The Design of Five-stage Pipeline Processor based on MIPS32
This document is a technical report for ACMCA2016 Summer Session Project. In this project, a simple five-stage mips32 pipeline processor is implemented in Verilog HDL. My processor has fully supported the integer instruction set of MIPS32 and successfully passed all test cases.
My implementation is mainly followed with this blog. It is designed in Harvard Structure rather than von Neumann Structure. All the instructions are stored in the ROM and all the data is stored in another separated memory (RAM), so that it easily prevents structure hazard.
Main features of my mips32 processor:
- five-stage integral pipeline, including: instruction fetch, instruction decode, execution, memory access, write back.
- 32-bit data and address bus width.
- 32 general purpose working registers.
- support multi-cycle instructions.
- support branch delay slots but not support load delay slot.
The source code is available on github: https://github.com/RunzheYang/mips32-cpu.
There are 17 worked modules in my implementation:
- Pipeline Stage Modules (3): decode.v, execute.v, memory.v
- Stage Switch Modules (4) : if_id.v, id_ex.v, ex_mem.v and mem_wb.v
- Register Modules (4) : pc_reg.v, reg_file.v, hilo_reg.v, llbit_reg.v
- Memory Manager Modules (2) : inst_rom.v, data_ram.v
- Control Module (1): ctrl.v
- Division Module (1): div.v
- Integration Modules (2): cpu.v, sopc.v
The function of integration module is connecting each module with wires. Control module is and clean related Stage Switch Module. And all the constants are defined in a single file const.v. I run tests with the file sopc_tb.v in experiment to observe the waveforms through the GTKWave.
As the figure shows, the minimal SOPC used in this experiment contains three components, the CPU, ROM and RAM. They are connected with several wires for data transfer. The ROM provides with instructions and there is no write operation for it. The RAM stores run-time data, and it should be able to support both read and write operations. clk and rst are two port accepting clock signal and reset signal.
Related modules: pc_reg.v, inst_rom.v, control.v.
The function of pc_reg.v is to compute and store the address of current instruction. inst_rom.v will read the address sent by pc_reg.v and then send the corresponding instruction to the if_id.v module. Since this mips32 processor supports multi-cycle instructions, in pc_reg.v I use a port stall to receive the command from control.v to keep PC unchanged if there is a multi-cycle instruction on the pipeline. Further, for supporting jump and branch, I preserve two ports branch_flag_in and branch_tar_addr_in to receive the signal from instruction decode stage. (An illustration will be post here soon)
Related modules: decode.v, reg_file.v, control.v.
The mission of this stage is to translate the instruction and extract the source operand from the instruction. The decode.v module parse the instruction to get the aluop (ALU operation type) and alusel (ALU result selection) and require the corresponding data form reg_file.v, then send all these information to the id_ex.v module. For supporting branch delay slot, the decode.v is designed with two ports in_delayslot_in and next_inst_delayslot_out. Since it doesn't support load delay slot, ex_aluop_in is a load instruction and share the same source operand register with current instruction, it will send a stall signal to the ctrl.v.
// solving load cause data hazard by a stall request assign stall_req = stall_src1_loadrelated | stall_src2_loadrelated;
(An illustration will be post here soon)
Related modules: execution.v, div.v, control.v.
The execution.v module works for calculation the result of instruction. To avoid Data Hard, I use dest_data_out will also send to decode.v module to achieve forwarding. Because madd, maddu, msub and msubu will occupy 2 cycles, then execution.v module also be able to send stall request to the ctrl.v module. div and divu will use 32 cycles to calculate the quotient and remainder, and store them in the hilo_reg. A div.v module is designed for accomplishing that. (An illustration will be post here soon)
Related modules: memory.v, llbit_reg.v, control.v.
The memory.v module is used to store or fetch data as store or load instruction command. The calculation of target address is also included in this part. Since the load instruction may cause a data hazard, memory.v module is also able to send the stall request. ll and sc are two special instructions which need a LLbit register to check whether the RMW series is undisturbed. (An illustration will be post here soon)
Related modules: reg_file.v, hilo_reg.v.
The write back stage is completed by the registers themselves. When the register receives a WriteEnable signal, it will update the specified address to new coming data.
Experiment and Result
Logical & Bitwise Shift
Data Transfer (move)
Conditional branch & Unconditional jump
Data Transfer (load, store)
- Exception and Interrupt Handling: Up to now my processor supports many instructions while most of them are used to do calculation. A real processor should be able to do more operations such like exception and interrupt handling, for running a modern operation system. To support exception and interrupt handling, there is a need to implement the co-processor CP0.
- Dynamic Scheduling: An out-of-order execution with Tomasulo algorithm is also in my plan, since multiple issue and multiple ALUs significantly improve the efficiency of the processor.
- Branch Predictor: A branch predictor tries to guess which way a branch will go before this is known for sure. It is a common way to help pipelined processor achieve high effective performance.
- 雷思磊. 自己动手写 CPU. 电子工业出版社, 2014.
- Yamin Li. Computer Principles and Design in Verilog HDL. Tsinghua University Press, 2011.
- J.L. Hennessy, D.A. Patterson, and K. Asanovi´c. Computer Architecture: A Quantitative Ap- proach. Computer Architecture: A Quantitative Approach. Morgan Kaufmann/Elsevier, 2012.
- molfat. MIPS processor with a 5-stage pipeline. Github. https://github.com/molfat/cpu-mips.
- Mips instruction set: https://en.wikipedia.org/wiki/MIPS_instruction_set
Runzhe Yang (ID: 5140309562)