diff --git a/rpcs3/Emu/Memory/vm_ptr.h b/rpcs3/Emu/Memory/vm_ptr.h
index a687eab5d4..9a7e150170 100644
--- a/rpcs3/Emu/Memory/vm_ptr.h
+++ b/rpcs3/Emu/Memory/vm_ptr.h
@@ -1,5 +1,7 @@
 #pragma once
 
+class PPUThread;
+
 namespace vm
 {
 	template<typename T, int lvl = 1, typename AT = u32>
@@ -319,7 +321,9 @@ namespace vm
 	public:
 		typedef RT(*type)(T...);
 
-		RT operator()(T... args) const; // defined in Callback.h (CB_FUNC.h)
+		RT call(PPUThread& CPU, T... args) const; // call using specified PPU thread context, defined in Callback.h (CB_FUNC.h)
+
+		RT operator()(T... args) const; // call using current PPU thread context, defined in Callback.h (CB_FUNC.h)
 
 		AT addr() const
 		{
diff --git a/rpcs3/Emu/SysCalls/CB_FUNC.h b/rpcs3/Emu/SysCalls/CB_FUNC.h
index 83d1de2c0b..9d35e3d885 100644
--- a/rpcs3/Emu/SysCalls/CB_FUNC.h
+++ b/rpcs3/Emu/SysCalls/CB_FUNC.h
@@ -3,51 +3,148 @@
 
 namespace vm
 {
-	template<typename TT>
-	struct _func_arg
+	enum _func_arg_type
 	{
-		static_assert(!std::is_floating_point<TT>::value, "TODO: Unsupported callback argument type (floating point)");
-		static_assert(!std::is_same<TT, u128>::value, "TODO: Unsupported callback argument type (vector)");
+		ARG_GENERAL,
+		ARG_FLOAT,
+		ARG_VECTOR,
+		ARG_STACK,
+	};
 
-		static_assert(sizeof(TT) <= 8, "Invalid callback argument type");
-		static_assert(!std::is_pointer<TT>::value, "Invalid callback argument type (pointer)");
-		static_assert(!std::is_reference<TT>::value, "Invalid callback argument type (reference)");
+	template<typename T, _func_arg_type type, int g_count, int f_count, int v_count>
+	struct _func_arg;
 
-		__forceinline static u64 get_value(const TT& arg)
+	template<typename T, int g_count, int f_count, int v_count>
+	struct _func_arg<T, ARG_GENERAL, g_count, f_count, v_count>
+	{
+		static_assert(sizeof(T) <= 8, "Invalid callback argument type for ARG_GENERAL");
+		
+		__forceinline static void set_value(PPUThread& CPU, const T arg)
 		{
-			u64 res = 0;
-			(TT&)res = arg;
-			return res;
+			(T&)CPU.GPR[g_count + 2] = arg;
 		}
 	};
 
+	template<typename T, int g_count, int f_count, int v_count>
+	struct _func_arg<T, ARG_FLOAT, g_count, f_count, v_count>
+	{
+		static_assert(sizeof(T) <= 8, "Invalid callback argument type for ARG_FLOAT");
+
+		__forceinline static void set_value(PPUThread& CPU, const T arg)
+		{
+			CPU.FPR[f_count] = arg;
+		}
+	};
+
+	template<typename T, int g_count, int f_count, int v_count>
+	struct _func_arg<T, ARG_VECTOR, g_count, f_count, v_count>
+	{
+		static_assert(std::is_same<T, u128>::value, "Invalid callback argument type for ARG_VECTOR");
+
+		__forceinline static void set_value(PPUThread& CPU, const T arg)
+		{
+			(T&)CPU.VPR[v_count + 1] = arg;
+		}
+	};
+
+	template<typename T, int g_count, int f_count, int v_count>
+	struct _func_arg<T, ARG_STACK, g_count, f_count, v_count>
+	{
+		static_assert(g_count <= 8, "TODO: Unsupported stack argument type (general)");
+		static_assert(f_count <= 12, "TODO: Unsupported stack argument type (float)");
+		static_assert(v_count <= 12, "TODO: Unsupported stack argument type (vector)");
+		static_assert(sizeof(T) <= 8, "Invalid callback argument type for ARG_STACK");
+
+		__forceinline static void set_value(PPUThread& CPU, const T arg)
+		{
+			// TODO
+		}
+	};
+
+	template<int g_count, int f_count, int v_count>
+	__forceinline static void _bind_func_args(PPUThread& CPU)
+	{
+		// terminator
+	}
+
+	template<int g_count, int f_count, int v_count, typename T1, typename... T>
+	__forceinline static void _bind_func_args(PPUThread& CPU, T1 arg1, T... args)
+	{
+		static_assert(!std::is_pointer<T1>::value, "Invalid callback argument type (pointer)");
+		static_assert(!std::is_reference<T1>::value, "Invalid callback argument type (reference)");
+		const bool is_float = std::is_floating_point<T1>::value;
+		const bool is_vector = std::is_same<T1, u128>::value;
+		const _func_arg_type t = is_float
+			? ((f_count >= 12) ? ARG_STACK : ARG_FLOAT)
+			: (is_vector ? ((v_count >= 12) ? ARG_STACK : ARG_VECTOR) : ((g_count >= 8) ? ARG_STACK : ARG_GENERAL));
+		const int g = g_count + (is_float || is_vector ? 0 : 1);
+		const int f = f_count + (is_float ? 1 : 0);
+		const int v = v_count + (is_vector ? 1 : 0);
+		
+		_func_arg<T1, t, g, f, v>::set_value(CPU, arg1);
+		_bind_func_args<g, f, v>(CPU, args...);
+	}
+
 	template<typename RT>
 	struct _func_res
 	{
-		static_assert(!std::is_floating_point<RT>::value, "TODO: Unsupported callback result type (floating point)");
-		static_assert(!std::is_same<RT, u128>::value, "TODO: Unsupported callback result type (vector)");
-
 		static_assert(sizeof(RT) <= 8, "Invalid callback result type");
 		static_assert(!std::is_pointer<RT>::value, "Invalid callback result type (pointer)");
 		static_assert(!std::is_reference<RT>::value, "Invalid callback result type (reference)");
 
-		__forceinline static RT get_value(const u64 res)
+		__forceinline static RT get_value(const PPUThread& CPU)
 		{
-			return (RT&)res;
+			if (std::is_floating_point<RT>::value)
+			{
+				return (RT)CPU.FPR[1];
+			}
+			else
+			{
+				return (RT&)CPU.GPR[3];
+			}
+		}
+	};
+
+	template<>
+	struct _func_res<u128>
+	{
+		__forceinline static u128 get_value(const PPUThread& CPU)
+		{
+			return CPU.VPR[2];
 		}
 	};
 
 	template<>
 	struct _func_res<void>
 	{
-		__forceinline static void get_value(const u64 res)
+		__forceinline static void get_value(const PPUThread& CPU)
 		{
 		}
 	};
 
-	template<typename AT, typename RT, typename... T>
-	RT _ptr_base<RT(*)(T...), 1, AT>::operator ()(T... args) const
+	template<typename RT, typename... T>
+	struct _func_caller
 	{
-		return _func_res<RT>::get_value(GetCurrentPPUThread().FastCall(vm::read32(m_addr), vm::read32(m_addr + 4), _func_arg<T>::get_value(args)...));
+		__forceinline static RT call(PPUThread& CPU, u32 pc, u32 rtoc, T... args)
+		{
+			_bind_func_args<0, 0, 0>(CPU, args...);
+			CPU.FastCall2(pc, rtoc);
+			return _func_res<RT>::get_value(CPU);
+		}
+	};
+
+	template<typename AT, typename RT, typename... T>
+	__forceinline RT _ptr_base<RT(*)(T...), 1, AT>::call(PPUThread& CPU, T... args) const
+	{
+		const u32 pc = vm::get_ref<be_t<u32>>(m_addr);
+		const u32 rtoc = vm::get_ref<be_t<u32>>(m_addr + 4);
+
+		return _func_caller<RT, T...>::call(CPU, pc, rtoc, args...);
+	}
+
+	template<typename AT, typename RT, typename... T>
+	__forceinline RT _ptr_base<RT(*)(T...), 1, AT>::operator ()(T... args) const
+	{
+		return call(GetCurrentPPUThread(), args...);
 	}
 }
\ No newline at end of file
diff --git a/rpcs3/Emu/SysCalls/Callback.h b/rpcs3/Emu/SysCalls/Callback.h
index 6402ca84b6..f5747e6ac5 100644
--- a/rpcs3/Emu/SysCalls/Callback.h
+++ b/rpcs3/Emu/SysCalls/Callback.h
@@ -1,8 +1,6 @@
 #pragma once
 #include "CB_FUNC.h"
 
-class CPUThread;
-
 class CallbackManager
 {
 	std::vector<std::function<s32()>> m_cb_list;
diff --git a/rpcs3/Emu/SysCalls/SC_FUNC.h b/rpcs3/Emu/SysCalls/SC_FUNC.h
index f01cc741d4..16e16cd2ab 100644
--- a/rpcs3/Emu/SysCalls/SC_FUNC.h
+++ b/rpcs3/Emu/SysCalls/SC_FUNC.h
@@ -59,7 +59,7 @@ namespace detail
 	{
 		static_assert(f_count <= 12, "TODO: Unsupported stack argument type (float)");
 		static_assert(v_count <= 12, "TODO: Unsupported stack argument type (vector)");
-		static_assert(sizeof(T) <= 8, "Invalid function argument type (ARG_STACK)");
+		static_assert(sizeof(T) <= 8, "Invalid function argument type for ARG_STACK");
 
 		static __forceinline T func(PPUThread& CPU)
 		{
@@ -126,6 +126,7 @@ namespace detail
 	template<int g_count, int f_count, int v_count>
 	static __forceinline std::tuple<> iterate(PPUThread& CPU)
 	{
+		// terminator
 		return std::tuple<>();
 	}
 
@@ -143,6 +144,7 @@ namespace detail
 		const int g = g_count + (is_float || is_vector ? 0 : 1);
 		const int f = f_count + (is_float ? 1 : 0);
 		const int v = v_count + (is_vector ? 1 : 0);
+
 		return std::tuple_cat(std::tuple<T>(bind_arg<T, t, g, f, v>::func(CPU)), iterate<g, f, v, A...>(CPU));
 	}