ih264d: Small optimizations and experiments with multi-threading

Using the multi-threaded decoder doesn't seem to be worth it but at least we have a way to enable it now

dependencies/ih264d/CMakeLists.txt

@@ -183,4 +183,10 @@ endif()
 
 if(MSVC)
 set_property(TARGET ih264d PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
+
+# tune settings for slightly better performance
+target_compile_options(ih264d PRIVATE $<$<CONFIG:Release,RelWithDebInfo>:/Oi>) # enable intrinsic functions
+target_compile_options(ih264d PRIVATE $<$<CONFIG:Release,RelWithDebInfo>:/Ot>) # favor speed
+target_compile_options(ih264d PRIVATE "/GS-") # disable runtime checks
+
 endif()

dependencies/ih264d/common/ithread.c

@@ -85,28 +85,59 @@ UWORD32 ithread_get_mutex_lock_size(void)
 	return sizeof(CRITICAL_SECTION);
 }
 
+struct _ithread_launch_param
+{
+	void (*startFunc)(void* argument);
+	void* argument;
+};
+
+DWORD WINAPI _ithread_WinThreadStartRoutine(LPVOID lpThreadParameter)
+{
+	struct _ithread_launch_param* param = (struct _ithread_launch_param*)lpThreadParameter;
+	typedef void *(*ThreadStartRoutineType)(void *);
+	ThreadStartRoutineType pfnThreadRoutine = (ThreadStartRoutineType)param->startFunc;
+	void* arg = param->argument;
+	free(param);
+	pfnThreadRoutine(arg);
+	return 0;
+}
+
 WORD32 ithread_create(void* thread_handle, void* attribute, void* strt, void* argument)
 {
-	//UNUSED(attribute);
-	//return pthread_create((pthread_t*)thread_handle, NULL, (void* (*)(void*)) strt, argument);
-	__debugbreak();
+	UNUSED(attribute);
+	struct _ithread_launch_param* param = malloc(sizeof(struct _ithread_launch_param));
+	param->startFunc = (void (*)(void*))strt;
+	param->argument = argument;
+	HANDLE *handle = (HANDLE*)thread_handle;
+	*handle = CreateThread(NULL, 0, _ithread_WinThreadStartRoutine, param, 0, NULL);
+	if(*handle == NULL)
+	{
+		return -1;
+	}
 	return 0;
 }
 
 WORD32 ithread_join(void* thread_handle, void** val_ptr)
 {
 	//UNUSED(val_ptr);
-	//pthread_t* pthread_handle = (pthread_t*)thread_handle;
-	//return pthread_join(*pthread_handle, NULL);
-
-	__debugbreak();
-	return 0;
+	HANDLE *handle = (HANDLE*)thread_handle;
+	DWORD result = WaitForSingleObject(*handle, INFINITE);
+	if(result == WAIT_OBJECT_0)
+	{
+		CloseHandle(*handle);
+		return 0;
+	}
+	else
+	{
+		return -1;
+	}
 }
 
 WORD32 ithread_get_mutex_struct_size(void)
 {
 	return sizeof(CRITICAL_SECTION);
 }
+
 WORD32 ithread_mutex_init(void* mutex)
 {
 	InitializeCriticalSection((LPCRITICAL_SECTION)mutex);
@@ -153,7 +184,6 @@ UWORD32 ithread_get_sem_struct_size(void)
 	//return(sizeof(sem_t));
 }
 
-
 WORD32 ithread_sem_init(void* sem, WORD32 pshared, UWORD32 value)
 {
 	__debugbreak();
@@ -168,7 +198,6 @@ WORD32 ithread_sem_post(void* sem)
 	//return sem_post((sem_t*)sem);
 }
 
-
 WORD32 ithread_sem_wait(void* sem)
 {
 	__debugbreak();
@@ -176,7 +205,6 @@ WORD32 ithread_sem_wait(void* sem)
 	//return sem_wait((sem_t*)sem);
 }
 
-
 WORD32 ithread_sem_destroy(void* sem)
 {
 	__debugbreak();

dependencies/ih264d/common/x86/ih264_platform_macros.h

@@ -79,10 +79,8 @@
 static inline int __builtin_clz(unsigned x)
 {
 	unsigned long n;
-	if (x == 0)
-		return 32;
 	_BitScanReverse(&n, x);
-	return 31 - n;
+	return n ^ 31;
 }
 
 static inline int __builtin_ctz(unsigned x) {

src/Cafe/OS/libs/h264_avc/H264Dec.cpp

@@ -254,6 +254,8 @@ namespace H264
 			m_codecCtx->pv_fxns = (void*)&ih264d_api_function;
 			m_codecCtx->u4_size = sizeof(iv_obj_t);
 
+			SetDecoderCoreCount(1);
+
 			m_isBufferedMode = isBufferedMode;
 
 			UpdateParameters(false);
@@ -278,6 +280,19 @@ namespace H264
 			m_codecCtx = nullptr;
 		}
 
+		void SetDecoderCoreCount(uint32 coreCount)
+		{
+			ih264d_ctl_set_num_cores_ip_t s_set_cores_ip;
+			ih264d_ctl_set_num_cores_op_t s_set_cores_op;
+			s_set_cores_ip.e_cmd = IVD_CMD_VIDEO_CTL;
+			s_set_cores_ip.e_sub_cmd = (IVD_CONTROL_API_COMMAND_TYPE_T)IH264D_CMD_CTL_SET_NUM_CORES;
+			s_set_cores_ip.u4_num_cores = coreCount; // valid numbers are 1-4
+			s_set_cores_ip.u4_size = sizeof(ih264d_ctl_set_num_cores_ip_t);
+			s_set_cores_op.u4_size = sizeof(ih264d_ctl_set_num_cores_op_t);
+			IV_API_CALL_STATUS_T status = ih264d_api_function(m_codecCtx, (void *)&s_set_cores_ip, (void *)&s_set_cores_op);
+			cemu_assert(status == IV_SUCCESS);
+		}
+
 		static bool GetImageInfo(uint8* stream, uint32 length, uint32& imageWidth, uint32& imageHeight)
 		{
 			// create temporary decoder
@@ -702,7 +717,6 @@ namespace H264
 			decodeResult = m_bufferedResults.front();
 			m_bufferedResults.erase(m_bufferedResults.begin());
 		}
-
 	private:
 		iv_obj_t* m_codecCtx{nullptr};
 		bool m_hasBufferSizeInfo{ false };