diff --git a/CHANGES b/CHANGES
index 9338b569f..98e6c4c4f 100644
--- a/CHANGES
+++ b/CHANGES
@@ -88,6 +88,8 @@ Other fixes:
  - SM83: Simplify register pair access on big endian
  - Wii: Fix pixelated filtering on interframe blending (fixes mgba.io/i/1830)
 Misc:
+ - GB: Allow pausing event loop while CPU is blocked
+ - GBA: Allow pausing event loop while CPU is blocked
  - Debugger: Keep track of global cycle count
  - FFmpeg: Add looping option for GIF/APNG
  - FFmpeg: Use range coder for FFV1 to reduce output size
diff --git a/CMakeLists.txt b/CMakeLists.txt
index ae2d1349c..b6fb9be82 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -13,7 +13,7 @@ set(CMAKE_C_STANDARD 99)
 if(NOT MSVC)
 	set(CMAKE_C_STANDARD_REQUIRED ON)
 	set(CMAKE_C_EXTENSIONS OFF)
-	if(SWITCH)
+	if(SWITCH OR 3DS)
 		set(CMAKE_C_STANDARD 11)
 		set(CMAKE_C_EXTENSIONS ON)
 	elseif(CMAKE_C_COMPILER_ID STREQUAL "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS "4.3")
@@ -297,6 +297,7 @@ include(CheckIncludeFiles)
 check_function_exists(strdup HAVE_STRDUP)
 check_function_exists(strndup HAVE_STRNDUP)
 check_function_exists(strlcpy HAVE_STRLCPY)
+check_function_exists(vasprintf HAVE_VASPRINTF)
 if(NOT DEFINED PSP2)
 	check_function_exists(localtime_r HAVE_LOCALTIME_R)
 endif()
@@ -379,6 +380,10 @@ if(HAVE_STRLCPY)
 	list(APPEND FUNCTION_DEFINES HAVE_STRLCPY)
 endif()
 
+if(HAVE_VASPRINTF)
+	list(APPEND FUNCTION_DEFINES HAVE_VASPRINTF)
+endif()
+
 if(HAVE_LOCALTIME_R)
 	list(APPEND FUNCTION_DEFINES HAVE_LOCALTIME_R)
 endif()
@@ -988,6 +993,9 @@ endif()
 if(NOT USE_CMOCKA)
 	set(BUILD_SUITE OFF)
 endif()
+if(BUILD_TEST OR BUILD_SUITE OR BUILD_CINEMA)
+	enable_testing()
+endif()
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/src/platform/test ${CMAKE_CURRENT_BINARY_DIR}/test)
 
 if(BUILD_EXAMPLE)
diff --git a/cinema/gb/mooneye-gb/acceptance/boot_hwio-dmg0/xbaseline_0000.png b/cinema/gb/mooneye-gb/acceptance/boot_hwio-dmg0/xbaseline_0000.png
new file mode 100644
index 000000000..691abb25d
Binary files /dev/null and b/cinema/gb/mooneye-gb/acceptance/boot_hwio-dmg0/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/acceptance/boot_regs-dmg0/xbaseline_0000.png b/cinema/gb/mooneye-gb/acceptance/boot_regs-dmg0/xbaseline_0000.png
new file mode 100644
index 000000000..7051512ec
Binary files /dev/null and b/cinema/gb/mooneye-gb/acceptance/boot_regs-dmg0/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/acceptance/oam_dma_start/xbaseline_0000.png b/cinema/gb/mooneye-gb/acceptance/oam_dma_start/xbaseline_0000.png
new file mode 100644
index 000000000..70c31f4a6
Binary files /dev/null and b/cinema/gb/mooneye-gb/acceptance/oam_dma_start/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/acceptance/push_timing/xbaseline_0000.png b/cinema/gb/mooneye-gb/acceptance/push_timing/xbaseline_0000.png
new file mode 100644
index 000000000..9fb710286
Binary files /dev/null and b/cinema/gb/mooneye-gb/acceptance/push_timing/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/acceptance/timer/rapid_toggle/xbaseline_0000.png b/cinema/gb/mooneye-gb/acceptance/timer/rapid_toggle/xbaseline_0000.png
new file mode 100644
index 000000000..bbe7e5937
Binary files /dev/null and b/cinema/gb/mooneye-gb/acceptance/timer/rapid_toggle/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/manual-only/sprite_priority/xbaseline_0000.png b/cinema/gb/mooneye-gb/manual-only/sprite_priority/xbaseline_0000.png
new file mode 100644
index 000000000..2201a6ac5
Binary files /dev/null and b/cinema/gb/mooneye-gb/manual-only/sprite_priority/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/misc/bits/unused_hwio-C/xbaseline_0000.png b/cinema/gb/mooneye-gb/misc/bits/unused_hwio-C/xbaseline_0000.png
new file mode 100644
index 000000000..f9b20ce12
Binary files /dev/null and b/cinema/gb/mooneye-gb/misc/bits/unused_hwio-C/xbaseline_0000.png differ
diff --git a/cinema/gb/mooneye-gb/misc/boot_hwio-C/xbaseline_0000.png b/cinema/gb/mooneye-gb/misc/boot_hwio-C/xbaseline_0000.png
new file mode 100644
index 000000000..cb37cb619
Binary files /dev/null and b/cinema/gb/mooneye-gb/misc/boot_hwio-C/xbaseline_0000.png differ
diff --git a/include/mgba-util/platform/posix/threading.h b/include/mgba-util/platform/posix/threading.h
index 64340c852..32d2ea4b1 100644
--- a/include/mgba-util/platform/posix/threading.h
+++ b/include/mgba-util/platform/posix/threading.h
@@ -24,6 +24,7 @@ typedef THREAD_ENTRY (*ThreadEntry)(void*);
 typedef pthread_t Thread;
 typedef pthread_mutex_t Mutex;
 typedef pthread_cond_t Condition;
+typedef pthread_key_t ThreadLocal;
 
 static inline int MutexInit(Mutex* mutex) {
 	return pthread_mutex_init(mutex, 0);
@@ -101,6 +102,18 @@ static inline int ThreadSetName(const char* name) {
 #endif
 }
 
+static inline void ThreadLocalInitKey(ThreadLocal* key) {
+	pthread_key_create(key, 0);
+}
+
+static inline void ThreadLocalSetKey(ThreadLocal key, void* value) {
+	pthread_setspecific(key, value);
+}
+
+static inline void* ThreadLocalGetValue(ThreadLocal key) {
+	return pthread_getspecific(key);
+}
+
 CXX_GUARD_END
 
 #endif
diff --git a/include/mgba-util/platform/psp2/threading.h b/include/mgba-util/platform/psp2/threading.h
index 36364ac95..96e21e851 100644
--- a/include/mgba-util/platform/psp2/threading.h
+++ b/include/mgba-util/platform/psp2/threading.h
@@ -17,6 +17,7 @@ typedef struct {
 } Condition;
 #define THREAD_ENTRY int
 typedef THREAD_ENTRY (*ThreadEntry)(void*);
+typedef int ThreadLocal;
 
 static inline int MutexInit(Mutex* mutex) {
 	Mutex id = sceKernelCreateMutex("mutex", 0, 0, 0);
@@ -143,4 +144,19 @@ static inline int ThreadSetName(const char* name) {
 	UNUSED(name);
 	return -1;
 }
+
+static inline void ThreadLocalInitKey(ThreadLocal* key) {
+	static int base = 0x90;
+	*key = __atomic_fetch_add(&base, 1, __ATOMIC_SEQ_CST);
+}
+
+static inline void ThreadLocalSetKey(ThreadLocal key, void* value) {
+	void** tls = sceKernelGetTLSAddr(key);
+	*tls = value;
+}
+
+static inline void* ThreadLocalGetValue(ThreadLocal key) {
+	void** tls = sceKernelGetTLSAddr(key);
+	return *tls;
+}
 #endif
diff --git a/include/mgba-util/platform/windows/threading.h b/include/mgba-util/platform/windows/threading.h
index 8ea73d3c1..d7a7b5532 100644
--- a/include/mgba-util/platform/windows/threading.h
+++ b/include/mgba-util/platform/windows/threading.h
@@ -16,6 +16,7 @@ typedef THREAD_ENTRY ThreadEntry(LPVOID);
 typedef HANDLE Thread;
 typedef CRITICAL_SECTION Mutex;
 typedef CONDITION_VARIABLE Condition;
+typedef DWORD ThreadLocal;
 
 static inline int MutexInit(Mutex* mutex) {
 	InitializeCriticalSection(mutex);
@@ -88,4 +89,16 @@ static inline int ThreadSetName(const char* name) {
 	return -1;
 }
 
+static inline void ThreadLocalInitKey(ThreadLocal* key) {
+	*key = TlsAlloc();
+}
+
+static inline void ThreadLocalSetKey(ThreadLocal key, void* value) {
+	TlsSetValue(key, value);
+}
+
+static inline void* ThreadLocalGetValue(ThreadLocal key) {
+	return TlsGetValue(key);
+}
+
 #endif
diff --git a/include/mgba-util/string.h b/include/mgba-util/string.h
index 306a35257..381869814 100644
--- a/include/mgba-util/string.h
+++ b/include/mgba-util/string.h
@@ -45,6 +45,7 @@ const char* hex4(const char* line, uint8_t* out);
 void rtrim(char* string);
 
 ssize_t parseQuotedString(const char* unparsed, ssize_t unparsedLen, char* parsed, ssize_t parsedLen);
+bool wildcard(const char* search, const char* string);
 
 CXX_GUARD_END
 
diff --git a/include/mgba-util/threading.h b/include/mgba-util/threading.h
index 3a2a4102f..779a533a0 100644
--- a/include/mgba-util/threading.h
+++ b/include/mgba-util/threading.h
@@ -11,6 +11,12 @@
 CXX_GUARD_START
 
 #ifndef DISABLE_THREADING
+#if __STDC_VERSION__ >= 201112L
+#define ThreadLocal _Thread_local void*
+#define ThreadLocalInitKey(X)
+#define ThreadLocalSetKey(K, V) K = V
+#define ThreadLocalGetValue(K) K
+#endif
 #ifdef USE_PTHREADS
 #include <mgba-util/platform/posix/threading.h>
 #elif defined(_WIN32)
@@ -40,6 +46,7 @@ typedef void* Thread;
 typedef void* Mutex;
 #endif
 typedef void* Condition;
+typedef int ThreadLocal;
 
 static inline int MutexInit(Mutex* mutex) {
 	UNUSED(mutex);
@@ -93,6 +100,20 @@ static inline int ConditionWake(Condition* cond) {
 	UNUSED(cond);
 	return 0;
 }
+
+static inline void ThreadLocalInitKey(ThreadLocal* key) {
+	UNUSED(key);
+}
+
+static inline void ThreadLocalSetKey(ThreadLocal key, void* value) {
+	UNUSED(key);
+	UNUSED(value);
+}
+
+static inline void* ThreadLocalGetValue(ThreadLocal key) {
+	UNUSED(key);
+	return NULL;
+}
 #endif
 
 CXX_GUARD_END
diff --git a/include/mgba/internal/gb/serialize.h b/include/mgba/internal/gb/serialize.h
index be5d17e2f..7d3a73be5 100644
--- a/include/mgba/internal/gb/serialize.h
+++ b/include/mgba/internal/gb/serialize.h
@@ -235,6 +235,8 @@ DECL_BIT(GBSerializedCpuFlags, Condition, 0);
 DECL_BIT(GBSerializedCpuFlags, IrqPending, 1);
 DECL_BIT(GBSerializedCpuFlags, DoubleSpeed, 2);
 DECL_BIT(GBSerializedCpuFlags, EiPending, 3);
+DECL_BIT(GBSerializedCpuFlags, Halted, 4);
+DECL_BIT(GBSerializedCpuFlags, Blocked, 5);
 
 DECL_BITFIELD(GBSerializedTimerFlags, uint8_t);
 DECL_BIT(GBSerializedTimerFlags, IrqPending, 0);
diff --git a/include/mgba/internal/gba/serialize.h b/include/mgba/internal/gba/serialize.h
index a5dabae2c..a8ea138e4 100644
--- a/include/mgba/internal/gba/serialize.h
+++ b/include/mgba/internal/gba/serialize.h
@@ -236,6 +236,7 @@ DECL_BITFIELD(GBASerializedMiscFlags, uint32_t);
 DECL_BIT(GBASerializedMiscFlags, Halted, 0);
 DECL_BIT(GBASerializedMiscFlags, POSTFLG, 1);
 DECL_BIT(GBASerializedMiscFlags, IrqPending, 2);
+DECL_BIT(GBASerializedMiscFlags, Blocked, 3);
 
 struct GBASerializedState {
 	uint32_t versionMagic;
diff --git a/src/core/thread.c b/src/core/thread.c
index 887b7fa3f..218dc92cf 100644
--- a/src/core/thread.c
+++ b/src/core/thread.c
@@ -16,23 +16,22 @@
 #ifndef DISABLE_THREADING
 
 static const float _defaultFPSTarget = 60.f;
+static ThreadLocal _contextKey;
 
 #ifdef USE_PTHREADS
-static pthread_key_t _contextKey;
 static pthread_once_t _contextOnce = PTHREAD_ONCE_INIT;
 
 static void _createTLS(void) {
-	pthread_key_create(&_contextKey, 0);
+	ThreadLocalInitKey(&_contextKey);
 }
 #elif _WIN32
-static DWORD _contextKey;
 static INIT_ONCE _contextOnce = INIT_ONCE_STATIC_INIT;
 
 static BOOL CALLBACK _createTLS(PINIT_ONCE once, PVOID param, PVOID* context) {
 	UNUSED(once);
 	UNUSED(param);
 	UNUSED(context);
-	_contextKey = TlsAlloc();
+	ThreadLocalInitKey(&_contextKey);
 	return TRUE;
 }
 #endif
@@ -144,12 +143,11 @@ static THREAD_ENTRY _mCoreThreadRun(void* context) {
 	struct mCoreThread* threadContext = context;
 #ifdef USE_PTHREADS
 	pthread_once(&_contextOnce, _createTLS);
-	pthread_setspecific(_contextKey, threadContext);
 #elif _WIN32
 	InitOnceExecuteOnce(&_contextOnce, _createTLS, NULL, 0);
-	TlsSetValue(_contextKey, threadContext);
 #endif
 
+	ThreadLocalSetKey(_contextKey, threadContext);
 	ThreadSetName("CPU Thread");
 
 #if !defined(_WIN32) && defined(USE_PTHREADS)
@@ -620,21 +618,14 @@ void mCoreThreadStopWaiting(struct mCoreThread* threadContext) {
 	MutexUnlock(&threadContext->impl->stateMutex);
 }
 
+struct mCoreThread* mCoreThreadGet(void) {
 #ifdef USE_PTHREADS
-struct mCoreThread* mCoreThreadGet(void) {
 	pthread_once(&_contextOnce, _createTLS);
-	return pthread_getspecific(_contextKey);
-}
 #elif _WIN32
-struct mCoreThread* mCoreThreadGet(void) {
 	InitOnceExecuteOnce(&_contextOnce, _createTLS, NULL, 0);
-	return TlsGetValue(_contextKey);
-}
-#else
-struct mCoreThread* mCoreThreadGet(void) {
-	return NULL;
-}
 #endif
+	return ThreadLocalGetValue(_contextKey);
+}
 
 static void _mCoreLog(struct mLogger* logger, int category, enum mLogLevel level, const char* format, va_list args) {
 	UNUSED(logger);
diff --git a/src/gb/gb.c b/src/gb/gb.c
index 3f9dda81a..7aea9eaad 100644
--- a/src/gb/gb.c
+++ b/src/gb/gb.c
@@ -681,7 +681,7 @@ void GBProcessEvents(struct SM83Core* cpu) {
 		nextEvent = cycles;
 		do {
 			nextEvent = mTimingTick(&gb->timing, nextEvent);
-		} while (gb->cpuBlocked);
+		} while (gb->cpuBlocked && !gb->earlyExit);
 		cpu->nextEvent = nextEvent;
 
 		if (cpu->halted) {
@@ -695,6 +695,9 @@ void GBProcessEvents(struct SM83Core* cpu) {
 		}
 	} while (cpu->cycles >= cpu->nextEvent);
 	gb->earlyExit = false;
+	if (gb->cpuBlocked) {
+		cpu->cycles = cpu->nextEvent;
+	}
 }
 
 void GBSetInterrupts(struct SM83Core* cpu, bool enable) {
diff --git a/src/gb/serialize.c b/src/gb/serialize.c
index 539f87c2a..3ae5cbf10 100644
--- a/src/gb/serialize.c
+++ b/src/gb/serialize.c
@@ -56,6 +56,8 @@ void GBSerialize(struct GB* gb, struct GBSerializedState* state) {
 	flags = GBSerializedCpuFlagsSetIrqPending(flags, gb->cpu->irqPending);
 	flags = GBSerializedCpuFlagsSetDoubleSpeed(flags, gb->doubleSpeed);
 	flags = GBSerializedCpuFlagsSetEiPending(flags, mTimingIsScheduled(&gb->timing, &gb->eiPending));
+	flags = GBSerializedCpuFlagsSetHalted(flags, gb->cpu->halted);
+	flags = GBSerializedCpuFlagsSetBlocked(flags, gb->cpuBlocked);
 	STORE_32LE(flags, 0, &state->cpu.flags);
 	STORE_32LE(gb->eiPending.when - mTimingCurrentTime(&gb->timing), 0, &state->cpu.eiPending);
 
@@ -173,6 +175,9 @@ bool GBDeserialize(struct GB* gb, const struct GBSerializedState* state) {
 	gb->cpu->condition = GBSerializedCpuFlagsGetCondition(flags);
 	gb->cpu->irqPending = GBSerializedCpuFlagsGetIrqPending(flags);
 	gb->doubleSpeed = GBSerializedCpuFlagsGetDoubleSpeed(flags);
+	gb->cpu->halted = GBSerializedCpuFlagsGetHalted(flags);
+	gb->cpuBlocked = GBSerializedCpuFlagsGetBlocked(flags);
+
 	gb->audio.timingFactor = gb->doubleSpeed + 1;
 
 	LOAD_32LE(gb->cpu->cycles, 0, &state->cpu.cycles);
diff --git a/src/gb/video.c b/src/gb/video.c
index fa5404c5a..85522ddf8 100644
--- a/src/gb/video.c
+++ b/src/gb/video.c
@@ -379,6 +379,7 @@ void _updateFrameCount(struct mTiming* timing, void* context, uint32_t cyclesLat
 	GBFrameEnded(video->p);
 	mCoreSyncPostFrame(video->p->sync);
 	++video->frameCounter;
+	video->p->earlyExit = true;
 
 	GBFrameStarted(video->p);
 }
diff --git a/src/gba/gba.c b/src/gba/gba.c
index 84ad8007e..05203cfc0 100644
--- a/src/gba/gba.c
+++ b/src/gba/gba.c
@@ -286,7 +286,7 @@ static void GBAProcessEvents(struct ARMCore* cpu) {
 			}
 #endif
 			nextEvent = mTimingTick(&gba->timing, cycles < nextEvent ? nextEvent : cycles);
-		} while (gba->cpuBlocked);
+		} while (gba->cpuBlocked && !gba->earlyExit);
 
 		cpu->nextEvent = nextEvent;
 		if (cpu->halted) {
@@ -305,11 +305,9 @@ static void GBAProcessEvents(struct ARMCore* cpu) {
 		}
 	}
 	gba->earlyExit = false;
-#ifndef NDEBUG
 	if (gba->cpuBlocked) {
-		mLOG(GBA, FATAL, "CPU is blocked!");
+		cpu->cycles = cpu->nextEvent;
 	}
-#endif
 }
 
 #ifdef USE_DEBUGGERS
diff --git a/src/gba/serialize.c b/src/gba/serialize.c
index 00a80c6b7..e8f1b215a 100644
--- a/src/gba/serialize.c
+++ b/src/gba/serialize.c
@@ -67,6 +67,7 @@ void GBASerialize(struct GBA* gba, struct GBASerializedState* state) {
 		miscFlags = GBASerializedMiscFlagsFillIrqPending(miscFlags);
 		STORE_32(gba->irqEvent.when - mTimingCurrentTime(&gba->timing), 0, &state->nextIrq);
 	}
+	miscFlags = GBASerializedMiscFlagsSetBlocked(miscFlags, gba->cpuBlocked);
 	STORE_32(miscFlags, 0, &state->miscFlags);
 
 	GBAMemorySerialize(&gba->memory, state);
@@ -185,6 +186,7 @@ bool GBADeserialize(struct GBA* gba, const struct GBASerializedState* state) {
 		LOAD_32(when, 0, &state->nextIrq);
 		mTimingSchedule(&gba->timing, &gba->irqEvent, when);		
 	}
+	gba->cpuBlocked = GBASerializedMiscFlagsGetBlocked(miscFlags);
 
 	GBAVideoDeserialize(&gba->video, state);
 	GBAMemoryDeserialize(&gba->memory, state);
diff --git a/src/gba/video.c b/src/gba/video.c
index 74377d13f..f600e99bf 100644
--- a/src/gba/video.c
+++ b/src/gba/video.c
@@ -196,6 +196,7 @@ void _startHdraw(struct mTiming* timing, void* context, uint32_t cyclesLate) {
 			video->frameskipCounter = video->frameskip;
 		}
 		++video->frameCounter;
+		video->p->earlyExit = true;
 		break;
 	case VIDEO_VERTICAL_TOTAL_PIXELS - 1:
 		video->p->memory.io[REG_DISPSTAT >> 1] = GBARegisterDISPSTATClearInVblank(dispstat);
diff --git a/src/platform/python/CMakeLists.txt b/src/platform/python/CMakeLists.txt
index 742ec69ca..7004519ae 100644
--- a/src/platform/python/CMakeLists.txt
+++ b/src/platform/python/CMakeLists.txt
@@ -54,9 +54,8 @@ add_custom_target(${BINARY_NAME}-py-bdist
                   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
                   DEPENDS ${BINARY_NAME}-py)
 
-file(GLOB BASE_TESTS ${CMAKE_CURRENT_SOURCE_DIR}/test_*.py)
-file(GLOB SUBTESTS ${CMAKE_CURRENT_SOURCE_DIR}/tests/*/test_*.py)
-foreach(TEST IN LISTS BASE_TESTS SUBTESTS)
+file(GLOB TESTS ${CMAKE_CURRENT_SOURCE_DIR}/tests/*/test_*.py)
+foreach(TEST IN LISTS TESTS)
 	if(APPLE)
 		set(PATH DYLD_LIBRARY_PATH)
 	elseif(WIN32)
@@ -64,7 +63,7 @@ foreach(TEST IN LISTS BASE_TESTS SUBTESTS)
 	else()
 		set(PATH LD_LIBRARY_PATH)
 	endif()
-	string(REGEX REPLACE "${CMAKE_CURRENT_SOURCE_DIR}/(tests/.*/)?test_" "" TEST_NAME "${TEST}")
+	string(REGEX REPLACE "${CMAKE_CURRENT_SOURCE_DIR}/tests/(.*/)?test_" "" TEST_NAME "${TEST}")
 	string(REPLACE ".py" "" TEST_NAME "${TEST_NAME}")
 	add_test(NAME python-${TEST_NAME}
                COMMAND ${PYTHON_EXECUTABLE} setup.py build -b ${CMAKE_CURRENT_BINARY_DIR} pytest --extras --addopts ${TEST}
diff --git a/src/platform/test/CMakeLists.txt b/src/platform/test/CMakeLists.txt
index 8882b8d94..c81b339f8 100644
--- a/src/platform/test/CMakeLists.txt
+++ b/src/platform/test/CMakeLists.txt
@@ -43,4 +43,5 @@ if(BUILD_CINEMA)
 	add_executable(${BINARY_NAME}-cinema ${CMAKE_CURRENT_SOURCE_DIR}/cinema-main.c)
 	target_link_libraries(${BINARY_NAME}-cinema ${BINARY_NAME} ${PLATFORM_LIBRARY})
 	set_target_properties(${BINARY_NAME}-cinema PROPERTIES COMPILE_DEFINITIONS "${OS_DEFINES};${FEATURE_DEFINES};${FUNCTION_DEFINES}")
+	add_test(cinema ${BINARY_NAME}-cinema -v)
 endif()
diff --git a/src/platform/test/cinema-main.c b/src/platform/test/cinema-main.c
index 1bc317755..4c18d40ae 100644
--- a/src/platform/test/cinema-main.c
+++ b/src/platform/test/cinema-main.c
@@ -13,6 +13,7 @@
 #include <mgba-util/png-io.h>
 #include <mgba-util/string.h>
 #include <mgba-util/table.h>
+#include <mgba-util/threading.h>
 #include <mgba-util/vector.h>
 #include <mgba-util/vfs.h>
 
@@ -32,21 +33,26 @@
 #include <sys/types.h>
 
 #define MAX_TEST 200
+#define MAX_JOBS 128
+#define LOG_THRESHOLD 1000000
 
 static const struct option longOpts[] = {
 	{ "base",       required_argument, 0, 'b' },
 	{ "diffs",      no_argument, 0, 'd' },
 	{ "help",       no_argument, 0, 'h' },
+	{ "jobs",       required_argument, 0, 'j' },
 	{ "dry-run",    no_argument, 0, 'n' },
 	{ "outdir",     required_argument, 0, 'o' },
 	{ "quiet",      no_argument, 0, 'q' },
 	{ "rebaseline", no_argument, 0, 'r' },
+	{ "rebaseline-missing", no_argument, 0, 'R' },
 	{ "verbose",    no_argument, 0, 'v' },
+	{ "xbaseline",  no_argument, 0, 'x' },
 	{ "version",    no_argument, 0, '\0' },
 	{ 0, 0, 0, 0 }
 };
 
-static const char shortOpts[] = "b:dhno:qrv";
+static const char shortOpts[] = "b:dhj:no:qRrvx";
 
 enum CInemaStatus {
 	CI_PASS,
@@ -57,6 +63,12 @@ enum CInemaStatus {
 	CI_SKIP
 };
 
+enum CInemaRebaseline {
+	CI_R_NONE = 0,
+	CI_R_FAILING,
+	CI_R_MISSING,
+};
+
 struct CInemaTest {
 	char directory[MAX_TEST];
 	char filename[MAX_TEST];
@@ -82,30 +94,111 @@ DEFINE_VECTOR(CInemaTestList, struct CInemaTest)
 DECLARE_VECTOR(ImageList, void*)
 DEFINE_VECTOR(ImageList, void*)
 
+struct StringBuilder {
+	struct StringList lines;
+	struct StringList partial;
+	unsigned repeat;
+
+};
+
+struct CInemaLogStream {
+	struct StringBuilder err;
+	struct StringBuilder out;
+};
+
 static bool showVersion = false;
 static bool showUsage = false;
 static char base[PATH_MAX] = {0};
 static char outdir[PATH_MAX] = {'.'};
 static bool dryRun = false;
 static bool diffs = false;
-static bool rebaseline = false;
+static enum CInemaRebaseline rebaseline = CI_R_NONE;
+static enum CInemaRebaseline xbaseline = CI_R_NONE;
 static int verbosity = 0;
 
+static struct Table configTree;
+static Mutex configMutex;
+
+static int jobs = 1;
+static size_t jobIndex = 0;
+static Mutex jobMutex;
+static Thread jobThreads[MAX_JOBS];
+static int jobStatus;
+static ThreadLocal logStream;
+static ThreadLocal currentTest;
+
 bool CInemaTestInit(struct CInemaTest*, const char* directory, const char* filename);
-void CInemaTestRun(struct CInemaTest*, struct Table* configTree);
+void CInemaTestRun(struct CInemaTest*);
 
 bool CInemaConfigGetUInt(struct Table* configTree, const char* testName, const char* key, unsigned* value);
 void CInemaConfigLoad(struct Table* configTree, const char* testName, struct mCore* core);
 
 static void _log(struct mLogger* log, int category, enum mLogLevel level, const char* format, va_list args);
 
+void CIflush(struct StringBuilder* list, FILE* file);
+
+static char* _compileStringList(struct StringList* list) {
+	size_t len = 0;
+	size_t i;
+	for (i = 0; i < StringListSize(list); ++i) {
+		len += strlen(*StringListGetPointer(list, i));
+	}
+	char* string = calloc(len + 1, sizeof(char));
+	char* cur = string;
+	for (i = 0; i < StringListSize(list); ++i) {
+		char* brick = *StringListGetPointer(list, i);
+		size_t portion = strlen(brick);
+		memcpy(cur, brick, portion);
+		free(brick);
+		cur += portion;
+	}
+	StringListClear(list);
+	return string;
+}
+
+static void _logToStream(FILE* file, const char* format, va_list args) {
+#ifdef HAVE_VASPRINTF
+	struct CInemaLogStream* stream = ThreadLocalGetValue(logStream);
+	if (!stream) {
+		vfprintf(file, format, args);
+	} else {
+		struct StringBuilder* builder = &stream->out;
+		if (file == stderr) {
+			builder = &stream->err;
+		}
+		if (StringListSize(&builder->lines) > LOG_THRESHOLD) {
+			CIflush(builder, file);
+		}
+		char** line = StringListAppend(&builder->partial);
+		vasprintf(line, format, args);
+		size_t len = strlen(*line);
+		if (len && (*line)[len - 1] == '\n') {
+			char* string = _compileStringList(&builder->partial);
+			size_t linecount = StringListSize(&builder->lines);
+			if (linecount && strcmp(string, *StringListGetPointer(&builder->lines, linecount - 1)) == 0) {
+				++builder->repeat;
+				free(string);
+			} else {
+				if (builder->repeat > 1) {
+					asprintf(StringListAppend(&builder->lines), "The previous message was repeated %u times.\n", builder->repeat);
+				}
+				*StringListAppend(&builder->lines) = string;
+				builder->repeat = 1;
+			}
+		}
+	}
+#else
+	vfprintf(file, format, args);
+#endif
+}
+
 ATTRIBUTE_FORMAT(printf, 2, 3) void CIlog(int minlevel, const char* format, ...) {
 	if (verbosity < minlevel) {
 		return;
 	}
 	va_list args;
 	va_start(args, format);
-	vprintf(format, args);
+	_logToStream(stdout, format, args);
 	va_end(args);
 }
 
@@ -115,10 +208,26 @@ ATTRIBUTE_FORMAT(printf, 2, 3) void CIerr(int minlevel, const char* format, ...)
 	}
 	va_list args;
 	va_start(args, format);
-	vfprintf(stderr, format, args);
+	_logToStream(stderr, format, args);
 	va_end(args);
 }
 
+void CIflush(struct StringBuilder* builder, FILE* out) {
+	if (StringListSize(&builder->partial)) {
+		*StringListAppend(&builder->lines) = _compileStringList(&builder->partial);
+	}
+#ifdef HAVE_VASPRINTF
+	if (builder->repeat > 1) {
+		asprintf(StringListAppend(&builder->lines), "The previous message was repeated %u times.\n", builder->repeat);
+	}
+#endif
+
+	char* string = _compileStringList(&builder->lines);
+	builder->repeat = 0;
+	fputs(string, out);
+	free(string);
+}
+
 static bool parseCInemaArgs(int argc, char* const* argv) {
 	int ch;
 	int index = 0;
@@ -142,6 +251,15 @@ static bool parseCInemaArgs(int argc, char* const* argv) {
 		case 'h':
 			showUsage = true;
 			break;
+		case 'j':
+			jobs = atoi(optarg);
+			if (jobs > MAX_JOBS) {
+				jobs = MAX_JOBS;
+			}
+			if (jobs < 1) {
+				jobs = 1;
+			}
+			break;
 		case 'n':
 			dryRun = true;
 			break;
@@ -153,11 +271,17 @@ static bool parseCInemaArgs(int argc, char* const* argv) {
 			--verbosity;
 			break;
 		case 'r':
-			rebaseline = true;
+			rebaseline = CI_R_FAILING;
+			break;
+		case 'R':
+			rebaseline = CI_R_MISSING;
 			break;
 		case 'v':
 			++verbosity;
 			break;
+		case 'x':
+			xbaseline = CI_R_FAILING;
+			break;
 		default:
 			return false;
 		}
@@ -167,15 +291,18 @@ static bool parseCInemaArgs(int argc, char* const* argv) {
 }
 
 static void usageCInema(const char* arg0) {
-	printf("usage: %s [-dhnqv] [-b BASE] [-o DIR] [--version] [test...]\n", arg0);
-	puts("  -b, --base [BASE]          Path to the CInema base directory");
+	printf("usage: %s [-dhnqrRv] [-j JOBS] [-b BASE] [-o DIR] [--version] [test...]\n", arg0);
+	puts("  -b, --base BASE            Path to the CInema base directory");
 	puts("  -d, --diffs                Output image diffs from failures");
 	puts("  -h, --help                 Print this usage and exit");
+	puts("  -j, --jobs JOBS            Run a number of jobs in parallel");
 	puts("  -n, --dry-run              List all collected tests instead of running them");
-	puts("  -o, --output [DIR]         Path to output applicable results");
+	puts("  -o, --output DIR           Path to output applicable results");
 	puts("  -q, --quiet                Decrease log verbosity (can be repeated)");
 	puts("  -r, --rebaseline           Rewrite the baseline for failing tests");
+	puts("  -R, --rebaseline-missing   Write missing baselines tests only");
 	puts("  -v, --verbose              Increase log verbosity (can be repeated)");
+	puts("  -x, --xbaseline            Rewrite the xfail baselines for failing tests");
 	puts("  --version                  Print version and exit");
 }
 
@@ -276,6 +403,84 @@ static void testToPath(const char* testName, char* path) {
 	path[i] = '\0';
 }
 
+static bool globTests(struct CInemaTestList* tests, const char* glob, const char* ancestors) {
+	bool success = true;
+	const char* next = strpbrk(glob, "*.");
+
+	char path[PATH_MAX];
+	if (!next) {
+		testToPath(glob, path);
+		return collectTests(tests, path);
+	} else if (next[0] == '.') {
+		char subtest[MAX_TEST];
+		if (!ancestors) {
+			strncpy(subtest, glob, next - glob);
+		} else {
+			size_t len = strlen(ancestors) + (next - glob) + 2;
+			if (len > sizeof(subtest)) {
+				len = sizeof(subtest);
+			}
+			snprintf(subtest, len, "%s.%s", ancestors, glob);
+		}
+		return globTests(tests, next + 1, subtest);
+	} else if (next[0] == '*') {
+		char globBuffer[MAX_TEST];
+		const char* subglob;
+
+		next = strchr(next, '.');
+		if (!next) {
+			subglob = glob;
+		} else {
+			size_t len = next - glob + 1;
+			if (len > sizeof(globBuffer)) {
+				len = sizeof(globBuffer);
+			}
+			strncpy(globBuffer, glob, len - 1);
+			subglob = globBuffer;
+		}
+		bool hasMoreGlobs = next && strchr(next, '*');
+
+		struct VDir* dir;
+		if (ancestors) {
+			testToPath(ancestors, path);
+			dir = VDirOpen(path);
+		} else {
+			dir = VDirOpen(base);
+		}
+		if (!dir) {
+			return false;
+		}
+
+		struct VDirEntry* dirent = dir->listNext(dir);
+		while (dirent) {
+			const char* name = dirent->name(dirent);
+			if (dirent->type(dirent) != VFS_DIRECTORY || strncmp(name, ".", 2) == 0 || strncmp(name, "..", 3) == 0) {
+				dirent = dir->listNext(dir);
+				continue;
+			}
+			if (wildcard(subglob, name)) {
+				char newgen[MAX_TEST];
+				if (ancestors) {
+					snprintf(newgen, sizeof(newgen), "%s.%s", ancestors, name);
+				} else {
+					strlcpy(newgen, name, sizeof(newgen));
+				}
+				if (next && hasMoreGlobs) {
+					globTests(tests, next + 1, newgen);
+				} else {
+					testToPath(newgen, path);
+					collectTests(tests, path);
+				}
+			}
+			dirent = dir->listNext(dir);
+		}
+
+		return true;
+	} else {
+		abort();
+	}
+}
+
 static void _loadConfigTree(struct Table* configTree, const char* testName) {
 	char key[MAX_TEST];
 	strlcpy(key, testName, sizeof(key));
@@ -307,7 +512,7 @@ static void _unloadConfigTree(const char* key, void* value, void* user) {
 	mCoreConfigDeinit(value);
 }
 
-static const char* _lookupValue(struct Table* configTree, const char* testName, const char* key) {
+static const char* CInemaConfigGet(struct Table* configTree, const char* testName, const char* key) {
 	_loadConfigTree(configTree, testName);
 
 	char testKey[MAX_TEST];
@@ -336,7 +541,7 @@ static const char* _lookupValue(struct Table* configTree, const char* testName,
 }
 
 bool CInemaConfigGetUInt(struct Table* configTree, const char* testName, const char* key, unsigned* out) {
-	const char* charValue = _lookupValue(configTree, testName, key);
+	const char* charValue = CInemaConfigGet(configTree, testName, key);
 	if (!charValue) {
 		return false;
 	}
@@ -400,9 +605,9 @@ bool CInemaTestInit(struct CInemaTest* test, const char* directory, const char*
 	return true;
 }
 
-static bool _loadBaselinePNG(struct VDir* dir, struct CInemaImage* image, size_t frame, enum CInemaStatus* status) {
+static bool _loadBaselinePNG(struct VDir* dir, const char* type, struct CInemaImage* image, size_t frame, enum CInemaStatus* status) {
 	char baselineName[32];
-	snprintf(baselineName, sizeof(baselineName), "baseline_%04" PRIz "u.png", frame);
+	snprintf(baselineName, sizeof(baselineName), "%s_%04" PRIz "u.png", type, frame);
 	struct VFile* baselineVF = dir->openFile(dir, baselineName, O_RDONLY);
 	if (!baselineVF) {
 		if (*status == CI_PASS) {
@@ -541,9 +746,9 @@ static void _writeDiff(const char* testName, const struct CInemaImage* image, si
 	dir->close(dir);
 }
 
-static void _writeBaseline(struct VDir* dir, const struct CInemaImage* image, size_t frame) {
+static void _writeBaseline(struct VDir* dir, const char* type, const struct CInemaImage* image, size_t frame) {
 	char baselineName[32];
-	snprintf(baselineName, sizeof(baselineName), "baseline_%04" PRIz "u.png", frame);
+	snprintf(baselineName, sizeof(baselineName), "%s_%04" PRIz "u.png", type, frame);
 	struct VFile* baselineVF = dir->openFile(dir, baselineName, O_CREAT | O_TRUNC | O_WRONLY);
 	if (baselineVF) {
 		_writeImage(baselineVF, image);
@@ -552,9 +757,180 @@ static void _writeBaseline(struct VDir* dir, const struct CInemaImage* image, si
 	}
 }
 
-void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
+static bool _updateInput(struct mCore* core, size_t frame, const char** input) {
+	if (!*input || !*input[0]) {
+		return false;
+	}
+	char* end;
+	uint32_t start = strtoul(*input, &end, 10);
+	if (end[0] != ':') {
+		return false;
+	}
+	if (start != frame) {
+		return true;
+	}
+	++end;
+	*input = end;
+	uint32_t keys = strtoul(*input, &end, 16);
+	if (end[0] == ',') {
+		++end;
+	}
+	*input = end;
+	core->setKeys(core, keys);
+	return true;
+}
+
+static bool _compareImages(struct CInemaTest* restrict test, const struct CInemaImage* restrict image, const struct CInemaImage* restrict expected, int* restrict max, uint8_t** restrict outdiff) {
+	const uint8_t* testPixels = image->data;
+	const uint8_t* expectPixels = expected->data;
+	uint8_t* diff = NULL;
+	size_t x;
+	size_t y;
+	bool failed = false;
+	for (y = 0; y < image->height; ++y) {
+		for (x = 0; x < image->width; ++x) {
+			size_t pix = expected->stride * y + x;
+			size_t tpix = image->stride * y + x;
+			int testR = testPixels[tpix * 4 + 0];
+			int testG = testPixels[tpix * 4 + 1];
+			int testB = testPixels[tpix * 4 + 2];
+			int expectR = expectPixels[pix * 4 + 0];
+			int expectG = expectPixels[pix * 4 + 1];
+			int expectB = expectPixels[pix * 4 + 2];
+			int r = expectR - testR;
+			int g = expectG - testG;
+			int b = expectB - testB;
+			if (r | g | b) {
+				failed = true;
+				if (outdiff && !diff) {
+					diff = calloc(expected->stride * expected->height, BYTES_PER_PIXEL);
+					*outdiff = diff;
+				}
+				test->status = CI_FAIL;
+				if (r < 0) {
+					r = -r;
+				}
+				if (g < 0) {
+					g = -g;
+				}
+				if (b < 0) {
+					b = -b;
+				}
+
+				if (diff) {
+					if (r > *max) {
+						*max = r;
+					}
+					if (g > *max) {
+						*max = g;
+					}
+					if (b > *max) {
+						*max = b;
+					}
+					diff[pix * 4 + 0] = r;
+					diff[pix * 4 + 1] = g;
+					diff[pix * 4 + 2] = b;
+				}
+
+				if (test) {
+					test->totalDistance += r + g + b;
+					++test->failedPixels;
+				}
+			}
+		}
+	}
+	return !failed;
+}
+
+void _writeDiffSet(struct CInemaImage* expected, const char* name, uint8_t* diff, int frame, int max, bool xfail) {
+	struct CInemaImage outdiff = {
+		.data = diff,
+		.width = expected->width,
+		.height = expected->height,
+		.stride = expected->stride,
+	};
+
+	if (xfail) {
+		_writeDiff(name, expected, frame, "xexpected");
+		_writeDiff(name, &outdiff, frame, "xdiff");
+	} else {
+		_writeDiff(name, expected, frame, "expected");
+		_writeDiff(name, &outdiff, frame, "diff");
+	}
+
+	size_t x;
+	size_t y;
+	for (y = 0; y < outdiff.height; ++y) {
+		for (x = 0; x < outdiff.width; ++x) {
+			size_t pix = outdiff.stride * y + x;
+			diff[pix * 4 + 0] = diff[pix * 4 + 0] * 255 / max;
+			diff[pix * 4 + 1] = diff[pix * 4 + 1] * 255 / max;
+			diff[pix * 4 + 2] = diff[pix * 4 + 2] * 255 / max;
+		}
+	}
+	if (xfail) {
+		_writeDiff(name, &outdiff, frame, "xnormalized");
+	}
+}
+
+#ifdef USE_FFMPEG
+static void _replayBaseline(struct CInemaTest* test, struct FFmpegEncoder* encoder, const struct CInemaImage* image, int frame) {
+	char baselineName[PATH_MAX];
+	snprintf(baselineName, sizeof(baselineName), "%s" PATH_SEP ".baseline.avi", test->directory);
+
+	if (!FFmpegEncoderOpen(encoder, baselineName)) {
+		CIerr(1, "Failed to save baseline video\n");
+		test->status = CI_ERROR;
+		return;
+	}
+
+	snprintf(baselineName, sizeof(baselineName), "%s" PATH_SEP "baseline.avi", test->directory);
+
+
+	struct CInemaImage buffer = {
+		.data = NULL,
+		.width = image->width,
+		.height = image->height,
+		.stride = image->width,
+	};
+	struct FFmpegDecoder decoder;
+	struct CInemaStream stream = {0};
+	stream.d.postVideoFrame = _cinemaVideoFrame;
+	stream.d.videoDimensionsChanged = _cinemaDimensionsChanged;
+	stream.status = &test->status;
+	stream.image = &buffer;
+
+	FFmpegDecoderInit(&decoder);
+	decoder.out = &stream.d;
+
+	if (!FFmpegDecoderOpen(&decoder, baselineName)) {
+		CIerr(1, "Failed to load baseline video\n");
+		test->status = CI_ERROR;
+		return;
+	}
+
+	int i;
+	for (i = 0; i < frame; ++i) {
+		while (!buffer.data) {
+			if (!FFmpegDecoderRead(&decoder)) {
+				CIlog(1, "Failed to read more frames. EOF?\n");
+				test->status = CI_FAIL;
+				break;
+			}
+		}
+		encoder->d.postVideoFrame(&encoder->d, buffer.data, buffer.stride);
+		free(buffer.data);
+		buffer.data = NULL;
+	}
+	FFmpegDecoderClose(&decoder);
+}
+#endif
+
+void CInemaTestRun(struct CInemaTest* test) {
 	unsigned ignore = 0;
-	CInemaConfigGetUInt(configTree, test->name, "ignore", &ignore);
+	MutexLock(&configMutex);
+	CInemaConfigGetUInt(&configTree, test->name, "ignore", &ignore);
+	MutexUnlock(&configMutex);
 	if (ignore) {
 		test->status = CI_SKIP;
 		return;
@@ -602,12 +978,16 @@ void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
 	unsigned skip = 0;
 	unsigned fail = 0;
 	unsigned video = 0;
+	const char* input = NULL;
 
-	CInemaConfigGetUInt(configTree, test->name, "frames", &limit);
-	CInemaConfigGetUInt(configTree, test->name, "skip", &skip);
-	CInemaConfigGetUInt(configTree, test->name, "fail", &fail);
-	CInemaConfigGetUInt(configTree, test->name, "video", &video);
-	CInemaConfigLoad(configTree, test->name, core);
+	MutexLock(&configMutex);
+	CInemaConfigGetUInt(&configTree, test->name, "frames", &limit);
+	CInemaConfigGetUInt(&configTree, test->name, "skip", &skip);
+	CInemaConfigGetUInt(&configTree, test->name, "fail", &fail);
+	CInemaConfigGetUInt(&configTree, test->name, "video", &video);
+	input = CInemaConfigGet(&configTree, test->name, "input");
+	CInemaConfigLoad(&configTree, test->name, core);
+	MutexUnlock(&configMutex);
 
 	struct VFile* save = VFileMemChunk(NULL, 0);
 	core->loadROM(core, rom);
@@ -631,30 +1011,39 @@ void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
 	struct FFmpegDecoder decoder;
 	struct FFmpegEncoder encoder;
 	struct CInemaStream stream = {0};
+
+	char baselineName[PATH_MAX];
+	snprintf(baselineName, sizeof(baselineName), "%s" PATH_SEP "baseline.avi", test->directory);
+	bool exists = access(baselineName, 0) == 0;
+
 	if (video) {
-		char fname[PATH_MAX];
-		snprintf(fname, sizeof(fname), "%s" PATH_SEP "baseline.mkv", test->directory);
-		if (rebaseline) {
-			FFmpegEncoderInit(&encoder);
-			FFmpegEncoderSetAudio(&encoder, NULL, 0);
-			FFmpegEncoderSetVideo(&encoder, "png", 0, 0);
-			FFmpegEncoderSetContainer(&encoder, "mkv");
-			FFmpegEncoderSetDimensions(&encoder, image.width, image.height);
-			if (!FFmpegEncoderOpen(&encoder, fname)) {
+		FFmpegEncoderInit(&encoder);
+		FFmpegDecoderInit(&decoder);
+
+		FFmpegEncoderSetAudio(&encoder, NULL, 0);
+		FFmpegEncoderSetVideo(&encoder, "zmbv", 0, 0);
+		FFmpegEncoderSetContainer(&encoder, "avi");
+		FFmpegEncoderSetDimensions(&encoder, image.width, image.height);
+
+		if (rebaseline && !exists) {
+			if (!FFmpegEncoderOpen(&encoder, baselineName)) {
 				CIerr(1, "Failed to save baseline video\n");
 			} else {
 				core->setAVStream(core, &encoder.d);
 			}
-		} else {
-			FFmpegDecoderInit(&decoder);
+		}
+
+		if (exists) {
 			stream.d.postVideoFrame = _cinemaVideoFrame;
 			stream.d.videoDimensionsChanged = _cinemaDimensionsChanged;
 			stream.status = &test->status;
 			decoder.out = &stream.d;
 
-			if (!FFmpegDecoderOpen(&decoder, fname)) {
+			if (!FFmpegDecoderOpen(&decoder, baselineName)) {
 				CIerr(1, "Failed to load baseline video\n");
 			}
+		} else if (!rebaseline) {
+			test->status = CI_FAIL;
 		}
 	}
 #else
@@ -664,13 +1053,18 @@ void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
 	}
 #endif
 
+	bool xdiff = false;
 	for (frame = 0; limit; ++frame, --limit) {
+		_updateInput(core, frame, &input);
 		core->runFrame(core);
 		++test->totalFrames;
 		unsigned frameCounter = core->frameCounter(core);
 		if (frameCounter <= minFrame) {
 			break;
 		}
+		if (test->status == CI_ERROR) {
+			break;
+		}
 		CIlog(3, "Test frame: %u\n", frameCounter);
 		core->desiredVideoDimensions(core, &image.width, &image.height);
 		uint8_t* diff = NULL;
@@ -684,7 +1078,7 @@ void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
 		if (video) {
 			baselineFound = false;
 #ifdef USE_FFMPEG
-			if (!rebaseline && FFmpegDecoderIsOpen(&decoder)) {
+			if (FFmpegDecoderIsOpen(&decoder)) {
 				stream.image = &expected;
 				while (!expected.data) {
 					if (!FFmpegDecoderRead(&decoder)) {
@@ -697,136 +1091,203 @@ void CInemaTestRun(struct CInemaTest* test, struct Table* configTree) {
 			}
 #endif
 		} else {
-			baselineFound = _loadBaselinePNG(dir, &expected, frame, &test->status);
+			baselineFound = _loadBaselinePNG(dir, "baseline", &expected, frame, &test->status);
 		}
+		if (test->status == CI_ERROR) {
+			break;
+		}
+		bool failed = false;
 		if (baselineFound) {
-			uint8_t* testPixels = image.data;
-			uint8_t* expectPixels = expected.data;
-			size_t x;
-			size_t y;
 			int max = 0;
-			bool failed = false;
-			for (y = 0; y < image.height; ++y) {
-				for (x = 0; x < image.width; ++x) {
-					size_t pix = expected.stride * y + x;
-					size_t tpix = image.stride * y + x;
-					int testR = testPixels[tpix * 4 + 0];
-					int testG = testPixels[tpix * 4 + 1];
-					int testB = testPixels[tpix * 4 + 2];
-					int expectR = expectPixels[pix * 4 + 0];
-					int expectG = expectPixels[pix * 4 + 1];
-					int expectB = expectPixels[pix * 4 + 2];
-					int r = expectR - testR;
-					int g = expectG - testG;
-					int b = expectB - testB;
-					if (r | g | b) {
-						failed = true;
-						if (diffs && !diff) {
-							diff = calloc(expected.stride * expected.height, BYTES_PER_PIXEL);
-						}
-						CIlog(3, "Frame %u failed at pixel %" PRIz "ux%" PRIz "u with diff %i,%i,%i (expected %02x%02x%02x, got %02x%02x%02x)\n",
-						    frameCounter, x, y, r, g, b,
-						    expectR, expectG, expectB,
-						    testR, testG, testB);
-						test->status = CI_FAIL;
-						if (r < 0) {
-							r = -r;
-						}
-						if (g < 0) {
-							g = -g;
-						}
-						if (b < 0) {
-							b = -b;
-						}
-
-						if (diff) {
-							if (r > max) {
-								max = r;
-							}
-							if (g > max) {
-								max = g;
-							}
-							if (b > max) {
-								max = b;
-							}
-							diff[pix * 4 + 0] = r;
-							diff[pix * 4 + 1] = g;
-							diff[pix * 4 + 2] = b;
-						}
-
-						test->totalDistance += r + g + b;
-						++test->failedPixels;
-					}
-				}
-			}
+			failed = !_compareImages(test, &image, &expected, &max, diffs ? &diff : NULL);
 			if (failed) {
 				++test->failedFrames;
+#ifdef USE_FFMPEG
+				if (video && exists && rebaseline && !FFmpegEncoderIsOpen(&encoder)) {
+					_replayBaseline(test, &encoder, &image, frame);
+					if (test->status == CI_ERROR) {
+						break;
+					}
+					encoder.d.postVideoFrame(&encoder.d, image.data, image.stride);
+					core->setAVStream(core, &encoder.d);
+				}
+#endif
 			}
 			test->totalPixels += image.height * image.width;
-			if (rebaseline && failed) {
-				_writeBaseline(dir, &image, frame);
+			if (rebaseline == CI_R_FAILING && !video && failed) {
+				_writeBaseline(dir, "baseline", &image, frame);
 			}
 			if (diff) {
 				if (failed) {
-					struct CInemaImage outdiff = {
-						.data = diff,
-						.width = expected.width,
-						.height = expected.height,
-						.stride = expected.stride,
-					};
-
 					_writeDiff(test->name, &image, frame, "result");
-					_writeDiff(test->name, &expected, frame, "expected");
-					_writeDiff(test->name, &outdiff, frame, "diff");
-
-					for (y = 0; y < outdiff.height; ++y) {
-						for (x = 0; x < outdiff.width; ++x) {
-							size_t pix = outdiff.stride * y + x;
-							diff[pix * 4 + 0] = diff[pix * 4 + 0] * 255 / max;
-							diff[pix * 4 + 1] = diff[pix * 4 + 1] * 255 / max;
-							diff[pix * 4 + 2] = diff[pix * 4 + 2] * 255 / max;
-						}
-					}
-					_writeDiff(test->name, &outdiff, frame, "normalized");
+					_writeDiffSet(&expected, test->name, diff, frame, max, false);
 				}
 				free(diff);
+				diff = NULL;
 			}
 			free(expected.data);
-		} else if (test->status == CI_ERROR) {
-			break;
 		} else if (rebaseline && !video) {
-			_writeBaseline(dir, &image, frame);
+			_writeBaseline(dir, "baseline", &image, frame);
 		} else if (!rebaseline) {
 			test->status = CI_FAIL;
 		}
-	}
 
-	if (fail) {
-		if (test->status == CI_FAIL) {
-			test->status = CI_XFAIL;
-		} else if (test->status == CI_PASS) {
-			test->status = CI_XPASS;
+		if (fail && failed) {
+			if (video) {
+				// TODO
+				baselineFound = false;
+			} else {
+				baselineFound = _loadBaselinePNG(dir, "xbaseline", &expected, frame, &test->status);
+			}
+
+			if (baselineFound) {
+				int max = 0;
+				failed = !_compareImages(test, &image, &expected, &max, diffs ? &diff : NULL);
+				if (diff) {
+					if (failed) {
+						_writeDiffSet(&expected, test->name, diff, frame, max, true);
+					}
+					free(diff);
+					diff = NULL;
+				}
+				if (failed) {
+					if (xbaseline == CI_R_FAILING && !video) {
+						_writeBaseline(dir, "xbaseline", &image, frame);
+					}
+					xdiff = true;
+				}
+				free(expected.data);
+			} else if (xbaseline && !video) {
+				_writeBaseline(dir, "xbaseline", &image, frame);
+			}
 		}
 	}
 
 #ifdef USE_FFMPEG
 	if (video) {
-		if (rebaseline) {
+		if (FFmpegEncoderIsOpen(&encoder)) {
 			FFmpegEncoderClose(&encoder);
-		} else {
+			if (exists && rebaseline) {
+				char tmpBaselineName[PATH_MAX];
+				snprintf(tmpBaselineName, sizeof(tmpBaselineName), "%s" PATH_SEP ".baseline.avi", test->directory);
+#ifdef _WIN32
+				MoveFileEx(tmpBaselineName, baselineName, MOVEFILE_REPLACE_EXISTING);
+#else
+				rename(tmpBaselineName, baselineName);
+#endif
+			}
+		}
+		if (FFmpegDecoderIsOpen(&decoder)) {
 			FFmpegDecoderClose(&decoder);
 		}
 	}
 #endif
 
+	if (fail) {
+		if (test->status == CI_FAIL && !xdiff) {
+			test->status = CI_XFAIL;
+		} else if (test->status == CI_PASS) {
+			test->status = CI_XPASS;
+		}
+	}
+
 	free(image.data);
 	mCoreConfigDeinit(&core->config);
 	core->deinit(core);
 	dir->close(dir);
 }
 
+static bool CInemaTask(struct CInemaTestList* tests, size_t i) {
+	bool success = true;
+	struct CInemaTest* test = CInemaTestListGetPointer(tests, i);
+	if (dryRun) {
+		CIlog(-1, "%s\n", test->name);
+	} else {
+		CIlog(1, "%s: ", test->name);
+		fflush(stdout);
+		ThreadLocalSetKey(currentTest, test);
+		CInemaTestRun(test);
+		ThreadLocalSetKey(currentTest, NULL);
+
+		switch (test->status) {
+		case CI_PASS:
+			CIlog(1, "pass\n");
+			break;
+		case CI_FAIL:
+			success = false;
+			CIlog(1, "fail\n");
+			break;
+		case CI_XPASS:
+			CIlog(1, "xpass\n");
+			break;
+		case CI_XFAIL:
+			CIlog(1, "xfail\n");
+			break;
+		case CI_SKIP:
+			CIlog(1, "skip\n");
+			break;
+		case CI_ERROR:
+			success = false;
+			CIlog(1, "error\n");
+			break;
+		}
+		if (test->failedFrames) {
+			CIlog(2, "\tfailed frames: %u/%u (%1.3g%%)\n", test->failedFrames, test->totalFrames, test->failedFrames / (test->totalFrames * 0.01));
+			CIlog(2, "\tfailed pixels: %" PRIu64 "/%" PRIu64 " (%1.3g%%)\n", test->failedPixels, test->totalPixels, test->failedPixels / (test->totalPixels * 0.01));
+			CIlog(2, "\tdistance: %" PRIu64 "/%" PRIu64 " (%1.3g%%)\n", test->totalDistance, test->totalPixels * 765, test->totalDistance / (test->totalPixels * 7.65));
+		}
+	}
+	return success;
+}
+
+static THREAD_ENTRY CInemaJob(void* context) {
+	struct CInemaTestList* tests = context;
+	struct CInemaLogStream stream;
+	StringListInit(&stream.out.lines, 0);
+	StringListInit(&stream.out.partial, 0);
+	stream.out.repeat = 0;
+	StringListInit(&stream.err.lines, 0);
+	StringListInit(&stream.err.partial, 0);
+	stream.err.repeat = 0;
+	ThreadLocalSetKey(logStream, &stream);
+
+	bool success = true;
+	while (true) {
+		size_t i;
+		MutexLock(&jobMutex);
+		i = jobIndex;
+		++jobIndex;
+		MutexUnlock(&jobMutex);
+		if (i >= CInemaTestListSize(tests)) {
+			break;
+		}
+		if (!CInemaTask(tests, i)) {
+			success = false;
+		}
+		CIflush(&stream.out, stdout);
+		CIflush(&stream.err, stderr);
+	}
+	MutexLock(&jobMutex);
+	if (!success) {
+		jobStatus = 1;
+	}
+	MutexUnlock(&jobMutex);
+
+	CIflush(&stream.out, stdout);
+	StringListDeinit(&stream.out.lines);
+	StringListDeinit(&stream.out.partial);
+
+	CIflush(&stream.err, stderr);
+	StringListDeinit(&stream.err.lines);
+	StringListDeinit(&stream.err.partial);
+}
+
 void _log(struct mLogger* log, int category, enum mLogLevel level, const char* format, va_list args) {
 	UNUSED(log);
+	if (level == mLOG_FATAL) {
+		struct CInemaTest* test = ThreadLocalGetValue(currentTest);
+		test->status = CI_ERROR;
+	}
 	if (verbosity < 0) {
 		return;
 	}
@@ -853,6 +1314,9 @@ void _log(struct mLogger* log, int category, enum mLogLevel level, const char* f
 }
 
 int main(int argc, char** argv) {
+	ThreadLocalInitKey(&logStream);
+	ThreadLocalSetKey(logStream, NULL);
+
 	int status = 0;
 	if (!parseCInemaArgs(argc, argv)) {
 		status = 1;
@@ -899,6 +1363,13 @@ int main(int argc, char** argv) {
 	if (argc > 0) {
 		size_t i;
 		for (i = 0; i < (size_t) argc; ++i) {
+			if (strchr(argv[i], '*')) {
+				if (!globTests(&tests, argv[i], NULL)) {
+					status = 1;
+					break;
+				}
+				continue;
+			}
 			char path[PATH_MAX + 1] = {0};
 			testToPath(argv[i], path);
 
@@ -918,48 +1389,33 @@ int main(int argc, char** argv) {
 		reduceTestList(&tests);
 	}
 
-	struct Table configTree;
 	HashTableInit(&configTree, 0, free);
+	MutexInit(&configMutex);
+	ThreadLocalInitKey(&currentTest);
+	ThreadLocalSetKey(currentTest, NULL);
 
-	size_t i;
-	for (i = 0; i < CInemaTestListSize(&tests); ++i) {
-		struct CInemaTest* test = CInemaTestListGetPointer(&tests, i);
-		if (dryRun) {
-			CIlog(-1, "%s\n", test->name);
-		} else {
-			CIlog(1, "%s: ", test->name);
-			fflush(stdout);
-			CInemaTestRun(test, &configTree);
-			switch (test->status) {
-			case CI_PASS:
-				CIlog(1, "pass\n");
-				break;
-			case CI_FAIL:
+	if (jobs == 1) {
+		size_t i;
+		for (i = 0; i < CInemaTestListSize(&tests); ++i) {
+			bool success = CInemaTask(&tests, i);
+			if (!success) {
 				status = 1;
-				CIlog(1, "fail\n");
-				break;
-			case CI_XPASS:
-				CIlog(1, "xpass\n");
-				break;
-			case CI_XFAIL:
-				CIlog(1, "xfail\n");
-				break;
-			case CI_SKIP:
-				CIlog(1, "skip\n");
-				break;
-			case CI_ERROR:
-				status = 1;
-				CIlog(1, "error\n");
-				break;
-			}
-			if (test->failedFrames) {
-				CIlog(2, "\tfailed frames: %u/%u (%1.3g%%)\n", test->failedFrames, test->totalFrames, test->failedFrames / (test->totalFrames * 0.01));
-				CIlog(2, "\tfailed pixels: %" PRIu64 "/%" PRIu64 " (%1.3g%%)\n", test->failedPixels, test->totalPixels, test->failedPixels / (test->totalPixels * 0.01));
-				CIlog(2, "\tdistance: %" PRIu64 "/%" PRIu64 " (%1.3g%%)\n", test->totalDistance, test->totalPixels * 765, test->totalDistance / (test->totalPixels * 7.65));
 			}
 		}
+	} else {
+		MutexInit(&jobMutex);
+		int i;
+		for (i = 0; i < jobs; ++i) {
+			ThreadCreate(&jobThreads[i], CInemaJob, &tests);
+		}
+		for (i = 0; i < jobs; ++i) {
+			ThreadJoin(&jobThreads[i]);
+		}
+		MutexDeinit(&jobMutex);
+		status = jobStatus;
 	}
 
+	MutexDeinit(&configMutex);
 	HashTableEnumerate(&configTree, _unloadConfigTree, NULL);
 	HashTableDeinit(&configTree);
 	CInemaTestListDeinit(&tests);
diff --git a/src/util/string.c b/src/util/string.c
index 102a8cfbc..6a60abf48 100644
--- a/src/util/string.c
+++ b/src/util/string.c
@@ -518,4 +518,34 @@ ssize_t parseQuotedString(const char* unparsed, ssize_t unparsedLen, char* parse
 		}
 	}
 	return -1;
+}
+
+bool wildcard(const char* search, const char* string) {
+	while (true) {
+		if (search[0] == '*') {
+			while (search[0] == '*') {
+				++search;
+			}
+			if (!search[0]) {
+				return true;
+			}
+			while (string[0]) {
+				if (string[0] == search[0] && wildcard(search, string)) {
+					return true;
+				}
+				++string;
+			}
+			return false;
+		} else if (!search[0]) {
+			return !string[0];
+		} else if (!string[0]) {
+			return false;
+		} else if (string[0] != search[0]) {
+			return false;
+		} else {
+			++search;
+			++string;
+		}
+	}
+	return false;
 }
\ No newline at end of file