RISC-V CPU Labs in Chisel

Note

Several code fragments are intentionally incomplete. They serve as lab exercises and must be filled in by students during the course.

This repository captures a three-part progression for building a RISC-V processor in Chisel. Each lab increases architectural complexity while preserving a common verification and simulation environment. The designs target the 32-bit base integer ISA (RV32I) and run real test programs compiled from C.

Toolchain note All Scala projects use SBT 1.10.7 with Chisel 3.6.1. The projects use chisel3.stage.ChiselStage.emitVerilog() to generate Verilog without requiring external CIRCT/firtool dependencies. This approach ensures compatibility with macOS ARM64 systems and simplifies the build process.

Repository Layout

• 0-minimal/ - Ultra-minimal RISC-V CPU with only 5 instructions (AUIPC, ADDI, LW, SW, JALR) designed to execute JIT self-modifying code demonstration. Educational example of focused processor design.
• 1-single-cycle/ - Baseline RV32I core that executes each instruction in one cycle; includes memory/peripheral models and software tests.
• 2-mmio-trap/ - Extends the single-cycle core with CSR support, a Core Local Interruptor (CLINT), and trap handling for external interrupts.
• 3-pipeline/ - Introduces multi-stage pipelines (3-stage and multiple 5-stage variants) with hazard management, forwarding, and the same CSR/interrupt features.
• tests/ - Test infrastructure including RISC-V compliance framework and other test suites

Lab Highlights

Minimal CPU

• Ultra-minimal RISC-V CPU with only 5 instructions: AUIPC, ADDI, LW, SW, JALR
• Designed specifically to execute JIT (Just-In-Time) self-modifying code demonstration
• Key features:
  • Addition-only ALU (no complex arithmetic operations)
  • Word-aligned memory access only (no byte/halfword operations)
  • No ECALL support (test verification via debug interface)
  • 60-byte test program (jit.asmbin) demonstrating: encode → copy → execute cycle
• Educational example showing how to build a focused, minimal processor for specific workloads
• All test verification performed via debug register reads, not system calls

Single-Cycle Core

• Implements the full RV32I instruction set with one-cycle latency per instruction.
• Simple Harvard-style memory interface with Verilator models in peripheral/.
• Nine ChiselTest unit and integration tests covering ALU, control flow, and sample programs.
• Key files: src/main/scala/riscv/core/*.scala, verilog/verilator/, and example programs under csrc/.

Interrupt-Capable Core

• Adds the Zicsr extension (CSRRW/CSRRS/CSRRC and immediate forms) and machine-level CSRs such as mstatus, mtvec, mepc, and mcause.
• Integrates a CLINT peripheral that raises timer and software interrupts; traps are dispatched through machine-mode handlers.
• Preserves single-cycle execution for non-trapping instructions while sequencing interrupt entry/return (mret).
• Eight dedicated tests exercise CSR access patterns and interrupt scenarios.
• Critical design: CLINT writes to CSRs take priority over CPU writes to ensure atomic trap entry.

Pipelined Cores

• Provides four pipeline implementations with increasing sophistication:
  1. ThreeStage: Simplified 3-stage pipeline (Fetch-Execute-WriteBack) for learning basics
  2. FiveStageStall: Classic 5-stage with stall-based hazard handling (no forwarding)
  3. FiveStageForward: 5-stage with full EX-to-EX and MEM-to-EX forwarding paths
  4. FiveStageFinal: Advanced 5-stage with ID-stage forwarding and early branch resolution (DEFAULT)
• Performance progression: ThreeStage (CPI ~2.5) → FiveStageStall (CPI ~1.8) → FiveStageForward (CPI ~1.3) → FiveStageFinal (CPI ~1.2)
• Includes hazard detection, pipeline flushing, branch handling, and comprehensive forwarding logic.
• Reuses CSR/CLINT infrastructure so both exceptions and asynchronous interrupts work in the presence of pipelining.
• Comprehensive test suite covers pipeline registers, hazard corner cases, and regression programs (src/test/scala/riscv/).

Build and Test Workflow

Required tools

• SBT 1.9.7+
• Chisel 3.6.1 with legacy FIRRTL compiler 1.6.0
• Verilator 5.042 or newer
• Optional: GNU RISC-V toolchain for assembling the C/asm test payloads

Build System Architecture

The projects use the legacy Scala FIRRTL compiler instead of CIRCT/firtool:

• Verilog Generation: chisel3.stage.ChiselStage.emitVerilog() generates Verilog directly
• No External Dependencies: No firtool or CIRCT tools required
• Platform Compatibility: Works on Linux, macOS (Intel and Apple Silicon), and other Unix-like systems
• Compilation Time: Legacy compiler is slower than CIRCT but fully functional

Each project's build.sbt includes:

libraryDependencies ++= Seq(
  "edu.berkeley.cs" %% "chisel3" % "3.6.1",
  "edu.berkeley.cs" %% "chiseltest" % "0.6.0" % "test",
  "edu.berkeley.cs" %% "firrtl" % "1.6.0",  // Legacy FIRRTL compiler
)

The VerilogGenerator in each src/main/scala/board/verilator/Top.scala uses:

import chisel3.stage.ChiselStage

object VerilogGenerator extends App {
  (new ChiselStage).emitVerilog(
    new Top(),
    Array("--target-dir", "verilog/verilator")
  )
}

Repository-wide targets (from top-level directory)

make clean         # Clean build artifacts from all projects
make distclean     # Deep clean: remove RISCOF results and all generated files

The distclean target removes all generated artifacts while preserving source files:

• RISCOF work directory and test results
• sbt build artifacts (target/, .bloop/, .metals/)
• Auto-generated compliance test files
• Verilator generated files and waveforms

Project-specific targets (from individual project directories)

make test       # Execute ChiselTest suite via sbt
make verilator  # Generate Verilog (via legacy FIRRTL compiler) and build Verilator C++ simulator
make sim        # Run Verilator simulation; generates waveforms in trace.vcd
make indent     # Format Scala and C++ sources (scalafmt + clang-format)
make clean      # Remove build artifacts
make compliance # Run RISCOF compliance tests for this project

The make verilator target performs two steps:

1. Generate Verilog from Chisel using sbt "runMain board.verilator.VerilogGenerator"
2. Compile the Verilator C++ simulator with VCD tracing support

The make sim target runs the compiled simulator with configurable parameters:

make sim SIM_ARGS="-instruction src/main/resources/fibonacci.asmbin" SIM_TIME=100000

VCD waveform generation can be controlled via the WRITE_VCD variable:

make sim                # Default: VCD enabled (generates trace.vcd)
WRITE_VCD=0 make sim    # Disable VCD for faster simulation
WRITE_VCD=1 make sim    # Explicit enable VCD

Note: VCD generation is enabled by default (WRITE_VCD=1). Set WRITE_VCD=0 to skip waveform generation when running large simulations or compliance tests where waveform analysis is not needed. This significantly improves simulation performance.

Learning Path and Architecture Overview

Recommended Study Sequence

1. Start with 0-minimal/ to understand focused processor design with minimal instruction set
2. Study 1-single-cycle/ to understand basic RISC-V instruction execution with full RV32I support
3. Explore 2-mmio-trap/ to learn privileged architecture and interrupt handling
4. Progress to 3-pipeline/ variants in order: ThreeStage → FiveStageStall → FiveStageForward → FiveStageFinal

Architecture Documentation

Comprehensive documentation is available in:

• Module Scaladoc: All core CPU modules have detailed Scaladoc comments
  • src/main/scala/riscv/core/CPU.scala - Top-level architecture overview
  • src/main/scala/riscv/core/Execute.scala - ALU operations and branch resolution
  • src/main/scala/riscv/core/InstructionDecode.scala - Instruction decoding and control signals
  • src/main/scala/riscv/core/CSR.scala (2-mmio-trap) - Privileged architecture and interrupt priority
• Compliance Test Documentation: src/test/scala/riscv/compliance/ComplianceTestBase.scala - RISCOF test framework

Key Architectural Decisions

Single-Cycle vs Pipeline:

• Single-cycle: CPI = 1, simple design, longer clock period
• Pipeline: CPI < 2, complex hazard handling, shorter clock period
• Trade-off: Hardware complexity vs performance

Hazard Handling Strategies:

• Stalling: Simple hardware, lower performance
• Forwarding: Complex hardware, eliminates most hazard stalls
• Early resolution: Reduces control hazard penalties

CSR Write Priority (2-mmio-trap):

• CLINT writes override CPU writes during trap entry
• Ensures atomic state updates for interrupt handling
• Critical for correct exception/interrupt behavior

Notes for Students

• The labs are designed to be read in order; later directories assume experience with modules introduced earlier.
• Comments marked with CA25: Exercise in the source indicate TODOs that must be completed as part of the exercises.
• When extending functionality, favor the existing Chisel module boundaries so the provided testbenches continue to apply.
• All source files include comprehensive Scaladoc - read the module documentation before modifying code.
• Use make test frequently to verify your changes pass all tests.