diff --git a/parallel-rsp/parallel.cpp b/parallel-rsp/parallel.cpp
index 614da2f..b3cf43d 100644
--- a/parallel-rsp/parallel.cpp
+++ b/parallel-rsp/parallel.cpp
@@ -83,7 +83,9 @@ extern "C"
 		RSP::cpu.invalidate_imem();
 
 		// Run CPU until we either break or we need to fire an IRQ.
-		cycles = RSP::cpu.get_state().pc = *RSP::rsp.SP_PC_REG & 0xfff;
+		RSP::cpu.get_state().pc = *RSP::rsp.SP_PC_REG & 0xfff;
+		RSP::cpu.get_state().instruction_count = 0;
+		RSP::cpu.get_state().last_instruction_type = RSP::VU_INSTRUCTION;
 
 #ifdef INTENSE_DEBUG
 		fprintf(stderr, "RUN TASK: %u\n", RSP::cpu.get_state().pc);
@@ -100,15 +102,11 @@ extern "C"
 				break;
 		}
 
-		if (RSP::cpu.get_state().pc > cycles)
-			cycles = RSP::cpu.get_state().pc - cycles;
-		else
-			cycles = 0;
 		*RSP::rsp.SP_PC_REG = 0x04001000 | (RSP::cpu.get_state().pc & 0xffc);
 
 		// From CXD4.
 		if (*RSP::rsp.SP_STATUS_REG & SP_STATUS_BROKE)
-			return cycles;
+			return RSP::cpu.get_state().instruction_count * 1.5; // Converting RCP clock rate to CPU clock rate
 		else if (*RSP::cpu.get_state().cp0.irq & 1)
 			RSP::rsp.CheckInterrupts();
 		else if (*RSP::rsp.SP_SEMAPHORE_REG != 0) // Semaphore lock fixes.
@@ -120,7 +118,7 @@ extern "C"
 		// CPU restarts with the correct SIGs.
 		*RSP::rsp.SP_STATUS_REG &= ~SP_STATUS_HALT;
 
-		return cycles;
+		return RSP::cpu.get_state().instruction_count * 1.5; // Converting RCP clock rate to CPU clock rate
 	}
 
 	EXPORT m64p_error CALL PluginGetVersion(m64p_plugin_type *PluginType, int *PluginVersion,
diff --git a/parallel-rsp/rsp_jit.cpp b/parallel-rsp/rsp_jit.cpp
index a0e1c36..9b1c554 100644
--- a/parallel-rsp/rsp_jit.cpp
+++ b/parallel-rsp/rsp_jit.cpp
@@ -305,6 +305,12 @@ extern "C"
 		printf("  ... Enter 0x%03x ...  ", unsigned(pc & 0xffcu));
 	}
 #endif
+	static void add_instruction_count(CPUState *rsp, uint32_t instruction_type)
+	{
+		if (instruction_type == rsp->last_instruction_type)
+			++rsp->instruction_count;
+		rsp->last_instruction_type = instruction_type;
+	}
 }
 
 void CPU::jit_save_indirect_register(jit_state_t *_jit, unsigned mips_register)
@@ -864,6 +870,11 @@ void CPU::jit_instruction(jit_state_t *_jit, uint32_t pc, uint32_t instr,
 	// VU
 	if ((instr >> 25) == 0x25)
 	{
+		regs.flush_register_window(_jit);
+		jit_begin_call(_jit);
+		jit_pushargr(JIT_REGISTER_STATE);
+		jit_pushargi(RSP::VU_INSTRUCTION);
+		jit_end_call(_jit, reinterpret_cast<jit_pointer_t>(add_instruction_count));
 		// VU instruction. COP2, and high bit of opcode is set.
 		uint32_t op = instr & 63;
 		uint32_t vd = (instr >> 6) & 31;
@@ -896,6 +907,14 @@ void CPU::jit_instruction(jit_state_t *_jit, uint32_t pc, uint32_t instr,
 		jit_end_call(_jit ,reinterpret_cast<jit_pointer_t>(vuop));
 		return;
 	}
+	else
+	{
+		regs.flush_register_window(_jit);
+		jit_begin_call(_jit);
+		jit_pushargr(JIT_REGISTER_STATE);
+		jit_pushargi(RSP::SU_INSTRUCTION);
+		jit_end_call(_jit, reinterpret_cast<jit_pointer_t>(add_instruction_count));
+	}
 
 	// TODO: Meaningful register allocation.
 	// For now, always flush register state to memory after an instruction for simplicity.
diff --git a/parallel-rsp/state.hpp b/parallel-rsp/state.hpp
index ce80dae..afb2ab0 100644
--- a/parallel-rsp/state.hpp
+++ b/parallel-rsp/state.hpp
@@ -117,6 +117,8 @@ struct alignas(64) CP2
 struct CPUState
 {
 	uint32_t pc = 0;
+	uint32_t instruction_count = 0;
+	uint32_t last_instruction_type = 0;
 	uint32_t dirty_blocks = 0;
 	static_assert(CODE_BLOCKS <= 32, "Code blocks must fit in 32-bit register.");
 
@@ -141,6 +143,12 @@ enum ReturnMode
 	MODE_CHECK_FLAGS = 4
 };
 
+enum InstructionType
+{
+	VU_INSTRUCTION = 0,
+	SU_INSTRUCTION = 1
+};
+
 } // namespace RSP
 
 #endif