32 y = origin_y + static_cast<s32>(offset_y);
     if (y < static_cast<s32>(g_drawing_area.top) || y > static_cast<s32>(g_drawing_area.bottom) ||
-        (cmd->params.interlaced_rendering && cmd->params.active_line_lsb == (Truncate8(static_cast<u32>(y)) & 1u)))
+        (cmd->interlaced_rendering &&
+         cmd->active_line_lsb == ConvertToBoolUnchecked(Truncate8(static_cast<u32>(y)) & 1u)))
     {
       continue;
     }
@@ -488,8 +489,8 @@ struct PixelVectors
     clip_left = GSVectorNi(g_drawing_area.left);
     clip_right = GSVectorNi(g_drawing_area.right);
 
-    mask_and = GSVectorNi(cmd->params.GetMaskAND());
+    mask_and = GSVectorNi(cmd->GetMaskAND());
-    mask_or = GSVectorNi(cmd->params.GetMaskOR());
+    mask_or = GSVectorNi(cmd->GetMaskOR());
 
     if constexpr (texture_enable)
     {
@@ -505,10 +506,11 @@ struct PixelVectors
 } // namespace
 
 template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-ALWAYS_INLINE_RELEASE static void
+ALWAYS_INLINE_RELEASE static void ShadePixel(const PixelVectors<texture_enable>& RESTRICT pv,
-ShadePixel(const PixelVectors<texture_enable>& pv, GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode,
+                                             GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode,
-           u32 start_x, u32 y, GSVectorNi vertex_color_rg, GSVectorNi vertex_color_ba, GSVectorNi texcoord_x,
+                                             u32 start_x, u32 y, GSVectorNi vertex_color_rg, GSVectorNi vertex_color_ba,
-           GSVectorNi texcoord_y, GSVectorNi preserve_mask, GSVectorNi dither)
+                                             GSVectorNi texcoord_x, GSVectorNi texcoord_y, GSVectorNi preserve_mask,
+                                             GSVectorNi dither)
 {
   static constexpr GSVectorNi coord_mask_x = GSVectorNi::cxpr(VRAM_WIDTH_MASK);
   static constexpr GSVectorNi coord_mask_y = GSVectorNi::cxpr(VRAM_HEIGHT_MASK);
@@ -692,7 +694,7 @@ ShadePixel(const PixelVectors<texture_enable>& pv, GPUTextureMode texture_mode,
 }
 
 template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-static void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd)
+static void DrawRectangle(const GPUBackendDrawRectangleCommand* RESTRICT cmd)
 {
   const s32 origin_x = cmd->x;
   const s32 origin_y = cmd->y;
@@ -717,7 +719,8 @@ static void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd)
   {
     const s32 y = origin_y + static_cast<s32>(offset_y);
     if (y >= static_cast<s32>(g_drawing_area.top) && y <= static_cast<s32>(g_drawing_area.bottom) &&
-        (!cmd->params.interlaced_rendering || cmd->params.active_line_lsb != (Truncate8(static_cast<u32>(y)) & 1u)))
+        (!cmd->interlaced_rendering ||
+         cmd->active_line_lsb != ConvertToBoolUnchecked(Truncate8(static_cast<u32>(y)) & 1u)))
     {
       const s32 draw_y = (y & VRAM_HEIGHT_MASK);
 
@@ -763,8 +766,8 @@ static void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd)
 
 // TODO: Vectorize line draw.
 template<bool shading_enable, bool transparency_enable>
-static void DrawLine(const GPUBackendDrawLineCommand* cmd, const GPUBackendDrawLineCommand::Vertex* p0,
+static void DrawLine(const GPUBackendDrawCommand* RESTRICT cmd, const GPUBackendDrawLineCommand::Vertex* RESTRICT p0,
-                     const GPUBackendDrawLineCommand::Vertex* p1)
+                     const GPUBackendDrawLineCommand::Vertex* RESTRICT p1)
 {
   static constexpr u32 XY_SHIFT = 32;
   static constexpr u32 RGB_SHIFT = 12;
@@ -817,7 +820,8 @@ static void DrawLine(const GPUBackendDrawLineCommand* cmd, const GPUBackendDrawL
     const s32 x = unfp_xy(curx);
     const s32 y = unfp_xy(cury);
 
-    if ((!cmd->params.interlaced_rendering || cmd->params.active_line_lsb != (Truncate8(static_cast<u32>(y)) & 1u)) &&
+    if ((!cmd->interlaced_rendering ||
+         cmd->active_line_lsb != ConvertToBoolUnchecked(Truncate8(static_cast<u32>(y)) & 1u)) &&
         x >= static_cast<s32>(g_drawing_area.left) && x <= static_cast<s32>(g_drawing_area.right) &&
         y >= static_cast<s32>(g_drawing_area.top) && y <= static_cast<s32>(g_drawing_area.bottom))
     {
@@ -968,8 +972,8 @@ struct TrianglePart
 #ifndef USE_VECTOR
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-static void DrawSpan(const GPUBackendDrawPolygonCommand* cmd, s32 y, s32 x_start, s32 x_bound, UVStepper uv,
+static void DrawSpan(const GPUBackendDrawCommand* RESTRICT cmd, s32 y, s32 x_start, s32 x_bound, UVStepper uv,
-                     const UVSteps& uvstep, RGBStepper rgb, const RGBSteps& rgbstep)
+                     const UVSteps& RESTRICT uvstep, RGBStepper rgb, const RGBSteps& RESTRICT rgbstep)
 {
   s32 width = x_bound - x_start;
   s32 current_x = TruncateGPUVertexPosition(x_start);
@@ -1008,9 +1012,10 @@ static void DrawSpan(const GPUBackendDrawPolygonCommand* cmd, s32 y, s32 x_start
 }
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCommand* cmd, const TrianglePart& tp,
+ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawCommand* RESTRICT cmd,
-                                                   const UVStepper& uv, const UVSteps& uvstep, const RGBStepper& rgb,
+                                                   const TrianglePart& RESTRICT tp, const UVStepper& RESTRICT uv,
-                                                   const RGBSteps& rgbstep)
+                                                   const UVSteps& RESTRICT uvstep, const RGBStepper& RESTRICT rgb,
+                                                   const RGBSteps& RESTRICT rgbstep)
 {
   static constexpr auto unfp_xy = [](s64 xfp) -> s32 { return static_cast<s32>(static_cast<u64>(xfp) >> 32); };
 
@@ -1051,7 +1056,8 @@ ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCo
         lrgb.StepY<true>(rgbstep);
 
       if (y > static_cast<s32>(g_drawing_area.bottom) ||
-          (cmd->params.interlaced_rendering && cmd->params.active_line_lsb == (static_cast<u32>(current_y) & 1u)))
+          (cmd->interlaced_rendering &&
+           cmd->active_line_lsb == ConvertToBoolUnchecked(static_cast<u32>(current_y) & 1u)))
       {
         continue;
       }
@@ -1082,7 +1088,8 @@ ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCo
         break;
       }
       if (y >= static_cast<s32>(g_drawing_area.top) &&
-          (!cmd->params.interlaced_rendering || cmd->params.active_line_lsb != (static_cast<u32>(current_y) & 1u)))
+          (!cmd->interlaced_rendering ||
+           cmd->active_line_lsb != ConvertToBoolUnchecked(static_cast<u32>(current_y) & 1u)))
       {
         DrawSpan<shading_enable, texture_enable, raw_texture_enable, transparency_enable>(
           cmd, y & VRAM_HEIGHT_MASK, unfp_xy(left_x), unfp_xy(right_x), luv, uvstep, lrgb, rgbstep);
@@ -1145,9 +1152,10 @@ struct TriangleVectors : PixelVectors<texture_enable>
 } // namespace
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_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* RESTRICT cmd, s32 y, s32 x_start, s32 x_bound,
-                                           UVStepper uv, const UVSteps& uvstep, RGBStepper rgb, const RGBSteps& rgbstep,
+                                           UVStepper uv, const UVSteps& RESTRICT uvstep, RGBStepper rgb,
-                                           const TriangleVectors<shading_enable, texture_enable>& tv)
+                                           const RGBSteps& RESTRICT rgbstep,
+                                           const TriangleVectors<shading_enable, texture_enable>& RESTRICT tv)
 {
   s32 width = x_bound - x_start;
   s32 current_x = TruncateGPUVertexPosition(x_start);
@@ -1195,7 +1203,7 @@ ALWAYS_INLINE_RELEASE static void DrawSpan(const GPUBackendDrawPolygonCommand* c
     dv = GSVectorNi::zero();
   }
 
-  const GSVectorNi dither = cmd->draw_mode.dither_enable ?
+  const GSVectorNi dither = cmd->dither_enable ?
                               GSVectorNi::broadcast128<false>(
                                 &VECTOR_DITHER_MATRIX[static_cast<u32>(y) & 3][(static_cast<u32>(current_x) & 3) * 2]) :
                               GSVectorNi::zero();
@@ -1250,9 +1258,10 @@ ALWAYS_INLINE_RELEASE static void DrawSpan(const GPUBackendDrawPolygonCommand* c
 }
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCommand* cmd, const TrianglePart& tp,
+ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawCommand* RESTRICT cmd,
-                                                   const UVStepper& uv, const UVSteps& uvstep, const RGBStepper& rgb,
+                                                   const TrianglePart& RESTRICT tp, const UVStepper& RESTRICT uv,
-                                                   const RGBSteps& rgbstep)
+                                                   const UVSteps& RESTRICT uvstep, const RGBStepper& RESTRICT rgb,
+                                                   const RGBSteps& RESTRICT rgbstep)
 {
   static constexpr auto unfp_xy = [](s64 xfp) -> s32 { return static_cast<s32>(static_cast<u64>(xfp) >> 32); };
 
@@ -1295,7 +1304,8 @@ ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCo
         lrgb.StepY<true>(rgbstep);
 
       if (y > static_cast<s32>(g_drawing_area.bottom) ||
-          (cmd->params.interlaced_rendering && cmd->params.active_line_lsb == (static_cast<u32>(current_y) & 1u)))
+          (cmd->interlaced_rendering &&
+           cmd->active_line_lsb == ConvertToBoolUnchecked(static_cast<u32>(current_y) & 1u)))
       {
         continue;
       }
@@ -1328,7 +1338,8 @@ ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCo
         break;
       }
       if (y >= static_cast<s32>(g_drawing_area.top) &&
-          (!cmd->params.interlaced_rendering || cmd->params.active_line_lsb != (static_cast<u32>(current_y) & 1u)))
+          (!cmd->interlaced_rendering ||
+           cmd->active_line_lsb != ConvertToBoolUnchecked(static_cast<u32>(current_y) & 1u)))
       {
         DrawSpan<shading_enable, texture_enable, raw_texture_enable, transparency_enable>(
           cmd, y & VRAM_HEIGHT_MASK, unfp_xy(left_x), unfp_xy(right_x), luv, uvstep, lrgb, rgbstep, tv);
@@ -1349,13 +1360,15 @@ ALWAYS_INLINE_RELEASE static void DrawTrianglePart(const GPUBackendDrawPolygonCo
 #endif // USE_VECTOR
 
 template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable>
-static void DrawTriangle(const GPUBackendDrawPolygonCommand* cmd, const GPUBackendDrawPolygonCommand::Vertex* v0,
+static void DrawTriangle(const GPUBackendDrawCommand* RESTRICT cmd,
-                         const GPUBackendDrawPolygonCommand::Vertex* v1, const GPUBackendDrawPolygonCommand::Vertex* v2)
+                         const GPUBackendDrawPolygonCommand::Vertex* RESTRICT v0,
+                         const GPUBackendDrawPolygonCommand::Vertex* RESTRICT v1,
+                         const GPUBackendDrawPolygonCommand::Vertex* RESTRICT v2)
 {
 #ifdef CHECK_VECTOR
-  const GPUBackendDrawPolygonCommand::Vertex* orig_v0 = v0;
+  const GPUBackendDrawPolygonCommand::Vertex* RESTRICT orig_v0 = v0;
-  const GPUBackendDrawPolygonCommand::Vertex* orig_v1 = v1;
+  const GPUBackendDrawPolygonCommand::Vertex* RESTRICT orig_v1 = v1;
-  const GPUBackendDrawPolygonCommand::Vertex* orig_v2 = v2;
+  const GPUBackendDrawPolygonCommand::Vertex* RESTRICT orig_v2 = v2;
 #endif
 
   // Sort vertices so that v0 is the top vertex, v1 is the bottom vertex, and v2 is the side vertex.
@@ -1410,8 +1423,8 @@ static void DrawTriangle(const GPUBackendDrawPolygonCommand* cmd, const GPUBacke
   const u32 ofi = BoolToUInt32(!right_facing);
 
   TrianglePart triparts[2];
-  TrianglePart& tpo = triparts[vo];
+  TrianglePart& RESTRICT tpo = triparts[vo];
-  TrianglePart& tpp = triparts[vo ^ 1];
+  TrianglePart& RESTRICT tpp = triparts[vo ^ 1];
   tpo.start_y = vertices[0 ^ vo]->y;
   tpo.end_y = vertices[1 ^ vo]->y;
   tpp.start_y = vertices[1 ^ vp]->y;
@@ -1462,7 +1475,7 @@ static void DrawTriangle(const GPUBackendDrawPolygonCommand* cmd, const GPUBacke
   // Undo the start of the vertex, so that when we add the offset for each line, it starts at the beginning value.
   UVStepper uv;
   RGBStepper rgb;
-  const GPUBackendDrawPolygonCommand::Vertex* top_left_vertex = vertices[tl];
+  const GPUBackendDrawPolygonCommand::Vertex* RESTRICT top_left_vertex = vertices[tl];
   if constexpr (texture_enable)
   {
     uv.Init(top_left_vertex->u, top_left_vertex->v);
@@ -1535,7 +1548,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
     {
       const u32 row = (y + yoffs) % VRAM_HEIGHT;
 
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
+      u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH + x];
       u32 xoffs = 0;
       for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
         GSVector4i::store<false>(row_ptr, fill);
@@ -1556,7 +1569,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
         if ((row & u32(1)) == active_field)
           continue;
 
-        u16* row_ptr = &g_vram[row * VRAM_WIDTH + x];
+        u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH + x];
         u32 xoffs = 0;
         for (; xoffs < aligned_width; xoffs += vector_width, row_ptr += vector_width)
           GSVector4i::store<false>(row_ptr, fill);
@@ -1572,7 +1585,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
         if ((row & u32(1)) == active_field)
           continue;
 
-        u16* row_ptr = &g_vram[row * VRAM_WIDTH];
+        u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH];
         for (u32 xoffs = 0; xoffs < width; xoffs++)
         {
           const u32 col = (x + xoffs) % VRAM_WIDTH;
@@ -1586,7 +1599,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
     for (u32 yoffs = 0; yoffs < height; yoffs++)
     {
       const u32 row = (y + yoffs) % VRAM_HEIGHT;
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH];
+      u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH];
       for (u32 xoffs = 0; xoffs < width; xoffs++)
       {
         const u32 col = (x + xoffs) % VRAM_WIDTH;
@@ -1615,7 +1628,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
       if ((row & u32(1)) == active_field)
         continue;
 
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH];
+      u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH];
       for (u32 xoffs = 0; xoffs < width; xoffs++)
       {
         const u32 col = (x + xoffs) % VRAM_WIDTH;
@@ -1628,7 +1641,7 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
     for (u32 yoffs = 0; yoffs < height; yoffs++)
     {
       const u32 row = (y + yoffs) % VRAM_HEIGHT;
-      u16* row_ptr = &g_vram[row * VRAM_WIDTH];
+      u16* RESTRICT row_ptr = &g_vram[row * VRAM_WIDTH];
       for (u32 xoffs = 0; xoffs < width; xoffs++)
       {
         const u32 col = (x + xoffs) % VRAM_WIDTH;
@@ -1639,12 +1652,13 @@ static void FillVRAMImpl(u32 x, u32 y, u32 width, u32 height, u32 color, bool in
 #endif
 }
 
-static void WriteVRAMImpl(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask)
+static void WriteVRAMImpl(u32 x, u32 y, u32 width, u32 height, const void* RESTRICT data, bool set_mask,
+                          bool check_mask)
 {
   // Fast path when the copy is not oversized.
   if ((x + width) <= VRAM_WIDTH && (y + height) <= VRAM_HEIGHT && !set_mask && !check_mask)
   {
-    const u16* src_ptr = static_cast<const u16*>(data);
+    const u16* RESTRICT src_ptr = static_cast<const u16*>(data);
     u16* dst_ptr = &g_vram[y * VRAM_WIDTH + x];
     for (u32 yoffs = 0; yoffs < height; yoffs++)
     {
@@ -1657,7 +1671,7 @@ static void WriteVRAMImpl(u32 x, u32 y, u32 width, u32 height, const void* data,
   {
     // Slow path when we need to handle wrap-around.
     // During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
-    const u16* src_ptr = static_cast<const u16*>(data);
+    const u16* RESTRICT src_ptr = static_cast<const u16*>(data);
     const u16 mask_and = check_mask ? 0x8000u : 0x0000u;
     const u16 mask_or = set_mask ? 0x8000u : 0x0000u;
 
@@ -1706,7 +1720,7 @@ static void WriteVRAMImpl(u32 x, u32 y, u32 width, u32 height, const void* data,
       for (; col < width;)
       {
         // TODO: Handle unaligned reads...
-        u16* pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
+        u16* RESTRICT pixel_ptr = &dst_row_ptr[(x + col++) % VRAM_WIDTH];
         if (((*pixel_ptr) & mask_and) == 0)
           *pixel_ptr = *(src_ptr++) | mask_or;
       }

src/core/gpu_thread.h

@@ -0,0 +1,106 @@
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: CC-BY-NC-ND-4.0
+
+#pragma once
+
+#include "common/types.h"
+
+#include <functional>
+#include <optional>
+
+class Error;
+struct WindowInfo;
+
+namespace Threading {
+class ThreadHandle;
+}
+
+enum class RenderAPI : u8;
+enum class GPUVSyncMode : u8;
+
+enum class GPURenderer : u8;
+enum class GPUBackendCommandType : u8;
+
+class GPUBackend;
+struct GPUThreadCommand;
+struct GPUBackendUpdateDisplayCommand;
+
+namespace GPUThread {
+using AsyncCallType = std::function<void()>;
+using AsyncBackendCallType = std::function<void(GPUBackend*)>;
+
+enum class RunIdleReason : u8
+{
+  NoGPUBackend = (1 << 0),
+  SystemPaused = (1 << 1),
+  FullscreenUIActive = (1 << 2),
+  LoadingScreenActive = (1 << 3),
+};
+
+/// Starts Big Picture UI.
+bool StartFullscreenUI(bool fullscreen, Error* error);
+bool IsFullscreenUIRequested();
+void StopFullscreenUI();
+
+/// Backend control.
+std::optional<GPURenderer> GetRequestedRenderer();
+bool CreateGPUBackend(std::string serial, GPURenderer renderer, bool upload_vram, bool fullscreen,
+                      bool force_recreate_device, Error* error);
+void DestroyGPUBackend();
+bool HasGPUBackend();
+bool IsGPUBackendRequested();
+void SetGameSerial(std::string serial);
+
+/// 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(bool fullscreen);
+
+/// Called when the window is resized.
+void ResizeDisplayWindow(s32 width, s32 height, float scale);
+
+/// Access to main window size from CPU thread.
+const WindowInfo& GetRenderWindowInfo();
+
+void UpdateSettings(bool gpu_settings_changed, bool device_settings_changed);
+
+bool IsOnThread();
+bool IsUsingThread();
+void RunOnThread(AsyncCallType func);
+void RunOnBackend(AsyncBackendCallType func, bool sync, bool spin_or_wake);
+void SetVSync(GPUVSyncMode mode, bool allow_present_throttle);
+
+// Should only be called on the GPU thread.
+bool GetRunIdleReason(RunIdleReason reason);
+void SetRunIdleReason(RunIdleReason reason, bool enabled);
+bool IsRunningIdle();
+bool IsSystemPaused();
+const std::string& GetGameSerial();
+
+GPUThreadCommand* AllocateCommand(GPUBackendCommandType command, u32 size);
+void PushCommand(GPUThreadCommand* cmd);
+void PushCommandAndWakeThread(GPUThreadCommand* cmd);
+void PushCommandAndSync(GPUThreadCommand* cmd, bool spin);
+void SyncGPUThread(bool spin);
+
+// NOTE: Only called by GPUBackend
+namespace Internal {
+const Threading::ThreadHandle& GetThreadHandle();
+void ProcessStartup();
+void SetThreadEnabled(bool enabled);
+void DoRunIdle();
+void RequestShutdown();
+void GPUThreadEntryPoint();
+void PresentFrame(bool allow_skip_present, u64 present_time);
+void RestoreContextAfterPresent();
+} // namespace Internal
+} // namespace GPUThread
+
+namespace Host {
+
+/// Called when the pause state changes, or fullscreen UI opens.
+void OnGPUThreadRunIdleChanged(bool is_active);
+
+} // namespace Host

src/core/gpu_thread_commands.h

@@ -0,0 +1,335 @@
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: CC-BY-NC-ND-4.0
+
+#pragma once
+
+#include "gpu_types.h"
+
+#include "common/align.h"
+
+#include <functional>
+#include <string>
+#include <vector>
+
+class Error;
+
+enum class GPUVSyncMode : u8;
+class MediaCapture;
+class StateWrapper;
+
+class GPUBackend;
+
+namespace System {
+struct MemorySaveState;
+}
+
+#ifdef _MSC_VER
+#pragma warning(push)
+#pragma warning(disable : 4200) // warning C4200: nonstandard extension used: zero-sized array in struct/union
+#endif
+
+enum class GPUBackendCommandType : u8
+{
+  Wraparound,
+  AsyncCall,
+  AsyncBackendCall,
+  Reconfigure,
+  Shutdown,
+  ClearVRAM,
+  ClearDisplay,
+  UpdateDisplay,
+  SubmitFrame,
+  BufferSwapped,
+  LoadState,
+  LoadMemoryState,
+  SaveMemoryState,
+  ReadVRAM,
+  FillVRAM,
+  UpdateVRAM,
+  CopyVRAM,
+  SetDrawingArea,
+  UpdateCLUT,
+  ClearCache,
+  DrawPolygon,
+  DrawPrecisePolygon,
+  DrawRectangle,
+  DrawLine,
+  DrawPreciseLine,
+};
+
+struct GPUThreadCommand
+{
+  u32 size;
+  GPUBackendCommandType type;
+
+  static constexpr u32 AlignCommandSize(u32 size)
+  {
+    // Ensure size is a multiple of 8 (minimum data size) so we don't end up with an unaligned command.
+    // NOTE: If we ever end up putting vectors in the command packets, this should be raised.
+    constexpr u32 COMMAND_QUEUE_ALLOCATION_ALIGNMENT = 8;
+    return Common::AlignUpPow2(size, COMMAND_QUEUE_ALLOCATION_ALIGNMENT);
+  }
+};
+
+struct GPUThreadReconfigureCommand : public GPUThreadCommand
+{
+  Error* error_ptr;
+  bool* out_result;
+  std::string game_serial;
+  std::optional<GPURenderer> renderer;
+  std::optional<bool> fullscreen;
+  std::optional<bool> start_fullscreen_ui;
+  GPUVSyncMode vsync_mode;
+  bool allow_present_throttle;
+  bool force_recreate_device;
+  bool upload_vram;
+};
+
+struct GPUThreadAsyncCallCommand : public GPUThreadCommand
+{
+  GPUThreadAsyncCallCommand(std::function<void()> func_) : func(std::move(func_)) {}
+
+  std::function<void()> func;
+};
+
+struct GPUThreadAsyncBackendCallCommand : public GPUThreadCommand
+{
+  GPUThreadAsyncBackendCallCommand(std::function<void(GPUBackend*)> func_) : func(std::move(func_)) {}
+
+  std::function<void(GPUBackend*)> func;
+};
+
+struct GPUBackendLoadStateCommand : public GPUThreadCommand
+{
+  u16 vram_data[VRAM_WIDTH * VRAM_HEIGHT];
+  u16 clut_data[GPU_CLUT_SIZE];
+  u32 texture_cache_state_version;
+  u32 texture_cache_state_size;
+  u8 texture_cache_state[0]; // texture_cache_state_size
+};
+
+struct GPUBackendDoMemoryStateCommand : public GPUThreadCommand
+{
+  System::MemorySaveState* memory_save_state;
+};
+
+struct GPUBackendFramePresentationParameters
+{
+  u32 frame_number;
+  u32 internal_frame_number;
+
+  u64 present_time;
+  MediaCapture* media_capture;
+
+  union
+  {
+    u8 bits;
+
+    BitField<u16, bool, 0, 1> allow_present_skip;
+    BitField<u16, bool, 1, 1> present_frame;
+    BitField<u16, bool, 2, 1> update_performance_counters;
+  };
+};
+
+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;
+  float display_pixel_aspect_ratio;
+
+  u16 X; // TODO: Can we get rid of this?
+
+  bool interlaced_display_enabled : 1;
+  bool interlaced_display_field : 1;
+  bool interlaced_display_interleaved : 1;
+  bool display_24bit : 1;
+  bool display_disabled : 1;
+  bool submit_frame : 1;
+  bool : 2;
+
+  GPUBackendFramePresentationParameters frame;
+};
+
+// Only used for runahead.
+struct GPUBackendSubmitFrameCommand : public GPUThreadCommand
+{
+  GPUBackendFramePresentationParameters frame;
+};
+
+struct GPUBackendReadVRAMCommand : public GPUThreadCommand
+{
+  u16 x;
+  u16 y;
+  u16 width;
+  u16 height;
+};
+
+struct GPUBackendFillVRAMCommand : public GPUThreadCommand
+{
+  u16 x;
+  u16 y;
+  u16 width;
+  u16 height;
+  u32 color;
+  bool interlaced_rendering;
+  u8 active_line_lsb;
+};
+
+struct GPUBackendUpdateVRAMCommand : public GPUThreadCommand
+{
+  u16 x;
+  u16 y;
+  u16 width;
+  u16 height;
+  bool set_mask_while_drawing;
+  bool check_mask_before_draw;
+  u16 data[0];
+};
+
+struct GPUBackendCopyVRAMCommand : public GPUThreadCommand
+{
+  u16 src_x;
+  u16 src_y;
+  u16 dst_x;
+  u16 dst_y;
+  u16 width;
+  u16 height;
+  bool set_mask_while_drawing;
+  bool check_mask_before_draw;
+};
+
+struct GPUBackendSetDrawingAreaCommand : public GPUThreadCommand
+{
+  GPUDrawingArea new_area;
+};
+
+struct GPUBackendUpdateCLUTCommand : public GPUThreadCommand
+{
+  GPUTexturePaletteReg reg;
+  bool clut_is_8bit;
+};
+
+struct GPUBackendDrawCommand : public GPUThreadCommand
+{
+  bool interlaced_rendering : 1;
+
+  /// Returns 0 if the currently-displayed field is on an even line in VRAM, otherwise 1.
+  bool active_line_lsb : 1;
+
+  bool set_mask_while_drawing : 1;
+  bool check_mask_before_draw : 1;
+
+  bool texture_enable : 1;
+  bool raw_texture_enable : 1;
+  bool transparency_enable : 1;
+  bool shading_enable : 1;
+  bool quad_polygon : 1;
+  bool dither_enable : 1;
+
+  bool valid_w : 1; // only used for precise polygons
+
+  // During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
+  ALWAYS_INLINE u16 GetMaskAND() const { return check_mask_before_draw ? 0x8000 : 0x0000; }
+  ALWAYS_INLINE u16 GetMaskOR() const { return set_mask_while_drawing ? 0x8000 : 0x0000; }
+
+  u16 num_vertices;
+  GPUDrawModeReg draw_mode;
+  GPUTexturePaletteReg palette;
+  GPUTextureWindow window;
+};
+
+struct GPUBackendDrawPolygonCommand : public GPUBackendDrawCommand
+{
+  struct Vertex
+  {
+    s32 x, y;
+    union
+    {
+      struct
+      {
+        u8 r, g, b, a;
+      };
+      u32 color;
+    };
+    union
+    {
+      struct
+      {
+        u8 u, v;
+      };
+      u16 texcoord;
+    };
+  };
+
+  Vertex vertices[0];
+};
+
+struct GPUBackendDrawPrecisePolygonCommand : public GPUBackendDrawCommand
+{
+  GPUBackendDrawCommand params;
+
+  struct Vertex
+  {
+    float x, y, w;
+    s32 native_x, native_y;
+    u32 color;
+    u16 texcoord;
+  };
+
+  Vertex vertices[0];
+};
+
+struct GPUBackendDrawRectangleCommand : public GPUBackendDrawCommand
+{
+  u16 width, height;
+  u16 texcoord;
+  s32 x, y;
+  u32 color;
+};
+
+struct GPUBackendDrawLineCommand : public GPUBackendDrawCommand
+{
+  struct Vertex
+  {
+    s32 x, y;
+    union
+    {
+      struct
+      {
+        u8 r, g, b, a;
+      };
+      u32 color;
+    };
+
+    ALWAYS_INLINE void Set(s32 x_, s32 y_, u32 color_)
+    {
+      x = x_;
+      y = y_;
+      color = color_;
+    }
+  };
+
+  Vertex vertices[0];
+};
+
+struct GPUBackendDrawPreciseLineCommand : public GPUBackendDrawCommand
+{
+  struct Vertex
+  {
+    float x, y, w;
+    s32 native_x, native_y;
+    u32 color;
+  };
+
+  Vertex vertices[0];
+};
+
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif

src/core/gpu_types.h

@@ -12,7 +12,6 @@
 #include "common/gsvector.h"
 
 #include <array>
-#include <string>
 
 enum : u32
 {
@@ -405,12 +404,17 @@ union GPUTexturePaletteReg
   ALWAYS_INLINE constexpr u32 GetYBase() const { return static_cast<u32>(y); }
 };
 
-struct GPUTextureWindow
+union GPUTextureWindow
 {
-  u8 and_x;
+  struct
-  u8 and_y;
+  {
-  u8 or_x;
+    u8 and_x;
-  u8 or_y;
+    u8 and_y;
+    u8 or_x;
+    u8 or_y;
+  };
+
+  u32 bits;
 
   ALWAYS_INLINE bool operator==(const GPUTextureWindow& rhs) const
   {
@@ -541,182 +545,3 @@ static constexpr s32 DITHER_MATRIX[DITHER_MATRIX_SIZE][DITHER_MATRIX_SIZE] = {{-
                                                                               {+2, -2, +3, -1},  // row 1
                                                                               {-3, +1, -4, +0},  // row 2
                                                                               {+3, -1, +2, -2}}; // row 3
-
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable : 4200) // warning C4200: nonstandard extension used: zero-sized array in struct/union
-#endif
-
-enum class GPUBackendCommandType : u8
-{
-  Wraparound,
-  Sync,
-  FillVRAM,
-  UpdateVRAM,
-  CopyVRAM,
-  SetDrawingArea,
-  UpdateCLUT,
-  DrawPolygon,
-  DrawRectangle,
-  DrawLine,
-};
-
-union GPUBackendCommandParameters
-{
-  u8 bits;
-
-  BitField<u8, bool, 0, 1> interlaced_rendering;
-
-  /// Returns 0 if the currently-displayed field is on an even line in VRAM, otherwise 1.
-  BitField<u8, u8, 1, 1> active_line_lsb;
-
-  BitField<u8, bool, 2, 1> set_mask_while_drawing;
-  BitField<u8, bool, 3, 1> check_mask_before_draw;
-
-  // During transfer/render operations, if ((dst_pixel & mask_and) == 0) { pixel = src_pixel | mask_or }
-  u16 GetMaskAND() const
-  {
-    // return check_mask_before_draw ? 0x8000 : 0x0000;
-    return Truncate16((bits << 12) & 0x8000);
-  }
-  u16 GetMaskOR() const
-  {
-    // return set_mask_while_drawing ? 0x8000 : 0x0000;
-    return Truncate16((bits << 13) & 0x8000);
-  }
-};
-
-struct GPUBackendCommand
-{
-  u32 size;
-  GPUBackendCommandType type;
-  GPUBackendCommandParameters params;
-};
-
-struct GPUBackendSyncCommand : public GPUBackendCommand
-{
-  bool allow_sleep;
-};
-
-struct GPUBackendFillVRAMCommand : public GPUBackendCommand
-{
-  u16 x;
-  u16 y;
-  u16 width;
-  u16 height;
-  u32 color;
-};
-
-struct GPUBackendUpdateVRAMCommand : public GPUBackendCommand
-{
-  u16 x;
-  u16 y;
-  u16 width;
-  u16 height;
-  u16 data[0];
-};
-
-struct GPUBackendCopyVRAMCommand : public GPUBackendCommand
-{
-  u16 src_x;
-  u16 src_y;
-  u16 dst_x;
-  u16 dst_y;
-  u16 width;
-  u16 height;
-};
-
-struct GPUBackendSetDrawingAreaCommand : public GPUBackendCommand
-{
-  GPUDrawingArea new_area;
-  s32 new_clamped_area[4];
-};
-
-struct GPUBackendUpdateCLUTCommand : public GPUBackendCommand
-{
-  GPUTexturePaletteReg reg;
-  bool clut_is_8bit;
-};
-
-struct GPUBackendDrawCommand : public GPUBackendCommand
-{
-  GPUDrawModeReg draw_mode;
-  GPURenderCommand rc;
-  GPUTexturePaletteReg palette;
-  GPUTextureWindow window;
-};
-
-struct GPUBackendDrawPolygonCommand : public GPUBackendDrawCommand
-{
-  u16 num_vertices;
-
-  struct Vertex
-  {
-    s32 x, y;
-    union
-    {
-      struct
-      {
-        u8 r, g, b, a;
-      };
-      u32 color;
-    };
-    union
-    {
-      struct
-      {
-        u8 u, v;
-      };
-      u16 texcoord;
-    };
-
-    ALWAYS_INLINE void Set(s32 x_, s32 y_, u32 color_, u16 texcoord_)
-    {
-      x = x_;
-      y = y_;
-      color = color_;
-      texcoord = texcoord_;
-    }
-  };
-
-  Vertex vertices[0];
-};
-
-struct GPUBackendDrawRectangleCommand : public GPUBackendDrawCommand
-{
-  s32 x, y;
-  u16 width, height;
-  u16 texcoord;
-  u32 color;
-};
-
-struct GPUBackendDrawLineCommand : public GPUBackendDrawCommand
-{
-  u16 num_vertices;
-
-  struct Vertex
-  {
-    s32 x, y;
-    union
-    {
-      struct
-      {
-        u8 r, g, b, a;
-      };
-      u32 color;
-    };
-
-    ALWAYS_INLINE void Set(s32 x_, s32 y_, u32 color_)
-    {
-      x = x_;
-      y = y_;
-      color = color_;
-    }
-  };
-
-  Vertex vertices[0];
-};
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif

src/core/guncon.cpp

@@ -207,12 +207,12 @@ void GunCon::UpdatePosition()
   float display_x, display_y;
   const auto& [window_x, window_y] = (m_has_relative_binds) ? GetAbsolutePositionFromRelativeAxes() :
                                                               InputManager::GetPointerAbsolutePosition(m_cursor_index);
-  g_gpu->ConvertScreenCoordinatesToDisplayCoordinates(window_x, window_y, &display_x, &display_y);
+  g_gpu.ConvertScreenCoordinatesToDisplayCoordinates(window_x, window_y, &display_x, &display_y);
 
   // are we within the active display area?
   u32 tick, line;
   if (display_x < 0 || display_y < 0 ||
-      !g_gpu->ConvertDisplayCoordinatesToBeamTicksAndLines(display_x, display_y, m_x_scale, &tick, &line) ||
+      !g_gpu.ConvertDisplayCoordinatesToBeamTicksAndLines(display_x, display_y, m_x_scale, &tick, &line) ||
       m_shoot_offscreen)
   {
     DEBUG_LOG("Lightgun out of range for window coordinates {:.0f},{:.0f}", window_x, window_y);
@@ -222,7 +222,7 @@ void GunCon::UpdatePosition()
   }
 
   // 8MHz units for X = 44100*768*11/7 = 53222400 / 8000000 = 6.6528
-  const double divider = static_cast<double>(g_gpu->GetCRTCFrequency()) / 8000000.0;
+  const double divider = static_cast<double>(g_gpu.GetCRTCFrequency()) / 8000000.0;
   m_position_x = static_cast<u16>(static_cast<float>(tick) / static_cast<float>(divider));
   m_position_y = static_cast<u16>(line);
   DEBUG_LOG("Lightgun window coordinates {:.0f},{:.0f} -> tick {} line {} 8mhz ticks {}", display_x, display_y, tick,

src/core/host.cpp

@@ -2,19 +2,13 @@
 // SPDX-License-Identifier: 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 "system_private.h"
 
 #include "scmversion/scmversion.h"
 
 #include "util/compress_helpers.h"
-#include "util/gpu_device.h"
-#include "util/imgui_manager.h"
-#include "util/input_manager.h"
 
 #include "common/assert.h"
 #include "common/error.h"
@@ -342,181 +336,3 @@ std::string Host::GetHTTPUserAgent()
 {
   return fmt::format("DuckStation for {} ({}) {}", TARGET_OS_STR, CPU_ARCH_STR, g_scm_tag_str);
 }
-
-bool Host::CreateGPUDevice(RenderAPI api, bool fullscreen, 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<GPUDevice::ExclusiveFullscreenMode> fullscreen_mode;
-  if (fullscreen && g_gpu_device && g_gpu_device->SupportsExclusiveFullscreen())
-  {
-    fullscreen_mode =
-      GPUDevice::ExclusiveFullscreenMode::Parse(Host::GetTinyStringSettingValue("GPU", "FullscreenMode", ""));
-  }
-  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_memory_import)
-    disabled_features |= GPUDevice::FEATURE_MASK_MEMORY_IMPORT;
-  if (g_settings.gpu_disable_raster_order_views)
-    disabled_features |= GPUDevice::FEATURE_MASK_RASTER_ORDER_VIEWS;
-  if (g_settings.gpu_disable_compute_shaders)
-    disabled_features |= GPUDevice::FEATURE_MASK_COMPUTE_SHADERS;
-  if (g_settings.gpu_disable_compressed_textures)
-    disabled_features |= GPUDevice::FEATURE_MASK_COMPRESSED_TEXTURES;
-
-    // Don't dump shaders on debug builds for Android, users will complain about storage...
-#if !defined(__ANDROID__) || defined(_DEBUG)
-  const std::string_view shader_dump_directory(EmuFolders::DataRoot);
-#else
-  const std::string_view shader_dump_directory;
-#endif
-
-  Error create_error;
-  std::optional<WindowInfo> wi;
-  if (!g_gpu_device ||
-      !(wi = Host::AcquireRenderWindow(api, fullscreen, fullscreen_mode.has_value(), &create_error)).has_value() ||
-      !g_gpu_device->Create(
-        g_settings.gpu_adapter, static_cast<GPUDevice::FeatureMask>(disabled_features), shader_dump_directory,
-        g_settings.gpu_disable_shader_cache ? std::string_view() : std::string_view(EmuFolders::Cache),
-        SHADER_CACHE_VERSION, g_settings.gpu_use_debug_device, wi.value(), System::GetEffectiveVSyncMode(),
-        System::ShouldAllowPresentThrottle(), fullscreen_mode.has_value() ? &fullscreen_mode.value() : nullptr,
-        exclusive_fullscreen_control, &create_error))
-  {
-    ERROR_LOG("Failed to create GPU device: {}", create_error.GetDescription());
-    if (g_gpu_device)
-      g_gpu_device->Destroy();
-    g_gpu_device.reset();
-    if (wi.has_value())
-      Host::ReleaseRenderWindow();
-
-    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_osd_margin, &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();
-    Host::ReleaseRenderWindow();
-    return false;
-  }
-
-  InputManager::SetDisplayWindowSize(ImGuiManager::GetWindowWidth(), ImGuiManager::GetWindowHeight());
-  return true;
-}
-
-void Host::UpdateDisplayWindow(bool fullscreen)
-{
-  if (!g_gpu_device)
-    return;
-
-  const GPUVSyncMode vsync_mode = System::GetEffectiveVSyncMode();
-  const bool allow_present_throttle = System::ShouldAllowPresentThrottle();
-  std::optional<GPUDevice::ExclusiveFullscreenMode> fullscreen_mode;
-  if (fullscreen && g_gpu_device->SupportsExclusiveFullscreen())
-  {
-    fullscreen_mode =
-      GPUDevice::ExclusiveFullscreenMode::Parse(Host::GetTinyStringSettingValue("GPU", "FullscreenMode", ""));
-  }
-  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);
-  }
-
-  g_gpu_device->DestroyMainSwapChain();
-
-  Error error;
-  std::optional<WindowInfo> wi =
-    Host::AcquireRenderWindow(g_gpu_device->GetRenderAPI(), fullscreen, fullscreen_mode.has_value(), &error);
-  if (!wi.has_value())
-  {
-    Host::ReportFatalError("Failed to get render window after update", error.GetDescription());
-    return;
-  }
-
-  // if surfaceless, just leave it
-  if (wi->IsSurfaceless())
-  {
-    DEV_LOG("Switching to surfaceless device");
-    if (!g_gpu_device->SwitchToSurfacelessRendering(&error))
-      ERROR_LOG("Failed to switch to surfaceless, rendering commands may fail: {}", error.GetDescription());
-
-    return;
-  }
-
-  if (!g_gpu_device->RecreateMainSwapChain(wi.value(), vsync_mode, allow_present_throttle,
-                                           fullscreen_mode.has_value() ? &fullscreen_mode.value() : nullptr,
-                                           exclusive_fullscreen_control, &error))
-  {
-    Host::ReportFatalError("Failed to change window after update", error.GetDescription());
-    return;
-  }
-
-  const float f_width = static_cast<float>(g_gpu_device->GetMainSwapChain()->GetWidth());
-  const float f_height = static_cast<float>(g_gpu_device->GetMainSwapChain()->GetHeight());
-  ImGuiManager::WindowResized(f_width, f_height);
-  InputManager::SetDisplayWindowSize(f_width, f_height);
-  System::DisplayWindowResized();
-}
-
-void Host::ResizeDisplayWindow(s32 width, s32 height, float scale)
-{
-  if (!g_gpu_device || !g_gpu_device->HasMainSwapChain())
-    return;
-
-  DEV_LOG("Display window resized to {}x{}", width, height);
-
-  Error error;
-  if (!g_gpu_device->GetMainSwapChain()->ResizeBuffers(width, height, scale, &error))
-  {
-    ERROR_LOG("Failed to resize main swap chain: {}", error.GetDescription());
-    UpdateDisplayWindow(Host::IsFullscreen());
-    return;
-  }
-
-  const float f_width = static_cast<float>(g_gpu_device->GetMainSwapChain()->GetWidth());
-  const float f_height = static_cast<float>(g_gpu_device->GetMainSwapChain()->GetHeight());
-  ImGuiManager::WindowResized(f_width, f_height);
-  InputManager::SetDisplayWindowSize(f_width, f_height);
-  System::DisplayWindowResized();
-}
-
-void Host::ReleaseGPUDevice()
-{
-  if (!g_gpu_device)
-    return;
-
-  ImGuiManager::DestroyAllDebugWindows();
-  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();
-
-  Host::ReleaseRenderWindow();
-}

src/core/host.h

@@ -75,10 +75,6 @@ std::span<const std::pair<const char*, const char*>> GetAvailableLanguageList();
 /// Refreshes the UI when the language is changed.
 bool ChangeLanguage(const char* new_language);
 
-/// Displays a loading screen with the logo, rendered with ImGui. Use when executing possibly-time-consuming tasks
-/// such as compiling shaders when starting up.
-void DisplayLoadingScreen(const char* message, int progress_min = -1, int progress_max = -1, int progress_value = -1);
-
 /// Safely executes a function on the VM thread.
 void RunOnCPUThread(std::function<void()> function, bool block = false);
 
@@ -96,21 +92,6 @@ bool IsFullscreen();
 /// Alters fullscreen state of hosting application.
 void SetFullscreen(bool enabled);
 
-/// Attempts to create the rendering device backend.
-bool CreateGPUDevice(RenderAPI api, bool fullscreen, Error* error);
-
-/// Handles fullscreen transitions and such.
-void UpdateDisplayWindow(bool fullscreen);
-
-/// 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();
-
 namespace Internal {
 
 /// Returns true if the host should use portable mode.

src/core/host_interface_progress_callback.cpp

@@ -1,88 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: CC-BY-NC-ND-4.0
-
-#include "host_interface_progress_callback.h"
-#include "host.h"
-
-#include "common/log.h"
-
-LOG_CHANNEL(Host);
-
-HostInterfaceProgressCallback::HostInterfaceProgressCallback() : ProgressCallback()
-{
-}
-
-void HostInterfaceProgressCallback::PushState()
-{
-  ProgressCallback::PushState();
-}
-
-void HostInterfaceProgressCallback::PopState()
-{
-  ProgressCallback::PopState();
-  Redraw(true);
-}
-
-void HostInterfaceProgressCallback::SetCancellable(bool cancellable)
-{
-  ProgressCallback::SetCancellable(cancellable);
-  Redraw(true);
-}
-
-void HostInterfaceProgressCallback::SetTitle(const std::string_view title)
-{
-  // todo?
-}
-
-void HostInterfaceProgressCallback::SetStatusText(const std::string_view text)
-{
-  ProgressCallback::SetStatusText(text);
-  Redraw(true);
-}
-
-void HostInterfaceProgressCallback::SetProgressRange(u32 range)
-{
-  u32 last_range = m_progress_range;
-
-  ProgressCallback::SetProgressRange(range);
-
-  if (m_progress_range != last_range)
-    Redraw(false);
-}
-
-void HostInterfaceProgressCallback::SetProgressValue(u32 value)
-{
-  u32 lastValue = m_progress_value;
-
-  ProgressCallback::SetProgressValue(value);
-
-  if (m_progress_value != lastValue)
-    Redraw(false);
-}
-
-void HostInterfaceProgressCallback::Redraw(bool force)
-{
-  if (m_last_progress_percent < 0 && m_open_time.GetTimeSeconds() < m_open_delay)
-    return;
-
-  const int percent =
-    static_cast<int>((static_cast<float>(m_progress_value) / static_cast<float>(m_progress_range)) * 100.0f);
-  if (percent == m_last_progress_percent && !force)
-    return;
-
-  m_last_progress_percent = percent;
-  Host::DisplayLoadingScreen(m_status_text.c_str(), 0, static_cast<int>(m_progress_range),
-                             static_cast<int>(m_progress_value));
-}
-
-void HostInterfaceProgressCallback::ModalError(const std::string_view message)
-{
-  ERROR_LOG(message);
-  Host::ReportErrorAsync("Error", message);
-}
-
-bool HostInterfaceProgressCallback::ModalConfirmation(const std::string_view message)
-{
-  INFO_LOG(message);
-  return Host::ConfirmMessage("Confirm", message);
-}

src/core/host_interface_progress_callback.h

@@ -1,34 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: CC-BY-NC-ND-4.0
-
-#pragma once
-
-#include "common/progress_callback.h"
-#include "common/timer.h"
-
-class HostInterfaceProgressCallback final : public ProgressCallback
-{
-public:
-  HostInterfaceProgressCallback();
-
-  ALWAYS_INLINE void SetOpenDelay(float delay) { m_open_delay = delay; }
-
-  void PushState() override;
-  void PopState() override;
-
-  void SetCancellable(bool cancellable) override;
-  void SetTitle(const std::string_view title) override;
-  void SetStatusText(const std::string_view text) override;
-  void SetProgressRange(u32 range) override;
-  void SetProgressValue(u32 value) override;
-
-  void ModalError(const std::string_view message) override;
-  bool ModalConfirmation(const std::string_view message) override;
-
-private:
-  void Redraw(bool force);
-
-  Timer m_open_time;
-  float m_open_delay = 1.0f;
-  int m_last_progress_percent = -1;
-};

src/core/hotkeys.cpp

@@ -8,6 +8,7 @@
 #include "fullscreen_ui.h"
 #include "gpu.h"
 #include "gpu_hw_texture_cache.h"
+#include "gpu_thread.h"
 #include "host.h"
 #include "imgui_overlays.h"
 #include "settings.h"
@@ -58,9 +59,8 @@ static void HotkeyModifyResolutionScale(s32 increment)
 
   if (System::IsValid())
   {
-    g_gpu->RestoreDeviceContext();
+    System::ClearMemorySaveStates(true);
-    g_gpu->UpdateSettings(old_settings);
+    GPUThread::UpdateSettings(true, false);
-    System::ClearMemorySaveStates();
   }
 }
 
@@ -375,11 +375,11 @@ DEFINE_HOTKEY("TogglePGXP", TRANSLATE_NOOP("Hotkeys", "Graphics"), TRANSLATE_NOO
               [](s32 pressed) {
                 if (!pressed && System::IsValid())
                 {
-                  Settings old_settings = g_settings;
+                  System::ClearMemorySaveStates(true);
+
                   g_settings.gpu_pgxp_enable = !g_settings.gpu_pgxp_enable;
-                  g_gpu->RestoreDeviceContext();
+                  GPUThread::UpdateSettings(true, false);
-                  g_gpu->UpdateSettings(old_settings);
+
-                  System::ClearMemorySaveStates();
                   Host::AddKeyedOSDMessage("TogglePGXP",
                                            g_settings.gpu_pgxp_enable ?
                                              TRANSLATE_STR("OSDMessage", "PGXP is now enabled.") :
@@ -426,13 +426,18 @@ DEFINE_HOTKEY("ToggleInternalPostProcessing", TRANSLATE_NOOP("Hotkeys", "Graphic
 DEFINE_HOTKEY("ReloadPostProcessingShaders", TRANSLATE_NOOP("Hotkeys", "Graphics"),
               TRANSLATE_NOOP("Hotkeys", "Reload Post Processing Shaders"), [](s32 pressed) {
                 if (!pressed && System::IsValid())
-                  PostProcessing::ReloadShaders();
+                {
+                  GPUThread::RunOnThread([]() {
+                    if (GPUThread::HasGPUBackend())
+                      PostProcessing::ReloadShaders();
+                  });
+                }
               })
 
 DEFINE_HOTKEY("ReloadTextureReplacements", TRANSLATE_NOOP("Hotkeys", "Graphics"),
               TRANSLATE_NOOP("Hotkeys", "Reload Texture Replacements"), [](s32 pressed) {
                 if (!pressed && System::IsValid())
-                  GPUTextureCache::ReloadTextureReplacements(true);
+                  GPUThread::RunOnThread([]() { GPUTextureCache::ReloadTextureReplacements(true); });
               })
 
 DEFINE_HOTKEY("ToggleWidescreen", TRANSLATE_NOOP("Hotkeys", "Graphics"), TRANSLATE_NOOP("Hotkeys", "Toggle Widescreen"),
@@ -448,12 +453,11 @@ DEFINE_HOTKEY("TogglePGXPDepth", TRANSLATE_NOOP("Hotkeys", "Graphics"),
                   if (!g_settings.gpu_pgxp_enable)
                     return;
 
-                  const Settings old_settings = g_settings;
+                  System::ClearMemorySaveStates(true);
-                  g_settings.gpu_pgxp_depth_buffer = !g_settings.gpu_pgxp_depth_buffer;
+
+                  g_settings.gpu_pgxp_depth_buffer = !g_settings.gpu_pgxp_depth_buffer;
+                  GPUThread::UpdateSettings(true, false);
 
-                  g_gpu->RestoreDeviceContext();
-                  g_gpu->UpdateSettings(old_settings);
-                  System::ClearMemorySaveStates();
                   Host::AddKeyedOSDMessage("TogglePGXPDepth",
                                            g_settings.gpu_pgxp_depth_buffer ?
                                              TRANSLATE_STR("OSDMessage", "PGXP Depth Buffer is now enabled.") :
@@ -469,12 +473,11 @@ DEFINE_HOTKEY("TogglePGXPCPU", TRANSLATE_NOOP("Hotkeys", "Graphics"), TRANSLATE_
                   if (!g_settings.gpu_pgxp_enable)
                     return;
 
-                  const Settings old_settings = g_settings;
+                  System::ClearMemorySaveStates(true);
+
+                  // GPU thread is unchanged
                   g_settings.gpu_pgxp_cpu = !g_settings.gpu_pgxp_cpu;
 
-                  g_gpu->RestoreDeviceContext();
-                  g_gpu->UpdateSettings(old_settings);
-                  System::ClearMemorySaveStates();
                   Host::AddKeyedOSDMessage("TogglePGXPCPU",
                                            g_settings.gpu_pgxp_cpu ?
                                              TRANSLATE_STR("OSDMessage", "PGXP CPU mode is now enabled.") :
@@ -584,29 +587,31 @@ 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();
+                  GPUThread::RunOnThread([]() {
-                  SaveStateSelectorUI::SelectNextSlot(false);
+                    SaveStateSelectorUI::SaveCurrentSlot();
+                    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 "performance_counters.h"
@@ -70,10 +72,10 @@ struct DebugWindowInfo
 } // namespace
 
 static void FormatProcessorStat(SmallStringBase& text, double usage, double time);
-static void DrawPerformanceOverlay(float& position_y, float scale, float margin, float spacing);
+static void DrawPerformanceOverlay(const GPUBackend* gpu, float& position_y, float scale, float margin, float spacing);
 static void DrawMediaCaptureOverlay(float& position_y, float scale, float margin, float spacing);
 static void DrawFrameTimeOverlay(float& position_y, float scale, float margin, float spacing);
-static void DrawEnhancementsOverlay();
+static void DrawEnhancementsOverlay(const GPUBackend* gpu);
 static void DrawInputsOverlay();
 
 #ifndef __ANDROID__
@@ -83,7 +85,7 @@ static constexpr const char* DEBUG_WINDOW_CONFIG_SECTION = "DebugWindows";
 static constexpr const std::array<DebugWindowInfo, NUM_DEBUG_WINDOWS> s_debug_window_info = {{
   {"SPU", "SPU State", ":icons/applications-system.png", &SPU::DrawDebugStateWindow, 800, 915},
   {"CDROM", "CD-ROM State", ":icons/applications-system.png", &CDROM::DrawDebugWindow, 800, 540},
-  {"GPU", "GPU State", ":icons/applications-system.png", [](float sc) { g_gpu->DrawDebugStateWindow(sc); }, 450, 550},
+  {"GPU", "GPU State", ":icons/applications-system.png", [](float sc) { g_gpu.DrawDebugStateWindow(sc); }, 450, 550},
   {"DMA", "DMA State", ":icons/applications-system.png", &DMA::DrawDebugStateWindow, 860, 180},
   {"MDEC", "MDEC State", ":icons/applications-system.png", &MDEC::DrawDebugStateWindow, 300, 350},
   {"Timers", "Timers State", ":icons/applications-system.png", &Timers::DrawDebugStateWindow, 800, 95},
@@ -119,93 +121,22 @@ static std::tuple<float, float> GetMinMax(std::span<const float> values)
   return std::tie(min, max);
 }
 
-void Host::DisplayLoadingScreen(const char* message, int progress_min /*= -1*/, int progress_max /*= -1*/,
+bool ImGuiManager::AreAnyDebugWindowsEnabled(const SettingsInterface& si)
-                                int progress_value /*= -1*/)
 {
-  if (!g_gpu_device || !g_gpu_device->HasMainSwapChain())
+#ifndef __ANDROID__
+  const bool block_all = Achievements::IsHardcoreModeActive();
+  if (block_all)
+    return false;
+
+  for (size_t i = 0; i < NUM_DEBUG_WINDOWS; i++)
   {
-    INFO_LOG("{}: {}/{}", message, progress_value, progress_max);
+    const DebugWindowInfo& info = s_debug_window_info[i];
-    return;
+    if (si.GetBoolValue(DEBUG_WINDOW_CONFIG_SECTION, info.name, false))
+      return true;
   }
+#endif
 
-  const auto& io = ImGui::GetIO();
+  return false;
-  const float scale = ImGuiManager::GetGlobalScale();
-  const float width = (400.0f * scale);
-  const bool has_progress = (progress_min < progress_max);
-
-  // eat the last imgui frame, it might've been partially rendered by the caller.
-  ImGui::EndFrame();
-  ImGui::NewFrame();
-
-  const float logo_width = 260.0f * scale;
-  const float logo_height = 260.0f * scale;
-
-  ImGui::SetNextWindowSize(ImVec2(logo_width, logo_height), ImGuiCond_Always);
-  ImGui::SetNextWindowPos(ImVec2(io.DisplaySize.x * 0.5f, (io.DisplaySize.y * 0.5f) - (50.0f * scale)),
-                          ImGuiCond_Always, ImVec2(0.5f, 0.5f));
-  if (ImGui::Begin("LoadingScreenLogo", nullptr,
-                   ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoMove |
-                     ImGuiWindowFlags_NoSavedSettings | ImGuiWindowFlags_NoScrollbar | ImGuiWindowFlags_NoNav |
-                     ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoFocusOnAppearing |
-                     ImGuiWindowFlags_NoBackground))
-  {
-    GPUTexture* tex = ImGuiFullscreen::GetCachedTexture("images/duck.png");
-    if (tex)
-      ImGui::Image(tex, ImVec2(logo_width, logo_height));
-  }
-  ImGui::End();
-
-  const float padding_and_rounding = 18.0f * scale;
-  ImGui::PushStyleVar(ImGuiStyleVar_WindowRounding, padding_and_rounding);
-  ImGui::PushStyleVar(ImGuiStyleVar_WindowPadding, ImVec2(padding_and_rounding, padding_and_rounding));
-  ImGui::SetNextWindowSize(ImVec2(width, (has_progress ? 90.0f : 55.0f) * scale), ImGuiCond_Always);
-  ImGui::SetNextWindowPos(ImVec2(io.DisplaySize.x * 0.5f, (io.DisplaySize.y * 0.5f) + (100.0f * scale)),
-                          ImGuiCond_Always, ImVec2(0.5f, 0.0f));
-  if (ImGui::Begin("LoadingScreen", nullptr,
-                   ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoMove |
-                     ImGuiWindowFlags_NoSavedSettings | ImGuiWindowFlags_NoScrollbar | ImGuiWindowFlags_NoNav |
-                     ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoFocusOnAppearing))
-  {
-    if (has_progress)
-    {
-      ImGui::TextUnformatted(message);
-
-      TinyString buf;
-      buf.format("{}/{}", progress_value, progress_max);
-
-      const ImVec2 prog_size = ImGui::CalcTextSize(buf.c_str(), buf.end_ptr());
-      ImGui::SameLine();
-      ImGui::SetCursorPosX(width - padding_and_rounding - prog_size.x);
-      ImGui::TextUnformatted(buf.c_str(), buf.end_ptr());
-      ImGui::SetCursorPosY(ImGui::GetCursorPosY() + 5.0f);
-
-      ImGui::ProgressBar(static_cast<float>(progress_value) / static_cast<float>(progress_max - progress_min),
-                         ImVec2(-1.0f, 0.0f), "");
-      INFO_LOG("{}: {}", message, buf);
-    }
-    else
-    {
-      const ImVec2 text_size(ImGui::CalcTextSize(message));
-      ImGui::SetCursorPosX((width - text_size.x) / 2.0f);
-      ImGui::TextUnformatted(message);
-      INFO_LOG(message);
-    }
-  }
-  ImGui::End();
-  ImGui::PopStyleVar(2);
-
-  ImGui::EndFrame();
-
-  // TODO: Glass effect or something.
-
-  GPUSwapChain* swap_chain = g_gpu_device->GetMainSwapChain();
-  if (g_gpu_device->BeginPresent(swap_chain) == GPUDevice::PresentResult::OK)
-  {
-    g_gpu_device->RenderImGui(swap_chain);
-    g_gpu_device->EndPresent(swap_chain, false);
-  }
-
-  ImGui::NewFrame();
 }
 
 bool ImGuiManager::UpdateDebugWindowConfig()
@@ -284,26 +215,28 @@ void ImGuiManager::DestroyAllDebugWindows()
 #endif
 }
 
-void ImGuiManager::RenderTextOverlays()
+void ImGuiManager::RenderTextOverlays(const GPUBackend* gpu)
 {
-  const System::State state = System::GetState();
+  // Don't draw anything with loading screen open, it'll be nonsensical.
-  if (state != System::State::Shutdown)
+  if (ImGuiFullscreen::IsLoadingScreenOpen())
-  {
+    return;
-    const float scale = ImGuiManager::GetGlobalScale();
-    const float f_margin = ImGuiManager::GetScreenMargin() * scale;
-    const float margin = ImCeil(ImGuiManager::GetScreenMargin() * scale);
-    const float spacing = ImCeil(5.0f * scale);
-    float position_y = ImFloor(f_margin);
-    DrawPerformanceOverlay(position_y, scale, margin, spacing);
-    DrawFrameTimeOverlay(position_y, scale, margin, spacing);
-    DrawMediaCaptureOverlay(position_y, scale, margin, spacing);
 
-    if (g_settings.display_show_enhancements && state != System::State::Paused)
+  const bool paused = GPUThread::IsSystemPaused();
-      DrawEnhancementsOverlay();
 
-    if (g_settings.display_show_inputs && state != System::State::Paused)
+  const float scale = ImGuiManager::GetGlobalScale();
-      DrawInputsOverlay();
+  const float f_margin = ImGuiManager::GetScreenMargin() * scale;
-  }
+  const float margin = ImCeil(ImGuiManager::GetScreenMargin() * scale);
+  const float spacing = ImCeil(5.0f * scale);
+  float position_y = ImFloor(f_margin);
+  DrawPerformanceOverlay(gpu, position_y, scale, margin, spacing);
+  DrawFrameTimeOverlay(position_y, scale, margin, spacing);
+  DrawMediaCaptureOverlay(position_y, scale, margin, spacing);
+
+  if (g_gpu_settings.display_show_enhancements && !paused)
+    DrawEnhancementsOverlay(gpu);
+
+  if (g_gpu_settings.display_show_inputs && !paused)
+    DrawInputsOverlay();
 }
 
 void ImGuiManager::FormatProcessorStat(SmallStringBase& text, double usage, double time)
@@ -317,12 +250,13 @@ void ImGuiManager::FormatProcessorStat(SmallStringBase& text, double usage, doub
     text.append_format("{:.1f}% ({:.2f}ms)", usage, time);
 }
 
-void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float margin, float spacing)
+void ImGuiManager::DrawPerformanceOverlay(const GPUBackend* gpu, float& position_y, float scale, float margin,
+                                          float spacing)
 {
-  if (!(g_settings.display_show_fps || g_settings.display_show_speed || g_settings.display_show_gpu_stats ||
+  if (!(g_gpu_settings.display_show_fps || g_gpu_settings.display_show_speed || g_gpu_settings.display_show_gpu_stats ||
-        g_settings.display_show_resolution || g_settings.display_show_cpu_usage ||
+        g_gpu_settings.display_show_resolution || g_gpu_settings.display_show_cpu_usage ||
-        (g_settings.display_show_status_indicators &&
+        (g_gpu_settings.display_show_status_indicators &&
-         (System::IsPaused() || System::IsFastForwardEnabled() || System::IsTurboEnabled()))))
+         (GPUThread::IsSystemPaused() || System::IsFastForwardEnabled() || System::IsTurboEnabled()))))
   {
     return;
   }
@@ -352,9 +286,9 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
   if (state == System::State::Running)
   {
     const float speed = PerformanceCounters::GetEmulationSpeed();
-    if (g_settings.display_show_fps)
+    if (g_gpu_settings.display_show_fps)
       text.append_format("G: {:.2f} | V: {:.2f}", PerformanceCounters::GetFPS(), PerformanceCounters::GetVPS());
-    if (g_settings.display_show_speed)
+    if (g_gpu_settings.display_show_speed)
     {
       text.append_format("{}{}%", text.empty() ? "" : " | ", static_cast<u32>(std::round(speed)));
 
@@ -377,33 +311,33 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
       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);
+      gpu->GetStatsString(text);
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
 
-      g_gpu->GetMemoryStatsString(text);
+      gpu->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)
     {
-      const u32 resolution_scale = g_gpu->GetResolutionScale();
+      const u32 resolution_scale = gpu->GetResolutionScale();
-      const auto [display_width, display_height] = g_gpu->GetFullDisplayResolution(); // wrong
+      const auto [display_width, display_height] = g_gpu.GetFullDisplayResolution(); // NOTE: Racey read.
-      const bool interlaced = g_gpu->IsInterlacedDisplayEnabled();
+      const bool interlaced = g_gpu.IsInterlacedDisplayEnabled();
-      const bool pal = g_gpu->IsInPALMode();
+      const bool pal = g_gpu.IsInPALMode();
       text.format("{}x{} {} {} [{}x]", display_width * resolution_scale, display_height * resolution_scale,
                   pal ? "PAL" : "NTSC", interlaced ? "Interlaced" : "Progressive", resolution_scale);
       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", PerformanceCounters::GetMinimumFrameTime(),
                   PerformanceCounters::GetAverageFrameTime(), PerformanceCounters::GetMaximumFrameTime());
@@ -454,11 +388,11 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
                           PerformanceCounters::GetCPUThreadAverageTime());
       DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
 
-      if (g_gpu->GetSWThread())
+      if (g_gpu_settings.gpu_use_thread)
       {
-        text.assign("SW: ");
+        text.assign("RNDR: ");
-        FormatProcessorStat(text, PerformanceCounters::GetSWThreadUsage(),
+        FormatProcessorStat(text, PerformanceCounters::GetGPUThreadUsage(),
-                            PerformanceCounters::GetSWThreadAverageTime());
+                            PerformanceCounters::GetGPUThreadAverageTime());
         DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
       }
 
@@ -472,14 +406,14 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
 #endif
     }
 
-    if (g_settings.display_show_gpu_usage && g_gpu_device->IsGPUTimingEnabled())
+    if (g_gpu_settings.display_show_gpu_usage && g_gpu_device->IsGPUTimingEnabled())
     {
       text.assign("GPU: ");
       FormatProcessorStat(text, PerformanceCounters::GetGPUUsage(), PerformanceCounters::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())
@@ -489,7 +423,7 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
       }
     }
   }
-  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_EMOJI_PAUSE);
@@ -499,12 +433,12 @@ void ImGuiManager::DrawPerformanceOverlay(float& position_y, float scale, float
 #undef DRAW_LINE
 }
 
-void ImGuiManager::DrawEnhancementsOverlay()
+void ImGuiManager::DrawEnhancementsOverlay(const GPUBackend* gpu)
 {
   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);
@@ -519,29 +453,29 @@ void ImGuiManager::DrawEnhancementsOverlay()
     text.append_format(" CDR={}x", g_settings.cdrom_read_speedup);
   if (g_settings.cdrom_seek_speedup != 1)
     text.append_format(" CDS={}x", g_settings.cdrom_seek_speedup);
-  if (g_settings.gpu_resolution_scale != 1)
+  if (g_gpu_settings.gpu_resolution_scale != 1)
-    text.append_format(" IR={}x", g_settings.gpu_resolution_scale);
+    text.append_format(" IR={}x", g_gpu_settings.gpu_resolution_scale);
-  if (g_settings.gpu_multisamples != 1)
+  if (g_gpu_settings.gpu_multisamples != 1)
   {
-    text.append_format(" {}x{}", g_settings.gpu_multisamples, g_settings.gpu_per_sample_shading ? "SSAA" : "MSAA");
+    text.append_format(" {}x{}", g_gpu_settings.gpu_multisamples, g_gpu_settings.gpu_per_sample_shading ? "SSAA" : "MSAA");
   }
-  if (g_settings.gpu_true_color)
+  if (g_gpu_settings.gpu_true_color)
     text.append(" TrueCol");
-  text.append_format(" DI={}", Settings::GetDisplayDeinterlacingModeName(g_settings.display_deinterlacing_mode));
+  text.append_format(" DI={}", Settings::GetDisplayDeinterlacingModeName(g_gpu_settings.display_deinterlacing_mode));
   if (g_settings.gpu_force_video_timing == ForceVideoTimingMode::NTSC && System::GetRegion() == ConsoleRegion::PAL)
     text.append(" PAL60");
   if (g_settings.gpu_force_video_timing == ForceVideoTimingMode::PAL && System::GetRegion() != ConsoleRegion::PAL)
     text.append(" NTSC50");
-  if (g_settings.gpu_texture_filter != GPUTextureFilter::Nearest)
+  if (g_gpu_settings.gpu_texture_filter != GPUTextureFilter::Nearest)
   {
-    if (g_settings.gpu_sprite_texture_filter != g_settings.gpu_texture_filter)
+    if (g_gpu_settings.gpu_sprite_texture_filter != g_gpu_settings.gpu_texture_filter)
     {
-      text.append_format(" {}/{}", Settings::GetTextureFilterName(g_settings.gpu_texture_filter),
+      text.append_format(" {}/{}", Settings::GetTextureFilterName(g_gpu_settings.gpu_texture_filter),
-                         Settings::GetTextureFilterName(g_settings.gpu_sprite_texture_filter));
+                         Settings::GetTextureFilterName(g_gpu_settings.gpu_sprite_texture_filter));
     }
     else
     {
-      text.append_format(" {}", Settings::GetTextureFilterName(g_settings.gpu_texture_filter));
+      text.append_format(" {}", Settings::GetTextureFilterName(g_gpu_settings.gpu_texture_filter));
     }
   }
   if (g_settings.gpu_widescreen_hack && g_settings.display_aspect_ratio != DisplayAspectRatio::Auto &&
@@ -549,8 +483,8 @@ void ImGuiManager::DrawEnhancementsOverlay()
   {
     text.append(" WSHack");
   }
-  if (g_settings.gpu_line_detect_mode != GPULineDetectMode::Disabled)
+  if (g_gpu_settings.gpu_line_detect_mode != GPULineDetectMode::Disabled)
-    text.append_format(" LD={}", Settings::GetLineDetectModeName(g_settings.gpu_line_detect_mode));
+    text.append_format(" LD={}", Settings::GetLineDetectModeName(g_gpu_settings.gpu_line_detect_mode));
   if (g_settings.gpu_pgxp_enable)
   {
     text.append(" PGXP");
@@ -626,7 +560,7 @@ void ImGuiManager::DrawMediaCaptureOverlay(float& position_y, float scale, float
 
 void ImGuiManager::DrawFrameTimeOverlay(float& position_y, float scale, float margin, float spacing)
 {
-  if (!g_settings.display_show_frame_times || System::IsPaused())
+  if (!g_settings.display_show_frame_times || GPUThread::IsSystemPaused())
     return;
 
   const float shadow_offset = std::ceil(1.0f * scale);
@@ -864,8 +798,6 @@ bool SaveStateSelectorUI::IsOpen()
 
 void SaveStateSelectorUI::Open(float open_time /* = DEFAULT_OPEN_TIME */)
 {
-  const std::string& serial = System::GetGameSerial();
-
   s_state.open_time = 0.0f;
   s_state.close_time = open_time;
 
@@ -876,7 +808,7 @@ void SaveStateSelectorUI::Open(float open_time /* = DEFAULT_OPEN_TIME */)
     s_state.placeholder_texture = ImGuiFullscreen::LoadTexture("no-save.png");
 
   s_state.is_open = true;
-  RefreshList(serial);
+  RefreshList();
   RefreshHotkeyLegend();
 }
 
@@ -889,7 +821,7 @@ void SaveStateSelectorUI::Close()
   s_state.next_legend = {};
 }
 
-void SaveStateSelectorUI::RefreshList(const std::string& serial)
+void SaveStateSelectorUI::RefreshList()
 {
   for (ListEntry& entry : s_state.slots)
   {
@@ -898,9 +830,7 @@ void SaveStateSelectorUI::RefreshList(const std::string& serial)
   }
   s_state.slots.clear();
 
-  if (System::IsShutdown())
+  const std::string& serial = GPUThread::GetGameSerial();
-    return;
-
   if (!serial.empty())
   {
     for (s32 i = 1; i <= System::PER_GAME_SAVE_STATE_SLOTS; i++)
@@ -956,6 +886,7 @@ void SaveStateSelectorUI::Clear()
 
 void SaveStateSelectorUI::ClearList()
 {
+  DebugAssert(GPUThread::IsOnThread());
   for (ListEntry& li : s_state.slots)
   {
     if (li.preview_texture)
@@ -1001,7 +932,7 @@ void SaveStateSelectorUI::SelectNextSlot(bool open_selector)
   s_state.current_slot++;
   if (s_state.current_slot >= total_slots)
   {
-    if (!System::GetGameSerial().empty())
+    if (!GPUThread::GetGameSerial().empty())
       s_state.current_slot_global ^= true;
     s_state.current_slot -= total_slots;
   }
@@ -1024,7 +955,7 @@ void SaveStateSelectorUI::SelectPreviousSlot(bool open_selector)
   s_state.current_slot--;
   if (s_state.current_slot < 0)
   {
-    if (!System::GetGameSerial().empty())
+    if (!GPUThread::GetGameSerial().empty())
       s_state.current_slot_global ^= true;
     s_state.current_slot +=
       s_state.current_slot_global ? System::GLOBAL_SAVE_STATE_SLOTS : System::PER_GAME_SAVE_STATE_SLOTS;
@@ -1245,7 +1176,7 @@ std::string SaveStateSelectorUI::GetCurrentSlotPath()
   std::string filename;
   if (!s_state.current_slot_global)
   {
-    if (const std::string& serial = System::GetGameSerial(); !serial.empty())
+    if (const std::string& serial = GPUThread::GetGameSerial(); !serial.empty())
       filename = System::GetGameSaveStateFileName(serial, s_state.current_slot + 1);
   }
   else
@@ -1258,18 +1189,22 @@ std::string SaveStateSelectorUI::GetCurrentSlotPath()
 
 void SaveStateSelectorUI::LoadCurrentSlot()
 {
+  DebugAssert(GPUThread::IsOnThread());
+
   if (std::string path = GetCurrentSlotPath(); !path.empty())
   {
     if (FileSystem::FileExists(path.c_str()))
     {
-      Error error;
+      Host::RunOnCPUThread([path = std::move(path)]() {
-      if (!System::LoadState(path.c_str(), &error, true))
+        Error error;
-      {
+        if (!System::LoadState(path.c_str(), &error, true))
-        Host::AddKeyedOSDMessage("LoadState",
+        {
-                                 fmt::format(TRANSLATE_FS("OSDMessage", "Failed to load state from slot {0}:\n{1}"),
+          Host::AddKeyedOSDMessage("LoadState",
-                                             GetCurrentSlot(), error.GetDescription()),
+                                   fmt::format(TRANSLATE_FS("OSDMessage", "Failed to load state from slot {0}:\n{1}"),
-                                 Host::OSD_ERROR_DURATION);
+                                               GetCurrentSlot(), error.GetDescription()),
-      }
+                                   Host::OSD_ERROR_DURATION);
+        }
+      });
     }
     else
     {
@@ -1289,14 +1224,16 @@ void SaveStateSelectorUI::SaveCurrentSlot()
 {
   if (std::string path = GetCurrentSlotPath(); !path.empty())
   {
-    Error error;
+    Host::RunOnCPUThread([path = std::move(path)]() {
-    if (!System::SaveState(std::move(path), &error, g_settings.create_save_state_backups, false))
+      Error error;
-    {
+      if (!System::SaveState(std::move(path), &error, g_settings.create_save_state_backups, false))
-      Host::AddIconOSDMessage("SaveState", ICON_EMOJI_WARNING,
+      {
-                              fmt::format(TRANSLATE_FS("OSDMessage", "Failed to save state to slot {0}:\n{1}"),
+        Host::AddIconOSDMessage("SaveState", ICON_EMOJI_WARNING,
-                                          GetCurrentSlot(), error.GetDescription()),
+                                fmt::format(TRANSLATE_FS("OSDMessage", "Failed to save state to slot {0}:\n{1}"),
-                              Host::OSD_ERROR_DURATION);
+                                            GetCurrentSlot(), error.GetDescription()),
-    }
+                                Host::OSD_ERROR_DURATION);
+      }
+    });
   }
 
   Close();
@@ -1323,7 +1260,7 @@ 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_state.is_open)
       SaveStateSelectorUI::Draw();

src/core/imgui_overlays.h

@@ -7,14 +7,23 @@
 
 #include <string>
 
+class SettingsInterface;
+
+class GPUBackend;
+
 namespace ImGuiManager {
-void RenderTextOverlays();
+
+static constexpr const char* LOGO_IMAGE_NAME = "images/duck.png";
+
+void RenderTextOverlays(const GPUBackend* gpu);
+bool AreAnyDebugWindowsEnabled(const SettingsInterface& si);
 void RenderDebugWindows();
 bool UpdateDebugWindowConfig();
 void DestroyAllDebugWindows();
 
 void RenderOverlayWindows();
 void DestroyOverlayTextures();
+
 } // namespace ImGuiManager
 
 namespace SaveStateSelectorUI {
@@ -23,7 +32,7 @@ static constexpr float DEFAULT_OPEN_TIME = 7.5f;
 
 bool IsOpen();
 void Open(float open_time = DEFAULT_OPEN_TIME);
-void RefreshList(const std::string& serial);
+void RefreshList();
 void Clear();
 void ClearList();
 void Close();

src/core/jogcon.cpp

@@ -376,6 +376,14 @@ bool JogCon::Transfer(const u8 data_in, u8* data_out)
         m_tx_buffer = {GetIDByte(), m_status_byte, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
         Poll();
       }
+      else if (m_configuration_mode && data_in == 0x44)
+      {
+        Assert(m_command_step == 0);
+        m_response_length = (GetResponseNumHalfwords() + 1) * 2;
+        m_command = Command::SetAnalogMode;
+        m_tx_buffer = {GetIDByte(), m_status_byte, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
+        ResetMotorConfig();
+      }
       else if (m_configuration_mode && data_in == 0x45)
       {
         m_response_length = (GetResponseNumHalfwords() + 1) * 2;
@@ -429,6 +437,25 @@ bool JogCon::Transfer(const u8 data_in, u8* data_out)
     }
     break;
 
+    case Command::SetAnalogMode:
+    {
+      if (m_command_step == 2)
+      {
+        DEV_LOG("analog mode val 0x{:02x}", data_in);
+
+        if (data_in == 0x00 || data_in == 0x01)
+          SetJogConMode(data_in == 0x01, true);
+      }
+      else if (m_command_step == 3)
+      {
+        DEV_LOG("analog mode lock 0x{:02x}", data_in);
+
+        if (data_in == 0x02 || data_in == 0x03)
+          WARNING_LOG("Unimplemented analog mode lock {}", (data_in == 0x03));
+      }
+    }
+    break;
+
     case Command::SetMode:
     {
       m_configuration_mode = (m_rx_buffer[2] == 1 && m_jogcon_mode);

src/core/jogcon.h

@@ -71,6 +71,7 @@ private:
     ReadPad,
     SetMode,
     GetAnalogMode,
+    SetAnalogMode,
     GetSetRumble,
     Command46,
     Command47,

src/core/justifier.cpp

@@ -215,12 +215,12 @@ void Justifier::UpdatePosition()
   float display_x, display_y;
   const auto [window_x, window_y] = (m_has_relative_binds) ? GetAbsolutePositionFromRelativeAxes() :
                                                              InputManager::GetPointerAbsolutePosition(m_cursor_index);
-  g_gpu->ConvertScreenCoordinatesToDisplayCoordinates(window_x, window_y, &display_x, &display_y);
+  g_gpu.ConvertScreenCoordinatesToDisplayCoordinates(window_x, window_y, &display_x, &display_y);
 
   // are we within the active display area?
   u32 tick, line;
   if (display_x < 0 || display_y < 0 ||
-      !g_gpu->ConvertDisplayCoordinatesToBeamTicksAndLines(display_x, display_y, m_x_scale, &tick, &line) ||
+      !g_gpu.ConvertDisplayCoordinatesToBeamTicksAndLines(display_x, display_y, m_x_scale, &tick, &line) ||
       m_shoot_offscreen)
   {
     DEV_LOG("Lightgun out of range for window coordinates {:.0f},{:.0f}", window_x, window_y);
@@ -234,11 +234,11 @@ void Justifier::UpdatePosition()
   m_irq_tick = static_cast<u16>(static_cast<TickCount>(tick) +
                                 System::ScaleTicksToOverclock(static_cast<TickCount>(m_tick_offset)));
   m_irq_first_line = static_cast<u16>(std::clamp<s32>(static_cast<s32>(line) + m_first_line_offset,
-                                                      static_cast<s32>(g_gpu->GetCRTCActiveStartLine()),
+                                                      static_cast<s32>(g_gpu.GetCRTCActiveStartLine()),
-                                                      static_cast<s32>(g_gpu->GetCRTCActiveEndLine())));
+                                                      static_cast<s32>(g_gpu.GetCRTCActiveEndLine())));
   m_irq_last_line = static_cast<u16>(std::clamp<s32>(static_cast<s32>(line) + m_last_line_offset,
-                                                     static_cast<s32>(g_gpu->GetCRTCActiveStartLine()),
+                                                     static_cast<s32>(g_gpu.GetCRTCActiveStartLine()),
-                                                     static_cast<s32>(g_gpu->GetCRTCActiveEndLine())));
+                                                     static_cast<s32>(g_gpu.GetCRTCActiveEndLine())));
 
   DEV_LOG("Lightgun window coordinates {},{} -> dpy {},{} -> tick {} line {} [{}-{}]", window_x, window_y, display_x,
           display_y, tick, line, m_irq_first_line, m_irq_last_line);
@@ -255,7 +255,7 @@ void Justifier::UpdateIRQEvent()
     return;
 
   u32 current_tick, current_line;
-  g_gpu->GetBeamPosition(&current_tick, &current_line);
+  g_gpu.GetBeamPosition(&current_tick, &current_line);
 
   u32 target_line;
   if (current_line < m_irq_first_line || current_line >= m_irq_last_line)
@@ -263,7 +263,7 @@ void Justifier::UpdateIRQEvent()
   else
     target_line = current_line + 1;
 
-  const TickCount ticks_until_pos = g_gpu->GetSystemTicksUntilTicksAndLine(m_irq_tick, target_line);
+  const TickCount ticks_until_pos = g_gpu.GetSystemTicksUntilTicksAndLine(m_irq_tick, target_line);
   DEBUG_LOG("Triggering IRQ in {} ticks @ tick {} line {}", ticks_until_pos, m_irq_tick, target_line);
   m_irq_event.Schedule(ticks_until_pos);
 }

src/core/negcon.cpp

@@ -10,6 +10,7 @@
 #include "common/assert.h"
 #include "common/bitutils.h"
 #include "common/log.h"
+#include "common/settings_interface.h"
 
 #include "IconsPromptFont.h"
 
@@ -261,14 +262,14 @@ std::unique_ptr<NeGcon> NeGcon::Create(u32 index)
 
 static const Controller::ControllerBindingInfo s_binding_info[] = {
 #define BUTTON(name, display_name, icon_name, button, genb)                                                            \
-  {                                                                                                                    \
+  {name, display_name, icon_name, static_cast<u32>(button), InputBindingInfo::Type::Button, genb}
-    name, display_name, icon_name, static_cast<u32>(button), InputBindingInfo::Type::Button, genb                      \
-  }
 #define AXIS(name, display_name, icon_name, halfaxis, genb)                                                            \
-  {                                                                                                                    \
+  {name,                                                                                                               \
-    name, display_name, icon_name, static_cast<u32>(NeGcon::Button::Count) + static_cast<u32>(halfaxis),               \
+   display_name,                                                                                                       \
-      InputBindingInfo::Type::HalfAxis, genb                                                                           \
+   icon_name,                                                                                                          \
-  }
+   static_cast<u32>(NeGcon::Button::Count) + static_cast<u32>(halfaxis),                                               \
+   InputBindingInfo::Type::HalfAxis,                                                                                   \
+   genb}
 
   // clang-format off
   BUTTON("Up", TRANSLATE_NOOP("NeGcon", "D-Pad Up"), ICON_PF_DPAD_UP, NeGcon::Button::Up, GenericInputBinding::DPadUp),

src/core/performance_counters.cpp

@@ -3,6 +3,8 @@
 
 #include "performance_counters.h"
 #include "gpu.h"
+#include "gpu_backend.h"
+#include "gpu_thread.h"
 #include "system.h"
 #include "system_private.h"
 
@@ -45,9 +47,9 @@ struct State
   float cpu_thread_usage;
   float cpu_thread_time;
 
-  u64 last_sw_time;
+  u64 last_gpu_thread_time;
-  float sw_thread_usage;
+  float gpu_thread_usage;
-  float sw_thread_time;
+  float gpu_thread_time;
 
   float average_gpu_time;
   float accumulated_gpu_time;
@@ -105,14 +107,14 @@ float PerformanceCounters::GetCPUThreadAverageTime()
   return s_state.cpu_thread_time;
 }
 
-float PerformanceCounters::GetSWThreadUsage()
+float PerformanceCounters::GetGPUThreadUsage()
 {
-  return s_state.sw_thread_usage;
+  return s_state.gpu_thread_usage;
 }
 
-float PerformanceCounters::GetSWThreadAverageTime()
+float PerformanceCounters::GetGPUThreadAverageTime()
 {
-  return s_state.sw_thread_time;
+  return s_state.gpu_thread_time;
 }
 
 float PerformanceCounters::GetGPUUsage()
@@ -150,17 +152,16 @@ void PerformanceCounters::Reset()
   s_state.last_frame_number = System::GetFrameNumber();
   s_state.last_internal_frame_number = System::GetInternalFrameNumber();
   s_state.last_cpu_time = System::GetCPUThreadHandle().GetCPUTime();
-  if (const Threading::Thread* sw_thread = g_gpu->GetSWThread(); sw_thread)
+  s_state.last_gpu_thread_time = GPUThread::Internal::GetThreadHandle().GetCPUTime();
-    s_state.last_sw_time = sw_thread->GetCPUTime();
-  else
-    s_state.last_sw_time = 0;
 
   s_state.average_frame_time_accumulator = 0.0f;
   s_state.minimum_frame_time_accumulator = 0.0f;
   s_state.maximum_frame_time_accumulator = 0.0f;
+
+  std::atomic_thread_fence(std::memory_order_release);
 }
 
-void PerformanceCounters::Update(u32 frame_number, u32 internal_frame_number)
+void PerformanceCounters::Update(GPUBackend* gpu, u32 frame_number, u32 internal_frame_number)
 {
   const Timer::Value now_ticks = Timer::GetCurrentValue();
 
@@ -177,7 +178,7 @@ void PerformanceCounters::Update(u32 frame_number, u32 internal_frame_number)
   // update fps counter
   const Timer::Value ticks_diff = now_ticks - s_state.last_update_time;
   const float time = static_cast<float>(Timer::ConvertValueToSeconds(ticks_diff));
-  if (time < PERFORMANCE_COUNTER_UPDATE_INTERVAL)
+  if (time < PERFORMANCE_COUNTER_UPDATE_INTERVAL || s_state.last_frame_number == frame_number)
     return;
 
   s_state.last_update_time = now_ticks;
@@ -202,18 +203,17 @@ void PerformanceCounters::Update(u32 frame_number, u32 internal_frame_number)
   s_state.fps = static_cast<float>(internal_frames_run) / time;
   s_state.speed = (s_state.vps / System::GetVideoFrameRate()) * 100.0f;
 
-  const Threading::Thread* sw_thread = g_gpu->GetSWThread();
   const u64 cpu_time = System::GetCPUThreadHandle().GetCPUTime();
-  const u64 sw_time = sw_thread ? sw_thread->GetCPUTime() : 0;
+  const u64 gpu_thread_time = GPUThread::Internal::GetThreadHandle().GetCPUTime();
   const u64 cpu_delta = cpu_time - s_state.last_cpu_time;
-  const u64 sw_delta = sw_time - s_state.last_sw_time;
+  const u64 gpu_thread_delta = gpu_thread_time - s_state.last_gpu_thread_time;
   s_state.last_cpu_time = cpu_time;
-  s_state.last_sw_time = sw_time;
+  s_state.last_gpu_thread_time = gpu_thread_time;
 
   s_state.cpu_thread_usage = static_cast<float>(static_cast<double>(cpu_delta) * pct_divider);
   s_state.cpu_thread_time = static_cast<float>(static_cast<double>(cpu_delta) * time_divider);
-  s_state.sw_thread_usage = static_cast<float>(static_cast<double>(sw_delta) * pct_divider);
+  s_state.gpu_thread_usage = static_cast<float>(static_cast<double>(gpu_thread_delta) * pct_divider);
-  s_state.sw_thread_time = static_cast<float>(static_cast<double>(sw_delta) * time_divider);
+  s_state.gpu_thread_time = static_cast<float>(static_cast<double>(gpu_thread_delta) * time_divider);
 
   if (MediaCapture* cap = System::GetMediaCapture())
     cap->UpdateCaptureThreadUsage(pct_divider, time_divider);
@@ -228,13 +228,13 @@ void PerformanceCounters::Update(u32 frame_number, u32 internal_frame_number)
   s_state.presents_since_last_update = 0;
 
   if (g_settings.display_show_gpu_stats)
-    g_gpu->UpdateStatistics(frames_run);
+    gpu->UpdateStatistics(frames_run);
 
-  VERBOSE_LOG("FPS: {:.2f} VPS: {:.2f} CPU: {:.2f} GPU: {:.2f} Avg: {:.2f}ms Min: {:.2f}ms Max: {:.2f}ms", s_state.fps,
+  VERBOSE_LOG("FPS: {:.2f} VPS: {:.2f} CPU: {:.2f} RNDR: {:.2f} GPU: {:.2f} Avg: {:.2f}ms Min: {:.2f}ms Max: {:.2f}ms",
-              s_state.vps, s_state.cpu_thread_usage, s_state.gpu_usage, s_state.average_frame_time,
+              s_state.fps, s_state.vps, s_state.cpu_thread_usage, s_state.gpu_thread_usage, s_state.gpu_usage,
-              s_state.minimum_frame_time, s_state.maximum_frame_time);
+              s_state.average_frame_time, s_state.minimum_frame_time, s_state.maximum_frame_time);
 
-  Host::OnPerformanceCountersUpdated();
+  Host::OnPerformanceCountersUpdated(gpu);
 }
 
 void PerformanceCounters::AccumulateGPUTime()

src/core/performance_counters.h

@@ -5,6 +5,8 @@
 
 #include "common/types.h"
 
+class GPUBackend;
+
 namespace PerformanceCounters
 {
 static constexpr u32 NUM_FRAME_TIME_SAMPLES = 150;
@@ -18,8 +20,8 @@ float GetMinimumFrameTime();
 float GetMaximumFrameTime();
 float GetCPUThreadUsage();
 float GetCPUThreadAverageTime();
-float GetSWThreadUsage();
+float GetGPUThreadUsage();
-float GetSWThreadAverageTime();
+float GetGPUThreadAverageTime();
 float GetGPUUsage();
 float GetGPUAverageTime();
 const FrameTimeHistory& GetFrameTimeHistory();
@@ -27,7 +29,7 @@ u32 GetFrameTimeHistoryPos();
 
 void Clear();
 void Reset();
-void Update(u32 frame_number, u32 internal_frame_number);
+void Update(GPUBackend* gpu, u32 frame_number, u32 internal_frame_number);
 void AccumulateGPUTime();
 
 } // namespace Host

src/core/playstation_mouse.cpp

@@ -10,6 +10,7 @@
 
 #include "common/assert.h"
 #include "common/log.h"
+#include "common/settings_interface.h"
 
 #include "IconsPromptFont.h"
 

src/core/settings.cpp

@@ -5,6 +5,7 @@
 #include "achievements.h"
 #include "controller.h"
 #include "host.h"
+#include "imgui_overlays.h"
 #include "system.h"
 
 #include "util/gpu_device.h"
@@ -28,7 +29,8 @@
 
 LOG_CHANNEL(Settings);
 
-Settings g_settings;
+ALIGN_TO_CACHE_LINE Settings g_settings;
+ALIGN_TO_CACHE_LINE GPUSettings g_gpu_settings;
 
 const char* SettingInfo::StringDefaultValue() const
 {
@@ -86,6 +88,8 @@ float SettingInfo::FloatStepValue() const
   return step_value ? StringUtil::FromChars<float>(step_value).value_or(fallback_value) : fallback_value;
 }
 
+GPUSettings::GPUSettings() = default;
+
 #ifdef DYNAMIC_HOST_PAGE_SIZE
 // See note in settings.h - 16K ends up faster with LUT because of nearby code/data.
 const CPUFastmemMode Settings::DEFAULT_CPU_FASTMEM_MODE =
@@ -206,6 +210,7 @@ void Settings::Load(const SettingsInterface& si, const SettingsInterface& contro
   gpu_disable_compressed_textures = si.GetBoolValue("GPU", "DisableCompressedTextures", 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_true_color = si.GetBoolValue("GPU", "TrueColor", true);
   gpu_scaled_dithering = si.GetBoolValue("GPU", "ScaledDithering", true);
@@ -250,6 +255,9 @@ void Settings::Load(const SettingsInterface& si, const SettingsInterface& contro
   gpu_pgxp_depth_buffer = si.GetBoolValue("GPU", "PGXPDepthBuffer", false);
   gpu_pgxp_disable_2d = si.GetBoolValue("GPU", "PGXPDisableOn2DPolygons", false);
   SetPGXPDepthClearThreshold(si.GetFloatValue("GPU", "PGXPDepthClearThreshold", DEFAULT_GPU_PGXP_DEPTH_THRESHOLD));
+  gpu_show_vram = si.GetBoolValue("Debug", "ShowVRAM");
+  gpu_dump_cpu_to_vram_copies = si.GetBoolValue("Debug", "DumpCPUToVRAMCopies");
+  gpu_dump_vram_to_cpu_copies = si.GetBoolValue("Debug", "DumpVRAMToCPUCopies");
   gpu_dump_fast_replay_mode = si.GetBoolValue("GPU", "DumpFastReplayMode", false);
 
   display_deinterlacing_mode =
@@ -421,13 +429,9 @@ void Settings::Load(const SettingsInterface& si, const SettingsInterface& contro
   achievements_leaderboard_duration =
     si.GetIntValue("Cheevos", "LeaderboardsDuration", DEFAULT_LEADERBOARD_NOTIFICATION_TIME);
 
-  debugging.show_vram = si.GetBoolValue("Debug", "ShowVRAM");
-  debugging.dump_cpu_to_vram_copies = si.GetBoolValue("Debug", "DumpCPUToVRAMCopies");
-  debugging.dump_vram_to_cpu_copies = si.GetBoolValue("Debug", "DumpVRAMToCPUCopies");
-
 #ifndef __ANDROID__
-  debugging.enable_gdb_server = si.GetBoolValue("Debug", "EnableGDBServer");
+  enable_gdb_server = si.GetBoolValue("Debug", "EnableGDBServer");
-  debugging.gdb_server_port = static_cast<u16>(si.GetUIntValue("Debug", "GDBServerPort", DEFAULT_GDB_SERVER_PORT));
+  gdb_server_port = static_cast<u16>(si.GetUIntValue("Debug", "GDBServerPort", DEFAULT_GDB_SERVER_PORT));
 #endif
 
   texture_replacements.enable_texture_replacements =
@@ -554,6 +558,7 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
   }
 
   si.SetBoolValue("GPU", "PerSampleShading", gpu_per_sample_shading);
+  si.SetUIntValue("GPU", "MaxQueuedFrames", gpu_max_queued_frames);
   si.SetBoolValue("GPU", "UseThread", gpu_use_thread);
   si.SetBoolValue("GPU", "UseSoftwareRendererForReadbacks", gpu_use_software_renderer_for_readbacks);
   si.SetBoolValue("GPU", "TrueColor", gpu_true_color);
@@ -581,6 +586,9 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
   si.SetBoolValue("GPU", "PGXPDepthBuffer", gpu_pgxp_depth_buffer);
   si.SetBoolValue("GPU", "PGXPDisableOn2DPolygons", gpu_pgxp_disable_2d);
   si.SetFloatValue("GPU", "PGXPDepthClearThreshold", GetPGXPDepthClearThreshold());
+  si.SetBoolValue("Debug", "ShowVRAM", gpu_show_vram);
+  si.SetBoolValue("Debug", "DumpCPUToVRAMCopies", gpu_dump_cpu_to_vram_copies);
+  si.SetBoolValue("Debug", "DumpVRAMToCPUCopies", gpu_dump_vram_to_cpu_copies);
   si.SetBoolValue("GPU", "DumpFastReplayMode", gpu_dump_fast_replay_mode);
 
   si.SetStringValue("GPU", "DeinterlacingMode", GetDisplayDeinterlacingModeName(display_deinterlacing_mode));
@@ -697,17 +705,10 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
   si.SetIntValue("Cheevos", "NotificationsDuration", achievements_notification_duration);
   si.SetIntValue("Cheevos", "LeaderboardsDuration", achievements_leaderboard_duration);
 
-  if (!ignore_base)
-  {
-    si.SetBoolValue("Debug", "ShowVRAM", debugging.show_vram);
-    si.SetBoolValue("Debug", "DumpCPUToVRAMCopies", debugging.dump_cpu_to_vram_copies);
-    si.SetBoolValue("Debug", "DumpVRAMToCPUCopies", debugging.dump_vram_to_cpu_copies);
-
 #ifndef __ANDROID__
-    si.SetBoolValue("Debug", "EnableGDBServer", debugging.enable_gdb_server);
+  si.SetBoolValue("Debug", "EnableGDBServer", enable_gdb_server);
-    si.SetUIntValue("Debug", "GDBServerPort", debugging.gdb_server_port);
+  si.SetUIntValue("Debug", "GDBServerPort", gdb_server_port);
 #endif
-  }
 
   si.SetBoolValue("TextureReplacements", "EnableTextureReplacements", texture_replacements.enable_texture_replacements);
   si.SetBoolValue("TextureReplacements", "EnableVRAMWriteReplacements",
@@ -950,7 +951,7 @@ std::string Settings::TextureReplacementSettings::Configuration::ExportToYAML(bo
                      comment_str, replacement_scale_linear_filter);    // ReplacementScaleLinearFilter
 }
 
-void Settings::FixIncompatibleSettings(bool display_osd_messages)
+void Settings::FixIncompatibleSettings(const SettingsInterface& si, bool display_osd_messages)
 {
   if (g_settings.disable_all_enhancements)
   {
@@ -1022,6 +1023,13 @@ void Settings::FixIncompatibleSettings(bool display_osd_messages)
     (g_settings.gpu_renderer != GPURenderer::Software && g_settings.gpu_texture_cache);
   g_settings.texture_replacements.enable_vram_write_replacements &= (g_settings.gpu_renderer != GPURenderer::Software);
 
+  // GPU thread should be disabled if any debug windows are active, since they will be racing to read CPU thread state.
+  if (g_settings.gpu_use_thread && g_settings.gpu_max_queued_frames > 0 && ImGuiManager::AreAnyDebugWindowsEnabled(si))
+  {
+    WARNING_LOG("Setting maximum queued frames to 0 because one or more debug windows are enabled.");
+    g_settings.gpu_max_queued_frames = 0;
+  }
+
 #ifndef ENABLE_MMAP_FASTMEM
   if (g_settings.cpu_fastmem_mode == CPUFastmemMode::MMap)
   {
@@ -1072,15 +1080,30 @@ void Settings::FixIncompatibleSettings(bool display_osd_messages)
     }
 
 #ifndef __ANDROID__
-    g_settings.debugging.enable_gdb_server = false;
+    g_settings.enable_gdb_server = false;
 #endif
 
-    g_settings.debugging.show_vram = false;
+    g_settings.gpu_show_vram = false;
-    g_settings.debugging.dump_cpu_to_vram_copies = false;
+    g_settings.gpu_dump_cpu_to_vram_copies = false;
-    g_settings.debugging.dump_vram_to_cpu_copies = false;
+    g_settings.gpu_dump_vram_to_cpu_copies = false;
   }
 }
 
+bool Settings::AreGPUDeviceSettingsChanged(const Settings& old_settings) const
+{
+  return (gpu_use_debug_device != old_settings.gpu_use_debug_device ||
+          gpu_disable_shader_cache != old_settings.gpu_disable_shader_cache ||
+          gpu_disable_dual_source_blend != old_settings.gpu_disable_dual_source_blend ||
+          gpu_disable_framebuffer_fetch != old_settings.gpu_disable_framebuffer_fetch ||
+          gpu_disable_texture_buffers != old_settings.gpu_disable_texture_buffers ||
+          gpu_disable_texture_copy_to_self != old_settings.gpu_disable_texture_copy_to_self ||
+          gpu_disable_memory_import != old_settings.gpu_disable_memory_import ||
+          gpu_disable_raster_order_views != old_settings.gpu_disable_raster_order_views ||
+          gpu_disable_compute_shaders != old_settings.gpu_disable_compute_shaders ||
+          gpu_disable_compressed_textures != old_settings.gpu_disable_compressed_textures ||
+          display_exclusive_fullscreen_control != old_settings.display_exclusive_fullscreen_control);
+}
+
 void Settings::SetDefaultLogConfig(SettingsInterface& si)
 {
   si.SetStringValue("Logging", "LogLevel", GetLogLevelName(DEFAULT_LOG_LEVEL));
@@ -1757,7 +1780,7 @@ const char* Settings::GetDisplayAspectRatioDisplayName(DisplayAspectRatio ar)
                                   "DisplayAspectRatio");
 }
 
-float Settings::GetDisplayAspectRatioValue() const
+float GPUSettings::GetDisplayAspectRatioValue() const
 {
   return s_display_aspect_ratio_values[static_cast<size_t>(display_aspect_ratio)];
 }

src/core/settings.h

@@ -8,8 +8,6 @@
 #include "util/audio_stream.h"
 
 #include "common/log.h"
-#include "common/settings_interface.h"
-#include "common/small_string.h"
 
 #include <array>
 #include <optional>
@@ -18,6 +16,8 @@
 #include <string_view>
 #include <vector>
 
+class SettingsInterface;
+
 enum class RenderAPI : u8;
 enum class MediaCaptureBackend : u8;
 
@@ -57,47 +57,17 @@ struct SettingInfo
   float FloatStepValue() const;
 };
 
-struct Settings
+struct GPUSettings
 {
-  Settings();
+  GPUSettings();
 
-  ConsoleRegion region = DEFAULT_CONSOLE_REGION;
+  std::string gpu_adapter;
-
-  CPUExecutionMode cpu_execution_mode = DEFAULT_CPU_EXECUTION_MODE;
-  CPUFastmemMode cpu_fastmem_mode = DEFAULT_CPU_FASTMEM_MODE;
-  bool cpu_overclock_enable : 1 = false;
-  bool cpu_overclock_active : 1 = false;
-  bool cpu_recompiler_memory_exceptions : 1 = false;
-  bool cpu_recompiler_block_linking : 1 = true;
-  bool cpu_recompiler_icache : 1 = false;
-  u32 cpu_overclock_numerator = 1;
-  u32 cpu_overclock_denominator = 1;
-
-  float emulation_speed = 1.0f;
-  float fast_forward_speed = 0.0f;
-  float turbo_speed = 0.0f;
-  bool sync_to_host_refresh_rate : 1 = false;
-  bool inhibit_screensaver : 1 = true;
-  bool pause_on_focus_loss : 1 = false;
-  bool pause_on_controller_disconnection : 1 = false;
-  bool save_state_on_exit : 1 = true;
-  bool create_save_state_backups : 1 = DEFAULT_SAVE_STATE_BACKUPS;
-  bool confim_power_off : 1 = true;
-  bool load_devices_from_save_states : 1 = false;
-  bool apply_compatibility_settings : 1 = true;
-  bool apply_game_settings : 1 = true;
-  bool disable_all_enhancements : 1 = false;
-  bool enable_discord_presence : 1 = false;
-
-  bool rewind_enable : 1 = false;
-  float rewind_save_frequency = 10.0f;
-  u16 rewind_save_slots = 10;
-  u8 runahead_frames = 0;
 
   GPURenderer gpu_renderer = DEFAULT_GPU_RENDERER;
-  std::string gpu_adapter;
   u8 gpu_resolution_scale = 1;
   u8 gpu_multisamples = 1;
+  u8 gpu_max_queued_frames = DEFAULT_GPU_MAX_QUEUED_FRAMES;
+
   bool gpu_use_thread : 1 = true;
   bool gpu_use_software_renderer_for_readbacks : 1 = false;
   bool gpu_use_debug_device : 1 = false;
@@ -117,6 +87,11 @@ struct Settings
   bool gpu_accurate_blending : 1 = false;
   bool gpu_widescreen_hack : 1 = false;
   bool gpu_texture_cache : 1 = false;
+  bool gpu_show_vram : 1 = false;
+  bool gpu_dump_cpu_to_vram_copies : 1 = false;
+  bool gpu_dump_vram_to_cpu_copies : 1 = false;
+  bool gpu_dump_fast_replay_mode : 1 = false;
+
   bool gpu_pgxp_enable : 1 = false;
   bool gpu_pgxp_culling : 1 = true;
   bool gpu_pgxp_texture_correction : 1 = true;
@@ -126,6 +101,7 @@ struct Settings
   bool gpu_pgxp_preserve_proj_fp : 1 = false;
   bool gpu_pgxp_depth_buffer : 1 = false;
   bool gpu_pgxp_disable_2d : 1 = false;
+
   ForceVideoTimingMode gpu_force_video_timing = DEFAULT_FORCE_VIDEO_TIMING_MODE;
   GPUTextureFilter gpu_texture_filter = DEFAULT_GPU_TEXTURE_FILTER;
   GPUTextureFilter gpu_sprite_texture_filter = DEFAULT_GPU_TEXTURE_FILTER;
@@ -175,62 +151,6 @@ struct Settings
   float gpu_pgxp_tolerance = -1.0f;
   float gpu_pgxp_depth_clear_threshold = DEFAULT_GPU_PGXP_DEPTH_THRESHOLD / GPU_PGXP_DEPTH_THRESHOLD_SCALE;
 
-  SaveStateCompressionMode save_state_compression = DEFAULT_SAVE_STATE_COMPRESSION_MODE;
-
-  u8 cdrom_readahead_sectors = DEFAULT_CDROM_READAHEAD_SECTORS;
-  CDROMMechaconVersion cdrom_mechacon_version = DEFAULT_CDROM_MECHACON_VERSION;
-  bool cdrom_region_check : 1 = false;
-  bool cdrom_subq_skew : 1 = false;
-  bool cdrom_load_image_to_ram : 1 = false;
-  bool cdrom_load_image_patches : 1 = false;
-  bool cdrom_mute_cd_audio : 1 = false;
-  u32 cdrom_read_speedup = 1;
-  u32 cdrom_seek_speedup = 1;
-
-  std::string audio_driver;
-  std::string audio_output_device;
-  u32 audio_output_volume = 100;
-  u32 audio_fast_forward_volume = 100;
-  AudioStreamParameters audio_stream_parameters;
-  AudioBackend audio_backend = AudioStream::DEFAULT_BACKEND;
-  bool audio_output_muted : 1 = false;
-
-  bool use_old_mdec_routines : 1 = false;
-  bool pcdrv_enable : 1 = false;
-  bool export_shared_memory : 1 = false;
-
-  // timing hacks section
-  TickCount dma_max_slice_ticks = DEFAULT_DMA_MAX_SLICE_TICKS;
-  TickCount dma_halt_ticks = DEFAULT_DMA_HALT_TICKS;
-  u32 gpu_fifo_size = DEFAULT_GPU_FIFO_SIZE;
-  TickCount gpu_max_run_ahead = DEFAULT_GPU_MAX_RUN_AHEAD;
-
-  // achievements
-  bool achievements_enabled : 1 = false;
-  bool achievements_hardcore_mode : 1 = false;
-  bool achievements_notifications : 1 = true;
-  bool achievements_leaderboard_notifications : 1 = true;
-  bool achievements_sound_effects : 1 = true;
-  bool achievements_overlays : 1 = true;
-  bool achievements_encore_mode : 1 = false;
-  bool achievements_spectator_mode : 1 = false;
-  bool achievements_unofficial_test_mode : 1 = false;
-  bool achievements_use_raintegration : 1 = false;
-  s32 achievements_notification_duration = DEFAULT_ACHIEVEMENT_NOTIFICATION_TIME;
-  s32 achievements_leaderboard_duration = DEFAULT_LEADERBOARD_NOTIFICATION_TIME;
-
-  struct DebugSettings
-  {
-#ifndef __ANDROID__
-    u16 gdb_server_port = DEFAULT_GDB_SERVER_PORT;
-    bool enable_gdb_server : 1 = false;
-#endif
-
-    bool show_vram : 1 = false;
-    bool dump_cpu_to_vram_copies : 1 = false;
-    bool dump_vram_to_cpu_copies : 1 = false;
-  } debugging;
-
   // texture replacements
   struct TextureReplacementSettings
   {
@@ -284,11 +204,143 @@ struct Settings
     bool operator!=(const TextureReplacementSettings& rhs) const;
   } texture_replacements;
 
+  float GetDisplayAspectRatioValue() const;
+
+  ALWAYS_INLINE bool IsUsingSoftwareRenderer() const { return (gpu_renderer == GPURenderer::Software); }
+  ALWAYS_INLINE bool IsUsingAccurateBlending() const { return (gpu_accurate_blending && !gpu_true_color); }
+
+  ALWAYS_INLINE bool UsingPGXPCPUMode() const { return gpu_pgxp_enable && gpu_pgxp_cpu; }
+  ALWAYS_INLINE bool UsingPGXPDepthBuffer() const { return gpu_pgxp_enable && gpu_pgxp_depth_buffer; }
+  ALWAYS_INLINE float GetPGXPDepthClearThreshold() const
+  {
+    return gpu_pgxp_depth_clear_threshold * GPU_PGXP_DEPTH_THRESHOLD_SCALE;
+  }
+  ALWAYS_INLINE void SetPGXPDepthClearThreshold(float value)
+  {
+    gpu_pgxp_depth_clear_threshold = value / GPU_PGXP_DEPTH_THRESHOLD_SCALE;
+  }
+
+  static constexpr GPURenderer DEFAULT_GPU_RENDERER = GPURenderer::Automatic;
+  static constexpr GPUTextureFilter DEFAULT_GPU_TEXTURE_FILTER = GPUTextureFilter::Nearest;
+  static constexpr GPULineDetectMode DEFAULT_GPU_LINE_DETECT_MODE = GPULineDetectMode::Disabled;
+  static constexpr GPUDownsampleMode DEFAULT_GPU_DOWNSAMPLE_MODE = GPUDownsampleMode::Disabled;
+  static constexpr GPUWireframeMode DEFAULT_GPU_WIREFRAME_MODE = GPUWireframeMode::Disabled;
+  static constexpr GPUDumpCompressionMode DEFAULT_GPU_DUMP_COMPRESSION_MODE = GPUDumpCompressionMode::ZstDefault;
+  static constexpr float DEFAULT_GPU_PGXP_DEPTH_THRESHOLD = 300.0f;
+  static constexpr float GPU_PGXP_DEPTH_THRESHOLD_SCALE = 4096.0f;
+
+  static constexpr DisplayDeinterlacingMode DEFAULT_DISPLAY_DEINTERLACING_MODE = DisplayDeinterlacingMode::Progressive;
+  static constexpr DisplayCropMode DEFAULT_DISPLAY_CROP_MODE = DisplayCropMode::Overscan;
+  static constexpr DisplayAspectRatio DEFAULT_DISPLAY_ASPECT_RATIO = DisplayAspectRatio::Auto;
+  static constexpr DisplayAlignment DEFAULT_DISPLAY_ALIGNMENT = DisplayAlignment::Center;
+  static constexpr DisplayRotation DEFAULT_DISPLAY_ROTATION = DisplayRotation::Normal;
+  static constexpr DisplayScalingMode DEFAULT_DISPLAY_SCALING = DisplayScalingMode::BilinearSmooth;
+  static constexpr ForceVideoTimingMode DEFAULT_FORCE_VIDEO_TIMING_MODE = ForceVideoTimingMode::Disabled;
+  static constexpr DisplayExclusiveFullscreenControl DEFAULT_DISPLAY_EXCLUSIVE_FULLSCREEN_CONTROL =
+    DisplayExclusiveFullscreenControl::Automatic;
+  static constexpr DisplayScreenshotMode DEFAULT_DISPLAY_SCREENSHOT_MODE = DisplayScreenshotMode::ScreenResolution;
+  static constexpr DisplayScreenshotFormat DEFAULT_DISPLAY_SCREENSHOT_FORMAT = DisplayScreenshotFormat::PNG;
+  static constexpr u8 DEFAULT_DISPLAY_SCREENSHOT_QUALITY = 85;
+  static constexpr float DEFAULT_DISPLAY_PRE_FRAME_SLEEP_BUFFER = 2.0f;
+  static constexpr float DEFAULT_OSD_SCALE = 100.0f;
+
+#ifndef __ANDROID__
+  static constexpr u8 DEFAULT_GPU_MAX_QUEUED_FRAMES = 2;
+#else
+  static constexpr u8 DEFAULT_GPU_MAX_QUEUED_FRAMES = 3;
+#endif
+};
+
+struct Settings : public GPUSettings
+{
+  Settings();
+
+  ConsoleRegion region = DEFAULT_CONSOLE_REGION;
+
+  CPUExecutionMode cpu_execution_mode = DEFAULT_CPU_EXECUTION_MODE;
+  CPUFastmemMode cpu_fastmem_mode = DEFAULT_CPU_FASTMEM_MODE;
+  bool cpu_overclock_enable : 1 = false;
+  bool cpu_overclock_active : 1 = false;
+  bool cpu_recompiler_memory_exceptions : 1 = false;
+  bool cpu_recompiler_block_linking : 1 = true;
+  bool cpu_recompiler_icache : 1 = false;
+  u32 cpu_overclock_numerator = 1;
+  u32 cpu_overclock_denominator = 1;
+
+  float emulation_speed = 1.0f;
+  float fast_forward_speed = 0.0f;
+  float turbo_speed = 0.0f;
+  bool sync_to_host_refresh_rate : 1 = false;
+  bool inhibit_screensaver : 1 = true;
+  bool pause_on_focus_loss : 1 = false;
+  bool pause_on_controller_disconnection : 1 = false;
+  bool save_state_on_exit : 1 = true;
+  bool create_save_state_backups : 1 = DEFAULT_SAVE_STATE_BACKUPS;
+  bool confim_power_off : 1 = true;
+  bool load_devices_from_save_states : 1 = false;
+  bool apply_compatibility_settings : 1 = true;
+  bool apply_game_settings : 1 = true;
+  bool disable_all_enhancements : 1 = false;
+  bool enable_discord_presence : 1 = false;
+
+  bool rewind_enable : 1 = false;
+  float rewind_save_frequency = 10.0f;
+  u16 rewind_save_slots = 10;
+  u8 runahead_frames = 0;
+
+  SaveStateCompressionMode save_state_compression = DEFAULT_SAVE_STATE_COMPRESSION_MODE;
+
+  u8 cdrom_readahead_sectors = DEFAULT_CDROM_READAHEAD_SECTORS;
+  CDROMMechaconVersion cdrom_mechacon_version = DEFAULT_CDROM_MECHACON_VERSION;
+  bool cdrom_region_check : 1 = false;
+  bool cdrom_subq_skew : 1 = false;
+  bool cdrom_load_image_to_ram : 1 = false;
+  bool cdrom_load_image_patches : 1 = false;
+  bool cdrom_mute_cd_audio : 1 = false;
+  u32 cdrom_read_speedup = 1;
+  u32 cdrom_seek_speedup = 1;
+
+  std::string audio_driver;
+  std::string audio_output_device;
+  u32 audio_output_volume = 100;
+  u32 audio_fast_forward_volume = 100;
+  AudioStreamParameters audio_stream_parameters;
+  AudioBackend audio_backend = AudioStream::DEFAULT_BACKEND;
+  bool audio_output_muted : 1 = false;
+
+  bool use_old_mdec_routines : 1 = false;
+  bool pcdrv_enable : 1 = false;
+  bool export_shared_memory : 1 = false;
+
+  // timing hacks section
+  TickCount dma_max_slice_ticks = DEFAULT_DMA_MAX_SLICE_TICKS;
+  TickCount dma_halt_ticks = DEFAULT_DMA_HALT_TICKS;
+  u32 gpu_fifo_size = DEFAULT_GPU_FIFO_SIZE;
+  TickCount gpu_max_run_ahead = DEFAULT_GPU_MAX_RUN_AHEAD;
+
+  // achievements
+  bool achievements_enabled : 1 = false;
+  bool achievements_hardcore_mode : 1 = false;
+  bool achievements_notifications : 1 = true;
+  bool achievements_leaderboard_notifications : 1 = true;
+  bool achievements_sound_effects : 1 = true;
+  bool achievements_overlays : 1 = true;
+  bool achievements_encore_mode : 1 = false;
+  bool achievements_spectator_mode : 1 = false;
+  bool achievements_unofficial_test_mode : 1 = false;
+  bool achievements_use_raintegration : 1 = false;
+  s32 achievements_notification_duration = DEFAULT_ACHIEVEMENT_NOTIFICATION_TIME;
+  s32 achievements_leaderboard_duration = DEFAULT_LEADERBOARD_NOTIFICATION_TIME;
+
+#ifndef __ANDROID__
+  u16 gdb_server_port = DEFAULT_GDB_SERVER_PORT;
+  bool enable_gdb_server : 1 = false;
+#endif
+
   bool bios_tty_logging : 1 = false;
   bool bios_patch_fast_boot : 1 = DEFAULT_FAST_BOOT_VALUE;
   bool bios_fast_forward_boot : 1 = false;
   bool enable_8mb_ram : 1 = false;
-  bool gpu_dump_fast_replay_mode : 1 = false;
 
   std::array<ControllerType, NUM_CONTROLLER_AND_CARD_PORTS> controller_types{};
   std::array<MemoryCardType, NUM_CONTROLLER_AND_CARD_PORTS> memory_card_types{};
@@ -305,33 +357,14 @@ struct Settings
   std::string pcdrv_root;
   bool pcdrv_enable_writes = false;
 
-  ALWAYS_INLINE bool IsUsingSoftwareRenderer() const { return (gpu_renderer == GPURenderer::Software); }
-  ALWAYS_INLINE bool IsUsingAccurateBlending() const { return (gpu_accurate_blending && !gpu_true_color); }
   ALWAYS_INLINE bool IsRunaheadEnabled() const { return (runahead_frames > 0); }
 
-  ALWAYS_INLINE PGXPMode GetPGXPMode()
-  {
-    return gpu_pgxp_enable ? (gpu_pgxp_cpu ? PGXPMode::CPU : PGXPMode::Memory) : PGXPMode::Disabled;
-  }
-
-  ALWAYS_INLINE bool UsingPGXPDepthBuffer() const { return gpu_pgxp_enable && gpu_pgxp_depth_buffer; }
-  ALWAYS_INLINE bool UsingPGXPCPUMode() const { return gpu_pgxp_enable && gpu_pgxp_cpu; }
-  ALWAYS_INLINE float GetPGXPDepthClearThreshold() const
-  {
-    return gpu_pgxp_depth_clear_threshold * GPU_PGXP_DEPTH_THRESHOLD_SCALE;
-  }
-  ALWAYS_INLINE void SetPGXPDepthClearThreshold(float value)
-  {
-    gpu_pgxp_depth_clear_threshold = value / GPU_PGXP_DEPTH_THRESHOLD_SCALE;
-  }
 
   ALWAYS_INLINE s32 GetAudioOutputVolume(bool fast_forwarding) const
   {
     return audio_output_muted ? 0 : (fast_forwarding ? audio_fast_forward_volume : audio_output_volume);
   }
 
-  float GetDisplayAspectRatioValue() const;
-
   ALWAYS_INLINE bool IsPort1MultitapEnabled() const
   {
     return (multitap_mode == MultitapMode::Port1Only || multitap_mode == MultitapMode::BothPorts);
@@ -378,7 +411,9 @@ struct Settings
   void Save(SettingsInterface& si, bool ignore_base) const;
   static void Clear(SettingsInterface& si);
 
-  void FixIncompatibleSettings(bool display_osd_messages);
+  void FixIncompatibleSettings(const SettingsInterface& si, bool display_osd_messages);
+
+  bool AreGPUDeviceSettingsChanged(const Settings& old_settings) const;
 
   /// Initializes configuration.
   static void SetDefaultLogConfig(SettingsInterface& si);
@@ -495,15 +530,7 @@ struct Settings
   static const char* GetPIODeviceTypeModeName(PIODeviceType type);
   static const char* GetPIODeviceTypeModeDisplayName(PIODeviceType type);
 
-  static constexpr GPURenderer DEFAULT_GPU_RENDERER = GPURenderer::Automatic;
-  static constexpr GPUTextureFilter DEFAULT_GPU_TEXTURE_FILTER = GPUTextureFilter::Nearest;
-  static constexpr GPULineDetectMode DEFAULT_GPU_LINE_DETECT_MODE = GPULineDetectMode::Disabled;
-  static constexpr GPUDownsampleMode DEFAULT_GPU_DOWNSAMPLE_MODE = GPUDownsampleMode::Disabled;
-  static constexpr GPUWireframeMode DEFAULT_GPU_WIREFRAME_MODE = GPUWireframeMode::Disabled;
-  static constexpr GPUDumpCompressionMode DEFAULT_GPU_DUMP_COMPRESSION_MODE = GPUDumpCompressionMode::ZstDefault;
   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;
 
   // Prefer recompiler when supported.
 #ifdef ENABLE_RECOMPILER
@@ -521,21 +548,6 @@ struct Settings
   static constexpr CPUFastmemMode DEFAULT_CPU_FASTMEM_MODE = CPUFastmemMode::LUT;
 #endif
 
-  static constexpr DisplayDeinterlacingMode DEFAULT_DISPLAY_DEINTERLACING_MODE = DisplayDeinterlacingMode::Progressive;
-  static constexpr DisplayCropMode DEFAULT_DISPLAY_CROP_MODE = DisplayCropMode::Overscan;
-  static constexpr DisplayAspectRatio DEFAULT_DISPLAY_ASPECT_RATIO = DisplayAspectRatio::Auto;
-  static constexpr DisplayAlignment DEFAULT_DISPLAY_ALIGNMENT = DisplayAlignment::Center;
-  static constexpr DisplayRotation DEFAULT_DISPLAY_ROTATION = DisplayRotation::Normal;
-  static constexpr DisplayScalingMode DEFAULT_DISPLAY_SCALING = DisplayScalingMode::BilinearSmooth;
-  static constexpr ForceVideoTimingMode DEFAULT_FORCE_VIDEO_TIMING_MODE = ForceVideoTimingMode::Disabled;
-  static constexpr DisplayExclusiveFullscreenControl DEFAULT_DISPLAY_EXCLUSIVE_FULLSCREEN_CONTROL =
-    DisplayExclusiveFullscreenControl::Automatic;
-  static constexpr DisplayScreenshotMode DEFAULT_DISPLAY_SCREENSHOT_MODE = DisplayScreenshotMode::ScreenResolution;
-  static constexpr DisplayScreenshotFormat DEFAULT_DISPLAY_SCREENSHOT_FORMAT = DisplayScreenshotFormat::PNG;
-  static constexpr u8 DEFAULT_DISPLAY_SCREENSHOT_QUALITY = 85;
-  static constexpr float DEFAULT_DISPLAY_PRE_FRAME_SLEEP_BUFFER = 2.0f;
-  static constexpr float DEFAULT_OSD_SCALE = 100.0f;
-
   static constexpr u8 DEFAULT_CDROM_READAHEAD_SECTORS = 8;
   static constexpr CDROMMechaconVersion DEFAULT_CDROM_MECHACON_VERSION = CDROMMechaconVersion::VC1A;
 
@@ -571,7 +583,8 @@ struct Settings
 #endif
 };
 
-extern Settings g_settings;
+ALIGN_TO_CACHE_LINE extern Settings g_settings;        // CPU thread copy.
+ALIGN_TO_CACHE_LINE extern GPUSettings g_gpu_settings; // GPU thread copy.
 
 namespace EmuFolders {
 extern std::string AppRoot;

src/core/system.h

@@ -165,7 +165,6 @@ std::unique_ptr<INISettingsInterface> GetGameSettingsInterface(const GameDatabas
 std::string GetInputProfilePath(std::string_view name);
 
 State GetState();
-void SetState(State new_state);
 bool IsRunning();
 bool IsPaused();
 bool IsShutdown();
@@ -384,7 +383,7 @@ s32 GetAudioOutputVolume();
 void UpdateVolume();
 
 /// Saves a screenshot to the specified file. If no file name is provided, one will be generated automatically.
-bool SaveScreenshot(const char* path = nullptr, DisplayScreenshotMode mode = g_settings.display_screenshot_mode,
+void SaveScreenshot(const char* path = nullptr, DisplayScreenshotMode mode = g_settings.display_screenshot_mode,
                     DisplayScreenshotFormat format = g_settings.display_screenshot_format,
                     u8 quality = g_settings.display_screenshot_quality, bool compress_on_thread = true);
 
@@ -400,7 +399,6 @@ MediaCapture* GetMediaCapture();
 
 /// Media capture (video and/or audio). If no path is provided, one will be generated automatically.
 bool StartMediaCapture(std::string path = {});
-bool StartMediaCapture(std::string path, bool capture_video, bool capture_audio);
 void StopMediaCapture();
 
 /// Toggle Widescreen Hack and Aspect Ratio
@@ -413,15 +411,11 @@ void ToggleSoftwareRendering();
 /// If the scale is set to 0, the internal resolution will be used, otherwise it is treated as a multiplier to 1x.
 void RequestDisplaySize(float scale = 0.0f);
 
-/// Renders the display.
-bool PresentDisplay(bool explicit_present, u64 present_time);
-void InvalidateDisplay();
-
 //////////////////////////////////////////////////////////////////////////
 // Memory Save States (Rewind and Runahead)
 //////////////////////////////////////////////////////////////////////////
 void CalculateRewindMemoryUsage(u32 num_saves, u32 resolution_scale, u64* ram_usage, u64* vram_usage);
-void ClearMemorySaveStates();
+void ClearMemorySaveStates(bool reallocate_resources);
 void SetRunaheadReplayFlag();
 
 /// Shared socket multiplexer, used by PINE/GDB/etc.

src/core/system_private.h

@@ -7,24 +7,32 @@
 
 #include <functional>
 
+class GPUBackend;
+struct GPUBackendFramePresentationParameters;
+
 namespace System {
 
 /// Memory save states - only for internal use.
 struct MemorySaveState
 {
-  std::unique_ptr<GPUTexture> vram_texture;
   DynamicHeapArray<u8> state_data;
   size_t state_size;
+
+  std::unique_ptr<GPUTexture> vram_texture;
+  DynamicHeapArray<u8> gpu_state_data;
+  size_t gpu_state_size;
 };
 
-bool SaveMemoryState(MemorySaveState* mss);
+MemorySaveState& AllocateMemoryState();
-bool LoadMemoryState(const MemorySaveState& mss);
+MemorySaveState& GetFirstMemoryState();
-
+MemorySaveState& PopMemoryState();
-/// Returns the maximum size of a save state, considering the current configuration.
+bool AllocateMemoryStates(size_t state_count);
-size_t GetMaxSaveStateSize();
+void FreeMemoryStateTextures();
-
+void FreeMemoryStateStorage();
-bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display, bool is_memory_state);
+void LoadMemoryState(MemorySaveState& mss, bool update_display);
+void SaveMemoryState(MemorySaveState& mss);
 
+bool IsRunaheadActive();
 void IncrementFrameNumber();
 void IncrementInternalFrameNumber();
 void FrameDone();
@@ -33,6 +41,10 @@ void FrameDone();
 GPUVSyncMode GetEffectiveVSyncMode();
 bool ShouldAllowPresentThrottle();
 
+/// Retrieves timing information for frame presentation on the GPU thread.
+/// Returns false if this frame should not be presented or the command buffer flushed.
+bool GetFramePresentationParameters(GPUBackendFramePresentationParameters* frame);
+
 /// Call when host display size changes.
 void DisplayWindowResized();
 
@@ -63,6 +75,7 @@ void IdlePollUpdate();
 /// Task threads, asynchronous work which will block system shutdown.
 void QueueTaskOnThread(std::function<void()> task);
 void RemoveSelfFromTaskThreads();
+void JoinTaskThreads();
 
 } // namespace System
 
@@ -89,11 +102,8 @@ void OnSystemPaused();
 /// Called when the VM is resumed after being paused.
 void OnSystemResumed();
 
-/// Called when the pause state changes, or fullscreen UI opens.
-void OnIdleStateChanged();
-
 /// Called when performance metrics are updated, approximately once a second.
-void OnPerformanceCountersUpdated();
+void OnPerformanceCountersUpdated(const GPUBackend* gpu_backend);
 
 /// Provided by the host; called when the running executable changes.
 void OnGameChanged(const std::string& disc_path, const std::string& game_serial, const std::string& game_name);

src/core/timers.cpp

@@ -315,8 +315,8 @@ u32 Timers::ReadRegister(u32 offset)
       if (timer_index < 2 && cs.external_counting_enabled)
       {
         // timers 0/1 depend on the GPU
-        if (timer_index == 0 || g_gpu->IsCRTCScanlinePending())
+        if (timer_index == 0 || g_gpu.IsCRTCScanlinePending())
-          g_gpu->SynchronizeCRTC();
+          g_gpu.SynchronizeCRTC();
       }
 
       s_state.sysclk_event.InvokeEarly();
@@ -329,8 +329,8 @@ u32 Timers::ReadRegister(u32 offset)
       if (timer_index < 2 && cs.external_counting_enabled)
       {
         // timers 0/1 depend on the GPU
-        if (timer_index == 0 || g_gpu->IsCRTCScanlinePending())
+        if (timer_index == 0 || g_gpu.IsCRTCScanlinePending())
-          g_gpu->SynchronizeCRTC();
+          g_gpu.SynchronizeCRTC();
       }
 
       s_state.sysclk_event.InvokeEarly();
@@ -365,8 +365,8 @@ void Timers::WriteRegister(u32 offset, u32 value)
   if (timer_index < 2 && cs.external_counting_enabled)
   {
     // timers 0/1 depend on the GPU
-    if (timer_index == 0 || g_gpu->IsCRTCScanlinePending())
+    if (timer_index == 0 || g_gpu.IsCRTCScanlinePending())
-      g_gpu->SynchronizeCRTC();
+      g_gpu.SynchronizeCRTC();
   }
 
   s_state.sysclk_event.InvokeEarly();

src/core/timing_event.cpp

@@ -10,6 +10,7 @@
 
 #include "common/assert.h"
 #include "common/log.h"
+#include "common/small_string.h"
 #include "common/thirdparty/SmallVector.h"
 
 LOG_CHANNEL(TimingEvents);

src/duckstation-qt/debuggerwindow.cpp

@@ -241,14 +241,14 @@ void DebuggerWindow::onBreakpointListItemChanged(QTreeWidgetItem* item, int colu
 
 void DebuggerWindow::onStepIntoActionTriggered()
 {
-  Assert(System::IsPaused());
+  Assert(QtHost::IsSystemPaused());
   saveCurrentState();
   g_emu_thread->singleStepCPU();
 }
 
 void DebuggerWindow::onStepOverActionTriggered()
 {
-  Assert(System::IsPaused());
+  Assert(QtHost::IsSystemPaused());
   if (!CPU::AddStepOverBreakpoint())
   {
     onStepIntoActionTriggered();
@@ -262,7 +262,7 @@ void DebuggerWindow::onStepOverActionTriggered()
 
 void DebuggerWindow::onStepOutActionTriggered()
 {
-  Assert(System::IsPaused());
+  Assert(QtHost::IsSystemPaused());
   if (!CPU::AddStepOutBreakpoint())
   {
     QMessageBox::critical(this, tr("Debugger"), tr("Failed to add step-out breakpoint, are you in a valid function?"));

src/duckstation-qt/graphicssettingswidget.cpp

@@ -293,6 +293,9 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
   // Debugging Tab
 
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.gpuThread, "GPU", "UseThread", true);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.maxQueuedFrames, "GPU", "MaxQueuedFrames",
+                                              Settings::DEFAULT_GPU_MAX_QUEUED_FRAMES);
+  connect(m_ui.gpuThread, &QCheckBox::checkStateChanged, this, &GraphicsSettingsWidget::onGPUThreadChanged);
 
   SettingWidgetBinder::BindWidgetToEnumSetting(
     sif, m_ui.gpuDumpCompressionMode, "GPU", "DumpCompressionMode", &Settings::ParseGPUDumpCompressionMode,
@@ -325,6 +328,7 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
   onMediaCaptureVideoEnabledChanged();
   onEnableTextureCacheChanged();
   onEnableAnyTextureReplacementsChanged();
+  onGPUThreadChanged();
   onShowDebugSettingsChanged(QtHost::ShouldShowDebugOptions());
 
   // Rendering Tab
@@ -610,8 +614,8 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
                              tr("Draws a wireframe outline of the triangles rendered by the console's GPU, either as a "
                                 "replacement or an overlay."));
   dialog->registerWidgetHelp(m_ui.gpuThread, tr("Threaded Rendering"), tr("Checked"),
-                             tr("Uses a second thread for drawing graphics. Currently only available for the software "
+                             tr("Uses a second thread for drawing graphics. Provides a significant speed improvement "
-                                "renderer, but can provide a significant speed improvement, and is safe to use."));
+                                "particularly with the software renderer, and is safe to use."));
 
   dialog->registerWidgetHelp(
     m_ui.useDebugDevice, tr("Use Debug Device"), tr("Unchecked"),
@@ -819,8 +823,6 @@ void GraphicsSettingsWidget::updateRendererDependentOptions()
   m_ui.blitSwapChain->setEnabled(render_api == RenderAPI::D3D11);
 #endif
 
-  m_ui.gpuThread->setEnabled(!is_hardware);
-
   m_ui.exclusiveFullscreenLabel->setEnabled(render_api == RenderAPI::D3D11 || render_api == RenderAPI::D3D12 ||
                                             render_api == RenderAPI::Vulkan);
   m_ui.exclusiveFullscreenControl->setEnabled(render_api == RenderAPI::Vulkan);
@@ -1181,6 +1183,13 @@ void GraphicsSettingsWidget::onEnableAnyTextureReplacementsChanged()
   m_ui.preloadTextureReplacements->setEnabled(any_replacements_enabled);
 }
 
+void GraphicsSettingsWidget::onGPUThreadChanged()
+{
+  const bool enabled = m_dialog->getEffectiveBoolValue("GPU", "UseThread", true);
+  m_ui.maxQueuedFrames->setEnabled(enabled);
+  m_ui.maxQueuedFramesLabel->setEnabled(enabled);
+}
+
 void GraphicsSettingsWidget::onTextureReplacementOptionsClicked()
 {
   QDialog dlg(QtUtils::GetRootWidget(this));

src/duckstation-qt/graphicssettingswidget.h

@@ -44,6 +44,8 @@ private Q_SLOTS:
   void onEnableAnyTextureReplacementsChanged();
   void onTextureReplacementOptionsClicked();
 
+  void onGPUThreadChanged();
+
 private:
   static constexpr int TAB_INDEX_RENDERING = 0;
   static constexpr int TAB_INDEX_ADVANCED = 1;

src/duckstation-qt/graphicssettingswidget.ui

@@ -1286,11 +1286,29 @@
           <item row="1" column="0" colspan="2">
            <layout class="QGridLayout" name="gridLayout_4">
             <item row="0" column="0">
-             <widget class="QCheckBox" name="gpuThread">
+             <layout class="QHBoxLayout" name="horizontalLayout_7" stretch="1,0,0">
-              <property name="text">
+              <item>
-               <string>Threaded Rendering</string>
+               <widget class="QCheckBox" name="gpuThread">
-              </property>
+                <property name="text">
-             </widget>
+                 <string>Threaded Rendering</string>
+                </property>
+               </widget>
+              </item>
+              <item>
+               <widget class="QLabel" name="maxQueuedFramesLabel">
+                <property name="text">
+                 <string>Max Queued Frames:</string>
+                </property>
+               </widget>
+              </item>
+              <item>
+               <widget class="QSpinBox" name="maxQueuedFrames">
+                <property name="maximum">
+                 <number>10</number>
+                </property>
+               </widget>
+              </item>
+             </layout>
             </item>
            </layout>
           </item>

src/duckstation-qt/mainwindow.cpp

@@ -85,6 +85,7 @@ static bool s_use_central_widget = false;
 // UI thread VM validity.
 static bool s_system_valid = false;
 static bool s_system_paused = false;
+static bool s_fullscreen_ui_started = false;
 static std::atomic_uint32_t s_system_locked{false};
 static QString s_current_game_title;
 static QString s_current_game_serial;
@@ -764,7 +765,7 @@ void MainWindow::recreate()
   {
     g_emu_thread->setSurfaceless(false);
     g_main_window->updateEmulationActions(false, System::IsValid(), Achievements::IsHardcoreModeActive());
-    g_main_window->onFullscreenUIStateChange(g_emu_thread->isRunningFullscreenUI());
+    g_main_window->onFullscreenUIStartedOrStopped(s_fullscreen_ui_started);
   }
 
   if (controller_settings_window_pos.has_value())
@@ -1258,8 +1259,9 @@ void MainWindow::onStartFullscreenUITriggered()
     g_emu_thread->startFullscreenUI();
 }
 
-void MainWindow::onFullscreenUIStateChange(bool running)
+void MainWindow::onFullscreenUIStartedOrStopped(bool running)
 {
+  s_fullscreen_ui_started = running;
   m_ui.actionStartFullscreenUI->setText(running ? tr("Stop Big Picture Mode") : tr("Start Big Picture Mode"));
   m_ui.actionStartFullscreenUI2->setText(running ? tr("Exit Big Picture") : tr("Big Picture"));
 }
@@ -2046,7 +2048,7 @@ void MainWindow::connectSignals()
   connect(g_emu_thread, &EmuThread::mediaCaptureStarted, this, &MainWindow::onMediaCaptureStarted);
   connect(g_emu_thread, &EmuThread::mediaCaptureStopped, this, &MainWindow::onMediaCaptureStopped);
   connect(g_emu_thread, &EmuThread::mouseModeRequested, this, &MainWindow::onMouseModeRequested);
-  connect(g_emu_thread, &EmuThread::fullscreenUIStateChange, this, &MainWindow::onFullscreenUIStateChange);
+  connect(g_emu_thread, &EmuThread::fullscreenUIStartedOrStopped, this, &MainWindow::onFullscreenUIStartedOrStopped);
   connect(g_emu_thread, &EmuThread::achievementsLoginRequested, this, &MainWindow::onAchievementsLoginRequested);
   connect(g_emu_thread, &EmuThread::achievementsChallengeModeChanged, this,
           &MainWindow::onAchievementsChallengeModeChanged);
@@ -2503,7 +2505,7 @@ bool MainWindow::requestShutdown(bool allow_confirm /* = true */, bool allow_sav
   // reshow the main window during display updates, because otherwise fullscreen transitions and renderer switches
   // would briefly show and then hide the main window. So instead, we do it on shutdown, here. Except if we're in
   // batch mode, when we're going to exit anyway.
-  if (!isRenderingToMain() && isHidden() && !QtHost::InBatchMode() && !g_emu_thread->isRunningFullscreenUI())
+  if (!isRenderingToMain() && isHidden() && !QtHost::InBatchMode() && !s_fullscreen_ui_started)
     updateWindowState(true);
 
   // Now we can actually shut down the VM.

src/duckstation-qt/mainwindow.h

@@ -168,7 +168,7 @@ private Q_SLOTS:
   void onCheatsActionTriggered();
   void onCheatsMenuAboutToShow();
   void onStartFullscreenUITriggered();
-  void onFullscreenUIStateChange(bool running);
+  void onFullscreenUIStartedOrStopped(bool running);
   void onRemoveDiscActionTriggered();
   void onScanForNewGamesTriggered();
   void onViewToolbarActionToggled(bool checked);

src/duckstation-qt/qthost.cpp

@@ -19,7 +19,9 @@
 #include "core/game_list.h"
 #include "core/gdb_server.h"
 #include "core/gpu.h"
+#include "core/gpu_backend.h"
 #include "core/gpu_hw_texture_cache.h"
+#include "core/gpu_thread.h"
 #include "core/host.h"
 #include "core/imgui_overlays.h"
 #include "core/memory_card.h"
@@ -223,7 +225,6 @@ bool QtHost::SaveGameSettings(SettingsInterface* sif, bool delete_if_empty)
   INISettingsInterface* ini = static_cast<INISettingsInterface*>(sif);
   Error error;
 
-
   // if there's no keys, just toss the whole thing out
   if (delete_if_empty && ini->IsEmpty())
   {
@@ -576,13 +577,8 @@ void Host::LoadSettings(const SettingsInterface& si, std::unique_lock<std::mutex
 void EmuThread::checkForSettingsChanges(const Settings& old_settings)
 {
   if (g_main_window)
-  {
     QMetaObject::invokeMethod(g_main_window, &MainWindow::checkForSettingChanges, Qt::QueuedConnection);
 
-    if (System::IsValid())
-      updatePerformanceCounters();
-  }
-
   // don't mess with fullscreen while locked
   if (!QtHost::IsSystemLocked())
   {
@@ -591,7 +587,7 @@ void EmuThread::checkForSettingsChanges(const Settings& old_settings)
     {
       m_is_rendering_to_main = render_to_main;
       if (g_gpu_device)
-        Host::UpdateDisplayWindow(m_is_fullscreen);
+        GPUThread::UpdateDisplayWindow(m_is_fullscreen);
     }
   }
 }
@@ -730,33 +726,24 @@ void EmuThread::startFullscreenUI()
     return;
   }
 
-  if (System::IsValid())
+  if (System::IsValid() || m_is_fullscreen_ui_started)
     return;
 
   // we want settings loaded so we choose the correct renderer
   // this also sorts out input sources.
   System::LoadSettings(false);
   m_is_rendering_to_main = shouldRenderToMain();
-  m_run_fullscreen_ui = true;
 
   // borrow the game start fullscreen flag
   const bool start_fullscreen =
     (s_start_fullscreen_ui_fullscreen || Host::GetBaseBoolSettingValue("Main", "StartFullscreen", false));
 
   Error error;
-  if (!Host::CreateGPUDevice(Settings::GetRenderAPIForRenderer(g_settings.gpu_renderer), start_fullscreen, &error) ||
+  if (!GPUThread::StartFullscreenUI(start_fullscreen, &error))
-      !FullscreenUI::Initialize())
   {
     Host::ReportErrorAsync("Error", error.GetDescription());
-    m_run_fullscreen_ui = false;
     return;
   }
-
-  emit fullscreenUIStateChange(true);
-
-  // poll more frequently so we don't lose events
-  stopBackgroundControllerPollTimer();
-  startBackgroundControllerPollTimer();
 }
 
 void EmuThread::stopFullscreenUI()
@@ -771,18 +758,8 @@ void EmuThread::stopFullscreenUI()
     return;
   }
 
-  setFullscreen(false, true);
+  if (m_is_fullscreen_ui_started)
-
+    GPUThread::StopFullscreenUI();
-  if (m_run_fullscreen_ui)
-  {
-    m_run_fullscreen_ui = false;
-    emit fullscreenUIStateChange(false);
-  }
-
-  if (!g_gpu_device)
-    return;
-
-  Host::ReleaseGPUDevice();
 }
 
 void EmuThread::bootSystem(std::shared_ptr<SystemBootParameters> params)
@@ -889,7 +866,7 @@ void EmuThread::onDisplayWindowMouseWheelEvent(const QPoint& delta_angle)
 
 void EmuThread::onDisplayWindowResized(int width, int height, float scale)
 {
-  Host::ResizeDisplayWindow(width, height, scale);
+  GPUThread::ResizeDisplayWindow(width, height, scale);
 }
 
 void EmuThread::redrawDisplayWindow()
@@ -900,10 +877,10 @@ void EmuThread::redrawDisplayWindow()
     return;
   }
 
-  if (!g_gpu_device || System::IsShutdown())
+  if (System::IsShutdown())
     return;
 
-  System::InvalidateDisplay();
+  GPUThread::PresentCurrentFrame();
 }
 
 void EmuThread::toggleFullscreen()
@@ -931,7 +908,7 @@ 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(fullscreen);
+  GPUThread::UpdateDisplayWindow(fullscreen);
 }
 
 bool Host::IsFullscreen()
@@ -960,7 +937,7 @@ void EmuThread::setSurfaceless(bool surfaceless)
     return;
 
   m_is_surfaceless = surfaceless;
-  Host::UpdateDisplayWindow(false);
+  GPUThread::UpdateDisplayWindow(false);
 }
 
 void EmuThread::requestDisplaySize(float scale)
@@ -1017,6 +994,7 @@ void Host::OnSystemStarting()
 void Host::OnSystemStarted()
 {
   g_emu_thread->stopBackgroundControllerPollTimer();
+  g_emu_thread->wakeThread();
 
   emit g_emu_thread->systemStarted();
 }
@@ -1034,6 +1012,7 @@ void Host::OnSystemResumed()
     g_emu_thread->setSurfaceless(false);
 
   emit g_emu_thread->systemResumed();
+  g_emu_thread->wakeThread();
 
   g_emu_thread->stopBackgroundControllerPollTimer();
 }
@@ -1045,9 +1024,14 @@ void Host::OnSystemDestroyed()
   emit g_emu_thread->systemDestroyed();
 }
 
-void Host::OnIdleStateChanged()
+void Host::OnFullscreenUIStartedOrStopped(bool started)
 {
-  g_emu_thread->wakeThread();
+  g_emu_thread->setFullscreenUIStarted(started);
+}
+
+void Host::OnGPUThreadRunIdleChanged(bool is_active)
+{
+  g_emu_thread->setGPUThreadRunIdle(is_active);
 }
 
 void EmuThread::reloadInputSources()
@@ -1291,7 +1275,12 @@ void EmuThread::reloadPostProcessingShaders()
   }
 
   if (System::IsValid())
-    PostProcessing::ReloadShaders();
+  {
+    GPUThread::RunOnThread([]() {
+      if (GPUThread::HasGPUBackend())
+        PostProcessing::ReloadShaders();
+    });
+  }
 }
 
 void EmuThread::updatePostProcessingSettings()
@@ -1303,7 +1292,12 @@ void EmuThread::updatePostProcessingSettings()
   }
 
   if (System::IsValid())
-    PostProcessing::UpdateSettings();
+  {
+    GPUThread::RunOnThread([]() {
+      if (GPUThread::HasGPUBackend())
+        PostProcessing::UpdateSettings();
+    });
+  }
 }
 
 void EmuThread::clearInputBindStateFromSource(InputBindingKey key)
@@ -1326,7 +1320,7 @@ void EmuThread::reloadTextureReplacements()
   }
 
   if (System::IsValid())
-    GPUTextureCache::ReloadTextureReplacements(true);
+    GPUThread::RunOnThread([]() { GPUTextureCache::ReloadTextureReplacements(true); });
 }
 
 void EmuThread::captureGPUFrameDump()
@@ -1679,7 +1673,8 @@ void Host::DestroyAuxiliaryRenderWindow(AuxiliaryRenderWindowHandle handle, s32*
     *height = size.height();
 
   // eat all pending events, to make sure we're not going to write input events back to a dead pointer
-  g_emu_thread->getEventLoop()->processEvents(QEventLoop::AllEvents);
+  if (g_emu_thread->isCurrentThread())
+    g_emu_thread->getEventLoop()->processEvents(QEventLoop::AllEvents);
 }
 
 void EmuThread::queueAuxiliaryRenderWindowInputEvent(Host::AuxiliaryRenderWindowUserData userdata,
@@ -1699,10 +1694,12 @@ void EmuThread::processAuxiliaryRenderWindowInputEvent(void* userdata, quint32 e
                                                        quint32 param3)
 {
   DebugAssert(isCurrentThread());
-  ImGuiManager::ProcessAuxiliaryRenderWindowInputEvent(userdata, static_cast<Host::AuxiliaryRenderWindowEvent>(event),
+  GPUThread::RunOnThread([userdata, event, param1, param2, param3]() {
-                                                       Host::AuxiliaryRenderWindowEventParam{.uint_param = param1},
+    ImGuiManager::ProcessAuxiliaryRenderWindowInputEvent(userdata, static_cast<Host::AuxiliaryRenderWindowEvent>(event),
-                                                       Host::AuxiliaryRenderWindowEventParam{.uint_param = param2},
+                                                         Host::AuxiliaryRenderWindowEventParam{.uint_param = param1},
-                                                       Host::AuxiliaryRenderWindowEventParam{.uint_param = param3});
+                                                         Host::AuxiliaryRenderWindowEventParam{.uint_param = param2},
+                                                         Host::AuxiliaryRenderWindowEventParam{.uint_param = param3});
+  });
 }
 
 void EmuThread::doBackgroundControllerPoll()
@@ -1731,7 +1728,7 @@ void EmuThread::startBackgroundControllerPollTimer()
     return;
 
   u32 poll_interval = BACKGROUND_CONTROLLER_POLLING_INTERVAL;
-  if (FullscreenUI::IsInitialized())
+  if (m_gpu_thread_run_idle)
     poll_interval = FULLSCREEN_UI_CONTROLLER_POLLING_INTERVAL;
   if (GDBServer::HasAnyClients())
     poll_interval = GDB_SERVER_POLLING_INTERVAL;
@@ -1747,6 +1744,37 @@ void EmuThread::stopBackgroundControllerPollTimer()
   m_background_controller_polling_timer->stop();
 }
 
+void EmuThread::setGPUThreadRunIdle(bool active)
+{
+  if (!isCurrentThread())
+  {
+    QMetaObject::invokeMethod(this, "setGPUThreadRunIdle", Qt::QueuedConnection, Q_ARG(bool, active));
+    return;
+  }
+
+  m_gpu_thread_run_idle = active;
+
+  // break out of the event loop if we're not executing a system
+  if (active && !g_settings.gpu_use_thread && !System::IsRunning())
+    m_event_loop->quit();
+
+  // adjust the timer speed to pick up controller input faster
+  if (!m_background_controller_polling_timer->isActive())
+    return;
+
+  g_emu_thread->stopBackgroundControllerPollTimer();
+  g_emu_thread->startBackgroundControllerPollTimer();
+}
+
+void EmuThread::setFullscreenUIStarted(bool started)
+{
+  if (m_is_fullscreen_ui_started == started)
+    return;
+
+  m_is_fullscreen_ui_started = started;
+  emit fullscreenUIStartedOrStopped(started);
+}
+
 void EmuThread::start()
 {
   AssertMsg(!g_emu_thread, "Emu thread does not exist");
@@ -1776,8 +1804,6 @@ void EmuThread::stopInThread()
 
 void EmuThread::run()
 {
-  Threading::SetNameOfCurrentThread("CPU Thread");
-
   m_event_loop = new QEventLoop();
   m_started_semaphore.release();
 
@@ -1796,6 +1822,9 @@ void EmuThread::run()
   createBackgroundControllerPollTimer();
   startBackgroundControllerPollTimer();
 
+  // kick off GPU thread
+  Threading::Thread gpu_thread(&EmuThread::gpuThreadEntryPoint);
+
   // main loop
   while (!m_shutdown_flag)
   {
@@ -1803,24 +1832,17 @@ void EmuThread::run()
     {
       System::Execute();
     }
+    else if (!GPUThread::IsUsingThread() && GPUThread::IsRunningIdle())
+    {
+      g_emu_thread->getEventLoop()->processEvents(QEventLoop::AllEvents);
+
+      // have to double-check the condition after processing events, because the events could shut us down
+      if (!GPUThread::IsUsingThread() && GPUThread::IsRunningIdle())
+        GPUThread::Internal::DoRunIdle();
+    }
     else
     {
-      // we want to keep rendering the UI when paused and fullscreen UI is enabled
+      m_event_loop->exec();
-      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::IdlePollUpdate();
-      if (g_gpu_device && g_gpu_device->HasMainSwapChain())
-      {
-        System::PresentDisplay(false, 0);
-        if (!g_gpu_device->GetMainSwapChain()->IsVSyncModeBlocking())
-          g_gpu_device->GetMainSwapChain()->ThrottlePresentation();
-      }
     }
   }
 
@@ -1828,13 +1850,25 @@ void EmuThread::run()
     System::ShutdownSystem(false);
 
   destroyBackgroundControllerPollTimer();
+
+  // tell GPU thread to exit
+  GPUThread::Internal::RequestShutdown();
+  gpu_thread.Join();
+
+  // and tidy up everything left
   System::CPUThreadShutdown();
 
   // move back to UI thread
   moveToThread(m_ui_thread);
 }
 
-void Host::FrameDone()
+void EmuThread::gpuThreadEntryPoint()
+{
+  Threading::SetNameOfCurrentThread("GPU Thread");
+  GPUThread::Internal::GPUThreadEntryPoint();
+}
+
+void Host::FrameDoneOnGPUThread(GPUBackend* gpu_backend, u32 frame_number)
 {
 }
 
@@ -1949,7 +1983,7 @@ void Host::OnInputDeviceConnected(std::string_view identifier, std::string_view
 {
   emit g_emu_thread->onInputDeviceConnected(std::string(identifier), std::string(device_name));
 
-  if (System::IsValid() || g_emu_thread->isRunningFullscreenUI())
+  if (System::IsValid() || GPUThread::IsFullscreenUIRequested())
   {
     Host::AddIconOSDMessage(fmt::format("ControllerConnected{}", identifier), ICON_FA_GAMEPAD,
                             fmt::format(TRANSLATE_FS("QtHost", "Controller {} connected."), identifier),
@@ -1975,7 +2009,7 @@ void Host::OnInputDeviceDisconnected(InputBindingKey key, std::string_view ident
     Host::AddIconOSDMessage(fmt::format("ControllerConnected{}", identifier), ICON_FA_GAMEPAD, std::move(message),
                             Host::OSD_WARNING_DURATION);
   }
-  else if (System::IsValid() || g_emu_thread->isRunningFullscreenUI())
+  else if (System::IsValid() || GPUThread::IsFullscreenUIRequested())
   {
     Host::AddIconOSDMessage(fmt::format("ControllerConnected{}", identifier), ICON_FA_GAMEPAD,
                             fmt::format(TRANSLATE_FS("QtHost", "Controller {} disconnected."), identifier),
@@ -2037,17 +2071,17 @@ void Host::ReleaseRenderWindow()
   g_emu_thread->releaseRenderWindow();
 }
 
-void EmuThread::updatePerformanceCounters()
+void EmuThread::updatePerformanceCounters(const GPUBackend* gpu_backend)
 {
-  const RenderAPI render_api = g_gpu_device ? g_gpu_device->GetRenderAPI() : RenderAPI::None;
+  const RenderAPI render_api = g_gpu_device->GetRenderAPI();
-  const bool hardware_renderer = g_gpu && g_gpu->IsHardwareRenderer();
+  const bool hardware_renderer = GPUBackend::IsUsingHardwareBackend();
   u32 render_width = 0;
   u32 render_height = 0;
 
-  if (g_gpu)
+  if (gpu_backend)
   {
-    const u32 render_scale = g_gpu->GetResolutionScale();
+    const u32 render_scale = gpu_backend->GetResolutionScale();
-    std::tie(render_width, render_height) = g_gpu->GetFullDisplayResolution();
+    std::tie(render_width, render_height) = g_gpu.GetFullDisplayResolution();
     render_width *= render_scale;
     render_height *= render_scale;
   }
@@ -2110,9 +2144,9 @@ void EmuThread::resetPerformanceCounters()
                             Q_ARG(const QString&, blank));
 }
 
-void Host::OnPerformanceCountersUpdated()
+void Host::OnPerformanceCountersUpdated(const GPUBackend* gpu_backend)
 {
-  g_emu_thread->updatePerformanceCounters();
+  g_emu_thread->updatePerformanceCounters(gpu_backend);
 }
 
 void Host::OnGameChanged(const std::string& disc_path, const std::string& game_serial, const std::string& game_name)
@@ -2209,8 +2243,8 @@ std::optional<WindowInfo> Host::GetTopLevelWindowInfo()
 
 EmuThread::SystemLock EmuThread::pauseAndLockSystem()
 {
-  const bool was_fullscreen = System::IsValid() && isFullscreen();
+  const bool was_fullscreen = QtHost::IsSystemValid() && isFullscreen();
-  const bool was_paused = System::IsPaused();
+  const bool was_paused = QtHost::IsSystemPaused();
 
   // We use surfaceless rather than switching out of fullscreen, because
   // we're paused, so we're not going to be rendering anyway.

src/duckstation-qt/qthost.h

@@ -44,6 +44,8 @@ class INISettingsInterface;
 enum class RenderAPI : u8;
 class GPUDevice;
 
+class GPUBackend;
+
 class MainWindow;
 class DisplayWidget;
 
@@ -93,7 +95,6 @@ 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; }
-  ALWAYS_INLINE bool isRunningFullscreenUI() const { return m_run_fullscreen_ui; }
 
   std::optional<WindowInfo> acquireRenderWindow(RenderAPI render_api, bool fullscreen, bool exclusive_fullscreen,
                                                 Error* error);
@@ -102,6 +103,7 @@ public:
 
   void startBackgroundControllerPollTimer();
   void stopBackgroundControllerPollTimer();
+  void setFullscreenUIStarted(bool started);
   void wakeThread();
 
   bool shouldRenderToMain() const;
@@ -109,7 +111,7 @@ public:
 
   void bootOrLoadState(std::string path);
 
-  void updatePerformanceCounters();
+  void updatePerformanceCounters(const GPUBackend* gpu_backend);
   void resetPerformanceCounters();
 
   /// Locks the system by pausing it, while a popup dialog is displayed.
@@ -147,7 +149,7 @@ Q_SIGNALS:
   void runningGameChanged(const QString& filename, const QString& game_serial, const QString& game_title);
   void inputProfileLoaded();
   void mouseModeRequested(bool relative, bool hide_cursor);
-  void fullscreenUIStateChange(bool running);
+  void fullscreenUIStartedOrStopped(bool running);
   void achievementsLoginRequested(Achievements::LoginRequestReason reason);
   void achievementsRefreshed(quint32 id, const QString& game_info_string);
   void achievementsChallengeModeChanged(bool enabled);
@@ -210,6 +212,7 @@ public Q_SLOTS:
   void clearInputBindStateFromSource(InputBindingKey key);
   void reloadTextureReplacements();
   void captureGPUFrameDump();
+  void setGPUThreadRunIdle(bool active);
 
 private Q_SLOTS:
   void stopInThread();
@@ -227,23 +230,23 @@ protected:
   void run() override;
 
 private:
-  using InputButtonHandler = std::function<void(bool)>;
-  using InputAxisHandler = std::function<void(float)>;
-
   void createBackgroundControllerPollTimer();
   void destroyBackgroundControllerPollTimer();
   void confirmActionIfMemoryCardBusy(const QString& action, bool cancel_resume_on_accept,
                                      std::function<void(bool)> callback) const;
 
+  static void gpuThreadEntryPoint();
+
   QThread* m_ui_thread;
   QSemaphore m_started_semaphore;
   QEventLoop* m_event_loop = nullptr;
   QTimer* m_background_controller_polling_timer = nullptr;
 
   bool m_shutdown_flag = false;
-  bool m_run_fullscreen_ui = false;
   bool m_is_rendering_to_main = false;
   bool m_is_fullscreen = false;
+  bool m_is_fullscreen_ui_started = false;
+  bool m_gpu_thread_run_idle = false;
   bool m_is_surfaceless = false;
   bool m_save_state_on_shutdown = false;
 

src/duckstation-regtest/regtest_host.cpp

@@ -5,7 +5,8 @@
 #include "core/controller.h"
 #include "core/fullscreen_ui.h"
 #include "core/game_list.h"
-#include "core/gpu.h"
+#include "core/gpu_backend.h"
+#include "core/gpu_thread.h"
 #include "core/host.h"
 #include "core/system.h"
 #include "core/system_private.h"
@@ -46,9 +47,11 @@ static void HookSignals();
 static bool SetFolders();
 static bool SetNewDataRoot(const std::string& filename);
 static std::string GetFrameDumpFilename(u32 frame);
+static void GPUThreadEntryPoint();
 } // namespace RegTestHost
 
 static std::unique_ptr<MemorySettingsInterface> s_base_settings_interface;
+static Threading::Thread s_gpu_thread;
 
 static u32 s_frames_to_run = 60 * 60;
 static u32 s_frames_remaining = 0;
@@ -281,12 +284,17 @@ void Host::OnSystemResumed()
   //
 }
 
-void Host::OnIdleStateChanged()
+void Host::OnGPUThreadRunIdleChanged(bool is_active)
 {
   //
 }
 
-void Host::OnPerformanceCountersUpdated()
+void Host::OnFullscreenUIStartedOrStopped(bool started)
+{
+  //
+}
+
+void Host::OnPerformanceCountersUpdated(const GPUBackend* gpu_backend)
 {
   //
 }
@@ -375,14 +383,10 @@ void Host::DestroyAuxiliaryRenderWindow(AuxiliaryRenderWindowHandle handle, s32*
 {
 }
 
-void Host::FrameDone()
+void Host::FrameDoneOnGPUThread(GPUBackend* gpu_backend, u32 frame_number)
 {
-  const u32 frame = System::GetFrameNumber();
+  if (s_frame_dump_interval > 0 && (s_frame_dump_interval == 1 || (frame_number % s_frame_dump_interval) == 0))
-  if (s_frame_dump_interval > 0 && (s_frame_dump_interval == 1 || (frame % s_frame_dump_interval) == 0))
+    gpu_backend->WriteDisplayTextureToFile(RegTestHost::GetFrameDumpFilename(frame_number));
-  {
-    std::string dump_filename(RegTestHost::GetFrameDumpFilename(frame));
-    g_gpu->WriteDisplayTextureToFile(std::move(dump_filename));
-  }
 }
 
 void Host::OpenURL(std::string_view url)
@@ -508,6 +512,12 @@ void RegTestHost::HookSignals()
   std::signal(SIGTERM, SignalHandler);
 }
 
+void RegTestHost::GPUThreadEntryPoint()
+{
+  Threading::SetNameOfCurrentThread("CPU Thread");
+  GPUThread::Internal::GPUThreadEntryPoint();
+}
+
 void RegTestHost::InitializeEarlyConsole()
 {
   const bool was_console_enabled = Log::IsConsoleOutputEnabled();
@@ -773,6 +783,7 @@ int main(int argc, char* argv[])
   }
 
   RegTestHost::HookSignals();
+  s_gpu_thread.Start(&RegTestHost::GPUThreadEntryPoint);
 
   Error error;
   int result = -1;
@@ -813,6 +824,12 @@ int main(int argc, char* argv[])
   result = 0;
 
 cleanup:
+  if (s_gpu_thread.Joinable())
+  {
+    GPUThread::Internal::RequestShutdown();
+    s_gpu_thread.Join();
+  }
+
   System::CPUThreadShutdown();
   System::ProcessShutdown();
   return result;

src/util/gpu_device.cpp

@@ -385,6 +385,17 @@ const char* GPUDevice::ShaderLanguageToString(GPUShaderLanguage language)
   }
 }
 
+const char* GPUDevice::VSyncModeToString(GPUVSyncMode mode)
+{
+  static constexpr std::array<const char*, static_cast<size_t>(GPUVSyncMode::Count)> vsync_modes = {{
+    "Disabled",
+    "FIFO",
+    "Mailbox",
+  }};
+
+  return vsync_modes[static_cast<size_t>(mode)];
+}
+
 bool GPUDevice::IsSameRenderAPI(RenderAPI lhs, RenderAPI rhs)
 {
   return (lhs == rhs || ((lhs == RenderAPI::OpenGL || lhs == RenderAPI::OpenGLES) &&

src/util/gpu_device.h

@@ -657,6 +657,9 @@ public:
   /// Returns a string representing the specified language.
   static const char* ShaderLanguageToString(GPUShaderLanguage language);
 
+  /// Returns a string representing the specified vsync mode.
+  static const char* VSyncModeToString(GPUVSyncMode mode);
+
   /// Returns a new device for the specified API.
   static std::unique_ptr<GPUDevice> CreateDeviceForAPI(RenderAPI api);
 

src/util/imgui_fullscreen.cpp

@@ -52,6 +52,8 @@ static void DrawChoiceDialog();
 static void DrawInputDialog();
 static void DrawMessageDialog();
 static void DrawBackgroundProgressDialogs(ImVec2& position, float spacing);
+static void DrawLoadingScreen(std::string_view image, std::string_view message, s32 progress_min, s32 progress_max,
+                              s32 progress_value, bool is_persistent);
 static void DrawNotifications(ImVec2& position, float spacing);
 static void DrawToast();
 static bool MenuButtonFrame(const char* str_id, bool enabled, float height, bool* visible, bool* hovered, ImRect* bb,
@@ -171,6 +173,13 @@ struct ALIGN_TO_CACHE_LINE UIState
 
   std::vector<BackgroundProgressDialogData> background_progress_dialogs;
   std::mutex background_progress_lock;
+
+  std::string loading_screen_image;
+  std::string loading_screen_message;
+  s32 loading_screen_min = 0;
+  s32 loading_screen_max = 0;
+  s32 loading_screen_value = 0;
+  bool loading_screen_open = false;
 };
 
 } // namespace
@@ -2916,6 +2925,145 @@ void ImGuiFullscreen::DrawBackgroundProgressDialogs(ImVec2& position, float spac
   ImGui::PopStyleColor(2);
 }
 
+void ImGuiFullscreen::RenderLoadingScreen(std::string_view image, std::string_view message, s32 progress_min /*= -1*/,
+                                          s32 progress_max /*= -1*/, s32 progress_value /*= -1*/)
+{
+  if (progress_min < progress_max)
+    INFO_LOG("{}: {}/{}", message, progress_value, progress_max);
+
+  if (!g_gpu_device || !g_gpu_device->HasMainSwapChain())
+    return;
+
+  // eat the last imgui frame, it might've been partially rendered by the caller.
+  ImGui::EndFrame();
+  ImGui::NewFrame();
+
+  DrawLoadingScreen(image, message, progress_min, progress_max, progress_value, false);
+
+  ImGui::EndFrame();
+
+  GPUSwapChain* swap_chain = g_gpu_device->GetMainSwapChain();
+  if (g_gpu_device->BeginPresent(swap_chain) == GPUDevice::PresentResult::OK)
+  {
+    g_gpu_device->RenderImGui(swap_chain);
+    g_gpu_device->EndPresent(swap_chain, false);
+  }
+
+  ImGui::NewFrame();
+}
+
+void ImGuiFullscreen::OpenOrUpdateLoadingScreen(std::string_view image, std::string_view message,
+                                                s32 progress_min /*= -1*/, s32 progress_max /*= -1*/,
+                                                s32 progress_value /*= -1*/)
+{
+  if (s_state.loading_screen_image != image)
+    s_state.loading_screen_image = image;
+  if (s_state.loading_screen_message != message)
+    s_state.loading_screen_message = message;
+  s_state.loading_screen_min = progress_min;
+  s_state.loading_screen_max = progress_max;
+  s_state.loading_screen_value = progress_value;
+  s_state.loading_screen_open = true;
+}
+
+bool ImGuiFullscreen::IsLoadingScreenOpen()
+{
+  return s_state.loading_screen_open;
+}
+
+void ImGuiFullscreen::RenderLoadingScreen()
+{
+  if (!s_state.loading_screen_open)
+    return;
+
+  DrawLoadingScreen(s_state.loading_screen_image, s_state.loading_screen_message, s_state.loading_screen_min,
+                    s_state.loading_screen_max, s_state.loading_screen_value, true);
+}
+
+void ImGuiFullscreen::CloseLoadingScreen()
+{
+  s_state.loading_screen_image = {};
+  s_state.loading_screen_message = {};
+  s_state.loading_screen_min = 0;
+  s_state.loading_screen_max = 0;
+  s_state.loading_screen_value = 0;
+  s_state.loading_screen_open = false;
+}
+
+void ImGuiFullscreen::DrawLoadingScreen(std::string_view image, std::string_view message, s32 progress_min,
+                                        s32 progress_max, s32 progress_value, bool is_persistent)
+{
+  const auto& io = ImGui::GetIO();
+  const float scale = ImGuiManager::GetGlobalScale();
+  const float width = (400.0f * scale);
+  const bool has_progress = (progress_min < progress_max);
+
+  const float logo_width = 260.0f * scale;
+  const float logo_height = 260.0f * scale;
+
+  ImGui::SetNextWindowSize(ImVec2(logo_width, logo_height), ImGuiCond_Always);
+  ImGui::SetNextWindowPos(ImVec2(io.DisplaySize.x * 0.5f, (io.DisplaySize.y * 0.5f) - (50.0f * scale)),
+                          ImGuiCond_Always, ImVec2(0.5f, 0.5f));
+  if (ImGui::Begin("LoadingScreenLogo", nullptr,
+                   ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoMove |
+                     ImGuiWindowFlags_NoSavedSettings | ImGuiWindowFlags_NoScrollbar | ImGuiWindowFlags_NoNav |
+                     ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoFocusOnAppearing |
+                     ImGuiWindowFlags_NoBackground))
+  {
+    GPUTexture* tex = GetCachedTexture(image);
+    if (tex)
+      ImGui::Image(tex, ImVec2(logo_width, logo_height));
+  }
+  ImGui::End();
+
+  const float padding_and_rounding = 18.0f * scale;
+  ImGui::PushStyleVar(ImGuiStyleVar_WindowRounding, padding_and_rounding);
+  ImGui::PushStyleVar(ImGuiStyleVar_WindowPadding, ImVec2(padding_and_rounding, padding_and_rounding));
+  ImGui::SetNextWindowSize(ImVec2(width, ((has_progress || is_persistent) ? 90.0f : 55.0f) * scale), ImGuiCond_Always);
+  ImGui::SetNextWindowPos(ImVec2(io.DisplaySize.x * 0.5f, (io.DisplaySize.y * 0.5f) + (100.0f * scale)),
+                          ImGuiCond_Always, ImVec2(0.5f, 0.0f));
+  if (ImGui::Begin("LoadingScreen", nullptr,
+                   ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoMove |
+                     ImGuiWindowFlags_NoSavedSettings | ImGuiWindowFlags_NoScrollbar | ImGuiWindowFlags_NoNav |
+                     ImGuiWindowFlags_AlwaysAutoResize | ImGuiWindowFlags_NoFocusOnAppearing))
+  {
+    if (has_progress || is_persistent)
+    {
+      if (!message.empty())
+        ImGui::TextUnformatted(message.data(), message.data() + message.size());
+
+      if (has_progress)
+      {
+        TinyString buf;
+        buf.format("{}/{}", progress_value, progress_max);
+
+        const ImVec2 prog_size = ImGui::CalcTextSize(buf.c_str(), buf.end_ptr());
+        ImGui::SameLine();
+        ImGui::SetCursorPosX(width - padding_and_rounding - prog_size.x);
+        ImGui::TextUnformatted(buf.c_str(), buf.end_ptr());
+      }
+
+      ImGui::SetCursorPosY(ImGui::GetCursorPosY() + 5.0f);
+
+      ImGui::ProgressBar(has_progress ?
+                           (static_cast<float>(progress_value) / static_cast<float>(progress_max - progress_min)) :
+                           static_cast<float>(-ImGui::GetTime()),
+                         ImVec2(-1.0f, 0.0f), "");
+    }
+    else
+    {
+      if (!message.empty())
+      {
+        const ImVec2 text_size(ImGui::CalcTextSize(message.data(), message.data() + message.size()));
+        ImGui::SetCursorPosX((width - text_size.x) / 2.0f);
+        ImGui::TextUnformatted(message.data(), message.data() + message.size());
+      }
+    }
+  }
+  ImGui::End();
+  ImGui::PopStyleVar(2);
+}
+
 //////////////////////////////////////////////////////////////////////////
 // Notifications
 //////////////////////////////////////////////////////////////////////////

src/util/imgui_fullscreen.h

@@ -327,6 +327,18 @@ void UpdateBackgroundProgressDialog(const char* str_id, std::string message, s32
 void CloseBackgroundProgressDialog(const char* str_id);
 bool IsBackgroundProgressDialogOpen(const char* str_id);
 
+/// Displays a loading screen with the logo, rendered with ImGui. Use when executing possibly-time-consuming tasks
+/// such as compiling shaders when starting up.
+void RenderLoadingScreen(std::string_view image, std::string_view message, s32 progress_min = -1, s32 progress_max = -1,
+                         s32 progress_value = -1);
+void OpenOrUpdateLoadingScreen(std::string_view image, std::string_view message, s32 progress_min = -1,
+                               s32 progress_max = -1, s32 progress_value = -1);
+bool IsLoadingScreenOpen();
+void CloseLoadingScreen();
+
+/// Renders a previously-configured loading screen.
+void RenderLoadingScreen();
+
 void AddNotification(std::string key, float duration, std::string title, std::string text, std::string image_path);
 void ClearNotifications();
 

src/util/page_fault_handler.cpp

@@ -30,12 +30,6 @@
 #include <mach/task.h>
 #endif
 
-namespace PageFaultHandler {
-static std::recursive_mutex s_exception_handler_mutex;
-static bool s_in_exception_handler = false;
-static bool s_installed = false;
-} // namespace PageFaultHandler
-
 #if defined(CPU_ARCH_ARM64)
 [[maybe_unused]] static bool IsStoreInstruction(const void* ptr)
 {
@@ -85,7 +79,11 @@ static bool s_installed = false;
 
 namespace PageFaultHandler {
 static LONG ExceptionHandler(PEXCEPTION_POINTERS exi);
-}
+
+static std::recursive_mutex s_exception_handler_mutex;
+static bool s_in_exception_handler = false;
+static bool s_installed = false;
+} // namespace PageFaultHandler
 
 LONG PageFaultHandler::ExceptionHandler(PEXCEPTION_POINTERS exi)
 {
@@ -140,6 +138,10 @@ bool PageFaultHandler::Install(Error* error)
 
 namespace PageFaultHandler {
 static void SignalHandler(int sig, siginfo_t* info, void* ctx);
+
+static std::recursive_mutex s_exception_handler_mutex;
+static bool s_in_exception_handler = false;
+static bool s_installed = false;
 } // namespace PageFaultHandler
 
 void PageFaultHandler::SignalHandler(int sig, siginfo_t* info, void* ctx)
@@ -210,7 +212,7 @@ void PageFaultHandler::SignalHandler(int sig, siginfo_t* info, void* ctx)
     result = HandlePageFault(exception_pc, exception_address, is_write);
     s_in_exception_handler = false;
   }
-  
+
   s_exception_handler_mutex.unlock();
 
   // Resumes execution right where we left off (re-executes instruction that caused the SIGSEGV).

src/util/postprocessing.cpp

@@ -12,7 +12,7 @@
 
 // TODO: Remove me
 #include "core/host.h"
-#include "core/host_interface_progress_callback.h"
+#include "core/fullscreen_ui.h"
 #include "core/settings.h"
 
 #include "IconsFontAwesome5.h"
@@ -406,7 +406,7 @@ void PostProcessing::Chain::LoadStages()
     return;
 
   Error error;
-  HostInterfaceProgressCallback progress;
+  LoadingScreenProgressCallback progress;
   progress.SetProgressRange(stage_count);
 
   for (u32 i = 0; i < stage_count; i++)
@@ -476,7 +476,7 @@ void PostProcessing::Chain::UpdateSettings(std::unique_lock<std::mutex>& setting
 
   m_stages.resize(stage_count);
 
-  HostInterfaceProgressCallback progress;
+  LoadingScreenProgressCallback progress;
   progress.SetProgressRange(stage_count);
 
   const GPUTexture::Format prev_format = m_target_format;

src/util/postprocessing_shader_fx.cpp

@@ -7,6 +7,7 @@
 #include "shadergen.h"
 
 // TODO: Remove me
+#include "core/gpu_thread.h"
 #include "core/host.h"
 #include "core/settings.h"
 
@@ -72,7 +73,7 @@ static std::tuple<std::unique_ptr<reshadefx::codegen>, GPUShaderLanguage> Create
   }
 
   // Should have a GPU device and be on the GPU thread.
-  Assert(g_gpu_device);
+  Assert(GPUThread::IsOnThread() && g_gpu_device);
 
   const bool debug_info = g_gpu_device->IsDebugDevice();
   const RenderAPI rapi = g_gpu_device->GetRenderAPI();

src/util/state_wrapper.h

@@ -34,6 +34,8 @@ public:
   ALWAYS_INLINE bool IsReading() const { return (m_mode == Mode::Read); }
   ALWAYS_INLINE bool IsWriting() const { return (m_mode == Mode::Write); }
   ALWAYS_INLINE u32 GetVersion() const { return m_version; }
+  ALWAYS_INLINE const u8* GetData() const { return m_data; }
+  ALWAYS_INLINE size_t GetDataSize() const { return m_size; }
   ALWAYS_INLINE size_t GetPosition() const { return m_pos; }
   ALWAYS_INLINE void SetPosition(size_t pos) { m_pos = pos; }