PicoRV32 - A Size-Optimized RISC-V CPU

PicoRV32 is a CPU core that implements the RISC-V RV32I Instruction Set.

Tools (gcc, binutils, etc..) can be obtained via the RISC-V Website.

PicoRV32 is free and open hardware licensed under the ISC license (a license that is similar in terms to the MIT license or the 2-clause BSD license).

Table of Contents

• Features and Typical Applications
• Files in this Repository
• Verilog Module Parameters
• Cycles per Instruction Performance
• PicoRV32 Native Memory Interface
• Pico Co-Processor Interface (PCPI)
• Custom Instructions for IRQ Handling
• Building a pure RV32I Toolchain
• Evaluation: Timing and Utilization on Xilinx 7-Series FPGAs

Features and Typical Applications

• Small (~1000 LUTs in a 7-Series Xilinx FPGA)
• High fMAX (~250 MHz on 7-Series Xilinx FPGAs)
• Selectable native memory interface or AXI4-Lite master
• Optional IRQ support (using a simple custom ISA)
• Optional Co-Processor Interface

This CPU is meant to be used as auxiliary processor in FPGA designs and ASICs. Due to its high fMAX it can be integrated in most existing designs without crossing clock domains. When operated on a lower frequency, it will have a lot of timing slack and thus can be added to a design without compromising timing closure.

For even smaller size it is possible disable support for registers x16..x31 as well as RDCYCLE[H], RDTIME[H], and RDINSTRET[H] instructions, turning the processor into an RV32E core.

Furthermore it is possible to choose between a dual-port and a single-port register file implementation. The former provides better performance while the latter results in a smaller core.

Note: In architectures that implement the register file in dedicated memory resources, such as many FPGAs, disabling the 16 upper registers and/or disabling the dual-port register file may not further reduce the core size.

The core exists in two variations: picorv32 and picorv32_axi. The former provides a simple native memory interface, that is easy to use in simple environments, and the latter provides an AXI-4 Lite Master interface that can easily be integrated with existing systems that are already using the AXI standard.

A separate core picorv32_axi_adapter is provided to bridge between the native memory interface and AXI4. This core can be used to create custom cores that include one or more PicoRV32 cores together with local RAM, ROM, and memory-mapped peripherals, communicating with each other using the native interface, and communicating with the outside world via AXI4.

The optional IRQ feature can be used to react to events from the outside, implement fault handlers, or catch instructions from a larger ISA and emulate them in software.

The optional Pico Co-Processor Interface (PCPI) can be used to implement non-branching instructions in an external coprocessor. An implementation of a core that implements the MUL[H[SU|U]] instructions is provided.

Files in this Repository

README.md

You are reading it right now.

picorv32.v

This Verilog file contains the following Verilog modules:

Module	Description
picorv32	The PicoRV32 CPU
picorv32_axi	The version of the CPU with AXI4-Lite interface
picorv32_axi_adapter	Adapter from PicoRV32 Memory Interface to AXI4-Lite
picorv32_pcpi_mul	A PCPI core that implements the `MUL[H[SU

Simply copy this file into your project.

Makefile and testbench.v

A basic test environment. Run make test, make test_sp and/or make test_axi to run the test firmware in different hardware configurations.

Note: The test bench is using Icarus Verilog. However, Icarus Verilog 0.9.7 (the latest release at the time of writing) has a few bugs that prevent the test bench from running. Upgrade to the latest github master of Icarus Verilog to run the test bench.

firmware/

A simple test firmware. This runs the basic tests from tests/, some C code, tests IRQ handling and the multiply PCPI core.

All the code in firmware/ is in the public domain. Simply copy whatever you can use.

tests/

Simple instruction-level tests from riscv-tests.

dhrystone/

Another simple test firmware that runs the Dhrystone benchmark.

scripts/

Various scripts and examples for different (synthesis) tools and hardware architectures.

Verilog Module Parameters

The following Verilog module parameters can be used to configure the PicoRV32 core.

ENABLE_COUNTERS (default = 1)

This parameter enables support for the RDCYCLE[H], RDTIME[H], and RDINSTRET[H] instructions. This instructions will cause a hardware trap (like any other unsupported instruction) if ENABLE_COUNTERS is set to zero.

Note: Strictly speaking the RDCYCLE[H], RDTIME[H], and RDINSTRET[H] instructions are not optional for an RV32I core. But chances are they are not going to be missed after the application code has been debugged and profiled. This instructions are optional for an RV32E core.

ENABLE_REGS_16_31 (default = 1)

This parameter enables support for registers the x16..x31. The RV32E ISA excludes this registers. However, the RV32E ISA spec requires a hardware trap for when code tries to access this registers. This is not implemented in PicoRV32.

ENABLE_REGS_DUALPORT (default = 1)

The register file can be implemented with two or one read ports. A dual ported register file improves performance a bit, but can also increase the size of the core.

LATCHED_MEM_RDATA (default = 0)

Set this to 1 if the mem_rdata is kept stable by the external circuit after a transaction. In the default configuration the PicoRV32 core only expects the mem_rdata input to be valid in the cycle with mem_valid && mem_ready and latches the value internally.

This parameter is only available for the picorv32 core. In the picorv32_axi core this is implicitly set to 0.

TWO_STAGE_SHIFT (default = 1)

By default shift operations are performed in two stages: first shift in units of 4 bits and then shift in units of 1 bit. This speeds up shift operations, but adds additional hardware. Set this parameter to 0 to disable the two-stage shift to further reduce the size of the core.

TWO_CYCLE_COMPARE (default = 0)

This relaxes the longest data path a bit by adding an additional FF stage at the cost of adding an additional clock cycle delay to the conditional branch instructions.

Note: Enabling this parameter will be most effective when retiming (aka "register balancing") is enabled in the synthesis flow.

TWO_CYCLE_ALU (default = 0)

This adds an additional FF stage in the ALU data path, improving timing at the cost of an additional clock cycle for all instructions that use the ALU.

Note: Enabling this parameter will be most effective when retiming (aka "register balancing") is enabled in the synthesis flow.

CATCH_MISALIGN (default = 1)

Set this to 0 to disable the circuitry for catching misaligned memory accesses.

CATCH_ILLINSN (default = 1)

Set this to 0 to disable the circuitry for catching illegal instructions.

ENABLE_PCPI (default = 0)

Set this to 1 to enable the Pico Co-Processor Interface (PCPI).

ENABLE_MUL (default = 0)

This parameter internally enables PCPI and instantiates the picorv32_pcpi_mul core that implements the MUL[H[SU|U]] instructions. The external PCPI interface only becomes functional when ENABLE_PCPI is set as well.

ENABLE_IRQ (default = 0)

Set this to 1 to enable IRQs. (see "Custom Instructions for IRQ Handling" below for a discussion of IRQs)

ENABLE_IRQ_QREGS (default = 1)

Set this to 0 to disable support for the getq and setq instructions. Without the q-registers, the irq return address will be stored in x3 (gp) and the IRQ bitmask in x4 (tp), the global pointer and thread pointer registers according to the RISC-V ABI. Code generated from ordinary C code will not interact with those registers.

Support for q-registers is always disabled when ENABLE_IRQ is set to 0.

ENABLE_IRQ_TIMER (default = 1)

Set this to 0 to disable support for the timer instruction.

Support for the timer is always disabled when ENABLE_IRQ is set to 0.

MASKED_IRQ (default = 32'h 0000_0000)

A 1 bit in this bitmask corresponds to a permanently disabled IRQ.

LATCHED_IRQ (default = 32'h ffff_ffff)

A 1 bit in this bitmask indicates that the corresponding IRQ is "latched", i.e. when the IRQ line is high for only one cycle, the interrupt will be marked as pending and stay pending until the interrupt handler is called (aka "pulse interrupts" or "edge-triggered interrupts").

Set a bit in this bitmask to 0 to convert an interrupt line to operate as "level sensitive" interrupt.

PROGADDR_RESET (default = 32'h 0000_0000)

The start address of the program.

PROGADDR_IRQ (default = 32'h 0000_0010)

The start address of the interrupt handler.

Cycles per Instruction Performance

A short reminder: This core is optimized for size, not performance.

Unless stated otherwise, the following numbers apply to a PicoRV32 with ENABLE_REGS_DUALPORT active and connected to a memory that can accommodate requests within one clock cycle.

The average Cycles per Instruction (CPI) is 4 to 5, depending on the mix of instructions in the code. The CPI numbers for the individual instructions can be found in the table below. The column "CPI (SP)" contains the CPI numbers for a core built without ENABLE_REGS_DUALPORT.

Instruction	CPI	CPI (SP)
direct jump (jal)	3	3
ALU reg + immediate	3	3
ALU reg + reg	3	4
branch (not taken)	3	4
memory load	5	5
memory store	5	6
branch (taken)	5	6
indirect jump (jalr)	6	6
shift operations	4-14	4-15

When ENABLE_MUL is activated, then a MUL instruction will execute in 40 cycles and a MULH[SU|U] instruction will execute in 72 cycles.

Dhrystone benchmark results: 0.309 DMIPS/MHz (544 Dhrystones/Second/MHz)

For the Dhrystone benchmark the average CPI is 4.167.

PicoRV32 Native Memory Interface

The native memory interface of PicoRV32 is a simple valid-ready interface that can run one memory transfer at a time:

output        mem_valid
output        mem_instr
input         mem_ready

output [31:0] mem_addr
output [31:0] mem_wdata
output [ 3:0] mem_wstrb
input  [31:0] mem_rdata

The core initiates a memory transfer by asserting mem_valid. The valid signal stays high until the peer asserts mem_ready. All core outputs are stable over the mem_valid period.

Read Transfer

In a read transfer mem_wstrb has the value 0 and mem_wdata is unused.

The memory reads the address mem_addr and makes the read value available on mem_rdata in the cycle mem_ready is high.

There is no need for an external wait cycle. The memory read can be implemented asynchronously with mem_ready going high in the same cycle as mem_valid, or mem_ready being tied to constant 1.

Write Transfer

In a write transfer mem_wstrb is not 0 and mem_rdata is unused. The memory write the data at mem_wdata to the address mem_addr and acknowledges the transfer by asserting mem_ready.

There is no need for an external wait cycle. The memory can acknowledge the write immediately with mem_ready going high in the same cycle as mem_valid, or mem_ready being tied to constant 1.

Look-Ahead Interface

The PicoRV32 core also provides a "Look-Ahead Memory Interface" that provides all information about the next memory transfer one clock cycle earlier than the normal interface.

output        mem_la_read
output        mem_la_write
output [31:0] mem_la_addr
output [31:0] mem_la_wdata
output [ 3:0] mem_la_wstrb

In the clock cycle before mem_valid goes high, this interface will output a pulse on mem_la_read or mem_la_write to indicate the start of a read or write transaction in the next clock cycles.

Note: The signals mem_la_read, mem_la_write, and mem_la_addr are driven by combinatorial circuits within the PicoRV32 core. It might be harder to achieve timing closure with the look-ahead interface than with the normal memory interface described above.

Pico Co-Processor Interface (PCPI)

The Pico Co-Processor Interface (PCPI) can be used to implement non-branching instructions in external cores:

output        pcpi_valid
output [31:0] pcpi_insn
output [31:0] pcpi_rs1
output [31:0] pcpi_rs2
input         pcpi_wr
input  [31:0] pcpi_rd
input         pcpi_wait
input         pcpi_ready

When an unsupported instruction is encountered and the PCPI feature is activated (see ENABLE_PCPI above), then pcpi_valid is asserted, the instruction word itself is output on pcpi_insn, the rs1 and rs2 fields are decoded and the values in those registers are output on pcpi_rs1 and pcpi_rs2.

An external PCPI core can then decode the instruction, execute it, and assert pcpi_ready when execution of the instruction is finished. Optionally a result value can be written to pcpi_rd and pcpi_wr asserted. The PicoRV32 core will then decode the rd field of the instruction and write the value from pcpi_rd to the respective register.

When no external PCPI core acknowledges the instruction within 16 clock cycles, then an illegal instruction exception is raised and the respective interrupt handler is called. A PCPI core that needs more than a couple of cycles to execute an instruction, should assert pcpi_wait as soon as the instruction has been decoded successfully and keep it asserted until it asserts pcpi_ready. This will prevent the PicoRV32 core from raising an illegal instruction exception.

Custom Instructions for IRQ Handling

Note: The IRQ handling features in PicoRV32 do not follow the RISC-V Privileged ISA specification. Instead a small set of very simple custom instructions is used to implement IRQ handling with minimal hardware overhead.

The following custom instructions are only supported when IRQs are enabled via the ENABLE_IRQ parameter (see above).

The PicoRV32 core has a built-in interrupt controller with 32 interrupt inputs. An interrupt can be triggered by asserting the corresponding bit in the irq input of the core.

When the interrupt handler is started, the eoi End Of Interrupt (EOI) signals for the handled interrupts go high. The eoi signals go low again when the interrupt handler returns.

The IRQs 0-2 can be triggered internally by the following built-in interrupt sources:

IRQ	Interrupt Source
0	Timer Interrupt
1	SBREAK or Illegal Instruction
2	BUS Error (Unalign Memory Access)

This interrupts can also be triggered by external sources, such as co-processors connected via PCPI.

The core has 4 additional 32-bit registers q0 .. q3 that are used for IRQ handling. When the IRQ handler is called, the register q0 contains the return address and q1 contains a bitmask of all IRQs to be handled. This means one call to the interrupt handler needs to service more than one IRQ when more than one bit is set in q1.

Registers q2 and q3 are uninitialized and can be used as temporary storage when saving/restoring register values in the IRQ handler.

All of the following instructions are encoded under the custom0 opcode. The f3 and rs2 fields are ignored in all this instructions.

See firmware/custom_ops.S for GNU assembler macros that implement mnemonics for this instructions.

See firmware/start.S for an example implementation of an interrupt handler assembler wrapper, and firmware/irq.c for the actual interrupt handler.

getq rd, qs

This instruction copies the value from a q-register to a general-purpose register.

0000000 ----- 000XX --- XXXXX 0001011
f7      rs2   qs    f3  rd    opcode

Example:

getq x5, q2

setq qd, rs

This instruction copies the value from a general-purpose register to a q-register.

0000001 ----- XXXXX --- 000XX 0001011
f7      rs2   rs    f3  qd    opcode

Example:

setq q2, x5

retirq

Return from interrupt. This instruction copies the value from q0 to the program counter and re-enables interrupts.

0000010 ----- 00000 --- 00000 0001011
f7      rs2   rs    f3  rd    opcode

Example:

retirq

maskirq

The "IRQ Mask" register contains a bitmask of masked (disabled) interrupts. This instruction writes a new value to the irq mask register and reads the old value.

0000011 ----- XXXXX --- XXXXX 0001011
f7      rs2   rs    f3  rd    opcode

Example:

maskirq x1, x2

The processor starts with all interrupts disabled.

An illegal instruction or bus error while the illegal instruction or bus error interrupt is disabled will cause the processor to halt.

waitirq

Pause execution until an interrupt becomes pending. The bitmask of pending IRQs is written to rd.

0000100 ----- 00000 --- XXXXX 0001011
f7      rs2   rs    f3  rd    opcode

Example:

waitirq x1

timer

Reset the timer counter to a new value. The counter counts down clock cycles and triggers the timer interrupt when transitioning from 1 to 0. Setting the counter to zero disables the timer. The old value of the counter is written to rd.

0000101 ----- XXXXX --- XXXXX 0001011
f7      rs2   rs    f3  rd    opcode

Example:

timer x1, x2

Building a pure RV32I Toolchain

The default settings in the riscv-tools build scripts will build a compiler, assembler and linker that can target any RISC-V ISA, but the libraries are built for RV32G and RV64G targets. Follow the instructions below to build a complete toolchain (including libraries) that target a pure RV32I CPU.

The following commands will build the RISC-V gnu toolchain and libraries for a pure RV32I target, and install it in /opt/riscv32i:

# Ubuntu packages needed:
sudo apt-get install autoconf automake autotools-dev curl libmpc-dev libmpfr-dev \
        libgmp-dev gawk build-essential bison flex texinfo gperf

sudo mkdir /opt/riscv32i
sudo chown $USER /opt/riscv32i

git clone https://github.com/riscv/riscv-gnu-toolchain riscv-gnu-toolchain-rv32i
cd riscv-gnu-toolchain-rv32i
git checkout 572033b

mkdir build; cd build
../configure --with-xlen=32 --with-arch=I --prefix=/opt/riscv32i
make -j$(nproc)

The commands will all be named using the prefix riscv32-unknown-elf-, which makes it easy to install them side-by-side with the regular riscv-tools, which are using the name prefix riscv64-unknown-elf- by default.

Note: This instructions are for git rev 572033b (2015-09-14) of riscv-gnu-toolchain.

Evaluation: Timing and Utilization on Xilinx 7-Series FPGAs

The following evaluations have been performed with Vivado 2015.1.

Timing on Xilinx 7-Series FPGAs

The picorv32_axi module with enabled TWO_CYCLE_ALU has been placed and routed for Xilinx Artix-7T (xc7a15t-fgg484), Xilinx Kintex-7T (xc7k70t-fbg676), and Xilinx Virtex-7T (xc7v585t-ffg1761) devices in all speed grades. A binary search is used to find the lowest clock period for which the design meets timing.

See make table.txt in scripts/vivado/.

Device	Speedgrade	Clock Period (Freq.)
Xilinx Artix-7T	-1	4.2 ns (238 MHz)
Xilinx Artix-7T	-2	3.5 ns (285 MHz)
Xilinx Artix-7T	-3	3.2 ns (312 MHz)
Xilinx Kintex-7T	-1	2.8 ns (357 MHz)
Xilinx Kintex-7T	-2	2.3 ns (434 MHz)
Xilinx Kintex-7T	-3	2.1 ns (476 MHz)
Xilinx Virtex-7T	-1	2.6 ns (384 MHz)
Xilinx Virtex-7T	-2	2.3 ns (434 MHz)
Xilinx Virtex-7T	-3	2.1 ns (476 MHz)

Utilization on Xilinx 7-Series FPGAs

The following table lists the resource utilization in area-optimized synthesis for the following three cores:

• PicoRV32 (small): The picorv32 module without counter instructions, without two-stage shifts, with externally latched mem_rdata, and without catching of misaligned memory accesses and illegal instructions.

• PicoRV32 (regular): The picorv32 module in its default configuration.

• PicoRV32 (large): The picorv32 module with enabled PCPI, IRQ and MUL features.

See make area in scripts/vivado/.

Core Variant	Slice LUTs	LUTs as Memory	Slice Registers
PicoRV32 (small)	751	48	422
PicoRV32 (regular)	901	48	564
PicoRV32 (large)	1718	88	1002